JP3845727B2 - Question answering system using statistical tests - Google Patents

Description

[0001]
BACKGROUND OF THE INVENTION
The present invention relates to a natural language information processing system using a computer, and more particularly to a question answering system that removes unnecessary co-occurrence words using a statistical test.
[0002]
For example, if you enter a question such as “Where is the capital of Japan?” Or “Where is the winning country of the 2002 World Cup”? Is a system that accurately outputs.
[0003]
Unlike information retrieval that requires searching for a solution from searched articles, the question answering system accurately outputs the solution itself, so that the user can obtain information on the solution more quickly. Further, since the question answering system automatically outputs the solution itself, it can also be used as a knowledge processing system inside another automatic knowledge processing system. The present invention relates to a question answering system that removes unnecessary co-occurrence words by using a statistical test among such question answering systems.
[0004]
[Prior art]
As an execution example of the question answering system, consider giving a question sentence “How many tigers were there at the beginning of the 20th century?”. The answer to the question sentence is “100,000 heads”. In order to obtain this, a general question answering system estimates that the quantity expression is the answer from the expression “how much”. In addition, from the expressions “how much was a tiger” and “how long have you been”, it is assumed that the quantity expression corresponding to “how much” would be a quantity expression with “head” as a unit expression such as “... head”. To do.
[0005]
Based on this information, for example, keyword extraction is performed, articles and documents are searched for using “first 20th century” and “tiger” as keywords, and a pattern of numeric expression + “head” is obtained from the articles and documents obtained by the search. By taking out, "100,000 heads" of solutions can be taken out.
[0006]
Conventionally, there has been a question answering system for extracting a solution including such unit expressions. (Refer nonpatent literature 1.).
[0007]
A method for selecting a word that characterizes a given document set has also existed conventionally (see Non-Patent Document 2).
[0008]
[Non-Patent Document 1]
Hiroshi Sasaki, Hideki Amagasaki, Hiroyuki Hira, Tsutomu Hirao, Hideto Kazawa, Jun Suzuki, Koji Kyori, Eisaku Maeda, SAIQA: Question Answering System Based on Mass Documents, Information Processing Society of Japan, Natural Language Processing Society 2001-NL-145, 2001
[Non-Patent Document 2]
Toru Hisamitsu, Yoshiki Niwa, Feature word selection method based on combinatorial probability model, Natural Language Processing Study Group, IPSJ, 140-12, 2000
[0009]
[Problems to be solved by the invention]
However, since the prior art described in Non-Patent Document 1 uses rules or tables that are described manually for the extraction of unit expressions, depending on the question sentence, it is appropriate to determine whether the unit expression is useful for solution extraction. In some cases, it is impossible to make a decision, and only a low accuracy rate can be obtained.
[0010]
The prior art described in Non-Patent Document 2 is intended to select feature words as keywords of sentences, and is not intended to extract unit expressions that are functional expressions of languages.
[0011]
The present invention solves the above-mentioned problems of the prior art, and when the answer to the question sentence is a quantity expression, it is possible to accurately determine whether or not the unit expression to be attached to the quantity expression is an appropriate unit expression. It is an object to provide a simple question answering method and question answering system.
[0012]
[Means for Solving the Problems]
In order to solve the above problems, the present invention is useful for extracting a solution by using a statistical test based on co-occurrence frequency information in digitized text information such as a newspaper corpus of main nouns and unit expressions in a question sentence. Only the useful unit representation is used for solution extraction, and the unit representation judged to be unnecessary is not used for solution extraction. The main noun refers to the main subject of the question in the question sentence. For example, in the question sentence “How many tigers were there at the beginning of the 20th century”, “tiger” is the main noun.
[0013]
Here, for example, the quantity expression corresponding to “how much” in the question sentence “How many tigers were there at the beginning of the 20th century” would be a quantity expression with “head” as a unit expression such as “... head”. When inferring, a large number of patterns such as ““ tiger is ”+ quantity expression” are extracted, and the unit expression “head” is extracted by extracting the unit expression attached to the quantity expression, or “quantity expression +“ now Similarly, the unit expression “head” can be extracted from a pattern such as “”.
[0014]
However, with such a method alone, for example, a unit expression “year” is extracted from a sentence such as “tiger is in 1992”. If “year” can also be used as a unit expression for solution extraction, the expression “1992” may be erroneously output as a solution.
[0015]
A method using co-occurrence frequency information in a corpus or the like is also conceivable. Count the number of unit expressions that co-occur with "tiger" and use only the unit expression with the largest number. In this case, “head” has the highest appearance frequency, so that “head” is used as a unit expression, and the problem of extracting “year” as a unit expression is solved.
[0016]
However, as a unit expression in the quantity expression equivalent to “how much” of the question sentence “How many tigers were there at the beginning of the 20th century”, the expression “animal” is also conceivable. If the expression of the solution is “100,000”, using only the “head” with the maximum frequency makes it impossible to extract “100,000” as a solution.
[0017]
Therefore, in the present invention, based on the co-occurrence frequency information in the corpus or the like of the main noun and the unit expression, using the statistical test, the unit noun and the co-occurrence among the unit expressions attached to the quantity expression. Thus, unit expressions that are unlikely to appear are determined to be unnecessary co-occurrence words and removed, and only unit expressions that are determined to be useful are used for solution extraction.
[0018]
In particular,
“Appearance probability of“ tiger ”appearing together with the unit expression = general probability of occurrence of“ tiger ”in the corpus
Through the test of the above hypothesis based on the number of occurrences of “tiger” and unit expression appearing together in corpus, etc., and the frequency of occurrence of “tiger” and unit expression, It is judged whether each unit expression such as “tiger”, “head”, “animal”, etc. co-occurs by chance or inevitably co-occurs. Based on the test result, if it is determined that a unit expression “tiger is” co-occurs by chance, it is determined that it is only an accidental unit expression and is an unnecessary unit expression. In addition, if it is judged that a unit expression is inevitably co-occurring with "tiger", it is inevitably co-occurring, so it is expected that the expression will be deeply related. Judged as a useful unit expression. This method makes it possible to determine whether each unit representation is useful or unnecessary as a unit representation used for solution extraction. Then, if the solution of the question answering system is judged using only the unit expressions judged to be useful, all the above problems are solved.
[0019]
That is, the present invention is a question answering method for inputting a question sentence in a natural language, generating a solution by collation with sentences in digitized text information stored in a database, and outputting the answer. A combination pattern of a main noun and a quantity expression of the question sentence appearing in the digitized text information using a statistical test when a question sentence input process to be performed and a solution of the input question sentence is a quantity expression The unit expression extracted from the unit sentence that is unlikely to appear together with the main noun in the question sentence is judged as an unnecessary co-occurrence word, and appears together with the main noun in the question sentence. An expression estimation process for determining a unit expression having a high possibility as a useful co-occurrence word and inferring a quantity expression as a solution of the question sentence using the unit expression determined as the useful co-occurrence word; Quantity expression And having a answer extraction step of extracting the solution of the question by using.
[0020]
Also, in the present invention, the expression estimation process is a quantity expression determination process for determining whether a solution of a question sentence is a quantity expression based on the question sentence information, and the main noun is recognized from the question sentence information. A combination in which the recognized main noun and the quantity expression are in a dependency relationship from the digitized text information when it is determined that the answer to the question sentence is a quantity expression in the main noun recognition process and the quantity expression judgment process. A unit expression extraction process for extracting one or a plurality of unit expressions attached to the quantity expression from the extracted combination pattern, the main noun in the digitized text information, and each extracted An appearance probability calculating step of calculating an appearance frequency of the unit expression and calculating a general appearance probability of the main noun based on the calculated appearance frequency information; Based on the general appearance probability of the main noun, the probability distribution of the number of times that the main noun appears together with each unit expression among the patterns in which each unit expression appears in the digitized text information is calculated. Probability distribution calculation process, and the number of times that the main noun in the calculated probability distribution co-occurs with each unit expression is attached to the quantity expression in the combination pattern in which the main noun and the quantity expression have a dependency relationship. A test probability that is a probability that is equal to or greater than the number of occurrences of each unit expression, and among the one or more unit expressions extracted in the unit expression extraction process, the calculated test probability is greater than or equal to a preset threshold value Unit expressions that are judged as unnecessary co-occurrence words that are unlikely to appear together with the main noun are removed, and the unit expression whose test probability is less than the threshold And unnecessary units represented removing process of determining a useful occurrence word, and having a quantity representation guess process to infer the quantity expressed as the solution of the question by using the unit representation is determined that the useful.
