JP5622310B2 - Mutual machine learning device, mutual machine learning method, and program - Google Patents

Description

  The present invention relates to a mutual machine learning device that performs mutual machine learning using two machine learnings.

  Conventionally, in machine learning using teacher data, a method called mutual machine learning in which a plurality of machine learnings are combined has been proposed (see, for example, Non-Patent Documents 1 and 2). In the technique of Non-Patent Document 1, a plurality of machine learners are generated by manually dividing the features used at the time of learning for the same word pair, and reliable results obtained from one machine learner are separated from each other. It is used as learning data for machine learners. In Non-Patent Document 2, a machine learning device is generated for each language for different languages, and a reliable result obtained from one machine learning device is used as learning data of another machine learning device.

Avrim Blum, Tom Mitchell, “Combining Labeled and Unlabeled Data with Co-Training”, In CALT '98: Proceedings of the evening concealment conference. 92-100, 1998 Jong-Hoon Oh, Kiyotaka Uchimoto, Kentaro Torisawa, “Bilingual Co-Training for Monopoly-Relation Acquisition”, In Proc of ACL-NLP9, In Proc. 432-440, 2009

  However, the method of Non-Patent Document 1 has a problem that the processing targets handled by a plurality of machine learners must be the same, and different processing targets cannot be handled. In addition, it has been desired to realize mutual machine learning with higher accuracy than conventional mutual machine learning.

  The present invention has been made to solve the above-described problems, and provides a mutual machine learning device and the like that can handle different processing targets by a plurality of machine learners and can realize highly accurate machine learning. With the goal.

  In order to achieve the above object, a mutual machine learning device according to the present invention provides a plurality of first relationship pair candidates that are candidates for a pair of language expressions having a semantic relationship extracted from a first corpus by a first method. And a common pair that is extracted from the second corpus by a second method different from the first method and is common to a plurality of second relationship pair candidates that are candidates for language expression pairs having a semantic relationship. Common to a certain common pair, a plurality of first unrelated pair candidates that are extracted from the first corpus, and that are candidate language expression pairs that have no semantic relationship, and a plurality of second relationship pair candidates A plurality of second unrelated pair candidates and a plurality of first relationship pair candidates extracted from the second common corpus, a virtual common pair that is a common pair, and language expression pairs that have no semantic relationship And A common pair storage unit that stores a virtual common pair that is a common pair that passes through, and first teacher data that is used in machine learning related to classification of whether or not the first relationship pair candidate has a semantic relationship. Machine learning is performed using the first learning data storage unit in which the learning data is stored, and the first learning data, and the genuine common pair and the virtual common pair have a semantic relationship using the result of the machine learning. A first classifying unit that classifies whether or not the second learning data is stored, and second learning data that is teacher data used in machine learning related to the classification of whether or not the second relationship pair candidate has a semantic relationship. Machine learning using the learning data storage unit and the second learning data, and using the result of the machine learning, the genuine common pair and the virtual common pair have a semantic relationship. A second classification unit that classifies whether or not the classification result of the common pair and the common pair are classified according to at least one of the classification result and the certainty of the common pair by the classification of the first classification unit In addition to the learning data, the common pair and the classification result related to the common pair are added to the first learning data according to at least one of the classification result and the certainty factor of the common pair by the classification of the second classification unit. An addition unit, and machine learning and classification by the first and second classification units and addition of learning data by the addition unit are repeatedly executed.

  With such a configuration, the first and second classification units can extract different processing targets, that is, pairs of language expressions extracted by the first method, and a second method different from the first method. It is possible to handle a pair of linguistic expressions, and to perform mutual machine learning using a virtual common pair, so that it is possible to realize machine learning with higher accuracy. As a result, linguistic expression pairs having a semantic relationship can be obtained with high accuracy by classifying linguistic expression pairs having a semantic relationship using the machine learning result.

The mutual machine learning device according to the present invention includes a plurality of first relationship pair candidates that are extracted from a first corpus by a first method and that are candidate language expression pairs having a semantic relationship. Genuine common that is a common pair that is extracted from a second corpus by a second method different from the above method and is common to a plurality of second relationship pair candidates that are candidates for pairs of language expressions having the semantic relationship A pair, a plurality of first relationship pair candidates, a plurality of second relationship pair candidates, and a pair of language expressions that are extracted from the first corpus and have no semantic relationship. Among the plurality of first irrelevant pair candidates and the plurality of second irrelevant pair candidates that are extracted from the second corpus and that are candidate language expression pairs that do not have the semantic relationship, the common common pair A common pair storage unit that stores a virtual common pair that is a common pair that is not a pair, and teacher data used in machine learning related to classification of whether the first relation pair candidate has the semantic relationship A first learning data storage unit that stores certain first learning data and machine learning using the first learning data, and using the result of the machine learning, the genuine common pair and the virtual common A first classification unit that classifies whether a pair has the semantic relationship, and teacher data used in machine learning regarding classification of whether the second relationship pair candidate has the semantic relationship Machine learning is performed using a second learning data storage unit in which second learning data is stored and the second learning data, and the result of the machine learning is used to perform the learning. Depending on at least one of a second classification unit that classifies whether the virtual common pair and the virtual common pair have the semantic relationship, and a classification result of the common pair and a certainty factor according to the classification of the first classification unit Then, the common pair and the classification result related to the common pair are added to the second learning data, and the common pair is determined according to at least one of the classification result and the certainty factor of the common pair by the classification of the second classification unit. And an addition unit that adds the classification result related to the common pair to the first learning data, and machine learning and classification by the first and second classification units and addition of the learning data by the addition unit are repeated. To be executed.
With such a configuration, different processing objects can be handled and machine learning with high accuracy can be realized as in the above-described mutual machine learning device. Further, it is possible to perform processing using more virtual common pairs than the above-described mutual machine learning device.

In the mutual machine learning device according to the present invention, the adding unit adds a common pair having a high certainty factor by classification of the first classification unit and a classification result related to the common pair to the second learning data, A common pair having a high certainty factor according to the classification of the second classification unit and a classification result related to the common pair may be added to the first learning data.
With such a configuration, it is considered that a common pair classified with a high degree of certainty by one of the classification units is reliable, and thus it is considered that the learning data can be appropriately increased by such a configuration.

  In the mutual machine learning device according to the present invention, the additional unit has a high certainty factor by the classification of the first classification unit and the classification related to the common pair and the common pair having the same classification result of the first and second classification units. The result is added to the second learning data, the confidence level by the classification of the second classification unit is high, and the classification result of the common pair and the common pair with the same classification result of the first and second classification units You may add to 1st learning data.

  A common pair that has the same classification result by the first and second classification units and is classified with high confidence by one classification unit is reliable regardless of the certainty by the classification of the other classification unit. it is conceivable that. Therefore, with such a configuration, it is considered that the learning data can be appropriately increased by adding the common pair to the learning data of the other classification unit.

  In the mutual machine learning device according to the present invention, the adding unit obtains a common pair having a high certainty factor by the classification of the first classification unit and a low certainty factor by the classification of the second classification unit and a classification result related to the common pair. In addition to the second learning data, a common pair having a high certainty factor by the classification of the second classification unit and a low certainty factor by the classification of the first classification unit and a classification result related to the common pair are used as the first learning data. May be added.

  A common pair that is classified with high confidence by one classification unit and classified with low confidence by the other classification unit is considered to be reliable for classification by the former classification unit. Accordingly, with such a configuration, it is considered that the learning data can be appropriately increased by adding the common pair to the learning data of the latter classification unit.

In the mutual machine learning device according to the present invention, a first relationship pair candidate storage unit that stores a plurality of first relationship pair candidates, and a first irrelevant pair candidate storage unit that stores a plurality of first irrelevant pair candidates; A second relationship pair candidate storage unit that stores a plurality of second relationship pair candidates; a second relationship pair candidate storage unit that stores a plurality of second unrelated pair candidates; and a plurality of first relationship pair candidates and a plurality Using the second relationship pair candidates, a genuine common pair is acquired and stored in the common pair storage unit, and a plurality of first relationship pair candidates, a plurality of second relationship pair candidates, and a plurality of first unrelated pair candidates An acquisition unit that acquires a virtual common pair using a plurality of second unrelated pair candidates and stores the virtual common pair in a common pair storage unit may be further provided.
With such a configuration, the mutual machine learning apparatus can also perform processing for acquiring a genuine common pair and a virtual common pair.

