JP6502807B2 - Information extraction apparatus, information extraction method and information extraction program - Google Patents

Description

  Embodiments relate to the extraction of information.

  For example, a technique for extracting relationship information between values of a plurality of attributes described in a document such as a web page (for example, relationship information between a material name and its characteristic value, relationship information between a product name and its price, etc.) It has been demanded. By using such a technology, it is possible to easily organize desired information from the huge amount of information described in the text. For example, it is possible to extract relationship information between a product name and its specification from a sentence, and to list the list in a short time in a table.

  However, the extracted relationship information may have ambiguity. For example, when extracting the relationship information of a material name and its mobility from a paper reporting the result of measuring the mobility of the material under different experimental conditions, a plurality of relationship information having the same material name but different mobility May be extracted. Alternatively, in a product such as a PC (Personal Computer) or an automobile, storage capacity such as an HDD (Hard Disk Drive) or an SSD (Solid State Drive), color, or exhaust for the same product name (brand name) Several models with different quantities may be prepared. Therefore, for example, when relation information between a product name and its price is extracted from a document in which the price of a PC is described, there is a possibility that a plurality of relation information items having the same product name (brand name) but different prices may be extracted. is there. On the other hand, even if a value corresponding to a model name is extracted as a product name, it is difficult to specify which brand the relation information extracted represents a relation between a model and a price.

JP, 2010-117797, A JP 2001-060199

  Embodiments aim to reduce the ambiguity of relationship information between attribute values extracted from a document.

  According to the embodiment, the information extraction device includes a related information extraction unit and a complementary information extraction unit. The relationship information extraction unit obtains a relationship information group by extracting from the document at least one relationship information indicating a relationship between the value of the first attribute and the value of the second attribute. The complementary information extraction unit, when it is determined according to the determination criteria that the first relation information belonging to the relation information group has ambiguity, the value of the first attribute forming the first relation information and the second information Extract complementary information related to at least one of the attribute values.

FIG. 1 is a block diagram illustrating an information extraction apparatus according to a first embodiment. FIG. 3 illustrates a document received by the input unit of FIG. 1; The flowchart which illustrates a series of relation information extraction processing performed by the relation information extraction part of Drawing 1, a learning part, and a calculation part. FIG. 7 is a view exemplifying candidate values of attributes extracted by the relationship information extraction unit, the learning unit, and the calculation unit of FIG. 1; FIG. 7 is a view exemplifying candidate values of attributes extracted by the relationship information extraction unit, the learning unit, and the calculation unit of FIG. 1; The figure which illustrates the related information group extracted by the related information extraction part of FIG. 1, a learning part, and a calculation part. FIG. 7 is a diagram exemplifying complementary information extracted by the complementary information extraction unit of FIG. 1; The block diagram which illustrates the information extraction device concerning a 2nd embodiment. FIG. 9 is a view exemplifying a related information evaluation UI realized by the output unit and the evaluation information input unit of FIG. 8; 5 is a flowchart illustrating the operation of the complementary information extraction unit of FIG. 1;

  Hereinafter, the description of the embodiments will be described with reference to the drawings. In the following, elements that are the same as or similar to the elements described above will be assigned the same or similar reference numerals, and overlapping descriptions will be basically omitted.

First Embodiment
An information extraction device 100 according to the first embodiment is illustrated in FIG. The information extraction apparatus 100 may be, for example, a terminal capable of inputting a document, or may provide an application or service for such a terminal.

  The information extraction apparatus 100 extracts relationship information between values of a plurality of items (hereinafter referred to as attributes) from a sentence. The information extraction apparatus 100 may specify a plurality of attributes to be extracted, or may specify an analysis purpose and determine a plurality of attributes to be extracted based on the analysis purpose.

  Documents are natural language-based text data such as Web pages and news articles. However, any type of data that can extract an element (for example, a word) that is a candidate for the value of the attribute based on its feature is available as a document.