[0021]
Also, in the present invention, the expression estimation process is a quantity expression determination process for determining whether a solution of a question sentence is a quantity expression based on the question sentence information, and the main noun is recognized from the question sentence information. A combination in which the recognized main noun and the quantity expression are in a dependency relationship from the digitized text information when it is determined that the answer to the question sentence is a quantity expression in the main noun recognition process and the quantity expression judgment process. A unit expression extraction process for extracting one or a plurality of unit expressions attached to the quantity expression from the extracted combination pattern, the main noun in the digitized text information, and each extracted An appearance probability calculating step of calculating an appearance frequency of the unit expression, and calculating a general appearance probability of each unit expression based on the calculated appearance frequency information; Probability distribution that calculates the probability distribution of the number of times each unit representation appears together with the main noun among the patterns in which the main noun appears in the digitized text information based on the general appearance probability of the position representation The number of times that each unit expression in the calculated probability distribution co-occurs with the main noun is added to the quantity expression in a combination pattern in which the main noun and the quantity expression have a dependency relationship. Calculating a test probability that is a probability that is equal to or greater than the number of appearances of the quantity expression appearing together with each of the unit expressions, and calculating the calculated one of the unit expressions extracted in the unit expression extraction process Unit expressions whose test probabilities are equal to or higher than a preset threshold are judged to be unnecessary co-occurrence words that are unlikely to appear together with the main noun and are removed. An unnecessary unit expression removing process for determining a unit expression having a probability less than a threshold as a useful co-occurrence word, and a quantity expression estimating process for estimating a quantity expression as a solution of the question sentence using the unit expression determined to be useful. It is characterized by having.
[0022]
Also, in the present invention, the expression estimation process is a quantity expression determination process for determining whether a solution of a question sentence is a quantity expression based on the question sentence information, and the main noun is recognized from the question sentence information. A combination in which the recognized main noun and the quantity expression are in a dependency relationship from the digitized text information when it is determined that the answer to the question sentence is a quantity expression in the main noun recognition process and the quantity expression judgment process. A unit expression extraction process for extracting one or a plurality of unit expressions attached to the quantity expression from the extracted combination pattern, the main noun in the digitized text information, and each extracted An appearance probability calculating step of calculating an appearance frequency of the unit expression and calculating a general appearance probability of the main noun based on the calculated appearance frequency information; Based on the general appearance probability of the main noun, the probability distribution of the number of times that the main noun appears together with each unit expression among the patterns in which each unit expression appears in the digitized text information is calculated. Probability distribution calculation process, and the number of times that the main noun in the calculated probability distribution co-occurs with each unit expression is attached to the quantity expression in the combination pattern in which the main noun and the quantity expression have a dependency relationship. A test probability that is a probability that is less than or equal to the number of occurrences of each unit expression being calculated, and among the one or more unit expressions extracted in the unit expression extraction process, the calculated test probability is less than a preset threshold value Unit expressions that are judged to be unnecessary co-occurrence words that are unlikely to appear together with the main noun are eliminated, and the unit expression whose test probability is equal to or greater than a threshold value And unnecessary units represented removing process of determining a useful occurrence word, and having a quantity representation guess process to infer the quantity expressed as the solution of the question by using the unit representation is determined that the useful.
[0023]
Also, in the present invention, the expression estimation process is a quantity expression determination process for determining whether a solution of a question sentence is a quantity expression based on the question sentence information, and the main noun is recognized from the question sentence information. A combination in which the recognized main noun and the quantity expression are in a dependency relationship from the digitized text information when it is determined that the answer to the question sentence is a quantity expression in the main noun recognition process and the quantity expression judgment process. A unit expression extraction process for extracting one or a plurality of unit expressions attached to the quantity expression from the extracted combination pattern, the main noun in the digitized text information, and each extracted An appearance probability calculating step of calculating an appearance frequency of the unit expression, and calculating a general appearance probability of each unit expression based on the calculated appearance frequency information; Probability distribution that calculates the probability distribution of the number of times each unit representation appears together with the main noun among the patterns in which the main noun appears in the digitized text information based on the general appearance probability of the position representation The number of times that each unit expression in the calculated probability distribution co-occurs with the main noun is added to the quantity expression in a combination pattern in which the main noun and the quantity expression have a dependency relationship. A test probability that is a probability that is equal to or less than the number of appearances of the quantity expression that co-occurs with each of the unit expressions, and the calculated one of the unit expressions extracted in the unit expression extraction process Unit expressions whose test probabilities are less than a preset threshold are judged to be unnecessary co-occurrence words that are unlikely to appear together with the main noun and are removed. An unnecessary unit expression removing process for determining a unit expression having a probability equal to or greater than a threshold value as a useful co-occurrence word, and a quantity expression estimating process for estimating a quantity expression as a solution of the question sentence using the unit expression determined to be useful. It is characterized by having.
[0024]
Also, in the present invention, the expression estimation process is a quantity expression determination process for determining whether a solution of a question sentence is a quantity expression based on the question sentence information, and the main noun is recognized from the question sentence information. A combination in which the recognized main noun and the quantity expression are in a dependency relationship from the digitized text information when it is determined that the answer to the question sentence is a quantity expression in the main noun recognition process and the quantity expression judgment process. A unit expression extraction process for extracting one or a plurality of unit expressions attached to the quantity expression from the extracted combination pattern, the main noun in the digitized text information, and each extracted Using the appearance frequency extraction process for calculating the appearance frequency of the unit representation and the hypergeometric distribution, the main noun in the hypergeometric distribution co-occurs with each unit representation. Calculating a test probability that is a probability that the number of appearances is equal to or greater than the number of appearances of each unit expression in the combination pattern in which the main noun and the quantity expression have a dependency relationship, and extracting the unit expression Among one or a plurality of unit expressions extracted in the process, a unit expression having a calculated test probability equal to or higher than a preset threshold is determined as an unnecessary co-occurrence word that is unlikely to appear together with the main noun. And removing unnecessary unit expressions that determine unit expressions with a test probability less than a threshold as useful co-occurrence words, and inferring quantity expressions as solutions to the question sentence using the unit expressions determined to be useful And a quantity expression inference process.
[0025]
Also, in the present invention, the expression estimation process is a quantity expression determination process for determining whether a solution of a question sentence is a quantity expression based on the question sentence information, and the main noun is recognized from the question sentence information. A combination in which the recognized main noun and the quantity expression are in a dependency relationship from the digitized text information when it is determined that the answer to the question sentence is a quantity expression in the main noun recognition process and the quantity expression judgment process. A unit expression extraction process for extracting one or a plurality of unit expressions attached to the quantity expression from the extracted combination pattern, the main noun in the digitized text information, and each extracted Using the appearance frequency extraction process for calculating the appearance frequency of the unit representation and the hypergeometric distribution, the main noun in the hypergeometric distribution co-occurs with each unit representation. Calculating a test probability that is a probability that the number of appearances is less than or equal to the number of appearances of each unit expression in the combination pattern in which the main noun and the quantity expression have a dependency relationship, and extracting the unit expression Among one or more unit expressions extracted in the process, a unit expression whose calculated test probability is less than a preset threshold is determined as an unnecessary co-occurrence word that is unlikely to appear along with the main noun. And removing unnecessary unit expressions that determine unit expressions with a test probability equal to or greater than a threshold value as useful co-occurrence words, and inferring quantity expressions as solutions to the question sentence using the unit expressions determined to be useful And a quantity expression inference process.
[0026]
  Further, the present invention is a question answering system for inputting a question sentence in a natural language, generating a solution by collation with sentences in electronic text information stored in a database, and outputting the answer. Included in the question text that appears in the digitized text information using a statistical test when the answer to the question text input means and the answer of the input question text is a quantitative expression,Using morphological analysis and sentence patternsA unit expression is extracted from a combination pattern of a specific noun and a quantity expression certified based on a predetermined rule, and the possibility that the extracted unit expression co-occurs with a specific noun in the question sentence is shown. Means for calculating a probability value, determining a unit expression used to determine a quantity expression as a solution of the question sentence according to whether the calculated probability value is equal to or greater than a threshold, and determining a quantity expression including the determined unit expression; Answer extraction means for extracting an answer to the question sentence using the determined quantity expression.