In the mutual machine learning device according to the present invention, a first corpus storage unit that stores a first corpus, a second corpus storage unit that stores a second corpus, and a plurality of first corpuses. First extraction of first relation pair candidates extracted and stored in the first relation pair candidate storage unit, and a plurality of first unrelated pair candidates extracted from the first corpus and stored in the first unrelated pair candidate storage unit And a plurality of second related pair candidates extracted from the second corpus and stored in a second related pair candidate storage unit, and a plurality of second unrelated pair candidates extracted from the second corpus A second extraction unit that accumulates in the candidate storage unit.
With such a configuration, the mutual machine learning apparatus can also perform processing for extracting the first relationship pair candidate and the like from the first and second corpora.

In the mutual machine learning device according to the present invention, the first classifying unit classifies the plurality of first relationship pair candidates after the machine learning and the repetition of the classification and the addition of the learning data, The classification unit may perform classification on a plurality of second relationship pair candidates after machine learning and the repetition of classification and addition of learning data.
With such a configuration, since the first relation pair candidate and the second relation pair candidate are classified using the result of machine learning of the learning data to which the common pair is added as described above, the classification with higher accuracy is performed. Will be able to do.

In the mutual machine learning device according to the present invention, the first corpus may be a structured corpus, and the second corpus may be an unstructured natural language sentence corpus.
With such a configuration, a pair of language expressions acquired from a structured corpus and a pair of language expressions acquired from an unstructured corpus can be handled as different processing targets.

  In the mutual machine learning device according to the present invention, the semantic relationship may be an upper-lower relationship.

  According to the mutual machine learning apparatus and the like according to the present invention, different processing objects can be handled, and machine learning with higher accuracy can be realized.

1 is a block diagram showing the configuration of a mutual machine learning device according to Embodiment 1 of the present invention. The flowchart which shows the operation | movement of the mutual machine learning apparatus by the embodiment The flowchart which shows the operation | movement of the mutual machine learning apparatus by the embodiment The figure for demonstrating the common pair in the embodiment The figure for demonstrating the structured corpus in the embodiment The figure which shows the experimental result in the same embodiment The figure which shows the experimental result in the same embodiment Schematic diagram showing an example of the appearance of the computer system in the embodiment The figure which shows an example of a structure of the computer system in the embodiment

  Hereinafter, a mutual machine learning apparatus according to the present invention will be described using embodiments. In the following embodiments, components and steps denoted by the same reference numerals are the same or equivalent, and repetitive description may be omitted.

(Embodiment 1)
A mutual machine learning apparatus according to Embodiment 1 of the present invention will be described with reference to the drawings.
FIG. 1 is a block diagram showing a configuration of a mutual machine learning device 1 according to the present embodiment. The mutual machine learning device 1 according to the present embodiment includes a first corpus storage unit 11, a second corpus storage unit 12, a first extraction unit 13, a second extraction unit 14, and a first relationship pair. The candidate storage unit 15, the first irrelevant pair candidate storage unit 16, the second relationship pair candidate storage unit 17, the second irrelevant pair candidate storage unit 18, the acquisition unit 19, the common pair storage unit 20, and the first Learning data storage unit 21, second learning data storage unit 22, first classification unit 23, second classification unit 24, addition unit 25, first relationship pair storage unit 26, second A relationship pair storage unit 27.

  The first corpus storage unit 11 stores the first corpus, and the second corpus storage unit 12 stores the second corpus. The first and second corpora may be different types of corpuses or the same type of corpora. In the former case, for example, the first corpus may be a structured corpus and the second corpus may be an unstructured natural language sentence corpus. In this embodiment, the case will be mainly described. Here, the structured corpus is a corpus in which a document included in the corpus has some structure such as a hierarchical structure or a tree structure. The structured corpus may be, for example, encyclopedia information or other information. Such information includes, for example, a hierarchical structure such as a title, a section, a subsection, a list, and a tree structure. As an encyclopedia corpus, for example, there is WIKIPEDIA (registered trademark) published on the web. An unstructured natural language corpus is a corpus that does not have a hierarchical or tree structure, such as newspaper information, novel information, or unstructured web information. Etc. Information including a document of a natural language sentence is usually a corpus of unstructured natural language sentences. Therefore, even a corpus having a structure can be used as an unstructured corpus by not using the structure. As unstructured web information, for example, there is a large-scale corpus of Japanese web documents targeted by “Search Engine Research Infrastructure TSUBAKI” or the like.

  The process of storing the corpus in the first corpus storage unit 11 and the second corpus storage unit 12 does not matter. For example, a corpus may be stored in the first corpus storage unit 11 or the like via a recording medium, or a corpus transmitted via a communication line or the like may be stored in the first corpus storage unit 11 or the like. It may be memorized.

  The first extraction unit 13 extracts a plurality of first relationship pair candidates from the first corpus stored in the first corpus storage unit 11 and accumulates them in the first relationship pair candidate storage unit 15. Further, the first extraction unit 13 extracts a plurality of first irrelevant pair candidates from the first corpus stored in the first corpus storage unit 11 and accumulates them in the first irrelevant pair candidate storage unit 16. The first relationship pair candidate is a linguistic expression pair candidate having a certain semantic relationship. The first relationship pair candidate is a candidate for a linguistic expression pair having the semantic relationship, and thus does not necessarily have the semantic relationship. The first irrelevant pair candidate is a linguistic expression pair candidate that does not have the semantic relationship. In addition, the first unrelated pair candidate is a candidate for a language expression pair that does not have the semantic relationship, and thus does not necessarily have the semantic relationship. The semantic relationship may be, for example, an upper-lower relationship (for example, drink and coffee), a causal relationship (for example, heavy rain and flood), or a whole-part relationship (for example, a person) Or hand, car and tire), rival or synonym (eg, up and down), product and manufacturer (eg, vacuum cleaner and company A), event It may be a relationship between the method and the method (for example, explosion and bomb), the relationship between the event and the tool (for example, lesson and textbook), or the relationship between the event and the prevention (for example, illness and medicine). It may be a relationship between materials and materials (for example, cans and aluminum), a relationship between famous places / buildings and places (for example, Nijo Castle and Kyoto), or other types of relationships. Good. The language expression may be, for example, a word (morpheme) or a phrase that is a sequence of words. Further, the linguistic expression may be a sequence of a plurality of words (for example, compound nouns). The first extraction unit 13 usually extracts a first relationship pair candidate that is a candidate for a pair of language expressions having any one semantic relationship. In the present embodiment, a case will be mainly described in which the first extraction unit 13 extracts a first relationship pair candidate that is a candidate for a pair of language expressions having upper and lower relationships. Further, the first extraction unit 13 may extract, as the first irrelevant pair candidate, a pair of language expressions that is considered not to have a semantic relationship that the first relation pair candidate has. Alternatively, a pair of language expressions considered to have a semantic relationship that is not the semantic relationship that the first relationship pair candidate has may be extracted as the first unrelated pair candidate. (A pair of linguistic expressions having a semantic relationship that is not a semantic relationship that the first relationship pair candidate has is a semantic relationship that the first relationship pair candidate has. Because it is thought that they would not have In the present embodiment, a case will be mainly described in which the first extraction unit 13 extracts a first irrelevant pair candidate that is a linguistic expression pair candidate having no upper-lower relationship.

  The second extraction unit 14 extracts a plurality of second relationship pair candidates from the second corpus stored in the second corpus storage unit 12 and accumulates them in the second relationship pair candidate storage unit 17. Further, the second extraction unit 14 extracts a plurality of second irrelevant pair candidates from the second corpus stored in the second corpus storage unit 12 and accumulates them in the second irrelevant pair candidate storage unit 18. The second relationship pair candidate is a linguistic expression pair candidate having the same semantic relationship as the semantic relationship considered to be possessed by the first relationship pair candidate. The second relationship pair candidate is a candidate for a linguistic expression pair having the semantic relationship, and thus does not necessarily have the semantic relationship. The second irrelevant pair candidate is a linguistic expression pair candidate that does not have the semantic relationship. Moreover, since the 2nd unrelated pair candidate is a candidate regarding the pair of the language expression which does not have the semantic relationship, it does not necessarily have the semantic relationship. The second extraction unit 14 typically extracts a second relationship pair candidate that is a linguistic expression pair candidate having any one semantic relationship. As described above, the semantic relationship is the same semantic relationship as the semantic relationship considered to be possessed by the first relationship pair candidate. In the present embodiment, a case will be mainly described in which the second extraction unit 14 extracts second relationship pair candidates that are candidates for pairs of language expressions having upper and lower relationships. Further, the second extraction unit 14 may extract, as the second unrelated pair candidate, a language expression pair that is considered not to have the semantic relationship that the second relation pair candidate has. Alternatively, a pair of language expressions that are considered to have a semantic relationship that is not the semantic relationship that the second relationship pair candidate has may be extracted as a second unrelated pair candidate. . In the present embodiment, a case will be mainly described in which the second extraction unit 14 extracts a second irrelevant pair candidate that is a candidate for a linguistic expression pair having a semantic relationship that is not an upper / lower relationship.