  As illustrated in FIG. 1, the information extraction apparatus 100 outputs the input unit 101, the relationship information extraction unit 102, the learning unit 103, the calculation unit 104, the determination unit 105, the complementary information extraction unit 106, and And a unit 107.

  The input unit 101 receives a document for which extraction of relationship information is performed. The input unit 101 may acquire a document via a recording medium or a network, or may directly input the document according to the user's operation. As an example of the document, an article on the price of a new PC model is shown in FIG. The input unit 101 outputs the document to the relationship information extraction unit 102.

  The relationship information extraction unit 102, the learning unit 103, and the calculation unit 104 perform a series of relationship information extraction processes exemplified in FIG. The relation information extraction unit 102 performs the processes of step S201, step S202, step S203, and step S204, the learning unit 103 performs the process of step S205, and the calculation unit 104 performs the process of step S206. These processes can be realized using, for example, machine learning.

  The relationship information extraction unit 102 receives a document from the input unit 101. In summary, the relationship information extraction unit 102 extracts candidates of relationship information between values of a plurality of attributes described in a document. The relationship information extraction unit 102 outputs the extracted relationship information candidates to the learning unit 103. Furthermore, as will be described later, the relationship information extraction unit 102 outputs the feature (also referred to as a feature amount) of the relationship information candidate and the feature of the plurality of attribute value candidates to the learning unit 103.

  Referring to FIG. 3, first, the relationship information extraction unit 102 performs text analysis as preprocessing on the document (step S201). The text analysis may be, for example, morphological analysis, specific expression extraction, syntax analysis, and the like.

  Next, the relation information extraction unit 102 extracts, from the document, candidates (typically, words such as nouns and numerical values) of respective values of a plurality of attributes to be extracted (for example, “product name” and “price”). (Step S202 and step S203).

  Specifically, the relation information extraction unit 102 searches the document for a name or proper noun registered in advance in the dictionary, or applies a pattern matching rule for an expression such as “○○ version” to the document. By doing so, candidates for the value of the attribute “product name” may be extracted. On the other hand, the relation information extraction unit 102 extracts the candidate of the value of the attribute “price” by applying to the document a pattern matching rule that targets expressions such as “(number) + (currency unit)”. Good.

  For example, the relation information extraction unit 102 can extract “Dbook”, “500 GB version”, “1 TB version”, “LNote” and the like as candidates for the value of the attribute “product name” from the document of FIG. On the other hand, the relation information extraction unit 102 can extract "50,000 yen", "100,000 yen", "30,000 yen", etc. as candidates for the value of the attribute "price" from the document of FIG. Furthermore, the relationship information extraction unit 102 derives the extracted feature value candidate candidate and outputs the feature to the learning unit 103.

  Next, the relationship information extraction unit 102 obtains a relationship information candidate by combining the extracted value candidates of the plurality of attributes (step S204). For example, the related information extraction unit 102 may set related information such as (“Dbook”, “50,000 yen”), (“Dbook”, “100,000 yen”), (“500 GB version”, “50,000 yen”), etc. You can get a candidate. Furthermore, the relation information extraction unit 102 derives the feature of the extracted relation information candidate (for example, a word appearing between the candidate of the value of “product name” and the candidate of the value of “price” in the document) Output to the learning unit 103.

  Here, the feature means a clue for characterizing an attribute value candidate or a relation information candidate. Specifically, the part of speech or meaning of candidates (words) of values of a plurality of attributes forming a candidate of relation information, the appearance positions of candidates of values of each attribute in a document, or the candidates of values of an attribute in a document The number of words appearing with the candidate of the value of another attribute can be used as a feature.

  In the example of FIG. 3, the relationship information extraction unit 102 extracts candidates for values of a plurality of attributes, and obtains candidates for relationship information by combining them. However, the relationship information extraction unit 102 may directly extract the relationship information candidate representing the relationship between the values of the plurality of attributes without passing through the step of extracting the plurality of attribute value candidates.