[0027]
  Further, the present invention is a question answering system for inputting a question sentence in a natural language, generating a solution by collation with sentences in electronic text information stored in a database, and outputting the answer. And a question expression input means for determining whether a solution of the question sentence is a quantity expression according to whether the question sentence includes a word for inquiring a quantity expression;Using morphological analysis and sentence patternsBased on a predetermined rule, when the means for certifying a specific noun included in the question sentence and the quantity expression determining means determine that the answer to the question sentence is a quantity expression, Unit expression extraction means for extracting a combination pattern in which a specific authorized noun and a quantity expression have a dependency relationship, and extracting one or a plurality of unit expressions attached to the quantity expression from the extracted combination pattern And an appearance probability for calculating the number of appearances of the specific noun and the extracted unit expressions in the digitized text information, and calculating a general appearance probability of the specific noun based on the calculated number of appearance information Based on the calculated general appearance probability of the specific noun, the specific text out of the patterns in which each unit expression appears in the digitized text information Probability distribution calculating means for calculating a probability distribution of the number of times that a noun appears together with each unit expression, and the number of times that the specific noun in the calculated probability distribution appears together with each unit expression Calculating a test probability that is a probability that the unit expression attached to the quantity expression in the combination pattern having a dependency relationship between the specific noun and the quantity expression is equal to or more than the number of appearances, and the unit expression extracting unit extracts Among the one or a plurality of unit expressions, an unnecessary unit expression removing means for removing a unit expression whose calculated test probability is equal to or higher than a preset threshold value and a quantity expression including a unit expression whose test probability is less than the threshold value are determined. Means for extracting a keyword from the question sentence, and using the extracted keyword to extract information describing a solution of the question sentence from the digitized text information; Using serial determined quantity expressed, characterized in that it comprises a answer extraction means for extracting the solution of the question message from the information solutions of the question is described.
[0028]
  Further, the present invention is a question answering system for inputting a question sentence in a natural language, generating a solution by collation with sentences in electronic text information stored in a database, and outputting the answer. And a question expression input means for determining whether a solution of the question sentence is a quantity expression according to whether the question sentence includes a word for inquiring a quantity expression;Using morphological analysis and sentence patternsBased on a predetermined rule, when the means for certifying a specific noun included in the question sentence and the quantity expression determining means determine that the answer to the question sentence is a quantity expression, Unit expression extraction means for extracting a combination pattern in which a specific authorized noun and a quantity expression have a dependency relationship, and extracting one or a plurality of unit expressions attached to the quantity expression from the extracted combination pattern And an appearance probability for calculating the number of appearances of the specific noun and the extracted unit representations in the digitized text information, and calculating a general appearance probability of the unit representations based on the calculated appearance number information Based on the calculation means and the general appearance probability of each unit representation, each unit representation among the patterns in which the specific noun appears in the digitized text information is A probability distribution calculating means for calculating a probability distribution of the number of times of occurrence co-occurring with a specific noun, and the number of times that each unit expression in the calculated probability distribution co-occurs with the specific noun The test probability is a probability that is equal to or more than the number of appearances of the quantity expression that co-occurs with the unit expressions attached to the quantity expression in the combination pattern in which the noun and the quantity expression have a dependency relationship And an unnecessary unit expression removing means for removing a unit expression having a calculated test probability equal to or higher than a preset threshold among one or a plurality of unit expressions extracted by the unit expression extracting means, and the test probability is less than the threshold value. Means for determining a quantity expression including a unit expression; a keyword is extracted from the question sentence; and a solution of the question sentence is described from the digitized text information using the extracted keyword. Means for retrieving the information, using the determined quantity expressed, characterized in that it comprises a answer extraction means for extracting the solution of the question message from the information solutions of the question is described.
[0029]
  Further, the present invention is a question answering system for inputting a question sentence in a natural language, generating a solution by collation with sentences in electronic text information stored in a database, and outputting the answer. And a question expression input means for determining whether a solution of the question sentence is a quantity expression according to whether the question sentence includes a word for inquiring a quantity expression;Using morphological analysis and sentence patternsBased on a predetermined rule, when the means for certifying a specific noun included in the question sentence and the quantity expression determining means determine that the answer to the question sentence is a quantity expression, Unit expression extraction means for extracting a combination pattern in which a specific authorized noun and a quantity expression have a dependency relationship, and extracting one or a plurality of unit expressions attached to the quantity expression from the extracted combination pattern And an appearance probability for calculating the number of appearances of the specific noun and the extracted unit expressions in the digitized text information, and calculating a general appearance probability of the specific noun based on the calculated number of appearance information Based on the calculated general appearance probability of the specific noun, the specific text out of the patterns in which each unit expression appears in the digitized text information Probability distribution calculating means for calculating a probability distribution of the number of times that a noun appears together with each unit expression, and the number of times that the specific noun in the calculated probability distribution appears together with each unit expression A test probability that is a probability that the number of appearances of each unit expression attached to the quantity expression in a combination pattern in which the specific noun and the quantity expression have a dependency relationship is equal to or less than the number of appearances is extracted, and the unit expression extraction unit extracts Among the one or a plurality of unit expressions, an unnecessary unit expression removing means for removing a unit expression whose calculated test probability is less than a preset threshold value and a quantity expression including a unit expression having the test probability equal to or higher than the threshold value are determined. Means for extracting a keyword from the question sentence, and using the extracted keyword to extract information describing a solution of the question sentence from the digitized text information; Using serial determined quantity expressed, characterized in that it comprises a answer extraction means for extracting the solution of the question message from the information solutions of the question is described.
[0030]
  Further, the present invention is a question answering system for inputting a question sentence in a natural language, generating a solution by collation with sentences in electronic text information stored in a database, and outputting the answer. And a question expression input means for determining whether a solution of the question sentence is a quantity expression according to whether the question sentence includes a word for inquiring a quantity expression;Using morphological analysis and sentence patternsBased on a predetermined rule, when the means for certifying a specific noun included in the question sentence and the quantity expression determining means determine that the answer to the question sentence is a quantity expression, Unit expression extraction means for extracting a combination pattern in which a specific authorized noun and a quantity expression have a dependency relationship, and extracting one or a plurality of unit expressions attached to the quantity expression from the extracted combination pattern And an appearance probability for calculating the number of appearances of the specific noun and the extracted unit representations in the digitized text information, and calculating a general appearance probability of the unit representations based on the calculated appearance number information Based on the calculation means and the general appearance probability of each unit representation, each unit representation among the patterns in which the specific noun appears in the digitized text information is A probability distribution calculating means for calculating a probability distribution of the number of times of occurrence co-occurring with a specific noun, and the number of times that each unit expression in the calculated probability distribution co-occurs with the specific noun The test probability is a probability that is equal to or less than the number of appearances of the quantity expression that co-occurs with the unit expressions attached to the quantity expression in the combination pattern in which the noun and the quantity expression have a dependency relationship And an unnecessary unit expression removing means for removing a unit expression whose calculated test probability is less than a preset threshold among one or a plurality of unit expressions extracted by the unit expression extracting means; Means for determining a quantity expression including a unit expression; a keyword is extracted from the question sentence; and a solution of the question sentence is described from the digitized text information using the extracted keyword. Means for retrieving the information, using the determined quantity expressed, characterized in that it comprises a answer extraction means for extracting the solution of the question message from the information solutions of the question is described.
[0031]
  Further, the present invention is a question answering system for inputting a question sentence in a natural language, generating a solution by collation with sentences in electronic text information stored in a database, and outputting the answer. And a question expression input means for determining whether a solution of the question sentence is a quantity expression according to whether the question sentence includes a word for inquiring a quantity expression;Using morphological analysis and sentence patternsBased on a predetermined rule, when the means for certifying a specific noun included in the question sentence and the quantity expression determining means determine that the answer to the question sentence is a quantity expression, Unit expression extraction means for extracting a combination pattern in which a specific authorized noun and a quantity expression have a dependency relationship, and extracting one or a plurality of unit expressions attached to the quantity expression from the extracted combination pattern And means for calculating the number of occurrences of the specific noun and the extracted unit representations in the digitized text information, and using the hypergeometric distribution, the specific noun in the hypergeometric distribution Number of occurrences of each unit expression attached to the quantity expression in the combination pattern in which the number of times of co-occurrence is a dependency relationship between the specific noun and the quantity expression An unnecessary unit expression removal that calculates a test probability that is a probability above and removes a unit expression having a calculated test probability equal to or higher than a preset threshold among one or a plurality of unit expressions extracted by the unit expression extraction unit Means for determining a quantity expression including a unit expression whose test probability is less than a threshold; a keyword is extracted from the question sentence, and a solution of the question sentence is described from the digitized text information using the extracted keyword And a means for extracting the answer of the question sentence from the information in which the answer of the question sentence is described using the determined quantity expression.
