JP4895645B2 - Information search apparatus and information search program - Google Patents

Description

  The present invention relates to an information retrieval technique based on computer processing that takes into account pairs of expressions (words, character strings, articles, etc.) having a binary relationship from text data using supervised machine learning processing.

  Known methods for extracting information from text databases, etc. include methods for extracting binary relationships using manually created patterns, and methods for extracting desired information by paying attention to the binary relationships of related words. ing. For example, in the method of Non-Patent Document 1, a pattern frame for extracting information to be obtained using a predicate term structure that is a result of syntax analysis is given, extracted from a corpus with a correct answer, and an inappropriate pattern among the extracted patterns The matching information is extracted using the selected pattern by eliminating the pattern. More specifically, this document introduces a technique for improving the accuracy of binary relation extraction processing by selecting a pattern against a learning corpus in order to improve the accuracy of the pattern.

In the method of Non-Patent Document 2, as a text (article) search method based on keywords without considering binary relations, information on the number of occurrences of keywords for each article and the number of articles in which keywords appear in the database A technique of obtaining an article having a high score as a search result by obtaining a score based on information on the information is shown.
Yakushiji Akane et al., “Extracting Medical and Biological Field Information Using Predicate Item Structure Patterns”, Proc. 93-96, The Language Processing Society of Japan, March 2005 Mata Murata et al., “Information Retrieval Using Location Information and Field Information”, Natural Language Processing Vol. 7, No. 2, p. 141-160, Language Processing Society of Japan, April 2000

  However, when a pattern is used as a binary relation extraction rule as in the method of Non-Patent Document 1, there is a problem that the pattern becomes complicated if the target problem becomes complicated. Therefore, there is a limit to the method using the pattern. There is also a problem that the performance of the extraction method does not increase.

  For example, if an AND search is performed using two words “Kyoto Univ.” And “President” as keywords, articles containing these two keywords “Kyoto Univ.” And “President” will be searched. The articles obtained in this way are not all articles related to the “President of Kyoto University”. Specifically, an article including the sentence “The President of the University of Tokyo attended the ceremony of Kyoto University” is not an article about “the President of Kyoto University” but an article about “the President of Tokyo University”. For this reason, such articles are considered inappropriate as search results. However, search methods that do not consider binary relations, such as the method of Non-Patent Document 2, give inappropriate search results as described above. There was a problem that it could not be excluded.

  In addition, if a user can specify that a keyword has a relationship when inputting a search keyword or outputting a search result, it is considered that a search result close to the result desired by the user can be obtained. This is not taken into account in known search methods, and the search results are left to the computer. For example, if it is desired to search by specifying the keyword “language processing” in addition to the keywords “Kyoto University” and “General Manager” as described above, text data including all three keywords in a simple AND search ( Articles) can only be obtained, but for users, “language processing” and other keywords may not be so closely related, but the results obtained by the conventional search method show that the user himself You have to find the article you want.

  Therefore, the present invention can be used for all problems that extract binary relations from text data, and uses binary relation extraction processing technology that can extract binary relations with high performance even for complex problems. An object of the present invention is to provide an information search apparatus capable of obtaining an appropriate search result in consideration. In addition, another object of the present invention is to consider the relation between keywords by extracting information related to the binary relation having the search keyword as an element as a feature, and machine-learning for each article, for example, using the feature. An object of the present invention is to provide an information search apparatus capable of obtaining appropriate search results. Another object of the present invention is to specify the presence or absence of a relationship between keywords by a user or a program at the time of inputting a search keyword using a binary relation extraction processing technique, so that two or more or three or more are specified. Information search device devised to obtain appropriate search results based on keywords, and refined search based on the fact that the user specified multiple relevant keywords when outputting the search results An object of the present invention is to provide an information search apparatus capable of obtaining a result. A further object of the present invention is to provide a program for causing a computer to function as an information retrieval apparatus using these binary relation extraction processes.

  The present invention is an information search apparatus that performs an information search process using a plurality of search keywords. 1) An example of a combination of a problem and a solution, where the problem has a binary relationship with search keywords as elements. The case is extracted from a teacher data storage unit in which teacher data including a binary relation to be extracted is stored, and predetermined information is extracted as a feature for each case, and the extracted feature In addition to generating a set and set of solutions, based on a predetermined machine learning algorithm, a set of features and a set of features for which the set of features is used as the solution A learning unit that performs learning processing and stores in the learning result storage unit information indicating what kind of feature set the solution is obtained as learning result information; and 2) a plurality of input search keywords An information search unit that generates a pair of input search keywords using a keyword, extracts text data from the text data to be searched based on the input search keyword by a predetermined method, and 3) the information search A candidate extraction unit which generates a pair composed of the input search keywords from each text data obtained by searching in the unit, and uses the generated pair as a binary relation candidate; and 4) performed by the learning unit The predetermined information is extracted as a feature for the binomial relationship candidate by an extraction process similar to the extraction process, and the binomial relationship candidate is based on the learning result information stored in the learning result storage unit. In the case of a set of features of, it is determined what solution should be taken, and the determination result indicates that the binary relation to be extracted for the binary relation candidate A binary relation selection unit that selects a binary relation to be extracted as a binary relation to be extracted, and 5) a text including the binary relation selected by the binary relation selection unit. A search result extraction unit that extracts data as a search result.

  In the above-described present invention, first, in the learning unit, teacher data including cases in which solution information indicating that the binary relation to be extracted is added to the binary relation having the search keyword as an element is stored in the teacher data A case is taken out from the section, predetermined information is extracted as a feature for each case, and a set of a set of extracted features and a solution is generated. Then, based on a predetermined machine learning algorithm, machine learning processing is performed on what kind of feature set results in a set of solutions and feature sets. The information indicating what kind of solution is obtained in the case of a set of “is stored in the learning result storage unit as learning result information.

  Thereafter, an input search keyword pair using a plurality of input search keywords is generated by the information search unit, and text data is extracted and acquired based on the input search keyword by a predetermined method. Here, “the plurality of input search keywords” is not necessarily a mode of accepting keyword input by the user, and may be, for example, input by a programming function (referred to as “input input” described below). The same applies to “multiple search keywords”). The “predetermined method” may be an AND search method including an input search keyword, or may be a score method (the same applies to an “information search unit” referred to below). . Further, a pair composed of input search keywords is generated from each text data obtained by searching by the candidate extraction unit, and the generated pair is set as a binary relation candidate. Then, the binary relation selection unit extracts predetermined information as features from the binary relation candidates by the extraction process similar to the extraction process performed by the learning unit, and further stores the learning result information stored in the learning result storage unit Based on the above, it is determined what kind of solution should be used in the case of a set of features of binomial relationship candidates, and as a result of the determination, a solution indicating that the binomial relationship is to be extracted for the binary relationship candidate is determined. If it is determined that the binary relation is to be selected, the binary relation candidate is selected as the binary relation to be extracted. Finally, the search result extraction unit extracts text data including the selected binary relation as a search result.

  The present invention is also an information search apparatus that performs information search processing using a plurality of search keywords. 1) An example of a combination of a problem and a solution, where the problem is a combination of a search keyword and text data. The case is extracted from the teacher data storage unit in which the teacher data including the text data to be extracted is stored, and the information related to the binary relation having at least the search keyword as an element is used as the feature for each case. A solution-feature pair extraction unit that extracts and generates a set of the extracted feature set and solution; and 2) any feature of the feature set and solution set based on a predetermined machine learning algorithm. Machine learning processing is performed to determine whether the solution is obtained in the case of a set of information, and information indicating whether the solution is obtained in the case of the feature set is learned as learning result information. A machine learning unit stored in the result storage unit and 3) a pair of input search keywords using a plurality of input search keywords are generated, and based on the input search keywords by a predetermined method from text data to be searched An information search unit for extracting and acquiring text data, and 4) extracting at least the search keyword for a set of the search keyword and the extracted text data by an extraction process similar to the extraction process performed by the solution-feature pair extraction unit. A feature extraction unit that extracts information related to binary relations as elements as features, and 5) a feature of a set of the search keyword and extracted text data based on the learning result information stored in the learning result storage unit A solution determination unit that determines what solution should be taken in the case of a set of 6), and 6) as a result of the determination, A search for selecting the text data as the text data to be extracted and extracting the selected text data as a search result when it is determined that the set is the text data to be extracted for the set of text data And a result extraction unit.

  In the present invention described above, first, teacher data including an example composed of a set of a problem and a solution, where the problem is a set of a search keyword and text data and the solution is text data to be extracted. Cases are extracted from the stored teacher data storage unit by the solution-feature pair extraction unit, and for each case, at least information related to the binary relations having the search keyword as an element is extracted as a feature. And a pair. Here, the “text data” may be an article or a document. Next, the machine learning unit performs machine learning processing based on a predetermined machine learning algorithm to determine what type of feature set results in a set of solution and feature set. , Information indicating “what kind of solution is to be obtained in the case of what feature set” is stored in the learning result storage unit as learning result information. Thereafter, an information search unit generates a pair of input search keywords using a plurality of input search keywords, and extracts text data based on the input search keywords from the text data to be searched by a predetermined method. Upon acquisition, the feature extraction unit extracts at least information related to the binary relationship having the search keyword as an element from the search keyword and the extracted text data set by the same extraction process as the solution-feature pair extraction unit. . Further, when the solution determining unit determines what solution should be made in the case of a set of features of the set of the search keyword and the extracted text data based on the learning result information stored in the learning result storage unit, When it is determined by the result extraction unit that the determination result is the text data to be extracted for the combination of the search keyword and the extracted text data, the text data is selected as an article to be extracted. The selected text data is extracted as a search result.

Here, an information retrieval apparatus as a reference example relating to the present invention will be described. An information search apparatus as a reference example is an information search apparatus that performs an information search process using a plurality of search keywords, and 1) a binary relation based on the supervised machine learning process or a processing method using a predetermined rule. A binary relation determination unit that determines whether or not the binary relation is to be extracted by using predetermined information relating to the elements of 2), 2) input of a plurality of search keywords, and a plurality of search keywords An input unit for receiving an input of relationship information given to the set of two search keywords, and 3) a plurality of input A pair of input search keywords using the search keywords is generated, and the text data is extracted from the text data to be searched based on the input search keywords by a predetermined method. And 4) two input search keywords to which the relationship information received by the input unit from each text data searched and acquired by the information search unit is added. A candidate extraction unit that generates a pair and uses the generated pair as a binary relationship candidate; and 5) based on the binary relationship determination unit, the binary relationship candidate is a candidate for the binary relationship. Using the predetermined information related to the binary relation element, it is determined whether or not the binary relation is to be extracted. If it is determined that the binary relation is to be extracted, the binary relation candidate is selected. a binary relation selecting unit which selects a binary relation to be extracted, 6) in which and a search result extraction unit that extracts the text data as a search result including the binary relations selected binary relationships selector There is .

In the reference example described above, the binary relation determination unit first uses the supervised machine learning process or the processing method using a predetermined rule to use the predetermined information related to the binary relation element, and It is determined whether the term relation is a binary relation to be extracted. Here, in order to determine whether or not a binary relation exists, in addition to the machine learning process, for example, a “pattern processing method using a predetermined rule” is used, for example, a keyword / keyword combination pattern to be extracted as a binary relation. And processing methods based on other rules. Further, “predetermined information related to binary relation elements” means information around the binary relation elements, such as character strings and words. Then, by the input unit, the input of a plurality of search keywords is given to the set of two search keywords, which indicates that there is a semantic relationship between any two search keyword sets among the plurality of search keywords. Accepts input of relationship information. Also in this case, the “input” is not necessarily a mode of accepting keyword input by the user, and may be input by a programming function, for example. Next, an information search unit generates a pair of input search keywords using a plurality of input search keywords, and extracts text data from the text data to be searched based on the input search keywords by a predetermined method. Get. And the candidate extraction part produces | generates the pair comprised by the two input search keywords to which the relationship information which received the input from each text data acquired as a result of search was provided, and this produced | generated pair is binomial relation. As a candidate for the binary relationship, the binary relationship selection unit uses the predetermined information related to the binary relationship element that is the candidate for the binary relationship, based on the binary relationship determination unit described above. Then, it is determined whether or not the binary relationship is to be extracted, and when it is determined that the binary relationship is to be extracted, a binary relationship candidate is selected as the binary relationship to be extracted. Finally, the search result extraction unit extracts text data including the selected binary relation as a search result. In addition, instead of text data including an input search keyword having a binary relationship, an aspect in which data having no binary relationship is extracted as a search result is also conceivable.

In this reference example , the two search units indicate that the input unit has a semantic relationship with respect to the input of a plurality of search keywords and a set of two arbitrary search keywords among the plurality of search keywords. By outputting an input interface that prompts input of relationship information given to a set of keywords, and receiving a plurality of input search keywords and relationship information input to the input interface, At the time of input, the user thinks which keyword is related to which keyword, and a search result having the relationship can be efficiently searched and extracted.