  The learning unit 103 receives, from the relationship information extraction unit 102, candidates for relationship information and their features, and features of candidates for a plurality of attribute values. The learning unit 103 learns the feature weight based on the feature of the relation information candidate and the feature of the plurality of attribute value candidates (step S205). Specifically, the learning unit 103 learns the importance (weight) of each feature based on the correct case (positive example) and the incorrect case (negative example) given in advance. The learning unit 103 outputs the relation information candidate and the (learned) feature weight to the calculation unit 104.

  The learning unit 103 can use any machine learning method or other similar technology that uses candidate of relation information, candidate of values of multiple attributes, correct case and incorrect case. Specifically, technologies such as Markov Logic Network, Support Vector Machine, Conditional Random Field, etc. can be used.

  In the example of FIG. 3, although the learning is sequentially performed by the learning unit 103, such sequential learning may be performed by using, for example, previously learned weights or weights set manually or by other techniques. It can be omitted.

  The calculating unit 104 receives the relation information candidate and the feature weight from the learning unit 103. The calculation unit 104 calculates (estimates) a score representing the certainty for each candidate of the relation information as illustrated in FIG. 6 based on the weight of the feature, and associates the score with the candidate of the relation information (step S206) ), To obtain a related information group including at least one related information. Each piece of related information may be assigned an ID for identification. The score is typically a probability value, but may be any indicator that represents the likelihood of an outcome. Note that the calculation unit 104 (or the determination unit 105 described later) may perform filtering processing to exclude a candidate whose score is less than the threshold from the related information group. According to such filtering processing, it is possible to exclude a candidate highly likely to represent an incorrect relationship. The calculation unit 104 outputs the relation information group to the determination unit 105.

  In order to use the calculation unit 104 to calculate the score representing the likelihood of the candidate of the relation information, as exemplified in FIGS. 4 and 5, the candidate of the values of the plurality of attributes forming the candidate of the relation information is A score representing certainty (e.g., product name, price, etc.) may be calculated.

  The determination unit 105 receives the related information group from the calculation unit 104. The determination unit 105 determines, for example, whether or not the relationship information belonging to the relationship information group has ambiguity (in other words, whether the relationship represented by the relationship information is clear or not), for example, a first determination criterion described later and The determination is made according to various determination criteria including two determination criteria. The determination unit 105 passes the relationship information determined to have ambiguity to the complementary information extraction unit 106, and passes the other relationship information to the output unit 107.

  For example, in the relation information group of FIG. 6, product names are the same (“Dbook”) and prices are different (“50,000 yen”, “100,000 yen”) plural pieces of relation information (relation information ID = 1, 2) is included. Although all of these pieces of related information have high scores, it can be estimated that they are not erroneous, but the contents of both are contradictory at first glance. Therefore, even if the related information is presented, the user (can also be called an analyst) can not judge whether the price of "Dbook" is "50,000 yen" or "100,000 yen" .

  Then, the candidate of the value of the first attribute (for example, "product name") forming any first relation information is the same as the candidate of the value of the first attribute forming another second relation information. And the candidate of the value of the second attribute (for example, "price") forming the first relation information is different from the candidate of the value of the second attribute forming the second relation information. The determining unit 105 may determine that the first relation information and the second relation information are ambiguous (first judgment criterion).

  On the other hand, the relationship information group in FIG. 6 includes relationship information (relationship information ID = 4) whose product name is a general noun (“1 TB version”). Although the related information is also high in score, it can be estimated that it is not an error, but the product name forming the related information does not indicate a specific product ("DBook" or "Lnote"). Therefore, even if the related information is presented, the user can not judge what "1TB version" of the product is "100,000 yen".

  Therefore, when the candidate of the value of the first attribute included in any relationship information corresponds to a specific type of word (for example, a general noun), the determination unit 105 determines that the relationship information is ambiguous. (The second judgment standard).

  The complementary information extraction unit 106 extracts complementary information for reducing ambiguity with respect to the related information notified from the determination unit 105, and complements the related information using the complementary information, and then the output unit 107 is performed. hand over. The complementary information relates to at least one of the plurality of attribute value candidates forming the relation information.