[0032]
  Further, the present invention is a question answering system for inputting a question sentence in a natural language, generating a solution by collation with sentences in electronic text information stored in a database, and outputting the answer. And a question expression input means for determining whether a solution of the question sentence is a quantity expression according to whether the question sentence includes a word for inquiring a quantity expression;Using morphological analysis and sentence patternsBased on a predetermined rule, when the means for certifying a specific noun included in the question sentence and the quantity expression determining means determine that the answer to the question sentence is a quantity expression, Unit expression extraction means for extracting a combination pattern in which a specific authorized noun and a quantity expression have a dependency relationship, and extracting one or a plurality of unit expressions attached to the quantity expression from the extracted combination pattern And means for calculating the number of occurrences of the specific noun and the extracted unit representations in the digitized text information, and using the hypergeometric distribution, the specific noun in the hypergeometric distribution Number of occurrences of each unit expression attached to the quantity expression in the combination pattern in which the number of times of co-occurrence is a dependency relationship between the specific noun and the quantity expression Unnecessary unit expression removal that calculates a test probability that is a lower probability and removes a unit expression whose calculated test probability is less than a preset threshold among one or more unit expressions extracted by the unit expression extraction unit Means for determining a quantity expression including a unit expression having a test probability equal to or greater than a threshold; a keyword is extracted from the question sentence, and a solution of the question sentence is described from the digitized text information using the extracted keyword And a means for extracting the answer of the question sentence from the information in which the answer of the question sentence is described using the determined quantity expression.
[0033]
The present invention also provides a question answering program using a statistical test for causing a computer to execute the question answering method using the statistical test.
[0034]
Further, the present invention is a recording medium on which a question answering program using a statistical test for causing a computer to execute the question answering method using the statistical test is recorded.
[0035]
By using the present invention, when a solution to a question sentence is a quantity expression, a method and a system capable of accurately determining whether or not a unit expression appropriate as a unit expression attached to the quantity expression is provided. It becomes possible.
[0036]
DETAILED DESCRIPTION OF THE INVENTION
Hereinafter, embodiments of the present invention will be described with reference to the drawings. FIG. 1 is a diagram showing an example of the configuration of a question answering system using the statistical test of the present invention. 1 is a question answering system using the statistical test of the present invention, 11 is an unnecessary co-occurrence word elimination system, 12 is a question sentence input unit for inputting question sentence information, and 13 is a keyword extraction or information search from a database 16. A keyword extraction / information retrieval unit, 14 an answer extraction unit for extracting a question sentence solution, 15 an answer output unit for outputting an answer, 16 a database storing digitized text information such as a newspaper corpus, and 17 a question sentence Reference numeral 18 denotes a unit expression database for storing unit expressions extracted from the database 16.
[0037]
FIG. 2 is an example of a configuration diagram of the unnecessary co-occurrence word removal system 11. 20 is a quantity expression judging means for judging whether the answer of the question sentence is a quantity expression based on the question sentence information input to the question sentence input unit 12, and 21 is a main noun authorization for certifying a main noun from the question sentence information. Means 22 extracts from the digitized text information stored in the database 16 a combination pattern in which the main noun recognized by the main noun recognition means 21 has a dependency relationship with the quantity expression, and the quantity is extracted from the extracted combination pattern. The unit expression extraction means 23 extracts one or more unit expressions attached to the expression and records them in the unit expression database 18, and the possibility that the one or more unit expressions appear together with the main noun The test means 24 for judging whether or not is high is a quantity expression estimation means 24 for estimating a quantity expression as a solution to the question sentence.
[0038]
In addition, 230 calculates the appearance frequency of the main noun and the extracted unit expressions in the computerized text information, and calculates the appearance probability of the main noun based on the calculated appearance frequency information. The means 231 is the number of times that the main noun appears co-occurring with the unit representations among the patterns in which the unit representations appear in the digitized text information based on the calculated general appearance probability of the main nouns. 232 is a probability distribution calculating means for calculating the probability distribution, and 232 is the number of times the main noun appears in the probability distribution calculated by the probability distribution calculating means 231 together with each unit expression. A test probability that is a probability that is equal to or greater than the number of occurrences of each unit expression attached to the quantity expression in the combination pattern that is a receiving relationship. Out of one or a plurality of unit expressions extracted by the unit expression extracting unit, it is unnecessary that a unit expression having a calculated test probability equal to or higher than a preset threshold is likely to appear together with the main noun. It is an unnecessary unit expression removing unit that determines that a co-occurrence word is removed from the unit expression database 18 and determines that a unit expression whose test probability is less than a threshold is a useful unit expression.
[0039]
Hereinafter, a first embodiment of the present invention will be described based on FIGS. 3 and 4 with reference to FIGS. 1, 2, and 8. FIG. In the first embodiment of the present invention, as will be described later, it is assumed that the appearance probability p 'in the corpus of the main noun when co-occurring with each unit expression is equal to the general appearance probability p of the main noun. Each unit expression w in which the main noun based on the general appearance probability p appears n times in the corpus_iThe number of occurrences that co-occur with each unit representation w in the corpus_iX, the frequency at which the main nouns co-occur_iBased on the probability that the number of times is greater than or equal to the number of times (test probability) P A unit expression that cannot be concluded as p ′> p is determined as an unnecessary unit expression.
[0040]
FIG. 3 is a diagram showing an example of a question response processing flow using the statistical test of the present invention. First, a question text is input to the question text input unit 12 (step S1). For example, a question sentence “How big is the land area of Japan?” Is input to the question sentence input unit 12. Next, the inputted question sentence information is transferred from the question sentence input unit 12 to the expression estimation unit 17 of the unnecessary co-occurrence word removal system 11 (step S2). Next, in the unnecessary co-occurrence word removal system 11, the expression estimation unit 17 uses a statistical test to determine whether a unit expression extracted from the corpus is useful for estimating an expression that is likely to be a solution or unnecessary. It is determined whether the unit expression is correct, and the unit expression determined to be unnecessary is removed (step S3).
[0041]
Next, the expression estimation unit 17 uses a unit expression determined to be useful to estimate an expression that is likely to be a solution (step S4). Specifically, the quantity expression estimation means 24 of the expression estimation unit 17 estimates a quantity expression that is likely to be a solution.
[0042]
Then, the keyword extraction / information retrieval unit 13 extracts a keyword from the question sentence passed from the expression estimation unit 17. Then, using the extracted keyword, an article group in which a solution is likely to be described is extracted from the database, and the extracted article group is passed to the answer extraction unit 14 (step S5).
[0043]
Next, the answer extraction unit 14 extracts an expression that matches the expression estimated by the expression estimation unit 17 from the extracted article group, and passes the extracted expression to the answer output unit 15 (step S6). Finally, the answer output unit 15 outputs the passed expression as an answer (step S7).
[0044]
Here, FIG. 4 shows an example of the details of the processing flow for removing unnecessary unit expressions in the first embodiment of the present invention. FIG. 4 is a diagram showing details of step S3 in FIG. First, the quantity expression judging means 20 judges whether the answer to the question sentence is a quantity expression based on the inputted question sentence information (step S21). For example, when the question sentence includes an expression such as “how much” as in “how much is the land area of Japan”, it is determined that the answer of the question sentence is a quantity expression. If it is determined that the answer to the question sentence is not a quantity expression, the process proceeds to step S5.
[0045]
When determining that the answer to the question sentence is a quantity expression, the quantity expression determining means 20 determines whether the question sentence has a unit expression (step S22). If the question sentence does not have a unit expression, the main noun recognition means 21 recognizes the main noun in the question sentence from the question sentence information (step S23). The main nouns are identified based on the rules using morphological analysis and sentence patterns from the pattern of “how much is X”. For example, the main noun in the above question sentence is recognized as “area”. If the question sentence has a unit expression, the process ends.
[0046]
Next, the appearance probability calculation means 230 of the test means 23 calculates a general appearance probability p of the main noun in the corpus (step S24). The general appearance probability p is obtained by dividing the number of appearances of the main noun in the corpus by the size of the corpus. If a large-scale newspaper corpus such as the Mainichi Shimbun is used, for example, the number of appearances in the “area” corpus is 7,933. If the size of the corpus is 409, 502, 077 characters, for example, the general appearance probability p of “area” is
p = 7,933 / 409,502,077
= 0.000019372307111530676
Is calculated.