Furthermore, in this reference example , the following modes can be adopted as the machine learning process. That is, the binary relation determination unit is a case composed of a combination of a problem and a solution, and the problem is a binary relation having a search keyword as an element and a binary relation to be extracted by the solution The example is taken out from the teacher data storage unit in which teacher data including is extracted, predetermined information is extracted as a feature for each case, a set of the extracted feature set and solution is generated, Based on a machine learning algorithm, a machine learning process is performed on what kind of feature set the set of generated feature sets and solutions are, and what kind of feature set In this case, the learning unit stores information indicating whether or not the solution is obtained as learning result information in a learning result storage unit. At this time, the binary relation selection unit extracts the predetermined information as a feature for the binary relation candidate by an extraction process similar to the extraction process performed by the binary relation determination unit, and the learning result Based on the learning result information stored in the storage unit, it is determined what solution should be taken in the case of a set of features of the binomial relationship candidate, and the binomial relationship candidate is extracted as the determination result What is necessary is just to select the said binomial relationship candidate as a binary relationship which should be extracted, when it determines with it being the solution which shows that it is the binary relationship which should be performed.

Further, in this reference example , in the case of performing binary relation extraction by pattern instead of machine learning processing, the binary relation determination unit patterns the pattern of keywords and keyword combinations to be extracted as binary relations. It is preferable to employ a configuration having a pattern storage unit stored in the storage unit. In this case, the binary relation selection unit matches the binary stored in the pattern storage unit with respect to the binary relation candidate. When it is determined that the binary relation is to be extracted, the binary relation candidate is preferably selected as the binary relation to be extracted.

Of the present invention and reference examples described above, except for the second-described invention, when outputting search results, information indicating which keywords and which keywords have a binary relationship is output. Then, the user may be able to grasp it. That is, the information search apparatus according to the present invention further includes an output unit that outputs the search result extracted by the search result extraction unit in addition to the above-described components. Information indicating the presence or absence of a binary relationship between the input search keywords may also be output.

  Also, in this case, regarding the information indicating the presence or absence of the binary relationship output by the output unit, the designation of either the information indicating the presence of the binary relationship or the information indicating the absence of the binary relationship is accepted and the received information is supported. By providing a re-output unit that re-outputs only the search results to the information search apparatus of the present invention, the search results are narrowed down by receiving the output of the search results and the user thinks that there is no relationship between keywords. be able to.

  The present invention is also an information search processing program using a binary relation determination result for executing a computer as the information search processing device.

  According to the present invention, by evaluating the relationship of search keywords appearing in articles of search results such as AND search processing using binary relation extraction processing, it was hit by including a search keyword. As a result, it is possible to eliminate articles whose contents are irrelevant in terms of search keywords and, as a result, are not related to the search intention. Furthermore, the performance of information retrieval processing can be improved by improving the accuracy of supervised machine learning.

  Also, by performing machine learning processing for each article, even if a keyword is included in the information search, articles that are not related to the keyword in terms of content are not searched, and only useful search results are presented to the user. Search performance is greatly improved.

  In addition, when entering keywords for information retrieval, it is possible to specify which keywords should be related by the user or program function, etc., and which keywords need not be related. Therefore, it is possible to efficiently perform information retrieval that is more flexible and necessary for the user.

  In addition, when outputting information search results, users can not only output search results but also specify which keywords must be related to each other and which keywords need not be related By configuring, necessary information can be obtained in a more narrowed form.

  DESCRIPTION OF EMBODIMENTS Hereinafter, embodiments of the present invention will be described with reference to the drawings. First, prior to description of each example described later, binary relation extraction processing that can be used in each example will be described.

  For example, when using machine learning, the binary relation extraction is realized by the binary relation extraction apparatus 1 that performs information processing according to a binary relation extraction program. This binary relation extraction device 1 uses a teacher data, which is text data with a tag indicating whether or not a binary relation to be extracted, to determine what pair of words is a binary relation to be extracted. It is a processing device that performs machine learning, acquires binary relation candidates from given text data 2, and extracts a binary relation 3 to be extracted.

  FIG. 1 shows a configuration example of a binary relation extraction apparatus 1 according to the present invention. The binary relation extraction apparatus 1 includes a teacher data storage unit 11, a solution-feature pair extraction unit 12, a machine learning unit 13, a learning result storage unit 14, a candidate extraction unit 15, a feature extraction unit 16, a solution estimation unit 17, and a binary A term relation extraction unit 18 is provided.

  The teacher data storage unit 11 is means for storing text data that is teacher data used in the machine learning process.

  Whether or not it is a binary relation that should be extracted as a problem for the binary relation elements (one element is the first element and the other element is the second element) appearing in the text data sentence as the teacher data Use the case where the information is the solution. Specifically, for a sentence that includes two or more binary relation elements in one sentence of text data, it is a pair (positive example) that should be extracted for a pair of elements that have a binary relation in that sentence. A tag indicating a solution of either a pair that should not be extracted (negative example) is manually added. When three or more binomial elements are included in one sentence, a tag is assigned to each pair that is a combination of all elements. As an example of teacher data, a binary relation to which only a solution indicating a pair to be extracted (positive example) is given may be used.

  The solution-feature pair extraction unit 12 is a processing unit that extracts a set of a solution and a set of features from cases in text data stored in the teacher data storage unit 11.

  The feature is information used in the machine learning process. The feature-feature pair extraction unit 12 uses, as features, for example, binary relation elements, words / characters appearing around the elements and their appearance positions and order, parts of speech information of the elements and surrounding words, and morphological analysis. Information such as information, parsing information, appearance distance between elements, presence / absence of other binary relation elements between elements, and the like are extracted.

  The machine learning unit 13 uses a supervised machine learning method to determine what type of solution is likely to be generated from a set of the solution extracted by the solution-feature pair extraction unit 12 and the feature set. Processing means for learning. The learning result is stored in the learning result storage unit 14.

  The feature extraction unit 16 is a processing unit that extracts a predetermined feature for a binary relation candidate extracted from the text data 2.

  The solution estimation unit 17 refers to the learning result stored in the learning result storage unit 14 and determines the degree of what kind of solution (classification destination) is likely to be obtained in the case of a set of features for each candidate of the binary relation. It is a processing means to estimate.

  Based on the estimation result of the solution estimation unit 17, the binary relationship extraction unit 18 estimates from the binomial relationship candidates that are estimated to have a high degree of solution indicating that it is a binary relationship to be extracted. This is a processing means for outputting as binary relation 3.

  In addition, the above-mentioned “solution estimation unit 17” estimates, for each of the binomial relationship candidates, the degree of what kind of solution (classification destination) is likely to be obtained in the case of a set of features. Based on the learning result information stored in the learning result storage unit 14, such a solution estimation unit 17 determines what kind of solution should be used in the case of a set of candidate features of the binomial relationship. Part ". In such a “solution determination unit”, not only “estimate the degree of what kind of solution (classification destination) is likely to be obtained” in the case of a set of features for each candidate of binary relations, , It can be determined whether or not it is a solution. In this case, the “binary relationship extraction unit 18” determines whether the solution should be used in the case of a set of features of binary relationship candidates based on the determination result of the “solution determination unit”. From the candidate for the term relation, not only the one that is estimated to have a high degree of solution indicating that it is the binary relation to be extracted is output as the term “binary relation 3,” "It is output as a binary relation 3 that is determined to be a solution indicating that it is a binary relation to be extracted."

  FIG. 2 shows a processing flow of the binary relation extraction apparatus 1.

  In the teacher data storage unit 11 of the binary relation extraction apparatus 1, whether the binary relation that is a pair of elements having a certain meaning is a binary relation to be extracted (positive) or should not be extracted as the teacher data. Text data 2 including a case to which “solution” information of either binary relation (negative) is added is stored.

  In addition, you may make it memorize | store the text data 2 containing the example which provided the predetermined solution only to the pair which should be extracted. In this case, the pair to which the solution of the text data 2 is given is regarded as being given a (positive) solution that is a binary relation to be extracted, and the remaining pairs to which no solution is given are extracted. Treated assuming that a binary (negative) solution that should not be given is given.

  First, the solution-feature pair extraction unit 12 extracts a predetermined feature for each case from the teacher data in the teacher data storage unit 11, and sets a set of the solution (information given by the tag) and the extracted feature set. Generate (step S1). The solution-feature pair extraction unit 12 extracts binary relations from text data that is teacher data using a predetermined tag, and morphological analysis processing, syntax analysis processing, element appearance positions and elements for the extracted binary relation elements A predetermined feature is extracted by performing a calculation process of the Sagano distance.

  The machine learning unit 13 is likely to be any solution (positive or negative) at any feature set from the set of the solution generated by the solution-feature pair extraction unit 12 and the feature set. Is learned by the machine learning method, and the learning result is stored in the learning result storage unit 14 (step S2). The machine learning unit 13 performs machine learning processing by using any one of techniques such as a k-nearest neighbor method, a simple Bayes method, a decision list method, a maximum entropy method, and a support vector machine method as a supervised machine learning method.

  Thereafter, the candidate extraction unit 15 inputs the text data 2 from which the binary relation is desired to be extracted, and extracts a binary relation candidate from the input text data 2 (step S3). The candidate extraction unit 15 divides the text data into sentence units, treats only sentences in which two or more binary relation elements appear in one sentence as processing targets, and extracts binary relation candidates from the sentence.

  The feature extraction unit 16 extracts features for each binary relation candidate extracted from the text data 2 by a process substantially similar to the process in the solution-feature pair extraction unit 12 (step S4).

  The solution estimation unit 17 determines, for each candidate, what kind of solution is likely to occur in the case of the set of features, that is, the degree of “prone to be positive” or “prone to be negative”. Is estimated based on (step S5). Then, the binary relation extraction unit 18 outputs a predetermined degree of candidates as binary relations 2 to be extracted from candidates estimated to be “prone to be positive” with a better degree (step S6).

  Next, a specific example of the binary relation extraction process of the present invention will be described. In this example, it is assumed that the binary relation extraction apparatus 1 extracts binary relations of protein expressions (protein names) having an interaction from a text database of biomedical papers. It is assumed that it is specified with accuracy of%.

  In addition, the elements constituting the binary relation appear in the same sentence. The elements constituting the binary relation may be elements that appear in the same paragraph or the same document.

  In the process of creating teacher data, when a specific expression such as a protein expression, a disease name and a treatment method is taken out as a binary relation element, for example, an expression that is a binary relation element is performed as follows.

  1) Extract an element using a rule. A pattern such as “NF-Kappa [AZ], where [AZ] is any character from A to Z” is defined manually, and the corresponding expression is extracted. Elements that are expression of protein names such as NF-Kappa A and NF-Kappa B are extracted based on this pattern.

  2) Extract an element using a dictionary. Use dictionaries that describe expressions such as disease names and treatment methods, and use exactly the same character strings as expressions (character strings, word strings, etc.) in those dictionaries as elements that are expressions of disease names and treatment methods. Extract.

  3) Extract elements by machine learning processing. Text data provided with a start position tag and an end position tag before and after the expression such as protein expression, disease name and treatment method is prepared as learning data. Then, perform machine learning processing using the learning data with the tag, and use the learning result to insert the tag at the start position and end position of the corresponding expression of the new text data without the tag. Specify the element with.

  4) Extract using information indicating a predetermined binary relationship. Using data in which a tag is assigned to an expression that can be a binary relation element in advance, an expression that is a binary relation element is extracted based on the tag.

  FIG. 3 shows an example of teacher data. As shown in FIG. 3A, English text data including a binary relation having an interactive protein expression as an element is used as teacher data. In this example, a tag indicating a solution (positive) is assigned to the teacher data only for the binary relation to be extracted. That is, in the machine learning process, teacher data including only positive cases is used.

  FIG. 3B shows an example of tags attached to teacher data. The teacher data includes two binary relation pairs P1 and P2. The binary relation (pair) P1 includes a first element p1 “delta-cetenin” and a second element p2 “presenilin 1”. The binary relation (pair) P2 includes a first element p1 “presenilin (PS) 1” and a second element p2 “delta-cetenin”.

The solution-feature pair extraction unit 12 extracts a set of a solution and a set of features from cases in the text data stored in the teacher data storage unit 11. For example, the following information is extracted as the feature.
1) Words or characters that appear around binary elements. For example, a predetermined number of words / characters in front of the first element (first element) of the binary relation, a predetermined number of words / characters behind the second element (second element), the first element and the second element A predetermined number of words / characters between
2) Appearance position, appearance order, etc. of words / characters appearing around binary relational elements;
3) Two elements of binary relations;
4) Part-of-speech information, morphological analysis information, etc .;
5) Parsing information of binary relational elements or surrounding words;
6) Appearance distance between the first and second elements of the binary relation;
7) Whether or not an element appears between the first element and the second element in the binary relation.