  Specifically, the complementary information extraction unit 106 analyzes or corrects a sentence such as dependency analysis, associative analysis, or paraphrase for at least one of the plurality of attribute value candidates forming the relationship information. , Complementary information may be extracted. Alternatively, the complementary information extraction unit 106 complements, for at least one of the plurality of attribute value candidates forming the relation information, a word corresponding to the upper concept or lower concept of the candidate or a word representing the details of the candidate It may be extracted as information.

  The complementary information extraction unit 106 may extract complementary information as shown in FIG. 10, for example. The complementary information extraction process of FIG. 10 is performed when there is a plurality of pieces of relationship information in which the values of the first attribute are the same and the values of the second attribute are different (step S401). For example, the complementary information extraction process of FIG. 10 is performed on the relationship information of the relationship information ID = 1, 2 in FIG. 6 and the relationship information of the relationship information ID = 3, 5.

  The complementary information extraction unit 106 searches the value of the first attribute related to the value of the second attribute for each of the values of the different second attribute as a candidate of the complementary information (step S402). For example, with respect to the relationship information of relationship information ID = 1, 2 in FIG. 6, the complementary information extraction unit 106 extracts the “500 GB version” as a candidate of the “product name” value related to “50,000 yen” The "1TB version" is extracted as a candidate for the "product name" value related to 100,000 yen. The extracted “500 GB version” and “1 TB version” information may complement the related information of the related information ID = 1, 2 to enable clearer distinction between the two.

  Note that the complementary information extraction unit 106 adds the total number of relational information including the value of the second attribute of interest, the type of candidate complementary information word, the score assigned to the relational information including the candidate for complementary information, A part of the total number of attributes with common values among multiple pieces of relationship information, the semantic upper / lower relationship between attribute values of the first The candidates for the complementary information to be searched may be narrowed down based on the whole. In addition, the complementary information extraction unit 106 may perform dependency analysis, anaphoric analysis, or paraphrase on the value of the first attribute included in each piece of relationship information to extract a candidate for complementary information.

  Next, the complementary information extraction unit 106 classifies the candidate of the complementary information found in step S402 according to the linguistic feature (step S403). Specifically, the complementary information extraction unit 106 may classify the candidate of the complementary information according to the part of speech, or according to the pattern matching rule used by the relationship information extraction unit 102 for extracting the candidate of the complementary information. The candidates for the complementary information may be classified. For example, "500 GB version" and "1 TB version" both have a part of speech of "[number] + [unit] + [noun]", and both have a pattern for the expression "○ version" It may have been extracted using matching rules. Therefore, the complementary information extraction unit 106 can classify these into the same candidate group.

  Next, the complementary information extraction unit 106 selects one of the candidate group obtained in step S403 with the largest number of elements, and handles each candidate belonging to the candidate group as complementary information (step S404). That is, in the example of FIG. 10, the complementary information can be regarded as the value of the third attribute separated from the first attribute. For example, as shown in FIG. 7, the complementary information extraction unit 106 performs complementary information of “500 GB version” related to “50,000 yen” with respect to related information including “DBook” and “50,000 yen”. Add or add complementary information “1 TB version” related to “100,000 yen” to related information including “DBook” and “100,000 yen”.

  The above-described determination unit 105 may make another determination after the completion of the complementary information extraction process of FIG. 10, and the complementary information extraction process may be repeated as necessary. In addition, in the case where relationship information having ambiguity still remains, the learning unit 103 may perform learning using the relationship information having no ambiguity as a correct case.

  The complementary information may not be able to completely remove the ambiguity of the related information, and may be, for example, as long as it contributes to the distinction between a plurality of related information. For example, the complementation information extraction unit 106 calculates how much a percentage of a plurality of pieces of relation information including the value of the same attribute can be distinguished by adding a candidate of certain complementation information, If the ratio exceeds the threshold, the candidate of the complementary information may be adopted.