[0047]
Next, the unit expression extracting means 22 extracts a combination pattern in which the main noun and the quantity expression are in a dependency relationship from the corpus, and one or a plurality of the combination expressions attached to the quantity expression from the extracted combination pattern. Unit representation w_iIs extracted and recorded in the unit representation database 18 (step S25). In the present embodiment, “area is” + quantity expression pattern is extracted from the corpus, and unit expression w attached to the quantity expression_iTake out. Each unit expression w_iIs extracted by using a morphological analysis result and a word dictionary. Each extracted unit representation w_iIs stored in the unit representation database 18. The pattern of quantity expression is extracted based on morphological analysis and rules using sentence patterns.
[0048]
Here, the following unit expression w is obtained from the quantity expression extracted from the corpus as “area is” + quantity expression pattern._iCan be extracted with the following frequency, for example.
[0049]
Square meter: 41 times, ha: 27 times, square kilometer: 5 times, double: 2 times, minute: 2 times, tatami mat: 1 time, number: 1 time, square centimeter: 1 time
The unit expression w before the colon_iThe number is the frequency. The frequency of the “area is” + quantity expression pattern is 80 times in total.
[0050]
Next, the appearance probability calculation means 230 uses the extracted unit representations w._iThe appearance frequency n in the corpus is calculated (step S26). Each unit representation w_iThe appearance frequency n of the corpus is, for example, square meters: 17,510 times, hectares: 8,088 times, square kilometers: 247 times, doubles: 41,686 times, minutes: 730, 790 times, tatami mats: 4,829 times, No .: 124,233 times, square centimeter: 20 times.
[0051]
In addition, the appearance probability calculation means 230 actually uses each unit representation w in the corpus._iAnd the frequency of occurrence of the main noun co-occurring x_iIs calculated (step S27). For example, when the main noun is “area”, the unit expression of the quantity expression in the pattern of “area is” + quantity expression is w_iSuppose that That is, each unit representation w in the corpus_iOf the n occurrence patterns, the quantity expression and the “area” co-occurred in the pattern ““ area is ”+ quantity expression” is x_iIt is. For example, in the case of a square meter, x_i= 41 times.
[0052]
Next, using the binomial distribution, we hypothesized that the occurrence probability p 'of the "area" when co-occurring with the "square meter" = the general appearance probability p of the "area" without such conditions Perform the test. That is, first, the probability distribution calculating means 231 uses the main noun and the unit representation w in the corpus based on the calculated general appearance probability p of the main noun._iThe probability distribution of the number of occurrences of and co-occurring is calculated (step S28). All individual trials are independent and a unit representation w_iWhen n appears n times, the probability distribution of the number of times the main noun with the general appearance probability p appears is obtained. Such a probability distribution is expressed as follows: r (r = 0, 1, 2, 3... N)
_nC_rp^r(1-p)^nr
And a binomial distribution as shown in FIG. 8 is obtained.
[0053]
Next, the unnecessary unit representation removing unit 232 determines the main noun and each unit representation w in the calculated probability distribution._iThe number of occurrences of and co-occurs is x_iThe probability (test probability) P that is greater than or equal to the number of times is calculated (step S29). In FIG. 8, the probability of the hatched portion is P. For example, in an appearance scene of n = 17510 times of “square meter”, an “area” with a general appearance probability p = about 0.00001937 is x_i= Find the test probability P that appears 41 times or more.
[0054]
In the embodiment of the present invention, the appearance probability p ′ of “area” when co-occurring with “square meter” and the general appearance probability p of “area” when there is no such condition may be inherently different. The occurrence probability p ′ of “area” when co-occurring with “square meter” is used as the general appearance probability p of “area” when there is no such condition. Yes.
[0055]
Therefore, if the value of the test probability P is sufficiently small, the hypothesis that p ′ = p is rejected, and it can be determined that p ′ and p are different. In the embodiment of the present invention, “x_iWhen the test probability P is obtained as in the case of “number of times or more”, only one region of the probability distribution is used. Furthermore, the area used is “x_iSince it is the right one-sided test using the region of the larger case such as “more than times”, if the value of P is sufficiently small, it can be concluded that p ′> p. However, this determination does not necessarily mean that the determination is correct, and only the probability P is a determination that may be erroneous.
[0056]
Unnecessary unit representation removing means 232 has a sufficiently small value of P, and the occurrence probability p ′ of “area” when co-occurring with “square meter” is a general appearance of “area” when there is no such condition. If it can be concluded that the probability is greater than p, “square meter” is determined to be a useful unit expression because the co-occurrence of “square meter” and “area” is not a coincidence but an inevitable co-occurrence.
[0057]
On the other hand, when the value of the test probability P is not sufficiently small and co-occurs with “square meter”, the appearance probability p ′ of “area” is the general appearance probability p of “area” when there is no such condition. If it is not possible to conclude that it is larger than this, the co-occurrence of “square meter” and “area” is not inevitable, and it is likely that it is a coincidence. Therefore, it should not be used in subsequent question answering system processing.
[0058]
That is, the unnecessary unit representation removing unit 232 uses each unit representation w._iIt is determined whether the value of the test probability P is less than a preset threshold value (step S30), a unit expression with a value of P less than the threshold value is determined as a useful unit expression (step S32), and the value of P is greater than or equal to the threshold value The unit expression is determined as an unnecessary unit expression and is removed from the unit expression database 18 (step S31), and the process is terminated.
[0059]
When the value of P is actually calculated in the above example, square meter: 0.00000000, hectare: 0.00000000, square kilometer: 0.00000000, double: 0.193992510, minute: 0.99999233, tatami mat: 0.089307732 : 0.90988811 and square centimeter: 0.00038737. This calculation can be performed by using the binomial distribution theory as described above. Here, calculation was performed using the subroutine pbinom of the free software Math-CDF-0.1.
[0060]
Here, for example, when 0.1 is used as a threshold value that means that the value of P is sufficiently small,
Square meter: 0.00000000
Ha: 0.00000000
Square kilometers: 0.00000000
Tatami mat: 0.08930732
Square centimeter: 0.00038737
Is a useful unit expression, ie a valid co-occurrence word,
Double: 0.19392510
Minute: 0.9999923
Number: 0.90988811
Is an unnecessary unit expression, that is, an invalid co-occurrence word.
[0061]
In fact, square meters: 0.00000000, hectares: 0.00000000, square kilometers: 0.00000000, tatami mats: 0.08930732, square centimeters: 0.00038737 can be used as a unit representation of area, but times: 0.19392510, minutes: 0.99999233, number: 0.90988811 is an illegal expression as a unit expression of area.
[0062]
In this way, the unit expression determined to be an appropriate unit expression is used to estimate the quantity expression as the solution to the question sentence.
[0063]
In the first embodiment of the present invention, it is assumed that the appearance probability p ′ of “area” when co-occurring with “square meter” is equal to the general appearance probability p of “area”, and the value of P A right-sided test is performed based on the size of, and it is concluded that the above hypothesis can be rejected, that is, p ′> p, and a unit expression that can be concluded as p ′> p is a useful unit expression, and p ′ Unit expressions that cannot be concluded as> p have been determined to be unnecessary unit expressions. However, in the present invention, the left one-sided test is performed as follows to conclude whether p ′ <p, and p ′ <p is concluded. The unit expression that can be attached is an unnecessary unit expression, and the unit expression that cannot be concluded as p ′ <p can also be determined as a useful unit expression.
[0064]
In the second embodiment of the present invention, first, the general appearance probability p of the expression “area” and the number of appearances n of each unit expression are obtained as in the first embodiment. In the example of “square meter”, for example, n = 17,510.
[0065]
Of these n patterns, the quantity expression and the “area” co-occurrence with the pattern ““ area is ”+ quantity expression” is expressed as x_iAnd For example, in the case of a square meter, x_i= 41.
[0066]
Next, assuming that each trial is all independent, the “area” of the general occurrence probability p is x_iThe probability (test probability) P that appears less than once is obtained. In other words, “square meter” appears n times, and the “area” that appears at a probability of p at one time is co-occurred with this “n square meters”._iThe probability (test probability) P that is less than or equal to the number of times is obtained.
[0067]
Here, the appearance probability p ′ of “area” when co-occurring with “square meter” and the general appearance probability p of “area” when there is no such condition are the same, although they may be inherently different. The general appearance probability p of “area” when there is no such condition is used as the appearance probability p ′ of “area” when co-occurring with “square meter”.