  It is assumed that the first and second extraction units 13 and 14 have different extraction methods in addition to different corpus from which language expression pairs are extracted. That is, the first extraction unit 13 extracts a plurality of first relationship pair candidates from the first corpus by the first method, and the second extraction unit 14 extracts a plurality of first relation pair candidates from the second corpus by the second method. A second relationship pair candidate is extracted. Note that the first method and the second method are different. Therefore, even if the types of the first and second corpora are the same, the method for extracting the first and second relationship pair candidates from each corpus is different, so the first relationship pair candidate and the second relationship pair candidate are Are of different kinds. In the present embodiment, the first method is a method for extracting first relationship pair candidates using the structure of the first corpus, and the second method is a lexico-tactic pattern (Lexico-syntactic pattern). ) Is used to extract a second related pair candidate. Details of these extraction methods will be described later.

  The obtaining unit 19 obtains a genuine common pair by using the plurality of first relationship pair candidates and the plurality of second relationship pair candidates, and accumulates them in the common pair storage unit 20. In addition, the acquisition unit 19 uses the plurality of first relationship pair candidates, the plurality of second relationship pair candidates, the plurality of first irrelevant pair candidates, and the plurality of second irrelevant pair candidates to generate a virtual common pair. Acquired and stored in the common pair storage unit 20. Genuine common pairs are a plurality of first relationship pair candidates stored in the first relationship pair candidate storage unit 15 and a plurality of second relationship pair candidates stored in the second relationship pair candidate storage unit 17. It is a common pair. Therefore, when a certain first relationship pair candidate and a certain second relationship pair candidate are pairs of the same language expression, the first relationship pair candidate (the second relationship pair candidate) becomes a genuine common pair. The virtual common pair is a plurality of first unrelated pair candidates stored in the first unrelated pair candidate storage unit 16 and a plurality of second relationship pair candidates stored in the second relationship pair candidate storage unit 17. And a plurality of second unrelated pair candidates stored in the second unrelated pair candidate storage unit 18 and a plurality of first relationship pair candidates stored in the first relationship pair candidate storage unit 15 And a common pair. Therefore, when a certain first unrelated pair candidate and a certain second related pair candidate are pairs of the same language expression, the first unrelated pair candidate (the second related pair candidate) is a virtual common pair. Further, when a certain second irrelevant pair candidate and a certain first related pair candidate are the same language expression pair, the second unrelated pair candidate (the first related pair candidate) is a virtual common pair. The genuine common pair and the virtual common pair are collectively referred to as a common pair.

The first learning data storage unit 21 stores first learning data that is teacher data used in machine learning related to the classification of whether or not the first relationship pair candidate has a semantic relationship.
The second learning data storage unit 22 stores second learning data that is teacher data used in machine learning related to the classification of whether or not the second relationship pair candidate has a semantic relationship.

  In addition, the process in which 1st learning data and 2nd learning data are memorize | stored in the 1st learning data storage part 21 and the 2nd learning data storage part 22 does not ask | require. For example, first learning data or the like may be stored in the first learning data storage unit 21 or the like via a recording medium, or first learning data transmitted via a communication line or the like. Or the like may be stored in the first learning data storage unit 21 or the like.

  The first classification unit 23 performs machine learning using the first learning data, and classifies whether the genuine common pair and the virtual common pair have a semantic relationship using the result of the machine learning. By the classification, the first classification unit 23 can obtain the classification result (whether it has a semantic relationship) and the certainty of the classification result. As will be described later, when the first learning data is added by the adding unit 25, the first classification unit 23 performs learning using the added first learning data as well. To do. The first classifying unit 23 classifies the plurality of first relationship pair candidates stored in the first relationship pair candidate storage unit 15 after repeating machine learning and classification and addition of learning data. Do. Then, the first classification unit 23 accumulates in the first relationship pair storage unit 26 the first relationship pair that is the first relationship pair candidate determined to have a semantic relationship.

  The second classification unit 24 performs machine learning using the second learning data, and classifies whether the genuine common pair and the virtual common pair have a semantic relationship using the result of the machine learning. By the classification, the second classification unit 24 can obtain the classification result (whether it has a semantic relationship) and the certainty of the classification result. As will be described later, when the second learning data is added by the adding unit 25, the second classification unit 24 also performs learning using the added second learning data. To do. The second classifying unit 24 classifies the plurality of second relation pair candidates after repeating machine learning and classification and addition of learning data. Then, the second classification unit 24 accumulates in the second relationship pair storage unit 27 the second relationship pair that is the second relationship pair candidate determined to have a semantic relationship.

  Here, the classification using machine learning by the first and second classification units 23 and 24 will be briefly described. The first and second classification units 23 and 24 classify the first and second relationship pair candidates into those having a semantic relationship and those not having them using machine learning. The machine learning inputs are first and second relationship pair candidates. The output of the machine learning is whether or not the first and second relationship pair candidates have a semantic relationship. In addition, learning data (training data) as teacher data used in the machine learning includes two language expression pairs and information indicating presence / absence of a semantic relationship between the pairs (that is, having a semantic relationship). Or not). After the learning using the learning data, when the first relationship pair candidate or the second relationship pair candidate to be classified is input, each feature value regarding the first relationship pair candidate is acquired, and the first relationship pair Whether the candidate or the like has a semantic relationship and its certainty are output. The features used in the machine learning will be described later.

  The first and second classification units 23 and 24 may use, for example, SVM (Support Vector Machine) or other things as machine learning. In the present embodiment, a case where SVM is used as machine learning will be described.

  The adding unit 25 adds a common pair having a high certainty factor according to the classification of the first classification unit 23 and a classification result related to the common pair to the second learning data. The adding unit 25 adds a common pair having a high certainty factor according to the classification of the second classification unit 24 and a classification result related to the common pair to the first learning data. Specifically, the adding unit 25 relates to a common pair having a high certainty factor by the classification of the first classification unit 23 and the same classification result of the first and second classification units 23 and 24, and the common pair. The classification result may be added to the second learning data, the confidence level by the classification of the second classification unit 24 is high, and the classification result of the first and second classification units 23 and 24 is the same. And the classification result regarding the common pair may be added to the first learning data. Further, the adding unit 25 performs the second learning on the common pair having a high certainty factor by the classification of the first classification unit 23 and a low certainty factor by the classification of the second classification unit 24 and the classification result regarding the common pair. It is possible to add to the data. The common pair having a high certainty factor by the classification of the second classification unit 24 and a low certainty factor by the classification of the first classification unit 23 and the classification result relating to the common pair You may add to learning data. Here, the virtual common pair is considered as a candidate for a semantic relationship in one corpus, but is unlikely to actually have a semantic relationship. Therefore, the virtual common pair is likely to be added as a negative example. On the other hand, the genuine common pair is considered to have a higher semantic relationship than the virtual common pair, and may be added as a positive example.

  To add a certain common pair and classification result to the first learning data may be to accumulate the common pair or the like in the first learning data storage unit 21 or to add the common pair or the like to the first learning data. One classification unit 23 may be set to be used as the first learning data. In the latter case, the first classification unit 23 reads the common pair added to the first learning data and stored in the common pair storage unit 20 and the classification result as the first learning data. It may be set as follows. Here, the classification result may be accumulated in the common pair storage unit 20 by the second classification unit 24. Further, adding a certain common pair and classification result to the second learning data may be to accumulate the common pair or the like in the second learning data storage unit 22, or to add the common pair or the like. Alternatively, the second classification unit 24 may be set to be used as the second learning data. In the latter case, the second classification unit 24 reads out the common pair added to the second learning data and stored in the common pair storage unit 20 and the classification result as the second learning data. It may be set as follows. Here, the classification result may be accumulated in the common pair storage unit 20 by the first classification unit 23. In the present embodiment, the adding unit 25 accumulates a common pair that is an addition target to the first learning data in the first learning data storage unit 21, and is a common addition target to the second learning data. The case where a pair etc. are accumulate | stored in the 2nd learning data storage part 22 is demonstrated. Note that the learning data stored in advance in the first learning data storage unit 21 and the second learning data storage unit 22 may be referred to as initial first learning data and initial second learning data, respectively. . The initial first learning data and the initial second learning data may be different from each other, or may be the same.