Among the features, for example, the part of speech information is acquired using an existing morphological analysis processing method such as a morphological analysis system “ChaSen” (see: http://chasen.aist-nara.ac.jp/index.html). .ja). In “ChaSen”, a Japanese sentence can be divided and the part of speech of each word can be estimated. For example, if you enter "go to school", you get the following results.
School Gakkou School Noun-General To He To particle-Case particle-General Go Iku Go Verb-independence
EOS
In this way, each line is divided so that one word is included, and reading and part-of-speech information are given to each word.

The part of speech information in the case of English text data is, for example, `` Transformation-Based Error-Driven Learning and Natural Language
Processing: A Case Study in Part-of-Speech Tagging ”(Eric Brill, Computational Linguistics, Vol. 21, No. 4, p. 543-565, 1995). By using this system, the part of speech of each word in the English sentence is estimated.

  Here, when a binary relation element appears in the same paragraph, information on whether or not the binary relation element crosses a sentence may be used as a feature. In addition, when a binary relation element appears in the same document, information on whether the binary relation element straddles a sentence or information on whether a paragraph straddles a paragraph may be used as a feature. .

The solution-feature pair extraction unit 12 extracts features from examples of teacher data to which tags as shown in FIG. 3B are assigned, and generates a set of feature sets and solutions. For example, for the case of the binary relation P2, as shown in FIG. 5, a set of a solution (positive) and the following feature set is generated.
“For,” “interaction,” and “with” appear in the first three words of the first element;
"And", "cloned", "the", "full", "-", "length", "cDNA", "of", "human" appear between elements;
“Which”, “encoded”, and “1225” appear in the last three words of the second element ”.

  Based on this solution and the set of features, the machine learning unit 13 performs machine learning processing on which feature set is likely to be a solution, and stores the learning result in the learning result storage unit 14. To do.

  The machine learning unit 13 uses a technique such as a k-nearest neighbor method, a simple Bayes method, a decision list method, a maximum entropy method, a support vector machine method, or the like as a supervised machine learning method.

  The k-nearest neighbor method is a method for obtaining a classification destination (solution) by using the k most similar cases instead of the most similar case, and by majority decision of the k cases. k is a predetermined integer number, and generally an odd number between 1 and 9 is used. The Simple Bayes method is a method of estimating the probability of each classification based on Bayes' theorem and determining the classification having the highest probability value as a classification destination.

  In the simple Bayes method, the probability of outputting the classification a in the context b is given by the following equation (1).

Here, the context b is a set of features f _j (εF, 1 ≦ j ≦ k) set in advance. p (b) is the appearance probability of the context b. Here, since it is independent of the classification a and is a constant, it is not calculated. P (a) (where P is a tilde at the top of p) and P (f _i | a) are the probabilities estimated from the teacher data, respectively. means the probability of having _i . When a value obtained by performing maximum likelihood estimation is used as P (f _i | a), the value often becomes zero, and it may be difficult to determine the classification destination because the value of Equation (2) is zero. Therefore, smoothing is performed. Here, what smoothed using the following formula | equation (3) is used.

However, freq _(f i, a), the number of cases has a feature _{f i} and classification is a, freq (a), the classification means the number of cases is a.

  The decision list method uses features and combinations of classification destinations as rules, stores them in the list in a predetermined priority order, and when input to be detected is given, from the highest priority in the list This is a method in which input data is compared with the feature of the rule, and the classification destination of the rule having the same feature is set as the classification destination of the input.

In the decision list method, the probability value of each classification is obtained using only one of the features f _j (εF, 1 ≦ j ≦ k) set in advance as a context. The probability of outputting classification a in a context b is given by

  p (a | b) = p (a | fmax) (4)

  However, fmax is given by the following equation.

Further, P (a _i | f _j ) (where P is a tilde at the top of p) is a rate of appearance of the classification a _i when the feature f _j is included in the context.

  In the maximum entropy method, assuming that a set of preset features fj (1≤j≤k) is F, expression (7) representing entropy is maximized while satisfying expression (6) below. This is a method of obtaining a probability distribution p (a, b) and determining a classification having the largest probability value among the classification probabilities obtained according to the probability distribution.

However, A and B mean a set of classifications and contexts, and g _j (a, b) is a function that is 1 if the context b has a feature f _j and the classification is a, and is 0 otherwise. means. Further, P (a _i | f _j ) (where P is a tilde at the top of p) means the rate of occurrence of (a, b) in the known data.

In equation (6), the expected value of the frequency of the output and feature pair is obtained by multiplying the probability p and the function g meaning the appearance of the pair of output and feature. And the expected value calculated based on the probability distribution calculated on the left side is the constraint, and entropy maximization (smoothing of the probability distribution) is performed to determine the probability distribution of the output and the context. For details of the maximum entropy method, see Reference 1 and Reference 2 below.
(Reference 1: Eric Sven
Ristad, Maximum Entropy Modeling for Natural Language, (ACL / EACL Tutorial
Program, Madrid, 1997);
(Reference 2: Eric Sven
Ristad, Maximum Entropy Modeling Toolkit, Release 1.6beta,
(http://www.mnemonic.com/software/memt,1998)

  The support vector machine method is a method of classifying data composed of two classifications by dividing a space by a hyperplane.

  FIG. 4 shows the concept of margin maximization in the support vector machine method. In FIG. 4, a white circle means a positive example, a black circle means a negative example, a solid line means a hyperplane that divides the space, and a broken line means a surface that represents the boundary of the margin area. 4A is a conceptual diagram when the interval between the positive example and the negative example is narrow (small margin), and FIG. 4B is a conceptual diagram when the interval between the positive example and the negative example is wide (large margin). is there.

  At this time, if the two classifications consist of positive and negative examples, the larger the interval (margin) between the positive and negative examples in the learning data, the less likely it is to make an incorrect classification with open data. As shown in FIG. 4B, a hyperplane that maximizes this margin is obtained, and classification is performed using it.

  Basically, it is as described above. Usually, an extension of the method that the training data may contain a small number of cases in the inner area of the margin, or an extension that makes the linear part of the hyperplane nonlinear ( The one with the introduction of the kernel function is used.

  This extended method is equivalent to classification using the following discriminant function, and the two classes can be discriminated depending on whether the output value of the discriminant function is positive or negative.

Where x is the context (set of features) to be identified, and x _i and y _j (i = 1,..., L, y _j ∈ {1, −1}) mean the context and classification destination of the learning data And the function sgn is
sgn (x) = 1 (x ≧ 0) (9)
-1 (otherwise)
In addition, each α _i is for maximizing equation (9) under the constraints of equations (10) and (11).

  The function K is called a kernel function, and various functions are used. In this embodiment, the following polynomial is used.

K (x, y) = (x · y + 1) d (12)
C and d are constants set experimentally. In a specific example to be described later, C is fixed to 1 through all the processes. Moreover, two types of 1 and 2 are tried for d. Here, x _i satisfying α _i > 0 is called a support vector, and the portion taking the sum of Expression (8) is normally calculated using only this case. That is, only actual cases called support vectors are used for actual analysis.

For details of the extended support vector machine method, see Reference 3 and Reference 4 below.
(Reference 4: Nello
Cristianini and John Shawe-Taylor, An Introduction to Support Vector Machines
and other kernel-based learning methods, (Cambridge University Press, 2000);
Reference 5: Taku
Kudoh, Tinysvm: Support Vector
machines, (http://cl.aist-nara.ac.jp/taku-ku//software/Tiny SVM / index.html, 2000))

  The support vector machine method handles data with two classifications. Therefore, when handling cases with three or more classifications, a pair-wise method or a one-VS rest method is usually used in combination with this.

  In the pairwise method, in the case of data having n classifications, every pair (n (n-1) / 2) of two different classification destinations is generated, and a binary classifier indicates which is better for each pair. That is, it is obtained by the support vector machine method processing module and finally obtains the classification destination by majority decision of the classification destination by n (n−1) / 2 binary classification.

  For example, when there are three classification destinations a, b, and c, the one VS rest method generates three sets of classification destination a and other, classification destination b and other, classification destination c and other, The learning process is performed on the set of the support vector machine method. Then, in the estimation process based on the learning result, the learning results of the three sets of support vector machines are used. Look at how the three support vector machines estimate the binomial candidates to be estimated. Of the three support vector machines, the other classification destination and the support vector This is a method for obtaining a classification destination of the one farthest from the separation plane of the machine. For example, when a candidate is farthest from the separation plane in the support vector machine created by the learning process of “classification destination a and others”, the candidate classification destination is estimated as a.

  Thereafter, the candidate extraction unit 15 extracts a binary relation candidate from the input new text data 2. Specifically, the text data 2 is divided into sentence units, and expressions (character strings) that are binary relation elements in each sentence are extracted. Then, it is checked whether there are two or more expressions that are binary relation elements in one sentence, and all two combinations (pairs) of binary relation elements in one sentence are considered as binary relation candidates. Generate.

  Also, the new text data 2 is divided into each paragraph, the expressions that are the binary relation elements in each paragraph are extracted, and all the two combinations (pairs) of the paragraph that has two or more elements from the same paragraph. May be generated as a binary relation candidate. Alternatively, expressions that are binary relation elements from within one document of text data 2 may be extracted, and all two combinations (pairs) may be generated as binary relation candidates.

  As a method for extracting an expression that is a binary relation element from the text data 2, the method described in the above-described teacher data generation method is used. For example, an expression that is extracted based on a learning result of supervised machine learning that extracts an expression that matches a description of a pattern or a dictionary is extracted.

  When two or more elements appear in one sentence of the text data 2, the pair of elements is set as a binary relation candidate. When three or more elements appear in one sentence, any combination of elements is considered as a binary relation candidate.

  Then, the feature extraction unit 16 extracts similar features from the binomial relationship candidates by the same processing as the solution-feature pair extraction unit 12.

  Based on the learning result stored in the learning result storage unit 14, the solution estimation unit 17 is likely to be a positive solution in the case of a set of candidate features for each binary relation candidate. Is estimated. Based on the estimation result of the solution estimation unit 17, the binary relation extraction unit 18 outputs, as a binary relation 2, a candidate having a high degree of estimation that is likely to be a positive solution.

  In this example, the above features are extracted and the support vector machine method is used as machine learning processing. When the accuracy was examined using 10-part cross validation, an accuracy of F value = 47.5% was obtained. The F value is a harmonic average of the recall and the precision. The recall is a ratio indicating how much of the binary relations to be extracted from the text data 2 can be output. The relevance ratio is a ratio indicating how much of the binary relations extracted by the binary relation extraction apparatus 1 is the binary relation to be extracted.

  In the binary relation extraction apparatus 1, the machine learning unit 13 uses a given teacher data to determine what kind of combination of a binary relation solution and a feature set is based on a given teacher data. Machine learning processing for what kind of solution is the case of a feature set, and learning result information as information indicating what kind of solution is for what kind of feature set The result is stored in the unit 14, and the solution estimation unit 17 estimates the degree of the solution that is likely to be the solution in the case of a set of binary candidate features based on the learning result information.

  When the k-nearest neighbor method is used as the machine learning method in the binary relation extraction apparatus 1, the machine learning unit 13 uses the feature ratios that overlap among the feature sets extracted from the case examples of the teacher data. It is defined as the similarity between cases based on (the ratio of how many same features), and the defined similarity and the case are stored in the learning result storage unit 14 as learning result information.

  Then, when new text data 2 is input, the solution estimation unit 17 refers to the similarity and the case defined by the learning result storage unit 14 and uses the binary relation candidates extracted from the text data 2 as follows. The k cases are selected from the cases in the learning result storage unit 14 in descending order of the similarity of the candidates, and the classification destination determined by the majority vote in the selected k cases is selected as the binomial candidate classification destination (solution ). That is, the solution estimation unit 17 determines the degree of likelihood of being a certain solution in the case of a set of binary candidate features, the number of votes of majority vote in the selected k cases, in this case, the classification “to be extracted” Is the number of votes obtained by. When the simple Bayes method is used as the machine learning method, the machine learning unit 13 stores, in the learning result storage unit 14, a set of the solution of the case and a set of features as learning result information for the case of the teacher data. Remember. Then, when new text data 2 is input, the solution estimation unit 17 extracts features based on the Bayes' theorem based on the set of the learning result information in the learning result storage unit 14 and the set of features. The probability of becoming each classification in the case of the feature set of the binomial relationship candidates acquired by the unit 16 is calculated, and the classification having the highest probability value is classified into the classification (solution) of the candidate of the binomial relationship Estimated. That is, in the solution estimation unit 17, the degree of easiness of becoming a solution in the case of a set of binary candidate features is set as the probability of becoming each classification, here, the probability of becoming “to be extracted”.