[0068]
Therefore, if the value of the test probability P is sufficiently small, the hypothesis that p ′ = p is rejected, and it can be determined that p ′ and p are different. In the embodiment of the present invention, “x_iWhen the test probability P is obtained as in the case of “no more times”, only one region of the probability distribution is used. Furthermore, the area used is “x_iSince it is a left one-sided test using the smaller region such as “less than or equal to times”, if the value of P is sufficiently small, it can be concluded that p ′ <p. However, this determination does not necessarily mean that the determination is correct, and only the probability P is a determination that may be erroneous.
[0069]
Unnecessary unit representation removing means 232 has a sufficiently small value of P, and the occurrence probability p ′ of “area” when co-occurring with “square meter” is a general appearance of “area” when there is no such condition. If it can be concluded that the probability is smaller than p, it is determined that the co-occurrence of “square meter” and “area” is a coincidence, and “square meter” is determined as an unnecessary unit representation, and is removed from the unit representation database 18.
[0070]
On the other hand, when the value of the test probability P is not sufficiently small and co-occurs with “square meter”, the appearance probability p ′ of “area” is the general appearance probability p of “area” when there is no such condition. If it is not possible to conclude that it is smaller than this, the co-occurrence of “square meter” and “area” is not a coincidence but is likely to be an inevitable co-occurrence. Then, using the unit expression judged to be useful, the quantity expression estimating means 24 estimates the quantity expression as a solution to the question sentence.
[0071]
In summary, the smaller the value of P, the more unnecessary unit expression is judged, and the larger the value of P, the more useful unit expression is judged.
[0072]
Actually calculating the value of P in the above example,
Square meter: 1.00000000
Hectare: 1.00000000
Square kilometers: 1.00000000
Double: 0.95148679
Minute: 0.00008198
Tatami mat: 0.99588861
No .: 0.30698656
Square centimeter: 0.99999993
It becomes.
[0073]
This calculation can be done using binomial distribution theory. Here, calculation was performed using the subroutine pbinom of the free software Math-CDF-0.1.
[0074]
Here, for example, when 0.99 is used as a threshold value that means that the value of P is sufficiently small, a unit expression having a value of P less than 0.99 is an unnecessary unit expression and an unnecessary co-occurrence word. Judging that there is a unit expression with a P value of 0.99 or more, it is highly likely that it will co-occur with the main noun “area”, and that it is a useful unit expression and is a valid co-occurrence word. To do. That is,
Square meter: 1.00000000
Hectare: 1.00000000
Square kilometers: 1.00000000
Tatami mat: 0.99588861
Square centimeter: 0.9999993
Is a valid co-occurrence word,
Double: 0.95148679
Minute: 0.00008198
No .: 0.30698656
Is considered an invalid co-occurrence word.
[0075]
Actually, square meter: 1.00000000, hectare: 1.00000000, square kilometer: 1.00000000, tatami mat: 0.99588861, square centimeter: 0.9999993 can be used as a unit representation of area, but times: 0.951448679, minutes: 0 00008198, No .: 0.30698656 is an illegal expression as a unit expression of area.
[0076]
Only the valid unit expressions obtained in this way are used to estimate the quantity expression as the solution of the question sentence.
[0077]
FIG. 5 is a diagram showing an example of the details of the processing flow for removing unnecessary unit expressions in the second embodiment of the present invention, and shows steps S41 to S52 which are details of step S3 in FIG. Is. In the second embodiment of the present invention, steps other than step 3 in FIG. 3 are the same as those in the first embodiment.
[0078]
As shown in FIG. 5, in the second embodiment of the present invention, in step S49, main nouns and unit representations w in the calculated probability distribution._iThe number of occurrences of and co-occurs is x_iA probability P (test probability) that is less than or equal to the number of times, and if the value of P is less than the threshold value in step S52, it is determined that the unit expression is unnecessary and is removed from the unit expression database; Unlike the first embodiment of the present invention shown in FIG. 4 in that it is determined that the unit expression is useful when value <threshold, the other steps in FIG. 5 are the same as in the first embodiment of the present invention. It is.
[0079]
Based on the gist of the present invention, various modifications are possible as follows.
[0080]
In the first embodiment and the second embodiment of the present invention, the general appearance probability is p, and the appearance probability of “area” as the main noun in the corpus is used. As a general appearance probability p in, the unit expression w_iA similar test can be performed even when n is the number of occurrences of the main noun “area” in the corpus.
[0081]
That is, the left one-sided test is performed with the hypothesis that the appearance probability p ′ of “square meter” when co-occurring with “area” is equal to the general appearance probability p of “square meter”, and “square meter” appears n times The number of times it co-occurs with the main noun “area” is x_iIf the value of the test probability P that is less than or equal to 41 times (for example, 41 times) is less than a threshold value (for example, 0.99), it is concluded that p ′ <p and “square meter” is determined to be an unnecessary unit expression. When the value of the test probability P is equal to or greater than the threshold value and it cannot be concluded that p ′ <p, it can be determined as a useful unit expression.
[0082]
For example, in the case of “square meter”,
p = 17510/409502077, n = 7933,
P = 1.00000000> 0.99
Since p ′ <p cannot be determined, “square meter” can be determined to be a useful unit expression.
[0083]
FIG. 6 shows an example of the details of the processing flow for removing unnecessary unit expressions according to the third embodiment of the present invention that performs such left-side test. FIG. 6 shows steps S81 to S92 which are details of step S3 in FIG. As shown in FIG. 6, in the third embodiment of the present invention, in step S84, each unit representation w in the corpus is displayed._iIs different from the second embodiment of the present invention shown in FIG. 5 in that the general appearance probability p is calculated and the appearance frequency n in the corpus of the main noun is calculated in step S86.
[0084]
Of course, in the present invention, a right one-sided test is performed with the hypothesis that the appearance probability p ′ of “square meter” when co-occurring with “area” is equal to the general appearance probability p of “square meter”. Is the number of times that appears co-occurring with the “area” that appears n times_iIf the value of the test probability P that is greater than or equal to the number of times is less than the threshold value, it is concluded that p ′> p and is a useful unit expression, the value of the test probability P is greater than the threshold value, and p If it cannot be concluded that '> p, it can also be determined as an unnecessary unit expression.
[0085]
FIG. 7 shows an example of the details of the processing flow for removing unnecessary unit expressions according to the fourth embodiment of the present invention that performs such right-side test. FIG. 7 shows steps S61 to S72 which are details of step S3 of FIG. As shown in FIG. 7, in the fourth embodiment of the present invention, each unit representation w in the corpus in step S64._i4 is different from the first embodiment of the present invention shown in FIG. 4 in that the general appearance probability p is calculated and the appearance frequency n in the corpus of the main noun is calculated in step S66.
[0086]
In the present invention, a test using the following hypergeometric distribution can also be performed.
[0087]
What is hypergeometric distribution?
hg (N, k, n, m) = C (K, m) × C (N−K, nm) / C (N, n)
The distribution can be expressed in the form of
[0088]
However, C is a symbol C (A, B) = A! / B! / (BA)! And hg (N, k, n, m) is "When there are K red balls in N balls, a red ball is exactly m in n balls taken out arbitrarily. “Probability of being included”.
[0089]
Here, hgs (N, K, n, k) = Σhg (N, K, n, m)
Then, hgs (N, K, n, k) is "When there are K red balls in N balls, there are k or more red balls in n balls taken out arbitrarily. Probability. Note that Σhg (N, K, n, m) means the sum of hg (N, k, n, m) for m where m ≧ k.
[0090]
Here, N, K, n, and k are interpreted as follows.
[0091]
N: Size of corpus
K: Appearance frequency of “area”
n: Appearance frequency of a certain unit expression
k: “Area” and the number of co-occurrence of a unit expression
According to the above interpretation, hgs (N, K, n, k) is obtained by extracting “n” unit representations of interest when there are K representations of “area” in an N-size corpus. , Of the n unit representations, it means “probability of occurrence of an event where the unit representation co-occurs” with “k” or more.
[0092]
In the embodiment of the present invention, the case where p = K / N is regarded as a general appearance probability of “area” is treated as being divided into K and N without putting them into K / N form. It corresponds to that. In the method using the hypergeometric distribution, there are few assumptions and the accuracy is high with a method with few approximations because K and N are not divided into p = K / N.
[0093]
In this method, the smaller the hgs (N, K, n, k), the more appropriate the co-occurrence of “area” and unit expression can be determined. The larger the hgs (N, K, n, k), the “ Co-occurrence of “area” and unit expression is a coincidence and can be determined to be invalid.
[0094]
Therefore, in the present invention, the unnecessary unit expression removing unit 232 determines that a unit expression whose hgs (N, K, n, k) is smaller than a certain threshold value is a useful unit expression, and represents the quantity expression as a solution to the question sentence. It is used for estimation, and hgs (N, K, n, k) is determined to be an unnecessary unit expression that is equal to or higher than a certain threshold and is removed.