  The machine learning and classification by the first and second classification units 23 and 24 and the addition of learning data by the adding unit 25 are repeatedly executed. During the repetition, the first and second classification units 23 and 24 perform machine learning using the learning data after the addition by the adding unit 25 is performed.

  The first corpus storage unit 11, the second corpus storage unit 12, the first related pair candidate storage unit 15, the first irrelevant pair candidate storage unit 16, the second related pair candidate storage unit 17, the second irrelevant pair candidate Storage in the storage unit 18, common pair storage unit 20, first learning data storage unit 21, second learning data storage unit 22, first relationship pair storage unit 26, and second relationship pair storage unit 27 is RAM or the like. It may be a temporary memory or a long-term memory. These storage units can be realized by a predetermined recording medium (for example, a semiconductor memory, a magnetic disk, an optical disk, etc.).

  Moreover, the 1st corpus memory | storage part 11, the 2nd corpus memory | storage part 12, the 1st related pair candidate memory | storage part 15, the 1st unrelated pair candidate memory | storage part 16, the 2nd related pair candidate memory | storage part 17, the 2nd unrelated pair candidate Any two or more of the storage unit 18, the common pair storage unit 20, the first learning data storage unit 21, the second learning data storage unit 22, the first relationship pair storage unit 26, and the second relationship pair storage unit 27 These storage units may be realized by the same recording medium or may be realized by separate recording media. In the former case, for example, an area storing the first corpus is the first corpus storage unit 11, and an area storing the second corpus is the second corpus storage unit 12.

Next, the operation of the mutual machine learning device 1 according to the present embodiment will be described using the flowchart of FIG. Here, the first corpus is “S”, the second corpus is “U”, the first related pair candidate set is “X _S ”, and the first unrelated pair candidate set is “R _S ”. , The set of second relationship pair candidates is “X _U ”, the set of second unrelated pair candidates is “R _U ”, the set of common pairs is “Y”, the set of genuine common pairs is “G”, A set of virtual common pairs is assumed to be “V”. X _S and _X U, G, the relationship V like is as shown in Figure 4. Note that Y = G∪V. Also, the first learning data stored in advance in the first learning data storage unit 21 is “L ⁰ _S ”, and the second learning data stored in advance in the second learning data storage unit 22 is “ L ⁰ _U ”.

(Step S <b> 101) The first extraction unit 13 extracts a plurality of first relationship pair candidate sets X _S from the first corpus S and accumulates them in the first relationship pair candidate storage unit 15.

(Step S102) the second extraction portion 14 will be accumulated in the second relationship pair candidate storage unit 17 extracts a set X _U of the second relationship pair candidates more of the second corpus U.

(Step S <b> 103) The first extraction unit 13 extracts a set R _S of a plurality of first irrelevant pair candidates from the first corpus S and accumulates them in the first irrelevant pair candidate storage unit 16. Note that X _S ∩R _S is an empty set.

(Step S104) The second extraction unit 14 stores the second independent pair candidate storage unit 18 extracts a set R _U in the plurality from the second corpus U second unrelated pair candidates. Note that X _U ∩R _U is an empty set.

(Step S105) The acquisition unit 19 includes a plurality of first relationship pair candidate sets X _S , a plurality of first unrelated pair candidate sets R _S , a plurality of second relationship pair candidate sets X _U , and a plurality of second unrelated relationships. using a set R _U pair candidate obtains a set G of genuine common pair and a set V of virtual common pair, storing them in a common pair storage unit 20. As shown in FIG. 4, the set G of genuine common pairs is a common part of the set X _S of first relationship pair candidates and the set X _U of second relationship pair candidates. That is, G = X _S ∩X _U. The virtual common pair set V includes a common part of the first relation pair candidate set X _S and the second unrelated pair candidate set R _U , the second relation pair candidate set X _U , It is a common part with the set R _S of irrelevant pair candidates. That is, V = (X _S ∩R _U ) ∪ (R _S ∩X _U ).

  (Step S106) The adding unit 25 sets the counter i to 0. This counter i is a counter for counting the cycles of steps S107 to S113.

(Step S107) The first classification unit 23 performs machine learning using the first learning data L ⁱ _S stored in the first learning data storage unit 21. _{Let the} classifier obtained by the machine learning be c ⁱ _S. Note that the features used when performing machine learning may be stored in advance in the first learning data storage unit 21 or may be acquired with reference to the first corpus, for example.

(Step S108) The second classification unit 24 performs machine learning using the second learning data L ⁱ _U stored in the second learning data storage unit 22. Let c ⁱ _U be the classifier obtained by the machine learning. Note that the features used when performing machine learning may be stored in advance in the second learning data storage unit 22, or may be acquired with reference to the second corpus, for example.

(Step S109) The first classifying unit 23 classifies each common pair included in the common pair set Y by using the classifier c ⁱ _S that is the result of the machine learning. As a result of this classification, it is possible to obtain a class label clε {yes, no} indicating whether or not there is a semantic relationship, and a certainty factor rεR ⁺ . The class label cl “yes” indicates that it is classified as having a semantic relationship, and the class label cl “no” indicates that it is classified as not having a semantic relationship. “R ⁺ ” is a non-negative real number. The classification result of yεY by the classifier c may be described as c (y) = (y, cl, r). In this classification, among the common pairs included in Y, the common pairs included in the first learning data L ⁱ _S or the second learning data L ⁱ _U need not be classified. Good. The features of each common pair used in this classification may be stored in advance in the common pair storage unit 26, or may be acquired with reference to the first corpus, for example.

(Step S110) The second classifying unit 24 classifies each common pair included in the common pair set Y by using the classifier c ⁱ _U that is the result of machine learning. As a result of this classification, the class label cl indicating whether or not there is a semantic relationship and the certainty factor r can be obtained as in the case of the first classification unit 23. In this classification, among the common pairs included in Y, the common pairs included in the first learning data L ⁱ _S or the second learning data L ⁱ _U need not be classified. Good. In addition, the features of each common pair used in this classification may be stored in advance in the common pair storage unit 26, or may be acquired with reference to the second corpus, for example.

(Step S111) Using the classification result, the adding unit 25 adds a common pair that satisfies a predetermined condition to the first learning data L ^{(i + 1)} _S and the second learning data L ^{(i + 1)} _U. The first learning data L ^{(i + 1)} _S includes all elements of L ⁱ _S , and the second learning data L ^{(i + 1)} _U includes all elements of L ⁱ _U. Details of the learning data addition process will be described later with reference to the flowchart of FIG.

  (Step S112) The adding unit 25 determines whether or not a condition for ending the cycle of steps S107 to S113 is satisfied. If the end condition is satisfied, the process proceeds to step S114. If not, the process proceeds to step S113.

The end condition is, for example, that the value of d ⁱ = | σ ⁱ −σ ⁽ⁱ⁻¹⁾ | / | σ ⁽ⁱ⁻¹⁾ | is a predetermined number of consecutive times (for example, it may be three times). Only a cycle may be less than a predetermined threshold “ε”. Note that σ ⁱ is the reliability of classification by the first classification unit 23 of each y∈Y classified in steps S109 and S110 in the cycle of the counter i, and is the reliability of classification by the second classification unit 24. Is r2,
σ ⁱ = Σ | r1-r2 |
This is the value indicated by. The sum is taken for all yεY. d ⁱ <ε means that the hyperplane that is the learning result has hardly changed compared to the previous cycle, that is, even if new learning data is added, the learning result hardly changes. Means not. It is preferable that the threshold ε is selected so as to be so. The threshold value ε may be, for example, 0.001.

  Needless to say, different termination conditions may be used. For example, according to an empirical rule, when the number of repetitions of the cycle of steps S107 to S113 reaches a predetermined number (for example, when counter i = A, where A is an integer of 1 or more), a new If it is known that the learning result has hardly changed even when learning data is added, the end condition may be that the counter i = A.

  Information indicating the end condition is stored in a recording medium (not shown), and the adding unit 25 may read information indicating the end condition from the recording medium and determine whether the end condition is satisfied. . Although the case where the adding unit 25 makes a determination regarding the end condition has been described here, it is needless to say that the component other than the adding unit 25 may make the determination.

  (Step S113) The adding unit 25 increments the counter i by 1. Then, the process returns to step S107.