  When the decision list method is used as the machine learning method, the machine learning unit 13 stores, in the learning result storage unit 14, a list in which rules of features and classification destinations are arranged in a predetermined priority order with respect to examples of teacher data. Remember. Then, when new text data 2 is input, the solution estimation unit 17 causes the features of the binomial relation candidates extracted from the text data 2 and the features of the rules in descending order of priority in the list of the learning result storage unit 14. And the classification destination of the rule with the same feature is estimated as the candidate classification destination (solution). That is, the solution estimation unit 17 classifies the degree of likelihood of being a solution in the case of a set of binary candidate features with a predetermined priority or a numerical value or scale corresponding thereto, in this case, “to be extracted”. Is a priority in the probability list.

  When the maximum entropy method is used as the machine learning method, the machine learning unit 13 specifies a class that can be a solution from the example of the teacher data, and maximizes an expression that satisfies a predetermined conditional expression and shows entropy. A probability distribution composed of a set of features and a binomial classification that can be a solution is obtained and stored in the learning result storage unit 14. Then, when new text data 2 is input, the solution estimation unit 17 uses the probability distribution of the learning result storage unit 14 to determine the set of features of the binary relation candidates extracted from the text data 2. A probability of a class that can be a solution is obtained, a class that can be a solution having the largest probability value is specified, and the specified class is estimated as a candidate solution. That is, in the solution estimation unit 17, the degree of easiness of becoming a solution in the case of a set of binary candidate features is set as the probability of becoming each classification, here, the probability of becoming “to be extracted”.

  When the support vector machine method is used as the machine learning method, the machine learning unit 13 specifies a class that can be a solution from the example of the teacher data, divides the class into a positive example and a negative example, A hyperplane that maximizes the interval between the positive and negative examples of a case and divides the positive and negative examples by a hyperplane in a space whose dimension is a set of case features according to a predetermined execution function using Is stored in the learning result storage unit 14. Then, when new text data 2 is input, the solution estimation unit 17 uses the hyperplane of the learning result storage unit 14, and the set of features of the binomial relationship candidates extracted from the text data 2 is superfluous. In the space divided by the plane, it is identified whether the space is on the positive example side or the negative example side, and the classification determined based on the identified result is estimated as the candidate solution. That is, the solution estimation unit 17 determines the degree of likelihood of being a certain solution in the case of a set of candidate features of binary relations by determining the distance from the separation plane to the space of the positive example (binary relations to be extracted). Say it. More specifically, when a binary relation to be extracted is a positive example and a binary relation that should not be extracted is a negative example, a case located in the space on the positive example side with respect to the separation plane The distance from the separation plane of the case is set as the degree of the case.

Further, the solution-feature pair extraction unit 12 may use, for example, “a word of two elements” as the feature. In addition, “first word / character string from the front of the element, second word / character string, first word / character string from the rear, second word / character string” is used as a feature. May be. In the case of FIG. 3A, the feature is
"First element is presenilin"
(PS) 1 ”;
The second element is "delta-catenin";
The first word of the first element is “presenilin”;
The second word is “(PS)”;
The second word from the end of the first element is “(PS)”;
The first word from the end is “1”;
The first word of the second element is "delta";
The second word is “-”;
The second word from the end of the second element is "-";
The first word from the end is “catenin”.

Or
“The first letter of the first element is“ p ”;
The first two letters are “pr”;
The first three letters are “pre”;
The last character is “1”;
The last two characters are “space, 1”;
The last three characters are “), space, 1”;
The first character of the second element is "d";
The first two letters are “de”;
The first three letters are “del”;
The last character is “n”;
The last two letters are “in”;
The last three characters are “ini”.

Also, if the feature itself is the two words before and after the element and its part of speech information,
"The word before the first element is" interaction ";
The part of speech of the previous two words is “noun”;
The previous word is “with”;
The part of speech of the previous word is “preposition”;
The next word is “and”;
The part of speech of the next word is “connective”;
The word after the second is “cloned”;
The part of speech of the second word is “verb”;
The word immediately before the second element is “of”;
The part of speech of the previous two words is “preposition”;
The previous word is “human”;
The part of speech of the previous word is “noun”;
The next word is “which”;
The part of speech of the next word is “pronoun”;
The second word is "encoced";
The part of speech of the next two words is “verb”.

  Further, when the number of words between the two elements is used as the feature as the distance between the two elements, the information “the distance between the two elements is“ 9 ”” is the feature.

  In addition, a state where the number of words between two elements is 0 to 1, “distance is small”, a state 2 to 4 is “medium”, a state 5 to 9 is “distance is large”, and 10 or more When each state having a state of “distance extra large” is a feature, information that “the distance between the two elements is“ distance large ”” is the feature.

  Further, when a feature is whether there is no other element between two elements, information that “there is no other element between the two elements” is the feature.

  Furthermore, when different terms are set as binary relation elements, the appearance order of elements may be used as a feature. For example, in the case of a binary relation between a disease name and a treatment method, “the first element is“ disease name ”and the second element is“ treatment method ”” or “the first element is“ treatment method ”and the second element is Information that “is a disease name” is a feature.

  The binary relation extraction apparatus 1 uses, as teacher data, a binary relation between a disease name and a treatment method, a binary relation between a disease name and a protein expression, a disease name and an organ (organ). ), Binomial relationship between disease name and animal species, binary relationship between disease name and chemical substance, binary relationship between protein expression and the experimental methods used so far By giving various binary relation examples, these corresponding binary relations can be extracted from the text data 2 of the biomedical paper.

For example, text data including the following binary relationship can be used as the teacher data.
"Oral
corticosteroids (element: cure) are
the preference of many for the treatment of CIDP (element: disease name), being much less expensive than IVIG (element: treatment method) infusion or TA (element: treatment method). "
“In the CIDP (element: disease name) patient, the IgG antibody (element: protein expression) titer to GD3 (element: chemical expression) was remarkably elevated (titer,
1: 10,000), indicating maximal avidity to the tetrasaccharide epitope
(-NeuAcalpha2-8NeuAcalpha2-3Galbeta1-4Glc-). ''
"Ciliated
metaplasia (CM) in the stomach (element: organ name) is mainly found in gastric mucosa (element: organ name) that harbours gastric cancer (element: disease name) "
"Variant
Creutzfeldt-Jakob disease (CJD) (element: disease name) is a transmissible spongiform encephalopathy believed to be caused
by the bovine (element: animal species)
spongiform encephalopathy agent, an abnormal isoform of the prion protein
(PrP (sc)) (element: protein expression). "
“AIDP (element: disease name) and CIDP (element: disease name) having specific antibodies to the carbohydrate epitope
(-NeuAcalpha2-8NeuAcalpha2-3Galbeta1-4Glc-) of gangliosides.
(Element: chemical substance expression)
"Gene expression
in archived frozen sural nerve biopsies of patients with chronic inflammatory
demyelinating polyneuropathy (CIDP) (element: disease name) was compared to that in vasculitic nerve biopsies (VAS) and to
normal nerve (NN) by DNA microarray technology.
"This novel
interaction was identified in a yeast two-hybrid screen (element: experimental method) using PrP (C) (element: protein expression) as bait and confirmed by an in
in vitro binding assay and co-immunoprecipitations "
"Comparative
study of the PrP (BSE) (element: protein expression) distribution in brains f (element: organ name) rom BSE (element: disease name)
field cases using rapid tests. "

  Further, for example, a pair of a company product name and a reputation for the product (for example, information such as good reputation or bad reputation) can be extracted as a binary relation.

  As described above, according to the binary relation extraction process using the binary relation extraction apparatus 1 described above, the evaluation (solution) of whether or not the binary relation to be extracted is the teacher data for the machine learning process. It is possible to automatically extract a binary relation estimated to be worth extracting from new text data simply by preparing text data to which is added. Thereby, it is possible to avoid the complexity of generating the pattern used for the binary relation extraction process. In addition, improvement in the performance of binary relation extraction processing can be expected by improving the accuracy of supervised machine learning.

  In the machine learning process described above, the case where there are two keywords has been described. However, even when there are three (or more) keywords, binary relations can be extracted. In the machine learning process described above, “degree” was introduced and the value of the product was large, but all binary relations should be extracted at least at one place in the text data. In this case, it may be taken as text data to be extracted.

In addition to the aspect using the machine learning process as described above, it is also possible to employ a method of extracting a binary relation to be extracted by a pattern-based processing method. For example, assuming the words “A” and “B” as keywords,
"A's B", "A is B", "A is B", "A's related B", "A dependent B"
Create a pattern such as If the text data (article) matches the pattern, the binary relation in the text data is determined to be a binary relation to be extracted. If the text data does not match, the binary relation is not to be extracted. Since it is determined, what is determined as a binary relation to be extracted may be extracted. Note that the binomial relation extraction process using such a pattern can also be applied to a case where there are three or more keywords. That is, when all binary relations are binary relations to be extracted at least at one place in the text data, the text data to be extracted may be taken.

  Next, the information search apparatus 4 which is 1st Example of this invention is demonstrated.

  This information search device 4 regards the relationship between two search keywords in AND search processing as a meaningful binary relationship, and is a relationship to be extracted for a binary relationship having this search keyword as an element (correct) Alternatively, in an article including two search keywords from the search text data 5 to be searched by using machine data with a teacher data to which a tag indicating any solution that is not to be extracted (negative) is attached. In this processing apparatus, the search keyword pair is determined to be a binary relation to be extracted, and a search result 6 is output.

  FIG. 6 shows a configuration example of the information search apparatus 4 according to the present invention. The information search apparatus 4 includes an information search unit 40, a learning unit 42, a candidate extraction unit 44, a binary relation selection unit 45, and a search result extraction unit 48. However, the teacher data storage unit 41 and the learning result storage unit 43 may be included in the information search device 4 as internal elements, or may be made available from the information search device 4 as external elements. In the figure, the learning unit 42 further includes a solution-feature pair extraction unit 421 and a machine learning unit 422, and the binary relation selection unit 45 further includes a feature extraction unit 451 and a solution determination unit 452. Is shown.

  In this information search device 4, a teacher data storage unit 41, a solution-feature pair extraction unit 421, a machine learning unit 422, a learning result storage unit 43, a candidate extraction unit 44, a feature extraction unit 451, and a solution determination unit 452 1, the teacher data storage unit 11, the solution-feature pair extraction unit 12, the machine learning unit 13, the learning result storage unit 14, the candidate extraction unit 15, the feature extraction unit 16, and the solution estimation unit 17. And processing means for performing similar processing. However, for the solution determination unit 452, in the case of performing a process of estimating the degree of the solution (classification destination) that is likely to be obtained in the case of the set of features for each candidate of the binary relation, Although the same processing as that of the solution estimation unit 17 is performed, as described in the section of the binary relation extraction device 1, when the solution estimation unit 17 is replaced with a “solution determination unit”, the solution determination of the information processing device 4 is performed. The unit 552 performs the same processing as the “solution determination unit”. The solution determination unit 452 to be described below mainly describes a case where the same processing as that of the solution estimation unit 17 is performed.

  The information search unit 40 searches the search text data 5 using the search keyword given in the AND search process, and acquires the corresponding article (text data). The information search unit 40 may employ a search process other than the AND search process, which will be described later.

  The candidate extraction unit 44 extracts a binary relation candidate having the same character string (word) pair as two search keywords included in the article acquired by the information search unit 40 as elements.

  Based on the determination result (estimation result) of the solution determination unit 452, the search result extraction unit 46 extracts an estimated positive solution (extracts from the binary relation candidates of the articles searched from the search text data 5. (A power binomial relationship) that is better than a predetermined level is extracted, and an article that includes the extracted binary relation candidate or information that identifies an article is output as a search result 6.

FIG. 7 shows a processing flow of the information search apparatus 4. In the teacher data storage unit 41 of the information search device 4, whether or not the binary relation to be extracted is the teacher data, that is, the binary relation having the two search keywords given by the AND search processing as elements (positive) or extraction. Text data including a case to which information of “solution”, which is a binary relationship that should not be performed or is negative, is stored.
First, in the learning unit 42, the solution-feature pair extraction unit 421 extracts a predetermined feature for each case from the teacher data in the teacher data storage unit 41, and sets a solution (information given by the tag) and the extracted features. Are generated (step S11). The solution-feature pair extraction unit 421 extracts binary relations from text data that is teacher data by a predetermined tag, and uses the morphological analysis process, the syntax analysis process, and the element of the extracted binary relation elements (search keywords). A predetermined feature is extracted by performing a process of calculating an appearance position and a distance between elements.

  Similarly, in the learning unit 42, the machine learning unit 422 determines what kind of solution (positive or negative) from the combination of the solution generated by the solution-feature pair extraction unit 421 and the set of features. The learning result is stored by the machine learning method and the learning result is stored in the learning result storage unit 43 (step S12). The machine learning unit 422 performs machine learning processing using any one of techniques such as a k-nearest neighbor method, a simple Bayes method, a decision list method, a maximum entropy method, a support vector machine method, and the like as a supervised machine learning method.

  Thereafter, the candidate extraction unit 44 generates all two combinations (pairs) using the two input search keywords given in the AND search process (step S13). The information search unit 40 AND-searches the search text data 5 using two pairs of input search keywords to extract articles (text data) including the input search keyword pairs, and the candidate extraction unit 44 performs search processing. All the two combinations (pairs) are extracted as binomial relationship candidates using the input search keyword appearing in the article extracted by (Step S14).