[0095]
In the fifth embodiment of the present invention, right-side test is performed using the hypergeometric distribution described above. FIG. 9 is a diagram showing an example of the details of the processing flow for removing unnecessary unit expressions in the fifth embodiment of the present invention, and shows steps S101 to S111 which are details of step S3 of FIG. It is. The other steps in FIG. 3 are the same as those in the first embodiment of the present invention.
[0096]
As shown in FIG. 9, in the fifth embodiment of the present invention, it is determined whether the answer to the question sentence is a quantity expression (step S101). If the answer is not a quantity expression, the process goes to step S5 in FIG. If the answer is a quantity expression, it is determined whether the question sentence has a unit expression (step S102).
[0097]
If the question sentence has a unit expression, the process is terminated and the process proceeds to step S4 in FIG. 3. If the question sentence does not have a unit expression, the main noun in the question sentence is determined from the question sentence information. Is certified (step S103).
[0098]
Next, the appearance frequency K of the main noun in the corpus is calculated (step S104). Then, a combination pattern in which the main noun and the quantity expression have a dependency relationship is extracted from the corpus, and one or a plurality of unit expressions w attached to the quantity expression from the extracted combination pattern._iIs extracted and recorded in the unit representation database 18 (step S105). And each extracted unit representation w_iThe appearance frequency n in the corpus is calculated (step S106).
[0099]
Next, each unit expression w_iAnd the number of times k and the main noun appear together are calculated (step S107). Next, the main nouns and each unit expression w in the hypergeometric distribution_iThe test probability P = hgs (N, K, n, k) that the number of occurrences of and co-occurs is k times or more is calculated (step S108). N is the size of the corpus.
[0100]
Then, it is determined whether P value <threshold value (step S109). If P value <threshold value, it is determined as a useful unit expression (step S111). The unit representation is determined to be removed from the unit representation database (step S110).
[0101]
In the present invention, the left one-sided test can also be performed using the hypergeometric distribution. FIG. 10 shows an example of the details of the processing flow for removing unnecessary unit expressions in the sixth embodiment of the present invention in which the left one-side test is performed using the hypergeometric distribution. FIG. 10 shows steps S121 to S131 which are details of step S3 in FIG. In the sixth embodiment of the present invention, the main noun and each unit expression w in the hypergeometric distribution in step S128 of FIG._iThis is useful when calculating the test probability P = hgs (N, K, n, k) where the number of occurrences of and co-occurrence is equal to or less than k times, and in step S130, if the value of P is not less than the threshold value. In the fifth embodiment of the present invention shown in FIG. 9, the unit representation is determined to be an unnecessary unit representation and removed from the unit representation database if the value of P <threshold in step S131. Different from form.
[0102]
In the fifth embodiment or the sixth embodiment of the present invention, K = the frequency of appearance of the main noun and n = the frequency of appearance of each unit expression. Yes, K = frequency of appearance of each unit expression, and n = number of appearances of main nouns.
[0103]
In all the embodiments of the present invention described above, this corresponds to performing a certain ratio test. Is the co-occurrence of the main noun and the unit expression appearing at a number greater than or greater than the number of co-occurrence of the main noun and the unit expression expected from the general occurrence probability of the main noun and the unit expression? I will test it.
[0104]
The present invention can also be realized by a statistical test method capable of performing a ratio test such as AIC or Z-score.
[0105]
It should be noted that the present invention can also be used for assisting the creation of rules or tables manually described in the prior art.
[0106]
【The invention's effect】
By using the present invention, when a solution to a question sentence is a quantity expression, a method and a system capable of accurately determining whether or not a unit expression appropriate as a unit expression attached to the quantity expression is provided. It becomes possible.
[Brief description of the drawings]
FIG. 1 is a diagram illustrating an example of a configuration of a question answering system using a statistical test.
FIG. 2 is an example of a configuration diagram of an unnecessary co-occurrence word removal system.
FIG. 3 is a diagram showing an example of a question answering process flow using a statistical test.
FIG. 4 is a diagram illustrating an example of details of an unnecessary unit expression removal processing flow;
FIG. 5 is a diagram illustrating an example of details of an unnecessary unit expression removal processing flow;
FIG. 6 is a diagram illustrating an example of details of an unnecessary unit expression removal processing flow;
FIG. 7 is a diagram illustrating an example of details of an unnecessary unit expression removal processing flow;
FIG. 8 is a diagram showing a binomial distribution.
FIG. 9 is a diagram illustrating an example of details of an unnecessary unit expression removal processing flow;
FIG. 10 is a diagram illustrating an example of details of an unnecessary unit representation removal processing flow;
[Explanation of symbols]
1 Question answering system using statistical tests
11 Unnecessary co-occurrence word removal system
12 Question text input part
13 Keyword extraction / information search section
14 answer extraction part
15 Answer output section
16 database
17 Expression guessing unit
18 Unit representation database
20 Quantity expression judging means
21 pronoun recognition means
22 Unit expression extraction means
23 Test means
24 Quantity expression guessing means
230 Appearance probability calculation means
231 Probability distribution calculation means
232 Means for removing unnecessary unit expressions

Claims (7)

A question answering system that inputs a question sentence in a natural language, generates a solution by collation with sentences in digitized text information stored in a database, and outputs it.
A question sentence input means for inputting the question sentence;
When the answer to the input question sentence is a quantity expression, a statistical test is applied to the predetermined rule using the morphological analysis and sentence pattern included in the question sentence that appears in the digitized text information. A unit expression is extracted from a combination pattern of a specific noun and quantity expression that are recognized based on it, and a probability value indicating the possibility that the extracted unit expression appears together with the specific noun in the question sentence is calculated. Means for determining a unit expression used to determine a quantity expression as a solution to the question sentence according to whether the calculated probability value is equal to or greater than a threshold, and determining a quantity expression including the determined unit expression;
A question answering system using a statistical test, comprising: answer extracting means for extracting a solution of the question sentence using the determined quantity expression. A question answering system that inputs a question sentence in a natural language, generates a solution by collation with sentences in digitized text information stored in a database, and outputs it.
A question sentence input means for inputting the question sentence;
A quantity expression judging means for judging whether or not a solution of the question sentence is a quantity expression according to whether or not the question sentence includes a word for inquiring a quantity expression;
Means for identifying a specific noun included in the question sentence based on a predetermined rule using morphological analysis and sentence patterns ;
When the quantity expression determining means determines that the answer to the question sentence is a quantity expression, a combination pattern in which the identified specific noun and the quantity expression are in a dependency relationship is extracted from the digitized text information; Unit expression extracting means for extracting one or a plurality of unit expressions attached to the quantity expression from the extracted combination pattern;
Appearance probability calculating means for calculating the number of appearances of the specific noun and the extracted unit expressions in the digitized text information, and calculating a general appearance probability of the specific noun based on the calculated number of appearance information When,
Based on the calculated general appearance probability of the specific noun, the probability of the number of appearances of the specific noun co-occurring with the unit representation in the digitized text information A probability distribution calculating means for calculating a distribution;
The number of times that the specific noun in the calculated probability distribution co-occurs with each unit expression is attached to the quantity expression in the combination pattern in which the specific noun and the quantity expression have a dependency relationship. A test probability that is a probability that is equal to or greater than the number of appearances of unit expressions is calculated, and unit expressions whose calculated test probabilities are equal to or higher than a preset threshold are removed from one or a plurality of unit expressions extracted by the unit expression extraction unit. Means for removing unnecessary unit expressions,
Means for determining a quantity expression including a unit expression whose test probability is less than a threshold;
Means for extracting a keyword from the question sentence, and extracting information describing a solution of the question sentence from the digitized text information using the extracted keyword;
A question answering system using a statistical test, comprising: answer extracting means for extracting the answer of the question sentence from information in which the answer of the question sentence is described using the determined quantity expression. A question answering system that inputs a question sentence in a natural language, generates a solution by collation with sentences in digitized text information stored in a database, and outputs it.