  (Step S114) The first classifying unit 23 classifies each first relationship pair candidate stored in the first relationship pair candidate storage unit 15 using the classifier that is the learning result at that time, The first relationship pair that is a first relationship pair candidate determined to have a semantic relationship by classification is stored in the first relationship pair storage unit 26. Note that the feature of each first relation pair candidate used in this classification may be stored in advance in the first relation pair candidate storage unit 15 or obtained by referring to the first corpus, for example. Also good.

  (Step S115) The second classifying unit 24 classifies each second relationship pair candidate stored in the second relationship pair candidate storage unit 17 using the classifier that is the learning result at that time, The second relationship pair that is the second relationship pair candidate determined to have a semantic relationship by classification is accumulated in the second relationship pair storage unit 27. Note that the feature of each second relationship pair candidate used in this classification may be stored in advance in the second relationship pair candidate storage unit 17 or obtained by referring to the second corpus, for example. Also good.

  In this way, a series of processes of mutual learning and classification using the learning result is completed. In the flowchart of FIG. 2, it goes without saying that the order of the processes from step S101 to S104 is not limited. Needless to say, the order of the processes in steps S107 and S108, the order of the processes in steps S109 and S110, and the order of the processes in steps S114 and S115 are not important. Needless to say, the processing that can be executed in parallel may be performed in parallel. As described above, in the flowchart of FIG. 2, various changes can be made within a range in which the purpose of a series of processing is achieved.

FIG. 3 is a flowchart showing details of the learning data addition process (step S111) in the flowchart of FIG.
(Step S201) The adding unit 25 specifies a set CR i S of classification results for common pairs not included in L i S ＳL i U from the classification result by the first classification unit 23 in step S109. CR i S is represented by the following equation. Incidentally, in step S109, if you are performing only classification for common pair that is not included in L i S ∪L i U is classified result itself of the first classification section 23, a CR i S. Note that the process of specifying the set CR i S is not particularly limited as long as it is a process that enables the specified CR to be distinguished from the unspecified one. The process of specifying the set CR i S may be, for example, storing the specified set CR i S in a recording medium (not shown), and a flag indicating that each element of the specified set CR i S is specified. Etc. may be set. This also applies to other specific processes.

(Step S202) The adding unit 25 specifies a set CR i U of classification results for common pairs not included in L i S ∪L i U from the classification result by the second classification unit 24 in step S110. CR i U is represented by the following equation. Note that, in step S110, if you are performing only classification for common pair that is not included in L i S ∪L i U is classified result itself of the second classification section 24, a CR i U.

(Step S203) addition unit 25, the classification results of the set ^CR _{i S,} identifies a set of descending order to the selected N pieces of classification results of confidence r TopN ^(CR _{i S).} N is a predetermined integer of 1 or more, and may be 900, for example. The adding unit 25 may sort the set of classification results CR ⁱ _S in descending order of the certainty factor r, and select the top N classification result sets TopN (CR ⁱ _S ).

  (Step S204) The adding unit 25 sets the counter j to 1.

(Step S205) The adding unit 25 performs the classification result (y ^j , cl ^j _S , r ^j _S ) ∈ of the j-th common pair y ^j included in the set TopN (CR ⁱ _S ) of the classification results specified in step S203. Using TopN (CR ⁱ _S ) and the classification result (y ^j , cl ^j _U , r ^j _U ) ∈CR ⁱ _U by the second classification unit 24 for the same common pair y ^j , the common pair y ^j is It is determined whether or not to be added to the second learning data L ^{(i + 1)} _U. If the common pair y ^j is to be added to the second learning data L ^{(i + 1)} _U , the process proceeds to step S206; otherwise, the process proceeds to step S207. The classification result (y ^j , cl ^j _S , r ^j _S ) of the j-th common pair y ^j included in the classification result set TopN (CR ⁱ _S ) specified in step S203 is the classification result set CR ^i. The classification result (y ^j , cl ^j _S , r ^j _S ) of the j-th common pair y ^j in the result of sorting _S in descending order of the certainty factor r may be used.

Specifically, when the following (Condition 1) or (Condition 2) is satisfied, the adding unit 25 determines to add the common pair y ^j to the second learning data L ^{(i + 1)} _U.
(Condition 1): r ^j _S > α and r ^j _U <β
(Condition 2): r ^j _S > α and cl ^j _S = cl ^j _U

  Condition 1 corresponds to a high certainty factor according to the classification of the first classification unit 23 and a low certainty factor according to the classification of the second classification unit 24. Condition 2 corresponds to a high degree of certainty by the first classification unit 23 and the same classification results of the first and second classification units 23 and 24. Note that the values of α and β are set appropriately in advance. In the present embodiment, a case is described in which the determination is performed using both the conditions 1 and 2, but the determination may be performed using only one of the conditions.

(Step S206) The adding unit 25 adds the common pair y ^j and the classification result cl ^j _S to the second learning data L ^{(i + 1)} _U used in the machine learning of the next cycle. That is,
L ^{(i + 1)} _U ← L ^{(i + 1)} _U ∪ (y ^j , cl ^j _S )
And It is assumed that L ^{(i + 1)} _U ← L ⁱ _U before the process of adding a common pair in the cycle of steps S205 to S208 is started.

  (Step S207) The adding unit 25 increments the counter j by 1.

(Step S208) The adding unit 25 includes the classification result (y ^j , cl ^j _S , r ^j _S ) of the j-th common pair y ^{j in} the classification result set TopN (CR ⁱ _S ) specified in step S203. Judge whether or not. If the jth classification result is included, the process returns to step S205; otherwise, the process proceeds to step S209. Since the set TopN (CR ⁱ _S ) of the classification result specified in step S203 includes N elements, the adding unit 25 determines whether j ≦ N, and when j ≦ N. May return to step S205, and if not, may proceed to step S209.

(Step S209) addition unit 25, the classification results of the set ^CR _{i U,} identifies a set of descending order to the selected N pieces of classification results of confidence r TopN ^(CR _{i U).}

  (Step S210) The adding unit 25 sets the counter j to 1.

(Step S211) The adding unit 25 classifies the classification result (y ^j , cl ^j _U , r ^j _U ) ε of the j-th common pair y ^j included in the classification result set TopN (CR ⁱ _U ) specified in step S209. Using TopN (CR ⁱ _U ) and the classification result (y ^j , cl ^j _S , r ^j _S ) ∈CR ⁱ _S by the first classification unit 23 for the same common pair y ^j , the common pair y ^j is It is determined whether or not to be added to the ^first learning data L ^{(i + 1)} _S. If the common pair y ^j is to be added to the second learning data L ^{(i + 1)} _S , the process proceeds to step S212. Otherwise, the process proceeds to step S213. The classification result (y ^j , cl ^j _U , r ^j _U ) of the j-th common pair y ^j included in the classification result set TopN (CR ⁱ _U ) specified in step S209 is the classification result set CR ^i. The classification result (y ^j , cl ^j _U , r ^j _U ) of the j-th common pair y ^j in the result of sorting _U in descending order of the certainty factor r may be used.

Specifically, when the following (Condition 3) or (Condition 4) is satisfied, the adding unit 25 determines to add the common pair y ^j to the first learning data L ^{(i + 1)} _S.
(Condition 3): r ^j _U > α and r ^j _S <β
(Condition 4): r ^j _U > α and cl ^j _U = cl ^j _S

  Condition 3 corresponds to a high certainty factor according to the classification of the second classification unit 24 and a low certainty factor according to the classification of the first classification unit 23. Condition 4 corresponds to a high degree of certainty by the second classification unit 24 and the same classification results of the first and second classification units 23 and 24. In the present embodiment, a case is described in which the determination is performed using both the conditions 3 and 4. However, the determination may be performed using only one of the conditions.

(Step S212) The adding unit 25 adds the common pair y ^j and the classification result cl ^j _U to the first learning data L ^{(i + 1)} _S used in the machine learning of the next cycle. That is,
L ^{(i + 1)} _S ← L ^{(i + 1)} _S ∪ (y ^j , cl ^j _U )
And Note that L ^{(i + 1)} _S ← L ⁱ _{S is} assumed before the process of adding a common pair in the cycle of steps S211 to S214 is started.

  (Step S213) The adding unit 25 increments the counter j by 1.

(Step S214) The adding unit 25 includes the classification result (y ^j , cl ^j _U , r ^j _U ) of the j-th common pair y ^{j in} the classification result set TopN (CR ⁱ _U ) specified in step S209. Judge whether or not. If the jth classification result is included, the process returns to step S211; otherwise, the process returns to the flowchart of FIG. Note that since the set TopN (CR ⁱ _U ) of the classification result specified in step S209 includes N elements, the adding unit 25 determines whether j ≦ N, and when j ≦ N. May return to step S211, or otherwise return to the flowchart of FIG.