  Then, in the binary relation selection unit 45, the feature extraction unit 451 performs predetermined processing for each binary relation candidate appearing in the searched article by a process substantially similar to the process in the solution-feature pair extraction unit 421. A set of features is extracted (step S15).

  Further, in the binary relation selection unit 45, the solution determination unit 452 determines, for each candidate, what kind of solution is likely to occur in the case of the set of features, that is, the degree of “prone to be positive” or “prone to be negative”. Is estimated based on the learning result of the learning result storage unit 43 (step S16). Then, the search result extraction unit 46 selects, as a binary relation to be extracted from the binary relation candidates that are estimated to be “prone to be positive” to a better degree than a predetermined degree, and includes this binary relation. An article or information specifying the article is selected as the search result 6 (step S17).

  Next, a specific example of information search processing by the information search device 4 will be described. In this example, the information search apparatus 4 uses, as teacher data, text data including a binary relation having elements of character strings that can be two search keywords used in the AND search process from the search text data 5. Then, a binary relation candidate having the input search keyword given in the AND search process as an element is created, and an article is extracted by performing a search using the binary relation candidate from the search text data 5. Estimate whether binary search candidates for the input search keywords included in the searched articles should be extracted, and search for articles containing binary search candidates that are estimated to be good. The search result 6 is output.

  It is assumed that “Kyoto University” and “General Manager” are set as search keywords for AND search. Further, it is determined by a person whether the binary relation of the search keyword is positive or negative, and a tag indicating a positive or negative solution is manually added. Therefore, teacher data including positive cases and negative cases is used in the machine learning process.

  FIGS. 8 to 10 show examples of teacher data stored in the teacher data storage unit 41 and examples of features extracted by the solution-feature pair extraction unit 42 from the teacher data. In this example, a tag indicating that the solution is positive for the binary relation to be extracted is assigned to the teacher data in FIGS. 8 and 9. In addition, the teacher data in FIG. 10 is provided with a tag indicating that the solution is negative for binary relations that should not be extracted.

  The teacher data of FIG. 8 includes a binary relation pair P3 which is a pair of two search keywords, and the binary relation (pair) P3 includes a first element p1 (search key K1) “Kyoto University”, The second element p2 (search key K2) is composed of “total length”, and a positive solution is given to the binary relation pair P3.

  Similarly, the teacher data of FIG. 9 includes a binary relation pair P4 which is a pair of two search keywords, and the binary relation (pair) P4 includes the first element p1 (search key K1) “K”. “Large” and the second element p2 (search key K2) “total length”, and a positive relation (positive) is given to the pair P4 of the binary relation. This is because it can be determined that the teacher data in FIGS. 8 and 9 is the content of “the total length of Kyoto University”.

  The teacher data in FIG. 10 includes a binary relation pair P5 that is a pair of two search keywords, and the binary relation (pair) P5 is the first element p1 (search key K1) “Kyoto Univ. ”, The second element p2 (search key K2)“ total length ”, and a negative relation (negative) is given to the pair P5 of the binary relation. This is because “Kyoto University” and “President” appear in the same data, but they are not related to each other and can be determined not to be the contents of “President of Kyoto University”.

The solution-feature pair extraction unit 421 extracts a set of a solution and a set of features from an example of teacher data stored in the teacher data storage unit 41. For example, as the feature, the two words before and after the element (search keyword) and the part of speech of the word are used as the feature. In this case, the feature is
“The word before the first element is“ Today ”;
The part of speech of the previous two words is “noun”;
The previous word is “,”;
The part of speech of the previous word is “reading”;
The next word is “de”;
The part of speech of the next word is “particle”;
The next word is “no”;
The part of speech of the next word is “particle”;
The second word before the second element is “de”;
The part of speech of the previous two words is “particle”;
The previous word is “,”;
The part of speech of the previous word is “reading”;
The next word is “ga”;
The part of speech of the next word is “particle”;
The second word is “attendance”;
The part of speech of the word after the second is “noun”.

  The solution-feature pair extraction unit 421 can extract information as described in the binary relation extraction process as a feature.

  Based on this solution and the set of features, the machine learning unit 422 performs machine learning processing to determine what type of solution (positive / negative) is likely to occur in what type of feature set. The learning result is stored in the learning result storage unit 43. The machine learning unit 422 uses the above-described processing methods such as the k-nearest neighbor method, the simple Bayes method, the decision list method, the maximum entropy method, and the support vector machine method as the supervised machine learning method.

  Thereafter, the information search unit 40 AND-searches the search text data 5 based on the given input search keywords “Kyoto Univ.” And “total length”, and acquires an article including the input search keyword. Then, the candidate extraction unit 44 extracts a binary relation candidate from the extracted article. Specifically, a binary relation candidate is extracted from an input search keyword included in an article as a search result of an AND search. Then, the feature extraction unit 451 extracts the same feature as the solution-feature pair extraction unit 421 from the binomial relationship candidates, and the solution determination unit 452 uses the learning result stored in the learning result storage unit 43. For each binomial relationship candidate, the degree of likelihood of being positive or negative is estimated in the case of a set of candidate features. The search result extraction unit 46 extracts a binary relationship that is likely to be an estimated positive solution from the binomial relationship candidates based on the estimation result of the solution determination unit 452, and obtains this binary relationship. The information including the article and the information specifying the article is output as the search result 6.

  For example, the candidate extraction unit 44 generates all combinations (pairs) of two input search keywords from a given input search keyword, and sets the generated pair as a binary relation candidate. Then, the information search unit 40 performs an AND search process using each binary relation candidate element (two input search keywords). Then, the feature extraction unit 451 extracts a set of predetermined features for the binomial relationship candidates appearing in the extracted article.

  Based on the learning result of the learning result storage unit 43, the solution determination unit 452 estimates the degree of ease of solution for each binary relation candidate in the case of a set of candidate features. If each of the binomial candidates that are pairs of the input search keywords appears only once in the searched article, all of the binomial candidates are positive (should be extracted). When it is estimated that the degree is good, the information specifying the article and the article is set as the search result 6.

  Also, when multiple binary relations that are pairs of input search keywords appear in the searched article, one of the multiple binary relation candidates appearing should be positive (extracted) )), And if each of the binomial candidates is estimated to be positive by satisfying all the above conditions, the article or article is identified. This information is designated as search result 6.

  Further, the candidate extraction unit 44 generates a pair of all two input search keywords from the given input search keyword, and sets the generated pair as a binary relation candidate. Then, the information search unit 40 performs an AND search process using each binary relation candidate element (two input search keywords). Then, the feature extraction unit 451 extracts a set of predetermined features for the binomial relationship candidates appearing in the extracted article.

  Based on the learning result of the learning result storage unit 43, the solution determination unit 452 estimates the degree of ease of solution for each binary relation candidate in the case of a set of candidate features. When each of the binary relation candidates that are a pair of input search keywords appears only once in the searched article, the degree of positiveness (to be extracted) for all the binary relation candidates Is multiplied by the positive degree estimated for all of the binomial relationship candidates, and is set as the positive degree of the article. Then, an article that is estimated to have a positive degree and information that identifies the article are set as a search result 6.

  In addition, when a plurality of binary relations that are pairs of input search keywords appear in the searched article, a positive degree is estimated for a plurality of binary relation candidates that appear, and the two binary relations are estimated. Among the estimated degrees of the term relation candidates, the degree with the best value is set as the degree of the binary relation candidate. Then, the degree of each binary relation is obtained, and the product obtained by multiplying the obtained degree is set as the positive degree of the article. Then, an article that is estimated to have a positive degree and information that identifies the article are set as a search result 6.

  As described above, according to the information search apparatus 4 of this embodiment, whether or not the binary relation to be extracted is the binary relation between the two search keywords of the AND search process as the teacher data for the machine learning process. It is possible to automatically extract articles including binary relations that are supposed to be extracted from the new search text data 5 simply by preparing text data with such an evaluation.

  The information search apparatus 4 of the present embodiment is hit by including a search keyword by evaluating the relationship of search keywords appearing in an article as a search result of AND search processing using a binary relation extraction process. However, it is possible to eliminate articles whose contents are irrelevant in terms of search, that is, contents that are not related to search intentions. In addition, improvement in the performance of information retrieval processing can be expected by improving the accuracy of supervised machine learning.

  In the above embodiment, an example of a binary relation composed of two elements has been described in the binary relation extraction process and the information search process. The present invention can also be applied to a ternary term relationship composed of three elements.

  For example, in the binary relation extraction apparatus 1, data including a ternary relation of three elements is prepared as teacher data. Then, the solution-feature pair extraction unit 12 determines the feature about the ternary relationship by, for example, the first two words of the first element (first appearing element) of the three elements, the third element (finally Machine learning unit by making the word information of the two words behind the appearing element), all the words between the first element and the second element (element appearing in the middle), and all the words between the second element and the third element 13 can learn the easiness of the solution based on the set of features of the ternary relationship, and the binary relationship extraction unit 18 can handle the extraction of the ternary relationship. Note that the solution given to the ternary relationship is the “ternary relationship that should be extracted” or “the ternary relationship that should not be extracted”, as in the case of the binary relationship.

  For example, in the binary relation extraction apparatus 1, data including a ternary relation of three elements is prepared as teacher data. Each processing means of the binary relation extraction device 1 then obtains the binary relation obtained by decomposing the ternary relation of the teacher data, the binary relation between the first element and the second element, and the second element and the second element. The binary relation between the three elements and the binary relation between the first element and the third element are treated as separate binary relations. Then, for each of the binomial relationships, the degree of solution as to whether it is a ternary relationship to be extracted is calculated, and a value obtained by multiplying the calculated degrees is set as the degree of the ternary relationship. And the thing with the big degree is taken out as a ternary relation which should be extracted.

  At this time, when the machine learning unit 13 uses the support vector machine method, there are two classification destinations (positive or negative), and therefore machine learning of the ternary relationship using the pairwise method or the one VS rest method. To do.

  In addition, the binary relation extraction unit 18 can obtain the certainty of extraction when extracting the binary relation 2. Then, as a certainty factor of the ternary relationship created by combining a plurality of binary relationships, a product having a high certainty factor of the ternary relationship is extracted using a product of the certainty factors of the combined binary relationships. As the certainty of the binary relation, the certainty calculated in the normal machine learning process is used.

  Such a ternary relation extraction process can be similarly performed in the information search apparatus 4. For example, when searching for articles related to “the president of Kyoto University in 2000”, the teacher data includes a ternary relationship based on three search keywords of “2000”, “Kyoto University”, and “total president”. Given data, a search result 6 of AND search using these three search keywords is output from the search text data 5.

  Moreover, in this example, as the information of the solution given to the binary relation or ternary relation of the case, “positive (is a binary relation that should be extracted)” or “negative (is a binary relation that should not be extracted)” However, the information of the solution to be assigned may be multi-category such as “having interaction”, “having reaction”, and “having no action”.

  In addition to the above-described machine learning processing, the following processing method can be employed. For example, using neural network (perceptron), HMM (Hidden Markov Model), CRF (Conditional Random Field), multiple regression analysis, decision tree learning, maximum likelihood estimation, etc. Also good.

  It should be noted that the binary relation extraction by the learning unit 42 of the information search device 4 is not limited to the above-described mode based on machine learning, but as described in the section of the binary relation extraction apparatus 1 described above. You may employ | adopt the method of taking out the binary relationship which should be extracted by the processing method based on.

  In particular, when there are three or more keywords, in the machine learning process, “degree” is introduced and the product value is large, but in the case of the pattern-based processing method, all two When the term relation becomes a binary relation to be extracted at least at one place in the text data, it can be taken as the text data to be extracted. This can be applied even in the case of machine learning processing. That is, when all binary relations become binary relations to be extracted at least at one place in the text data by the machine learning method, the text data to be extracted may be taken.

  Moreover, although the information search part 40 mentioned above demonstrated performing AND search process, the aspect which performs the search process by a score other than that can be employ | adopted. Hereinafter, an example of search processing based on scores will be described.

  As an example of a search process using scores, the TF / IDF method is mainly used for automatic extraction of important keywords. TF represents the number of appearances of the word in the document, and IDF represents the reciprocal of the number of documents that appear in a number of documents that the word has in advance. In general, a word having a larger product of TF and IDF is more appropriate as a key phrase. By the way, since there are a plurality of documents in the IDF, when reading a book, there is only one book, so there is no plurality of documents. Therefore, for example, a method is considered in which a plurality of books are taken and each ID is regarded as a document to obtain an IDF. Also, there is a method in which the IDF is obtained from data other than book data and used as the book data. That is, not only books but also document data is collected, and IDFs are obtained considering each document as an individual document. Therefore, the IDF obtained can be used when the IDF is used in the present invention. Since TF is the number of appearances in the document, when applied to the present invention, the number of appearances in the entire electronic document may be used.