A question sentence input means for inputting the question sentence;
A quantity expression judging means for judging whether or not a solution of the question sentence is a quantity expression according to whether or not the question sentence includes a word for inquiring a quantity expression;
Means for identifying a specific noun included in the question sentence based on a predetermined rule using morphological analysis and sentence patterns ;
When the quantity expression determining means determines that the answer to the question sentence is a quantity expression, a combination pattern in which the identified specific noun and the quantity expression are in a dependency relationship is extracted from the digitized text information; Unit expression extracting means for extracting one or a plurality of unit expressions attached to the quantity expression from the extracted combination pattern;
Appearance probability calculating means for calculating the number of appearances of the specific noun and the extracted unit expressions in the digitized text information, and calculating a general appearance probability of the unit expressions based on the calculated number of appearance information When,
Based on the general appearance probability of each unit expression, a probability distribution of the number of times each unit expression appears together with the specific noun among patterns in which the specific noun appears in the digitized text information A probability distribution calculating means for calculating;
The number of times that each unit expression in the calculated probability distribution co-occurs with the specific noun is attached to the quantity expression in a combination pattern in which the specific noun and the quantity expression have a dependency relationship. A test probability that is a probability that is equal to or greater than the number of appearances of the quantity expression that co-occurs with each unit expression is calculated, and the calculated test probability is one or more of the unit expressions extracted by the unit expression extraction unit Unnecessary unit representation removing means for removing unit representations that are equal to or greater than a preset threshold value;
Means for determining a quantity expression including a unit expression whose test probability is less than a threshold;
Means for extracting a keyword from the question sentence, and extracting information describing a solution of the question sentence from the digitized text information using the extracted keyword;
A question answering system using a statistical test, comprising: answer extracting means for extracting the answer of the question sentence from information in which the answer of the question sentence is described using the determined quantity expression. A question answering system that inputs a question sentence in a natural language, generates a solution by collation with sentences in digitized text information stored in a database, and outputs it.
A question sentence input means for inputting the question sentence;
A quantity expression judging means for judging whether or not a solution of the question sentence is a quantity expression according to whether or not the question sentence includes a word for inquiring a quantity expression;
Means for identifying a specific noun included in the question sentence based on a predetermined rule using morphological analysis and sentence patterns ;
When the quantity expression determining means determines that the answer to the question sentence is a quantity expression, a combination pattern in which the identified specific noun and the quantity expression are in a dependency relationship is extracted from the digitized text information; Unit expression extracting means for extracting one or a plurality of unit expressions attached to the quantity expression from the extracted combination pattern;
Appearance probability calculating means for calculating the number of appearances of the specific noun and the extracted unit expressions in the digitized text information, and calculating a general appearance probability of the specific noun based on the calculated number of appearance information When,
Based on the calculated general appearance probability of the specific noun, the probability of the number of appearances of the specific noun co-occurring with the unit representation in the digitized text information A probability distribution calculating means for calculating a distribution;
The number of times that the specific noun in the calculated probability distribution co-occurs with each unit expression is attached to the quantity expression in the combination pattern in which the specific noun and the quantity expression have a dependency relationship. A test probability that is a probability that the number of occurrences of unit expressions is less than or equal to the number of occurrences is calculated, and unit expressions whose calculated test probabilities are less than a preset threshold are removed from one or more unit expressions extracted by the unit expression extraction unit Means for removing unnecessary unit expressions,
Means for determining a quantity expression including a unit expression having a test probability equal to or greater than a threshold;
Means for extracting a keyword from the question sentence, and extracting information describing a solution of the question sentence from the digitized text information using the extracted keyword;
A question answering system using a statistical test, comprising: answer extracting means for extracting the answer of the question sentence from information in which the answer of the question sentence is described using the determined quantity expression. A question answering system that inputs a question sentence in a natural language, generates a solution by collation with sentences in digitized text information stored in a database, and outputs it.
A question sentence input means for inputting the question sentence;
A quantity expression judging means for judging whether or not a solution of the question sentence is a quantity expression according to whether or not the question sentence includes a word for inquiring a quantity expression;
Means for identifying a specific noun included in the question sentence based on a predetermined rule using morphological analysis and sentence patterns ;
When the quantity expression determining means determines that the answer to the question sentence is a quantity expression, a combination pattern in which the identified specific noun and the quantity expression are in a dependency relationship is extracted from the digitized text information; Unit expression extracting means for extracting one or a plurality of unit expressions attached to the quantity expression from the extracted combination pattern;
Appearance probability calculating means for calculating the number of appearances of the specific noun and the extracted unit expressions in the digitized text information, and calculating a general appearance probability of the unit expressions based on the calculated number of appearance information When,
Based on the general appearance probability of each unit expression, a probability distribution of the number of times each unit expression appears together with the specific noun among patterns in which the specific noun appears in the digitized text information A probability distribution calculating means for calculating;
The number of times that each unit expression in the calculated probability distribution co-occurs with the specific noun is attached to the quantity expression in a combination pattern in which the specific noun and the quantity expression have a dependency relationship. A test probability that is a probability that is equal to or less than the number of appearances of the quantity expression appearing together with each unit expression is calculated, and the calculated test probability is one or more of the unit expressions extracted by the unit expression extraction unit. Unnecessary unit representation removing means for removing unit representations of which is less than a preset threshold;
Means for determining a quantity expression including a unit expression having a test probability equal to or greater than a threshold;
Means for extracting a keyword from the question sentence, and extracting information describing a solution of the question sentence from the digitized text information using the extracted keyword;
A question answering system using a statistical test, comprising: answer extracting means for extracting the answer of the question sentence from information in which the answer of the question sentence is described using the determined quantity expression. A question answering system that inputs a question sentence in a natural language, generates a solution by collation with sentences in digitized text information stored in a database, and outputs it.
A question sentence input means for inputting the question sentence;
A quantity expression judging means for judging whether or not a solution of the question sentence is a quantity expression according to whether or not the question sentence includes a word for inquiring a quantity expression;
Means for identifying a specific noun included in the question sentence based on a predetermined rule using morphological analysis and sentence patterns ;
When the quantity expression determining means determines that the answer to the question sentence is a quantity expression, a combination pattern in which the identified specific noun and the quantity expression are in a dependency relationship is extracted from the digitized text information; Unit expression extracting means for extracting one or a plurality of unit expressions attached to the quantity expression from the extracted combination pattern;
Means for calculating the number of appearances of the specific noun and the extracted unit expressions in the digitized text information;
Using the hypergeometric distribution, the number of occurrences of the specific noun in the hypergeometric distribution co-occurring with the unit expressions is the quantity expression in the combination pattern in which the specific noun and the quantity expression have a dependency relationship. A test probability that is a probability that is equal to or greater than the number of appearances of each unit expression attached is calculated, and the calculated test probability is equal to or greater than a preset threshold among one or a plurality of unit expressions extracted by the unit expression extraction unit. Means for removing unnecessary unit expressions for removing unit expressions of
Means for determining a quantity expression including a unit expression whose test probability is less than a threshold;
Means for extracting a keyword from the question sentence, and extracting information describing a solution of the question sentence from the digitized text information using the extracted keyword;
A question answering system using a statistical test, comprising: answer extracting means for extracting the answer of the question sentence from information in which the answer of the question sentence is described using the determined quantity expression. A question answering system that inputs a question sentence in a natural language, generates a solution by collation with sentences in digitized text information stored in a database, and outputs it.
A question sentence input means for inputting the question sentence;
A quantity expression judging means for judging whether or not a solution of the question sentence is a quantity expression according to whether or not the question sentence includes a word for inquiring a quantity expression;
Means for identifying a specific noun included in the question sentence based on a predetermined rule using morphological analysis and sentence patterns ;
When the quantity expression determining means determines that the answer to the question sentence is a quantity expression, a combination pattern in which the identified specific noun and the quantity expression are in a dependency relationship is extracted from the digitized text information; Unit expression extracting means for extracting one or a plurality of unit expressions attached to the quantity expression from the extracted combination pattern;
Means for calculating the number of appearances of the specific noun and the extracted unit expressions in the digitized text information;
Using the hypergeometric distribution, the number of occurrences of the specific noun in the hypergeometric distribution co-occurring with the unit expressions is the quantity expression in the combination pattern in which the specific noun and the quantity expression have a dependency relationship. A test probability that is a probability that is less than or equal to the number of appearances of each unit expression attached is calculated, and the calculated test probability is less than a preset threshold among one or more unit expressions extracted by the unit expression extraction unit Means for removing unnecessary unit expressions for removing unit expressions of
Means for determining a quantity expression including a unit expression having a test probability equal to or greater than a threshold;
Means for extracting a keyword from the question sentence, and extracting information describing a solution of the question sentence from the digitized text information using the extracted keyword;
A question answering system using a statistical test, comprising: answer extracting means for extracting the answer of the question sentence from information in which the answer of the question sentence is described using the determined quantity expression.

Images