  Here, when the first corpus is structured and the second corpus is unstructured, the first and second relationship pair candidates and the first and second unrelated pairs A method for extracting candidates will be described. It is assumed that the semantic relationship is an upper / lower relationship. Here, a case where an encyclopedia corpus is used as the structured first corpus will be described. In the encyclopedia corpus, for example, as shown in FIG. 5A, for the title “Tiger”, there are sections “Range” and “Taxonomy”, and for the section “Taxonomy”, the subsection “ "Subspecies", and for the subsection "Subspecies", there are lists "Bengal tiger", "Malayan tiger", and "Siberian tiger". If the encyclopedia corpus is described in a markup language such as HTML or XML, specify the title tag that identifies the title, the section tag that identifies the section, the subsection tag that identifies the subsection, and the list By using the list tag or the like, the title, section, subsection, list, etc. can be specified, and their tree structure can be obtained as shown in FIG. The tree structure can be acquired as follows, for example. For example, a tag in a lower hierarchy than a section existing after the section tag A (for example, a tag in a subsection or a tag in a list), and a tag in a hierarchy higher than the subsection tag or the subsection. Among the tags that appear before (for example, title Doug) appear, the tag closest to the tag A of the section (referred to as “tag B”) is one level lower than the tag A node of the section. Node. In addition, a tag that exists after the tag B and appears before a tag of a different hierarchy from the tag B appears (this tag is in the same hierarchy as the tag B) is also included in the tag A of the section. Only one layer connected to the node is a lower node. It is assumed that a title, a section, a subsection, and a list are determined in advance from the upper layer. When the tree structure shown in FIG. 5B is obtained in this way, a pair of an upper node and a direct or indirect lower node with respect to the upper node becomes a first relation pair candidate. In the case of FIG. 5B, for example, (Tiger, Range), (Tiger, Taxonomy), (Tiger, Subspecies), (Tiger, Bengal tiger), (Taxonomic, Subspecies), (Taxonomy, Bengalti, etc.) Becomes the first relationship pair candidate. In the first relation pair candidate, (higher language expression, lower language expression) is set. Also, in the encyclopedia corpus tree structure, a pair that is not a pair of an upper node and a direct or indirect lower node with respect to the upper node, for example, a pair of nodes having the same parent node is a first unrelated pair. Be a candidate. In the case of FIG. 5B, for example, (Range, Taxonomy), (Bengal tiger, Malayan tiger), etc. are the first unrelated pair candidates. The first extraction unit 13 separates the first corpus portion used to extract the first related pair candidate and the first corpus portion used to extract the first unrelated pair candidate. May be. For example, refer to the above-mentioned Non-Patent Document 2 for a method of extracting upper and lower relations from Japanese WIKI IPEDIA (registered trademark). Also, by using the method described in that document, it is possible to know the tree structure of WIKIPEDIA (registered trademark), so that it is possible to extract a relationship that is not a higher-order relationship as described above.

Next, a method for extracting the second related pair candidate and the second unrelated pair candidate from the second corpus will be described. Here, a case where web information is used as the unstructured second corpus will be described. The second extraction unit 14 can extract the second related pair candidate and the second unrelated pair candidate by using the lexicosyntactic pattern. In the second corpus, for example, the second extraction unit 14 identifies a location that matches the lexicosyntactic pattern “B of A”, “B of A”, and the like corresponding to the upper and lower relationships, By extracting the linguistic expressions A and B, the second relationship pair candidate (A, B) can be extracted. In addition, the second extraction unit 14 uses, for example, a lexicosyntactic pattern “D caused by C”, “C” corresponding to a relationship other than upper and lower levels (for example, a causal relationship) in the second corpus. The second unrelated pair candidate (C, D) can be extracted by identifying the part that matches “D” used for the above and extracting the language expressions C, D. For a method of extracting a linguistic expression pair candidate having a semantic relationship using a lexicosyntactic pattern in this way, refer to the following document, for example.
Literature: Maya Ando, Satoshi Sekin, Shun Shizaki, “Automatic extraction of hypothems from Japan newspace using lexico-intactic tactics”. of LREC '04, 2004

  In addition, when the first corpus is structured, the second corpus is unstructured, and the semantic relationship is an upper-lower relationship, the first and second classification units The features used in the machine learning of 23 and 24 will be described. Here, a case where the first corpus is WIKIPEDIA (registered trademark) and the second corpus is web text will be described.

  The features of machine learning by the first classification unit 23 include the first relation pair candidate and two language expressions included in the learning data (this language expression is referred to as language expressions A and B), and each language expression A. , B morphemes and parts of speech. Moreover, the morpheme of the main word of each language expression A and B may be included. A headword is a major morpheme connected to another part. In Japanese, the last morpheme is usually the main word. For example, in the case of the language expression “XYZ University”, the last morpheme “University” is the main word. The distance between the language expressions A and B (hierarchy depth) in the tree structure may be included in the feature. For example, in the case of FIG. 5, the distance between “Tiger” and “Rnage” is “1”, and the distance between “Tiger” and “Bengal tiger” is “3”. Information on whether the language expressions A and B match any of a plurality of patterns (for example, “list of”, “list of”, etc.) such as a list of items and a section in which the list appears, etc. The language expressions A and B are pre-determined in advance (for example, WIKIPEDIA (registered trademark)) in terms of WIKIPEDIA (registered trademark) headwords (including titles, section titles, subsection titles, not including lists). Those that appear more often than the frequency (for example, “references”, “external links”, etc.), the layout type (eg title, section, list, etc.) of language expressions A and B, There are node types (for example, a root node, a leaf node, an intermediate node, etc.) of a tree structure of language expressions A and B. "iger" is a root node, "Bengal tiger" is a leaf node, and "Range" is an intermediate node), and any one or more of the parent nodes and child nodes of the language expressions A and B are features May be included. In addition, an attribute or attribute value obtained from the Infobox of WIKIPEDIA (registered trademark) may be included in the feature. For these features, refer to the aforementioned Non-Patent Document 2.

Machine learning by the second classifying unit 24 includes two language expressions included in the second relation pair candidate and the learning data (this language expression is referred to as language expressions A and B), and each language expression A. , B morphemes and parts of speech. Also, an identifier of the lexicosyntactic pattern used to acquire the two language expressions, a PMI (point-wise mutual information) score between the two language expressions and the patterns, and language expressions A and B Any one or more of the PMI scores and the noun classes of the language expressions A and B may be included in the feature. The noun class is obtained by dividing 5 × 10 ⁵ nouns into 500 classes by EM-based clustering shown in the following document. For example, the noun class C ₃₁₁ is a class having nouns related to biology and chemistry such as “polysaccharide” and “organic compound”.
Literature: Jun'ichi Kazama, Kentaro Torisawa, “Inducting gazeteters for named entity recognition by large-scale crushing of dependency Pro. 407-415, 2008

  Needless to say, the features used in machine learning and classification by the first and second classification units 23 and 24 are not limited to those described above. Other features may be used, and some of the features described above may not be used. In addition, when the semantic relationship is not a higher-order or lower-order relationship, it is preferable that an appropriate feature corresponding to the semantic relationship is used.

[Experimental example]
Next, an experimental example of the mutual machine learning device 1 according to the present embodiment will be described. Here, as a first corpus, a Japanese version of WIKIPEDIA (registered trademark) in July 2009 was used. Then, 1.9 × 10 ⁷ first relationship pair candidates were acquired. The extraction of the first relationship pair candidate was performed from the main article of WIKIPEDIA (registered trademark). It should be noted that 24,000 first relationship pair candidates were randomly extracted therefrom, and it was manually determined whether or not they had a semantic relationship (upper-lower relationship). Of these, 20000 pairs were used as learning data, and the remaining 4000 pairs were equally divided into development data and test data. The development data is used for selecting an optimum parameter. The test data is used for evaluation. Each of the learning data, development data, and test data includes a positive example and a negative example that are manually determined.

Further, as the second corpus, 5 × 10 ⁷ pages of web text from the above-mentioned TSUBAKI was used. Then, 6 × 10 ⁶ second relationship pair candidates were obtained from these web texts using lexicosyntactic patterns of upper and lower relationships. In addition, 9500 second relationship pair candidates were randomly extracted from the web text, and it was manually determined whether or not they had a semantic relationship (higher-lower relationship). Of these, 7500 pairs were used as learning data, and the remaining 2000 pairs were divided equally into development data and test data. Each of the learning data, development data, and test data includes a positive example and a negative example that are manually determined.
In this experimental example, the ratio of candidates having a semantic relationship (higher-lower relationship) in the first relationship pair candidate set or the second relationship pair candidate set was not so high, and was about 25 to 30%. .