  Strictly speaking, the IDF is calculated not only by reciprocal TF but also by the following equation (12).

  In the basic TF / IDF method, the score is obtained by the following equation (13).

ただし、ｗ∈Ｗで加算する。Ｗは入力されたキーワードの集合、ｔｆ（ｗ，Ｄ）は文書Ｄでのｗの出現回数、ｄｆ（ｗ）は全文書でｗが出現した文書の数、Ｎは文書の総数である。そして、ｓｃｏｒｅ（Ｄ）が高い文書を検索結果として出力する。

However, addition is performed with wεW. W is a set of input keywords, tf (w, D) is the number of occurrences of w in document D, df (w) is the number of documents in which w appears in all documents, and N is the total number of documents. Then, a document with a high score (D) is output as a search result.

As a search process based on scores that are considered to be particularly good, Robertson's 2-Poisson model (Robertson, SE and Walker, S.
(1994). “Some Simple Effective Approximations to the 2-Poisson Model for
Probabilistic Weighted Retrieval. ”In Proceedings of the Seventeenth Annual
International ACM SIGIR Conference on Research and Development in Information
Retrieval.). Robertson
These methods are the score given by the following formula (14) for each article.
Is calculated, and the top score is output as the search result (Score (d) below is the score of article d).

  However, the keyword here is an input keyword. In Expression (14), TF (d, t) is the number of appearances of the keyword t in the article d. DF (t) is the number of articles in which the keyword t appears in all databases. N is an article existing in all databases, and length (d) is the length of the article d (character string unit). Δ is the average length of articles in all databases. In addition, the A part of the equation (14) is referred to as a TF term as a term relating to TF, and the B part is referred to as an IDF term as a term relating to IDF (the reciprocal of DF). Unlike the TF used as a part of the weight in the general vector space method, the TF term in the equation (14) is like the A part.

  When treating Score as an additional point by keyword, if there is a keyword with a large TF value even if the influence of TF is large, a sufficiently large score will be obtained even when the relevance that almost no other keyword exists is low I will take it. This formula helps to prevent this. In this expression, even if TF becomes infinite, it has only a value of 1. For this reason, all keywords will be evaluated uniformly. Furthermore, there is a complicated part of the left side of + in the denominator of the A part, but this is a term for correcting the longer article because the value of TF becomes larger. Furthermore, what added several reinforcement terms to Formula (14) is also considered.

  This method is also called the Okapi weighting method.

  Also, in equation (14), the product of the TF term and the IDF term before taking the product by Σ can be used as the word weight.

Also, create a vector with each word as a dimension and the score of each word as an element, change the article vector to a vector (vector _x) using the words contained in the article, and change the input keyword group vector to The words included in the input keyword group may be used as a vector (vector_y), and the value of the cosine (cos (vector_x, vector_y)) of the vectors may be used as the article similarity. To calculate the score of each word, tf / idf or okapi should be used. The expression after the Σ of those expressions is the expression for calculating the score. The value of the expression is the score for each word.
tf (d, t) * log (N / df (t)) for tf ・ idf (15)
For okapi, tf (d, t) / (tf (d, t) + length (d) / delta)
* log (N / df (t)) (16)
Is the formula.

Further, the following information search can be performed.
(For Okapi, reference 6; SE Robertson, S. Walker, S. Jones, MM Hancock-Beaulieu, and M. Gatford Okapi at TREC-3, TREC-3, 1994)
(For SMART, reference 7; Amit Singhal
AT & T at TREC-6, TREC-6, 1997)
These Okapi and SMART formulas may be used as more advanced information retrieval methods. Furthermore, the equation of the deformation may be used.

  In addition, by using the protein expression as a search keyword and extracting the binary relation of the protein expression to be extracted as described above, the binary relation of the protein expression of the search keyword should be extracted in the text data to be searched. An information search device that outputs only the extracted text data as a search result may be configured. In addition, this is not limited to the binary relationship of protein expression, but targets a binary relationship consisting of a pair of the company's product name and the reputation of the product (for example, information such as good or bad reputation). By using a binary relation element as a search keyword and extracting the binary relation, only the text data extracted that the binary relation of the search keyword should be extracted in the search target text data is output as the search result. An information search apparatus may be configured.

  Furthermore, in the case of English words, the process of returning from the plural form, the past form, and the third person singular present form "s" to the basic form, for example, "studies" is changed to "study" is called "steming" As for the word used for the machine learning feature, the word after the “steming” may be used for the feature.

  Next, an information retrieval apparatus 7 that is a second embodiment of the present invention will be described.

  This information search device 7 performs machine learning on the article using teacher data to which a tag indicating a solution that is either an article to be extracted (positive) or not an article to be extracted (negative) is attached. The processing device extracts articles including two search keywords from the search text data 5 to be searched, and outputs, as search results 6, those articles that are estimated to be articles to be extracted.

  FIG. 11 shows a configuration example of the information search apparatus 7 according to the present invention. The information search device 7 includes an information search unit 70, a solution-feature pair extraction unit 72, a machine learning unit 73, a feature extraction unit 75, a solution determination unit 76, and a search result extraction unit 77. However, as in the first embodiment, the teacher data storage unit 71 and the learning result storage unit 74 may be included in the information search device 7 as internal elements, or can be used from the information search device 7 as external elements. It is good.

  In particular, in this embodiment, machine learning is performed in units of articles, and the “problem” mentioned in the machine learning process described above is replaced with “binary relationship” instead of “search keyword and article pair”. For “solution”, instead of “binary relationship to be extracted or not to be extracted”, “article to be extracted or not to be extracted”, and for “target extraction target” to “binary In place of “relation”, “search keyword and article pair” are each read.

  Based on this, the teacher data storage unit 71, the solution-feature pair extraction unit 72, the machine learning unit 73, the feature extraction unit 75, and the solution determination unit 76 in the information search apparatus 7 of this embodiment are shown in FIG. Processing that performs substantially the same processing as the teacher data storage unit 11, the solution-feature pair extraction unit 12, the machine learning unit 13, the learning result storage unit 14, the feature extraction unit 16, and the solution estimation unit 17 of the term relation extraction device 1. Means. The information search unit 70 in the information search device 7 is a processing unit that performs substantially the same processing as the information search unit 40 in the information search device 4 of the first embodiment. Further, the information search device 7 does not include a processing unit corresponding to the candidate extraction unit 15 or a processing unit corresponding to the binary relation selection unit 45 of the information search device 7 of the first embodiment. However, as in the case of the first embodiment, for the solution determination unit 76, what kind of solution (classification destination) is likely to be obtained in the case of a set of features for each candidate of binary relations? When performing the process of estimating the degree, the same process as the solution estimation unit 17 described above is performed. However, when the solution estimation unit 17 is replaced with a “solution determination unit”, the solution determination unit 76 of the information processing apparatus 7 is performed. Performs the same processing as the “solution determination unit”.

  FIG. 12 shows a processing flow of the information search apparatus 7. In the teacher data storage unit 71 of the information search device 7, the teacher data is either an article that should be extracted (positive) or an article that should not be extracted (negative) as a set of search keywords and articles. Text data including a case to which the “solution” information is added is stored.

First, the solution-feature pair extraction unit 72 extracts, from the teacher data storage unit 71, a case in which the problem is a combination of a search keyword and an article, and an article from which a solution is to be extracted. The information related to the binary relation is extracted as a feature, and a set of these features and a set of solutions are generated (step S21). For example, the following information is extracted as the feature.
1) A word or a character that appears around a binary relation element having the input keyword as an element. For example, a predetermined number of words / characters before the first element (first element) of the binary relation, a predetermined number of words / characters behind the second element (second element), the first element and the second element A predetermined number of words / characters between
2) Appearance position, appearance order, etc. of words / characters appearing around binary-related elements whose elements are input keywords;
3) Two elements of binary relations with the input keyword as an element;
4) Binary element or part-of-speech information, morphological analysis information, etc.
5) parse information of binary relation elements or surrounding words with the input keyword as an element;
6) Appearance distance between the first element and the second element of the binary relation having the input keyword as an element;
7) Presence / absence of an element between the first and second elements of the binary relation having the input keyword as an element;
8) Extract all binary relations that have the input keyword as an element from the article, and check whether these binary relations are meaningful binary relations by either a machine learning method or a pattern-based method. Whether one was determined to be a meaningful binary relationship;
9) When an input keyword appears in an article, words or characters that appear around the keyword;
Of the above examples, the example of the feature 1) -7) is the same as that described in the binary relation extraction apparatus 1 described above.

  Examples of 8) include the following. That is, when the keywords are “Kyoto University” and “Governor”, and the article is “... Kyoto University President ...”, in Article C, the keywords A and B are machine learning or pattern methods. When it is determined that the binary relation is significant, the feature is “significant binary relation”.

  Examples of 9) include the following. In other words, if the keywords are “Kyoto University” and “General Manager” and the article is “… Today, the President of Kyoto University is…”, and if only the immediately preceding and following words are taken as surrounding words, the feature ( Immediately before), there are two reading marks "," and "no", and the features (immediately after) are two reading marks "no" and "ga".

Next, the machine learning unit 73 tends to be any solution (positive or negative) at any feature set from the feature set and solution set generated by the solution-feature pair extraction unit 72. (Or what kind of solution is to be obtained) is learned by the machine learning method, and the learning result is stored in the learning result storage unit 74 (step S22). Examples of learning data stored in the learning result storage unit 74 by the machine learning unit 73 are as follows, for example.
Problem 1: Keywords “Kyoto University” “President”, article “… Kyodai President”
Answer 1: Problem to be extracted 2: Keywords "Kyoto University""President", article "... The President of the University of Tokyo has come to Kyoto University ..."
Answer 2: If you learn that this should not be extracted, it is likely to be extracted when the appearance distance of the first element and the second element of the binary relation having the input keyword as an element is small. Learn what should not be extracted if it is far away.

  After that, the information search unit 70 searches the search text data 5 using the two search keyword pairs given in the AND search process, and acquires the corresponding article (step S23). As described in the first embodiment, the information search unit 70 can adopt a search process other than the AND search process.

  Then, the feature extraction unit 75, in the same manner as the solution-feature pair extraction unit 72, has a binary relation with the search keyword as an element for the set of the article and the search keyword extracted from the text data 5 by the information search unit 70. The related information is extracted as a feature (step S24). In this respect, the processing content of the feature extraction unit 75 is different from the feature extraction unit 16 of the binary relation extraction device 1 and the feature extraction unit 451 of the information search device 4.

Based on the learning data stored in the learning result storage unit 74, the solution determination unit 76 determines what kind of solution should be taken for the combination of the search keyword and the extracted article (step S25). A specific example will be described using the example of learning data described above.
Problem as a new problem: Keywords "Osaka University""President", article "... President of Osaka University ..."
Enter
6) Appearance distance between the first element and the second element of the binary relation having the input keyword as an element;
It is determined that the solution to be extracted should be extracted because the distance is small this time, using learning results such as being easy to be extracted if it is small and not being extracted if it is far. Here, the feature of distance 6) is taken as an example, but learning using features of other features is also possible. For example, when the word between the first element and the second element of the binary relation is “no”, it may be learned that it is likely to be extracted.

  Based on the determination result of the solution determination unit 76, the search result extraction unit 77 extracts, as the search result 6, an article determined to be extracted from the articles searched from the search text data 5 ( Step S26).

  As described above, according to the information search device 7 of the present embodiment, as the teacher data for machine learning processing, information on “solution” indicating whether or not an article to be extracted is included in a set of a search keyword and an article. Text data including given examples Text data Text data is prepared, and by machine learning for each article, articles that are considered to be extracted are automatically extracted from new search text data 5 Is possible.

  In the information search, the information search device 7 of the present embodiment can determine whether there is a semantic relationship between a plurality of given keywords, and can perform an information search using the determination result. For example, in an existing WEB keyword search service, an AND search is performed using “Kyoto University” and “General Manager” as keywords. Then, a page containing two keywords “Kyoto University” and “President” will be searched and come out. However, the obtained pages are not all pages related to the president of Kyoto University. For example, a page describing a sentence “The President of the University of Tokyo attended the Kyoto University ceremony” is not a page related to the president of Kyoto University, but a page related to the president of the University of Tokyo. For this reason, such a page should not be extracted as a result of a search. On the other hand, the information search apparatus 7 of this embodiment uses machine learning to enable each page to determine whether “Kyoto University” and “General Manager” are closely related pages. Can be taken out and presented to the user. That is, more specifically, as an example of an article containing two keywords “Kyoto University” and “President” as the teacher data, an article describing the desired article, that is, “Kyoto University President”, Is created as a negative example, machine learning can be performed to determine whether the article is desired or not, and the result can be used to obtain a search result. Therefore, although many unnecessary search results are included in the conventional search technique, only the useful search results can be presented to the user by using the information search device 7 of this embodiment, and the search performance is high. Improve dramatically.