  The output unit 107 receives related information (which may include complementary information) from the determination unit 105 or the complementary information extraction unit 106. The output unit 107 presents the relationship information to the user. For example, the output unit 107 may display the related information in a tabular form on the display, or may read out the content of the related information from a speaker using a voice synthesis technology. In addition to the relationship information, the output unit 107 may further present information (e.g., quotation of the description of the vicinity of the occurrence location) regarding the appearance location in the document of the value of each attribute forming the relationship information. .

  As described above, the information extraction apparatus according to the first embodiment extracts, from the document, relation information representing the relation between values of a plurality of attributes, and if the relation information has ambiguity, the relation information is extracted. Further, complementary information related to the value of any attribute to be formed is further extracted, and the relevant information is complemented using the complementary information. Therefore, according to this information extraction apparatus, it is possible to reduce or eliminate the ambiguity of the related information, and to present the information contributing to the analysis of the document (for example, the unambiguous related information and the judgment material of its validity) .

Second Embodiment
As illustrated in FIG. 8, the information extraction apparatus 300 according to the second embodiment includes the input unit 101, the relationship information extraction unit 102, the learning unit 103, the calculation unit 104, the determination unit 105, and the complementation. An information extraction unit 106, an output unit 107, an evaluation information input unit 308, and an evaluation result storage unit 309 are included. That is, the information extraction device 300 corresponds to the information extraction device 100 of FIG. 1 to which the evaluation information input unit 308 and the evaluation result storage unit 309 are added.

  The evaluation information input unit 308 receives an evaluation result of validity from the user (which may be called an evaluator) with respect to the related information presented by the output unit 107. The evaluation information input unit 308 stores the received evaluation result in the evaluation result storage unit 309 in association with the relationship information.

  The evaluation information input unit 308 cooperates with the output unit 107 to function as a related information evaluation UI (User Interface). For example, as shown in FIG. 9, the relation information evaluation UI forms relation information (for example, in the form of a table) to be an object of evaluation and information to be a judgment material for evaluating the relevance information (for example, forms relation information). ) And a GUI part for evaluation input (for example, an ボ タ ン button and an ボ タ ン button) may be displayed.

  The evaluation result storage unit 309 stores the evaluated relationship information and the evaluation result thereof. Relational information whose evaluation result is good (for example, equal to or higher than the threshold) may be used as a correct case of learning performed by the learning unit 103, for example. Alternatively, the information stored in the evaluation result storage unit 309 may be used by the complementary information extraction unit 106 to determine the appropriateness of the candidate for the complementary information. For example, the complementary information extraction unit 106 determines that the candidate for the complementary information is appropriate, on the condition that the number or the ratio of the related information having a good evaluation result including the candidate for the given complementary information is equal to or more than the threshold. May be adopted.

  As described above, the information extraction device according to the second embodiment receives feedback from the evaluator about the validity of the extracted relationship information (which may include complementary information). Therefore, the information extraction apparatus can improve the extraction accuracy of the related information or the complementary information based on the evaluation result.

  The information extraction apparatus according to the present embodiment may be implemented by a single hardware device, or part of the functions of the information extraction apparatus may be executed on an external server connected to a network. The information extraction device also includes a control device such as a CPU, a storage device such as a memory, a ROM, and a RAM, an external storage device such as an HDD, a display device such as a display device, and an input device such as a keyboard and a mouse. It can also be implemented by a common computer with

  The various functional units described in the above embodiments may be realized by using a circuit. The circuit may be a dedicated circuit that implements a specific function or may be a general-purpose circuit such as a processor.

  At least a part of the processing in each of the above-described embodiments can also be realized by using a general-purpose computer as basic hardware. The program for realizing the above process may be provided by being stored in a computer readable recording medium. The program is stored in the recording medium as an installable file or an executable file. The recording medium may be a magnetic disk, an optical disk (CD-ROM, CD-R, DVD or the like), a magneto-optical disk (MO or the like), a semiconductor memory or the like. The recording medium may store the program and may be any computer readable one. Further, the program for realizing the above processing may be stored on a computer (server) connected to a network such as the Internet, and may be downloaded to the computer (client) via the network.