  Moreover, the 1st unrelated pair candidate was extracted using the category system of WIKIPEDIA (registered trademark). That is, since the first related pair candidate is extracted from the main article, and the first unrelated pair candidate is extracted using the category system, the first corpus portion used for extracting the first related pair candidate and the first This is different from the portion of the first corpus used to extract irrelevant pair candidates. Note that, as described above, the first irrelevant pair candidate is one in which one language expression included in the pair is not an ancestor of the other language expression. Moreover, the 2nd unrelated pair candidate was extracted from the TSUBAKI page using lexicosyntactic patterns such as a causal relationship.

  Further, in this experimental example, TinySVM (http://chasen.org/˜taku/software/TinySVM/) with a primary kernel d = 2 was used as the first and second classification units 23 and 24. In addition, the parameters α, β, and N were determined by experiments using development data. In this experimental example, α = 1.0, β = 0.3, and N = 900. Moreover, evaluation was performed using the precision (P), the recall (R), and the F value (F).

  In this experimental example, six systems were compared. Three of them are B1, B2, and B3, and are intended to show the effect of different feature sets and different learning data. In B1 and B2, two classifiers were machine-learned separately, whereas in B3, an integrated feature set and learning data were used for machine learning of a single classifier.

  B1 is composed of a completely independent classification unit. The S and U classifiers were learned and evaluated using their own features and learning data, respectively. That is, the features and learning data of WIKIPEDIA (registered trademark) were used for the S classification unit, and the web features and learning data were used for the U classification unit.

  B2 is the same as B1 except that the two classification units are learned using the integrated learning data. That is, each of the two classification units performed machine learning with 27500 pieces of learning data. Note that the features used in the machine learning are different. Here, in the machine learning of the classification unit of U, when learning data acquired from web text is learned, the learning such as distance does not exist, but learning is performed.

  B3 is obtained by adding a master classification unit to B1. As with B2, learning was performed using the integrated learning data. We also performed machine learning using all available features. That is, the same feature was used in both of the two classification parts. Furthermore, the SVM score for each pair obtained by the two classifiers of B1 was also included in the feature.

  The other three systems, BICO, Co-B, and Co-STAR (mutual machine learning apparatus 1 according to the present embodiment), are bilingual mutual machine learning (BICO), mutual machine learning (Co-B and Co- STAR). In particular, Co-B and Co-STAR are compared to evaluate whether or not a virtual common pair is used. For Co-B and Co-STAR, initial learning data similar to B1 and B2 was used. Note that those using the initial learning data similar to B1 are denoted as Co-B and Co-STAR, and those using the initial learning data similar to B2 are Co-B * and Co-STAR *. It is written. That is, for Co-B and Co-STAR, the first learning data is 20000 pieces (extracted from WIKIPEDIA (registered trademark)), and the second learning data is 7500 pieces (extracted from the web). However, for Co-B * and Co-STAR *, 27500 first and second learning data (extracted from WIKIPEDIA (registered trademark) and extracted from the web) were combined. Things).

  BICO uses a bilingual mutual machine learning algorithm described in Non-Patent Document 2 described above. In the algorithm, the upper and lower semantic relationships of two languages are acquired cooperatively by two processes. For the BICO, 20000 pieces of English learning data and 20000 pieces of Japanese learning data were prepared. Note that the 20,000 pieces of Japanese learning data are the same as the learning data acquired from the aforementioned WIKIPEDIA (registered trademark).

  Co-B is a modification of the mutual machine learning apparatus 1 (Co-STAR) according to the present embodiment, and uses only a genuine common pair as described above. In this experimental example, 67000 common common pairs were used.

  Co-STAR is a mutual machine learning apparatus 1 according to the present embodiment, and uses both a genuine common pair and a virtual common pair. The total number of these common pairs was 643000.

The experimental result is as shown in FIG. The experimental results are based on the initial learning data and features (all the features described above) as described above. Machine learning, common pair classification, and learning data are added until the end condition is satisfied. This is a result of classifying the test data after repeating the above. The correct answer and the incorrect answer were determined depending on whether the result of the classification of the test data is the same as the result of manual judgment. As the termination condition, adopts the value of d ⁱ is less than 3 consecutive times 0.001. In addition, since BICO performs processing using data of WIKIPEDIA (registered trademark) in two languages, there is no result for web data. In the experimental results shown in FIG. 6, WebSet is an experimental result using web text (that is, a result of classification by the second classification unit 24), and WikiSet is an experimental result using WIKIPEDIA (registered trademark) (that is, The result of classification by the first classification unit 23).

  By comparing B1 to B3, it can be seen that B2 and B3 are superior to B1 in terms of the F value. Since B2 and B3 used more learning data (27500), the results were better than B1 using a smaller number of learning data (7500 and 20000). B2 and B3 differ in the number of classification units, and the F values of both are the same, although the classification unit has learned using different features and learning data.

  It can be seen that Co-STAR has better performance than B1 to B3. It can also be seen that Co-STAR has better performance with less learning data than BICO. The total learning data for Co-STAR is 27500, and the total learning data for BICO is 40000. The difference in performance between Co-B and Co-STAR shows the effect of using or not using a virtual common pair. Since Co-STAR has a higher F value than Co-B, it can be seen that using the virtual common pair together with the genuine common pair can realize more effective cooperation between the two classification units.

As described above, the mutual machine learning device 1 (Co-STAR) according to the present embodiment has an F value that is 1.4 to 8.5% higher than other methods, and has higher performance than other methods. It can be seen that The mutual machine learning device 1 that has performed machine learning in this manner can acquire 4.3 × 10 ⁵ first relationship pairs (higher and lower relationship pairs) from the web text, and can obtain WIKIPEDIA ( It was possible to obtain 4.6 × 10 ⁶ second relationship pairs (pairs of upper and lower relationships) from (registered trademark). In addition, by setting the SVM threshold to 0.23 for web data and 0.1 for WIKIPEDIA (registered trademark), it was possible to obtain a 90% conformance rate. .

  Next, an experimental example for evaluating the robustness of the mutual machine learning device 1 according to the present embodiment will be described. In this experimental example, for the structured corpus (WIKIPEDIA (registered trademark)), manual learning data similar to the above-described experimental example is used, and for the unstructured corpus (web text), automatically. The acquired learning data with a lot of noise (that is, not necessarily highly accurate) was used. The learning data will be briefly described. The positive example of the learning data was acquired as follows. First, from the definition sentence of WIKIPEDIA (registered trademark) (the sentence at the beginning of the article of WIKIPEDIA (registered trademark)), “(lower language expression) is (upper language expression)” and “(lower language expression) are A pair of upper and lower relations acquired using a pattern such as “a kind of (upper language expression)” and a pair of upper and lower relations acquired using a category of WIKIPEDIA (registered trademark) were acquired. When acquiring a pair of upper and lower relations using a category of WIKIPEDIA (registered trademark), first, a pair having a title as a lower language expression and a category of the title as an upper language expression was acquired. Then, a pair of upper and lower relations whose title is a lower language expression can be obtained from the definition sentence of WIKIPEDIA (registered trademark) using a pattern, and the upper language of the pair of the upper and lower relations When the main words of the expression and the higher-level language expression that is the category match, the pair whose title is the lower-level language expression and whose title category is the higher-level language expression is the upper-lower relationship pair . For example, from the title “New influenza” and the definition sentence “New influenza is an influenza infection whose pathogen is a virus that has a new human-to-human transmission ability among influenza viruses”. Using patterns, upper and lower relationships (influenza infections, new influenza) are acquired as positive pairs. In addition, if there is “virus infection” in the category of “new influenza”, the main language expression “infectious disease” of the category is the higher-level language expression of the upper-lower relationship acquired using the definition sentence pattern In order to agree with the main word, “virus infection” acquired from this category is also acquired as a positive example of the language expression above “new influenza”. That is, from the category, the upper-lower relationship (virus infection, new influenza) is acquired as a positive pair.

Among the upper and lower relations acquired in this way, the common relation with the second relation pair candidate was used as positive example learning data. Negative examples of learning data were obtained from virtual common pairs, particularly R _S ＲX _U. Since the size of the learning data acquired in this way becomes very large, 7500 pieces of learning data were randomly selected so as to be the same as the above-described experimental example. At that time, selection was made so that the ratio of positive example to negative example was 1: 4.