Next, an information search apparatus 8 that is a reference example of the present invention will be described.

  The information search device 8 allows the user to input information indicating the relationship between two keywords when inputting a plurality of search keywords, or to input the information by a program function. Considering a relationship of two or more search keywords as a meaningful binary relationship, the binary relationship having this search keyword as an element is a relationship to be extracted (positive) or not to be extracted (negative) ) Machine learning using teacher data to which a tag indicating one of the solutions is attached, and an article including two search keywords is extracted from the search text data 5 to be searched, and the search keyword pair is extracted. Among those determined to be binary relations to be performed, the search result 6 is the one in which the search keyword pair is designated as related at the time of input It is a processing device that force.

FIG. 13 shows a configuration example of the information search apparatus 8 according to this reference example . The information search device 8 includes an information search unit 80, a binary relationship determination unit 82, an input unit 84, a candidate extraction unit 85, a binary relationship selection unit 86, and a search result extraction unit 87. Further, in the information search apparatus 8 shown in the figure, the binary relation determination unit 82 further includes a learning unit 820 (including a solution-feature pair extraction unit 821 and a machine learning unit 822), and binary relation selection. The unit 86 further includes a feature extraction unit 861 and a solution determination unit 862. In this case, the information search device 8 can use the teacher data storage unit 81 and the learning result storage unit 83. The teacher data storage unit 81 and the learning result storage unit 83 may be provided as internal elements in the information search device 4. Other configurations of the binomial relationship determination unit 82 will be described later.

  In this information search device 8, the binary relation determination unit 82 {(learning unit 820 (solution-feature pair extraction unit 821 and machine learning unit 822)}) is the same as the learning unit 42 in the information search device 4 of the first embodiment. The information search unit 80, candidate extraction unit 85, binary relation selection unit 86 (including a feature extraction unit 861 and a solution determination unit 862), and a search result extraction unit 87 Processing means for performing the same processing as the information search unit 40, candidate extraction unit 44, binary relation selection unit 45 (including feature extraction unit 451 and solution determination unit 452), and search result extraction unit 46 of the search device 4. The teacher data storage unit 81 and the learning result storage unit 83 are the same as the teacher data storage unit 41 and the learning result storage unit 43 described in the first embodiment.

In the present reference example, the input unit 84 indicates that the input of a plurality of search keywords and the two search keywords indicating that a set of two arbitrary search keywords among the plurality of search keywords has a semantic relationship. It is a processing means which performs the process which receives the input of the relationship information provided to this group. More specifically, the input unit 84 indicates the input of a plurality of search keywords and the two searches indicating that a set of arbitrary two search keywords among the plurality of search keywords has a semantic relationship. An input interface that prompts input of relationship information given to the set of keywords is output, and a plurality of input search keywords and relationship information input to the input interface are received. The input unit 84 will be described with respect to a mode in which an input interface is temporarily output and the user is input relationship information. For example, a search function and relevance information can be automatically input by a program function. Also, it is possible to adopt a mode in which the search keyword is input by the user and the relevance information is input by a program function.

  Corresponding to such an input unit 84, the candidate extraction unit 85 generates a pair of two search keywords that are input with the relationship information given by the input unit 84, and uses the pair as a binary relationship. Candidate for

FIG. 14 shows a processing flow of the information search apparatus 8. As in the case of the first embodiment, the teacher data storage unit 41 of the information search device 4 extracts, as teacher data, a binary relation having two search keywords given by AND search processing as elements. Text data including a case to which information of “solution”, which is either a term relationship (positive) or a binary relationship that should not be extracted (negative), is stored.
First, in the learning unit 820 of the binary relation determination unit 82, the solution-feature pair extraction unit 821 extracts a predetermined feature for each case from the teacher data in the teacher data storage unit 81, and determines the solution (information given by the tag). ) And the extracted feature set (step S31). The solution-feature pair extraction unit 821 extracts a binary relation from text data that is teacher data using a predetermined tag, and uses the morphological analysis process, the syntax analysis process, and the element of the extracted binary relation element (search keyword). A predetermined feature is extracted by performing a process of calculating an appearance position and a distance between elements.

  Similarly, in the learning unit 820, the machine learning unit 822 determines which solution (positive or negative) from the combination of the solution generated by the solution-feature pair extraction unit 821 and the feature set. The learning result is stored by the machine learning method, and the learning result is stored in the learning result storage unit 83 (step S32). The machine learning unit 822 performs machine learning processing using any one of techniques such as a k-nearest neighbor method, a simple Bayes method, a decision list method, a maximum entropy method, a support vector machine method, and the like as a supervised machine learning method. Is the same as in the first embodiment.

  On the other hand, the input unit 84 accepts input of a search keyword and relationship information between any two search keywords (step S33). The candidate extraction unit 85 generates all two combinations (pairs) using the two input search keywords given in the AND search process (step S34). The information search unit 80 performs AND search processing on the search text data 5 using two pairs of input search keywords to extract articles (text data) including the input search keyword pairs, and the candidate extraction unit 44 performs search processing. All the two combinations (pairs) are extracted as binomial relationship candidates using the input search keyword appearing in the article extracted by (step S35).

  Then, in the binary relation selection unit 86, the feature extraction unit 861 performs predetermined processing on each candidate of the binary relation that appears in the searched article by a process substantially similar to the process in the solution-feature pair extraction unit 821. A set of features is extracted (step S36).

  Further, in the binary relation selection unit 86, the solution determination unit 862 determines, for each candidate, what kind of solution is likely to occur in the case of the set of features, that is, the degree of “probably positive” or “prone to be negative” Is estimated based on the learning result of the learning result storage unit 83 (step S37). Then, the search result extraction unit 87 selects, as a binary relation to be extracted from candidates of the binary relation that are estimated to be “positive” to a better degree than a predetermined degree, and includes this binary relation. An article or information specifying the article is selected as the search result 6 (step S38).

  Here, an example of an input interface will be described in the case where three search keywords are input as a plurality of search keywords. FIG. 15A shows an input interface in which a keyword input window is arranged on the circumference. In the illustrated example, it is assumed that the words “Kyoto University”, “General Manager”, and “Language Processing” are input to three keyword input windows, respectively. In this input interface, any two keyword input windows can be connected with a line as relevance information indicating two keywords that the user wants to have a semantic relationship. In the illustrated example, the keyword input windows of “Kyoto University” and “Governor” are connected by a line representing the relationship information. In other words, the user wishes to search on the condition that “Kyoto University” and “General Manager” are related and “Language Processing” may not be so related. ing. If it is desired to use four or more search keywords, it is preferable to increase the keyword input windows on the circumference by appropriate user operations. Further, the “circle” may be a perfect circle as shown in the figure, or may be an ellipse.

  FIG. 15B shows an input interface in which one keyword input window is shown. In the keyword input window, keywords input by the user are separated by “,”. Then, when it is desired to have a semantic relationship between keywords as the relationship information, the user inputs rel (A, B) as shown in the example (rel is a function meaning “relation”). is there). As shown in the example, when “C, rel (A, B)” is entered in the keyword input window, the user can use the keyword “A” and “B” on the condition that the keywords “A” and “B” are related. “C” means that the user wishes to search under the condition that the relationship may not be so much. The same is true for two or more keywords.

  FIG. 15C shows an example of a tabular input interface in which rows are formed in the horizontal direction and columns are formed in the vertical direction. In this input interface, each column of the first row is a keyword input window, and a keyword to be given a semantic relationship with the keyword of each column input to the first row is input below the second row. To do. In the illustrated example, the keywords “A”, “B”, and “C” are input to the keyword input window in each column of the first row, the keyword “B” is input in the second row and first column, and the second row and second column. The case where the keyword “A” is input is shown. In this example, the user wishes to search on the condition that the keywords “A” and “B” are related and the keyword “C” may not be so related. I mean. However, keyword input in the second row and below is automatically input if the keyword in any column in the first row is designated by the input device, or other than that column is designated in each column after the second row. By taking measures such as automatically popping up keywords, it is possible to reduce the time and effort of manual input by the user. The same applies to two or four or more keywords.

  FIG. 15D is another example of a tabular input interface. In this example, each column at the top of the table is a keyword input window, and all the keywords are displayed in the top and bottom in the bottom of each column at the top, and there is a relationship with the keyword in that column next to the keyword. For example, a radio button (instead of this, a check box or the like) is arranged as the relationship information indicating this. It should be noted that radio buttons cannot be selected for the same keyword as that column. The illustrated example shows a state where the user has turned on the radio button of the keyword “B” for the keyword “A”. In this case, if the radio button of the keyword “A” for the keyword “B” is automatically turned ON in conjunction with the ON input of the radio button, the user's input is saved. In this example, the user wishes to search on the condition that the keywords “A” and “B” are related and the keyword “C” may not be so related. I mean. The same applies to two or four or more keywords.

Further, as shown in FIG. 16 , the information search device 8 ′, which is another example of the reference example, includes a pattern storage unit 820 ′ instead of the configuration of the binary relation determination unit 82 including the learning unit 820 described above. A binomial relationship determination unit 82 'can also be employed. In this case, the binary relation selection unit 86 ′ has a processing function different from that of the binary relation selection unit 86 described above. Further, the pattern storage unit 820 ′ and the binary relation selection unit 86 (particularly the solution determination unit 862) use a pattern storage unit 83 ′ instead of the learning result storage unit 83 shown in FIG. Further, in the modification of the information search device 8, the teacher data storage unit 81 is not used.

  In this case, the pattern storage unit 820 'stores in advance in the pattern storage unit 83' a keyword and a combination of keywords to be extracted as a binary relationship as a pattern. When the binary relation selection unit 86 ′ determines that the binary relation candidate extracted by the candidate extraction unit 85 matches the pattern in the pattern storage unit 83 ′ and is to be extracted, The binary relation candidate is selected as a binary relation to be extracted. The selected binary relation is extracted by the search result extraction unit 87.

Also, when specifying the relationship between keywords at the time of input, if there are three or more keywords, all binary relationships should be extracted if they become binary relationships that should be extracted at least in one place in the text data You may change the process in the said reference example so that it may take as text data.

As described above, the information search devices 8 and 8 ′ of the present reference example have which keywords must be related by the user (or by a preset program), and which keywords are not related. It is possible to specify whether or not the information can be specified, and a more flexible information search can be realized. As a specific example, there has conventionally been a method of determining whether or not there is a semantic relationship between a plurality of given keywords in information retrieval and performing information retrieval using the determination result. For example, in an existing WEB keyword search service, an AND search is performed using “Kyoto University” and “General Manager” as keywords. Then, a page containing two keywords “Kyoto University” and “President” will be searched and come out. However, the obtained pages are not all pages related to the president of Kyoto University. For example, a page describing a sentence “The President of the University of Tokyo attended the Kyoto University ceremony” is not a page related to the president of Kyoto University, but a page related to the president of the University of Tokyo. For this reason, such a page is not desired for the user as a result of the search. On the other hand, in this reference example, a search is made by designating at the time of input so that only the case where “Kyoto University” and “total length” are related is taken as a search result. In this method, it is considered that a keyword “language processing” is specified and searched in addition to these two keywords. In this case, it is assumed that “language processing” and other keywords need not be so closely related. In this case, the user inputs the keywords “Kyoto University”, “General Manager”, and “Language Processing”, but the other is not so much, provided that “Kyoto University” and “General Manager” are related. Search on the condition that it is good. In this way, the information search devices 8 and 8 ' of the present reference example can specify the condition of the relationship between the keywords when inputting the keywords.

Next, the information search device 9 of the third embodiment of the present invention will be described.

The information search device 9 includes an output unit 91 and a re-output unit 92 that can be applied to the information search devices 4 , 7 , 8, 8 ′ of the first and second embodiments and the reference example described above. FIG. 17 shows an example of the configuration (the processing means upstream from the search result extraction unit is omitted), and FIG. 18 shows the flow of processing.

  The output unit 91 is a processing unit that outputs the search results extracted by the search result extraction units 46, 77, and 87 in the information search devices 4, 7, and 8. Specifically, for example, the search result is displayed and output on a predetermined screen of the display. At the time of the output, the output unit 91 also outputs information indicating the presence or absence of the binary relationship between the search keywords for each search result (step S41). FIG. 19 shows an example of a search result output screen displayed and output by the output unit 91. In this example, the search result is displayed as “1, 2, 3, 4, 5,. For example, “AB”, “BC”, and “CA” displayed on the upper side of each search result are “two” for all two of the input keywords “A”, “B”, and “C”. This is a designation button indicating that “A and B are related”, “B and C are related”, and “B and C are related”. A search result with a mark (O in the illustrated example) beside each search result means that it corresponds to one of “AB”, “BC”, and “CA”. Instead of or in addition to such a display output format, a button or the like indicating that there is no relationship between the two keywords may be displayed and output so that the user can select it.