  While certain embodiments of the present invention have been described, these embodiments have been presented by way of example only, and are not intended to limit the scope of the invention. These novel embodiments can be implemented in various other forms, and various omissions, substitutions, and modifications can be made without departing from the scope of the invention. These embodiments and modifications thereof are included in the scope and the gist of the invention, and are included in the invention described in the claims and the equivalent scope thereof.

100, 300 ... information extraction device 101 ... input unit 102 ... relationship information extraction unit 103 ... learning unit 104 ... calculation unit 105 ... determination unit 106 ... complementary information extraction unit 107 ... Output unit 308 ... Evaluation information input unit 309 ... Evaluation result storage unit

Claims (13)

A relationship information extraction unit for obtaining a relationship information group by extracting from a document at least one relationship information representing a relationship between the value of the first attribute and the value of the second attribute;
When it is determined according to the determination criteria that the first relation information belonging to the relation information group has ambiguity, at least the value of the first attribute and the value of the second attribute forming the first relation information And a complementary information extraction unit that extracts complementary information related to one side.   The value of the first attribute forming the first relation information is the same as the value of the first attribute forming the second relation information belonging to the relation information group, and the first relation information It is determined that the first relationship information and the second relationship information have an ambiguity when the value of the second attribute that forms the second relationship information is different from the value of the second attribute that forms the second relationship information. The information extraction device according to claim 1, further comprising a determination unit.   The information processing apparatus further comprises a determination unit that determines that the first relation information has an ambiguity when the value of the first attribute included in the first relation information corresponds to a specific type of word. The information extraction device according to 1).   The complementary information extraction unit, when it is determined that the first relation information has an ambiguity, at least one of the value of the first attribute and the value of the second attribute forming the first relation information. 3. The information extraction device according to claim 2, wherein dependency analysis, analytic analysis or paraphrase is performed to extract the complementary information.   The complementary information extraction unit, when it is determined that the first relation information has an ambiguity, a word corresponding to a high-level concept or a low-level concept of the value of the first attribute forming the first relation information The information extraction device according to claim 2, wherein a word representing details of the value of the first attribute is extracted as the complementary information.   The complementary information extraction unit, when it is determined that the first relation information has ambiguity, at least one third relation information including the value of the second attribute that is the same as the first relation information. The information extraction apparatus according to claim 2, wherein any one of values of first attributes forming the information is extracted as the complementary information.   When it is determined that the first relationship information has ambiguity, the complementary information extraction unit extracts the complementary information, and complements the first relationship information using the complementary information. The information extraction device according to 1).   The information extraction device according to claim 1, further comprising an output unit that outputs the first relation information. The evaluation input unit according to claim 8, further comprising: an evaluation input unit that receives an evaluation result of validity from an evaluator for the first relation information; and a storage unit that stores the evaluation result in association with the first relation information. Information extraction device.   The information extraction device according to claim 8, wherein the output unit outputs the first relation information in a tabular form. A learning unit that learns the weight of the feature of the first relation information;
The information extraction device according to claim 1, further comprising: a calculation unit that calculates a score representing the likelihood of the first relation information based on the weight of the feature. An information extraction method implemented by a computer, comprising
Obtaining a group of relationship information by extracting from the document at least one relationship information representing a relationship between the value of the first attribute and the value of the second attribute;
When it is determined according to the determination criteria that the first relation information belonging to the relation information group has ambiguity, at least the value of the first attribute and the value of the second attribute forming the first relation information An information extraction method comprising extracting complementary information related to one side. Computer,
Means for obtaining a group of relationship information by extracting from the document at least one relationship information representing a relationship between the value of the first attribute and the value of the second attribute;
When it is determined according to the determination criteria that the first relation information belonging to the relation information group has ambiguity, at least the value of the first attribute and the value of the second attribute forming the first relation information An information extraction program to function as a means for extracting complementary information related to one.