  The re-output unit 92 receives the designation input of the presence / absence of the relationship between keywords by the user as described above, extracts the search result corresponding to the input from the search result output by the output unit 91, and outputs it again Means. In other words, the re-output unit 92, according to the display example described above, allows the user to input a designation input of one or more of the designation buttons “AB”, “BC”, and “CA” using the input device. Receiving (step S42), a search result corresponding to the presence / absence of the specified binary relation is extracted and re-output (step S43). In addition, only the search result may be displayed on the screen displayed and output by the re-output unit 92, and the presence / absence of the relationship between the search result and the search keyword is expressed as in the screen example shown in FIG. Both information may be displayed. In the latter case, the search results can be further narrowed down.

  In addition, when the function of the re-output unit 92 as described above is not required, an information search device 9 that includes only the output unit 91 in addition to the information search devices 4, 7, and 8 is also realized. Can do.

  As described above, with the information search device 9 of the present embodiment, the user can specify whether or not there is a relationship between search keywords when outputting information search. For example, an AND search is performed with the three keywords “Kyoto University”, “General Manager”, and “Language Processing”, and a button indicating the relationship of each keyword is displayed together with the search result in the search result, and the user himself / herself relates to each keyword. Specify whether or not you want to have a button, check whether there is a relationship between each keyword based on the specified result, perform a search based on the specified result, and return the result to the user By presenting, it is possible to provide a search result that has been narrowed down beneficially for the user.

  In addition, this invention is not limited to embodiment mentioned above.

For example, in the information search apparatus 8 described in the reference example, a configuration in which the input unit 84 related to the relationship information between keywords is excluded can be considered. In this case, the information retrieval apparatus uses the predetermined information relating to the binary relation element based on the supervised machine learning process or the processing method using a predetermined rule. A binary relation determination unit that determines whether or not a term relation exists and an input search keyword pair using a plurality of input search keywords are generated, and the input search keyword is generated from text data to be searched by a predetermined method. An information search unit that extracts and acquires text data based on the input search keyword, and two input search keywords to which the relationship information received by the input unit from each text data searched and acquired by the information search unit is added And a candidate extraction unit that uses the generated pair as a binary relation candidate and a binary relation determination unit, When it is determined that the binary relationship to be extracted is determined using predetermined information related to the binary relationship element that is a candidate for the binary relationship, A binary relation selection unit for selecting the binary relation as a binary relation to be extracted, and a search result extraction unit for extracting text data including the binary relation selected by the binary relation determination unit as a search result; It has a configuration comprising “

Further, since there are various processing methods for determining the text data to be extracted, as a concept information search apparatus including the concept of the information search apparatus of the first and third embodiments, “ Based on a predetermined method, a text data determination unit that determines whether the text data is a binary relationship to be extracted by using predetermined information related to a binary relationship element included in the text data is input An information search unit that generates a pair of input search keywords using a plurality of search keywords, extracts text data based on the input search keywords by a predetermined method from text data to be searched, and the information Two input search keywords to which the relationship information received by the input unit from each text data retrieved by the search unit is added Based on the candidate extraction unit that generates the configured pair and uses the generated pair as a binary relation candidate, and the text data determination unit, the binary relation candidate becomes the binary relation candidate. Using the predetermined information related to the binary relation element, it is determined whether or not the binary relation is to be extracted. If it is determined that the binary relation is to be extracted, the binary relation candidate is selected. Information including a binary relation selection unit that is selected as a binary relation to be extracted and a search result extraction unit that extracts text data including the binary relation selected by the binary relation determination unit as a search result " A processing device is also conceivable.

  In addition, in the processing referred to above, regarding the processing of “taking a larger value”, for example, “processing that takes a value that is greater than or equal to a threshold” or “processing that has a larger value than the number of predetermined values” The processing can be replaced with “taken in descending order” processing or “take a value obtained by multiplying the maximum value of the extracted values by a predetermined ratio and take a value that is equal to or greater than the obtained value”. Further, these threshold values and predetermined values can be determined in advance, and the user can change and set the values as appropriate.

  In addition, the specific configuration of each part is not limited to the above embodiment, and various modifications can be made without departing from the spirit of the present invention.

It is a figure which shows the structural example of the binary relation extraction apparatus utilized by this invention. It is a figure which shows the flow of a process of a binary relation extraction apparatus. It is a figure which shows the example of teacher data. It is a figure which shows the concept of margin maximization of a support vector machine method. It is a figure which shows the example of a set with the set of the features of a binary relation shown in FIG. It is a figure which shows the structural example of the information search device which concerns on 1st Example of this invention. It is a figure which shows the flow of a process of the information search device. It is a figure which shows the example of the group with the set of teacher data and the feature of the binary relation in the Example. It is a figure which shows the example of the group with the set of teacher data and the feature of the binary relation in the Example. It is a figure which shows the example of a group with teacher data and the set of the feature of the binary relation. It is a figure which shows the structural example of the information search device which concerns on 2nd Example of this invention. It is a figure which shows the flow of a process of the information search device. It is a figure which shows the structural example of the information search device which concerns on the reference example of this invention. It is a figure which shows the flow of a process of the information search device. It is a figure which shows the example of the input interface which the same information search device outputs. It is a figure which shows the other structural example of the information search device. It is a figure which shows the example of a partial structure of the information search device based on 3rd Example of this invention. It is a figure which shows the flow of a process of the information search device. It is a figure which shows the example of the search result output screen which the same information search device outputs.

Explanation of symbols

DESCRIPTION OF SYMBOLS 1 ... Binary relation extraction apparatus 11 ... Teacher data storage part 12 ... Solution-feature pair extraction part 13 ... Machine learning part 14 ... Learning result storage part 15 ... Candidate extraction part 16 ... Feature extraction part 17 ... Solution estimation part 18 ... Two Term relation extraction unit 2 ... text data 3 ... binary relation 4 ... information search device 40 ... information search unit 41 ... teacher data storage unit 42 ... learning unit 421 ... solution-feature pair extraction unit 422 ... machine learning unit 43 ... learning result Storage unit 44 ... Candidate extraction unit 45 ... Binary relation selection unit 451 ... Feature extraction unit 452 ... Solution determination unit 46 ... Search result extraction unit 5 ... Search text data 6 ... Search result 7 ... Information search device 70 ... Information search unit 71 ... Teacher data storage unit 72 ... Solution-feature pair extraction unit 73 ... Machine learning unit 74 ... Learning result storage unit 75 ... Feature extraction unit 76 ... Solution determination unit 77 ... Search result extraction unit 8, 8 '... Information search device 81 ... teacher data storage units 82, 82 ' Binary relationship determination unit 820 ... learning unit 821 ... solution-feature pair extraction unit 822 ... machine learning unit 820 '... pattern storage unit 83 ... learning result storage unit 83' ... pattern storage unit 84 ... input unit 85 ... candidate extraction unit 86 , 86 '... binary relation selection unit 861 ... feature extraction unit 862 ... solution determination unit 87 ... search result extraction unit 9 ... information search device 91 ... output unit 92 ... re-output unit

Claims (6)

An information search device that performs an information search process using a plurality of search keywords,
Teacher data in which teacher data is stored, including cases that consist of pairs of problems and solutions, where the problem is a binary relationship with the search keyword as an element and the solution is a binary relationship to be extracted Taking out the case from the storage unit, extracting predetermined information as a feature for each case, generating a set of the extracted feature set and solution, and generating the set based on a predetermined machine learning algorithm For a set of features and a solution, machine learning processing is performed to determine what kind of feature set the solution is, and indicates what type of feature set the solution is A learning unit that stores information in the learning result storage unit as learning result information;
An information search unit that generates a pair of input search keywords using a plurality of input search keywords, and extracts and acquires text data based on the input search keywords from a text data to be searched by a predetermined method; ,
A candidate extraction unit that generates a pair composed of the input search keyword from each text data searched and acquired in the information search unit, and uses the generated pair as a binary relation candidate;
The extraction process similar to the extraction process performed in the learning unit extracts the predetermined information as a feature for the binomial relationship candidates, and based on the learning result information stored in the learning result storage unit, In the case of the set of features of the binomial relationship candidates, it is determined what solution should be taken, and the determination result should be a solution indicating that the binomial relationship is to be extracted for the binomial relationship candidate. A binary relation selection unit that, when determined, selects the binary relation candidate to be extracted as a binary relation;
A search result extraction unit that extracts text data including the binary relationship selected by the binary relationship selection unit as a search result;
An information retrieval apparatus comprising: An output unit for outputting the search result extracted by the search result extraction unit;
The information search device according to claim 1, wherein the output unit also outputs information indicating the presence or absence of a binary relationship between the input search keywords for each search result. Regarding the information indicating the presence or absence of the binary relationship output by the output unit, the search result corresponding to the received information is accepted by accepting any designation of information indicating the presence or absence of the binary relationship. The information search device according to claim 2, further comprising a re-output unit that re-outputs only the information. An information search device that performs an information search process using a plurality of search keywords,
From a teacher data storage unit in which teacher data is stored that includes examples of a combination of a problem and a solution, where the problem is a combination of a search keyword and text data and the solution is text data to be extracted A solution-feature pair extraction unit that takes out the case, extracts information related to a binary relationship having at least a search keyword as an element for each case, and generates a set of the extracted feature set and solution; ,
Based on a predetermined machine learning algorithm, with respect to the set of features and solution, machine learning processing is performed to determine what kind of feature set results in the solution. A machine learning unit that stores information indicating whether the solution is obtained as learning result information in a learning result storage unit;
An information search unit that generates a pair of input search keywords using a plurality of input search keywords, and extracts and acquires text data based on the input search keywords from a text data to be searched by a predetermined method; ,
Feature extraction that extracts information related to a binary relation having at least the search keyword as an element for a set of the search keyword and the extracted text data by an extraction process similar to the extraction process performed by the solution-feature pair extraction unit And
Based on the learning result information stored in the learning result storage unit, a solution determination unit that determines what solution should be taken in the case of a set of features of the set of the search keyword and the extracted text data;
As a result of the determination, when it is determined that the solution is text data to be extracted for a set of the search keyword and extracted text data, the text data is selected as an article to be extracted, and the selected A search result extraction unit that extracts text data as a search result;
An information retrieval apparatus comprising: A program for causing a computer to perform information search processing using a plurality of search keywords,
Teacher data in which teacher data is stored, including cases that consist of pairs of problems and solutions, where the problem is a binary relationship with the search keyword as an element and the solution is a binary relationship to be extracted Taking out the case from the storage unit, extracting predetermined information as a feature for each case, generating a set of the extracted feature set and solution, and generating the set based on a predetermined machine learning algorithm For a set of features and a solution, machine learning processing is performed to determine what kind of feature set the solution is, and indicates what type of feature set the solution is A learning step of storing information in the learning result storage unit as learning result information;
An information search step of generating a pair of input search keywords using a plurality of input search keywords, and extracting and acquiring text data from the text data to be searched based on the input search keywords by a predetermined method; ,
A candidate extraction step of generating a pair composed of the input search keyword from each text data searched and acquired in the information search step, and using the generated pair as a binary relation candidate;
The extraction process similar to the extraction process performed in the learning step extracts the predetermined information as a feature for the binomial relationship candidates, and based on the learning result information stored in the learning result storage unit, In the case of the set of features of the binomial relationship candidates, it is determined what solution should be taken, and the determination result should be a solution indicating that the binomial relationship is to be extracted for the binomial relationship candidate. A binary relation selection step of selecting, as a binary relation, to be extracted as a binary relation to be extracted;
A search result extraction step of extracting text data including the binary relationship selected in the binary relationship selection step as a search result;
An information retrieval program characterized by causing A program for causing a computer to perform information search processing using a plurality of search keywords,
From a teacher data storage unit in which teacher data is stored that includes examples of a combination of a problem and a solution, where the problem is a combination of a search keyword and text data and the solution is text data to be extracted A solution-feature pair extraction step of taking out the case, extracting information related to a binary relation having at least a search keyword as an element for each case as a feature, and generating a set of the extracted feature set and solution; ,
Based on a predetermined machine learning algorithm, with respect to the set of features and solution, machine learning processing is performed to determine what kind of feature set results in the solution. A machine learning step of storing information indicating whether the solution is to be stored in the learning result storage unit as learning result information;
An information search step of generating a pair of input search keywords using a plurality of input search keywords, and extracting and acquiring text data from the text data to be searched based on the input search keywords by a predetermined method; ,
Feature extraction that extracts information related to a binary relation having at least the search keyword as an element for a set of the search keyword and the extracted text data by an extraction process similar to the extraction process performed by the solution-feature pair extraction step Process,
Based on the learning result information stored in the learning result storage unit, a solution determination step for determining what solution should be taken in the case of a set of features of the set of the search keyword and the extracted text data;
As a result of the determination, when it is determined that the solution is text data to be extracted for a set of the search keyword and extracted text data, the text data is selected as an article to be extracted, and the selected A search result extraction process for extracting text data as a search result;
An information retrieval program characterized by causing

Images