  Using such learning data, an experiment similar to the above-described experimental example was performed, and the result was as shown in FIG. From the result of FIG. 7, it can be seen that the mutual machine learning device 1 (Co-STAR) according to the present embodiment has robustness with respect to learning data with low accuracy. Moreover, although the mutual machine learning apparatus 1 (Co-STAR) by this Embodiment has a little performance fall compared with the result of FIG. 6, it has a better performance with respect to B1-B3. I understand that. Therefore, it can be understood that the labor for preparing the learning data manually can be reduced when the performance may be slightly reduced.

  As described above, according to the mutual machine learning device 1 according to the present embodiment, a mutual machine with higher performance can be obtained by performing mutual machine learning using a common pair, that is, a genuine common pair and a virtual common pair. Learning can be realized. Since the non-patent documents 1 and 2 described above use only the genuine common pair, the mutual machine learning device 1 according to the present embodiment performs the non-patent document by cooperating two machine learnings using the virtual common pair. This means that higher performance can be realized than the methods 1 and 2. Moreover, it turns out that the mutual machine learning apparatus 1 by this Embodiment has robustness also with respect to learning data with low precision. Therefore, it is possible to reduce manual work when preparing learning data. Further, according to the mutual machine learning device 1 according to the present embodiment, different processing targets are used for the first and second classification units 23 and 24, such as structured data and unstructured data. Can also be handled.

  Thus, the semantic relationship acquired using the mutual machine learning device 1 according to the present embodiment can be used in, for example, an information search system such as a web search, a machine translation system, or the like. Specifically, in web search, it is possible to expand a query using a semantic relationship. For example, when an unknown word that is not registered in the dictionary is input, if the upper-lower relationship with the unknown word as a lower word is acquired by the mutual machine learning device 1 according to the present embodiment, It is possible to perform a search using a broader term of the unknown word. Also in the machine translation system, it is possible to select a more appropriate translated word by having a semantic relationship, and to perform flexible translation such as translating using the broader word without registering the translated word. It can be carried out. It should be noted that the method of using the semantic relationship acquired using the mutual machine learning apparatus 1 according to the present embodiment is not limited to these, and it goes without saying that there are various other utilization methods.

In the present embodiment, the virtual common pair may be expanded. That is, the set V of virtual common pairs may also include common pairs other than the region of V shown in FIG. For example, the virtual common pair is not a genuine common pair among a plurality of first relationship pair candidates, a plurality of second relationship pair candidates, a plurality of first irrelevant pair candidates, and a plurality of second irrelevant pair candidates. It may be a common pair that is a pair. In that case, the set V of virtual common pairs also includes portions of X _S , R _S , X _U , and R _U that are not shaded in FIG. Strictly speaking, a pair of non-shaded portions of X _S , R _S , X _U , and R _U is not a common pair, but since the virtual common pair is expanded here, the pair is not shown for convenience. Is also called a common pair. That is, in this case, the set V of virtual common pairs is a set of pairs including the original common pair (a pair common to two sets).

  Moreover, in this Embodiment, although the addition part 25 demonstrated the case where the common pair with the high reliability by the classification | category of the 1st and 2nd classification | category parts 23 and 24 and its classification result were added to learning data, The adding unit 25 may perform other processes. That is, the adding unit 25 converts the common pair and the classification result related to the common pair into the second learning data according to at least one of the classification result and the certainty factor of the common pair by the classification of the first classification unit 23. And adding the common pair and the classification result related to the common pair to the first learning data according to at least one of the classification result of the common pair and the certainty factor by the classification of the second classification unit 24. Good. Here, “adding the common pair and the classification result relating to the common pair to the learning data according to at least one of the classification result and the certainty of the common pair” means that the classification result of the common pair and the certainty If at least one of the criteria satisfies the predetermined condition, the common pair and the classification result are added to the learning data, and at least one of the classification result and the certainty of the common pair does not satisfy the predetermined condition Means that the common pair or the like is not added to the learning data. The predetermined condition may relate to only the classification result, may relate to only the certainty degree, or may relate to both the classification result and the certainty degree, for example. For example, when the classification result of the first and second classification units 23 and 24 is the same for a certain common pair, the adding unit 25 determines that the common pair and the classification result of the common pair are the first and second. It may be added to the learning data. Further, for example, the adding unit 25 adds the randomly selected common pair and the classification result to the second learning data among the common pairs having high certainty by the classification of the first classification unit 23, and the second learning data Among the common pairs having high certainty by the classification of the second classification unit 24, a randomly selected common pair and the classification result may be added to the first learning data. Further, for example, the adding unit 25 adds the common pair that is a positive example and the classification result to the second learning data among the common pairs having high certainty by the classification of the first classification unit 23, and the second learning data Among the common pairs having a high certainty factor by the classification of the classification unit 24, a common pair that is a positive example and the classification result may be added to the first learning data. Further, for example, the adding unit 25 adds the common pair of the positive example based on the classification of the first classification unit 23 and the classification result to the second learning data, and adds the positive example based on the classification of the second classification unit 24. The common pair and the classification result may be added to the first learning data.

  In the present embodiment, the case where the first corpus has a structure and the second corpus has no structure has been mainly described, but this need not be the case. For example, both may have a structure. Even in that case, for example, the first relationship pair candidate or the like is extracted from the first corpus using the structure in the same manner as in the present embodiment, and from the second corpus as in the present embodiment. The second relationship pair candidate or the like may be extracted using a lexicosyntactic pattern. If the semantic relationship is not an upper-lower relationship, it is preferable to extract the first relationship pair candidate, the second relationship pair candidate, etc. using a structure or the like according to the semantic relationship. is there. For example, the first relationship pair candidate or the like may be extracted using a document structure, an inter-document structure, a table structure, or the like.

  In the present embodiment, the first extraction unit 13 extracts the first relationship pair candidate and the first irrelevant pair candidate, and the second extraction unit 14 extracts the second relationship pair candidate and the second irrelevant pair candidate. Although the case where it does is demonstrated, it does not need to be so. In that case, the mutual machine learning device 1 may not include the first corpus storage unit 11, the second corpus storage unit 12, the first extraction unit 13, and the second extraction unit 14. In that case, the first relation pair candidate, the first irrelevant pair candidate, the second relation pair candidate, and the second irrelevant pair candidate are extracted in the same manner as the extraction of the first and second extraction units 13 and 14. May have been extracted, or may be extracted by other methods (for example, a manual method). In addition, the process of storing the first related pair candidates and the like in the first related pair candidate storage unit 15, the first unrelated pair candidate storage unit 16, the second related pair candidate storage unit 17, and the second unrelated pair candidate storage unit 18 It doesn't matter. For example, the first relationship pair candidate or the like may be stored in the first relationship pair candidate storage unit 15 or the like via a recording medium, or the first relationship pair candidate transmitted via a communication line or the like Or the like may be stored in the first relationship pair candidate storage unit 15 or the like. Note that, when machine learning and classification are performed, a feature is necessary. Therefore, if the mutual machine learning device 1 does not include the first and second corpus storage units 11 and 12, the first relationship pair candidate It is preferable that feature information is previously associated with a pair of language expressions such as. The first and second classification units 23 and 24 can perform machine learning and classification by using information on the features.

  In the present embodiment, the first and second relationship pairs in which the first and second relationship pairs, which are the results of the first and second classification units 23 and 24 classifying the first and second relationship pair candidates, are accumulated. Although the case where the relationship pair storage units 26 and 27 are provided has been described, the first and second classification units 23 and 24 do not classify the first and second relationship pair candidates, or the first and second classifications. When the units 23 and 24 do not accumulate the first and second relationship pairs (for example, flags the first and second relationship pairs stored in the first and second relationship pair candidate storage units 15 and 17) 1), the mutual machine learning device 1 may not include the first and second relationship pair storage units 26 and 27.

  Moreover, although this Embodiment demonstrated the case where the acquisition part 19 acquires a common pair, it does not need to be so. In that case, the mutual machine learning device 1 includes the first relationship pair candidate storage unit 15, the first irrelevant pair candidate storage unit 16, the second relationship pair candidate storage unit 17, the second irrelevant pair candidate storage unit 18, and the acquisition unit. 19 may not be provided. In this case, the process of storing the common pair (genuine common pair or virtual common pair) in the common pair storage unit 20 is not limited. For example, a common pair may be stored in the common pair storage unit 20 via a recording medium, or a common pair transmitted via a communication line or the like may be stored in the common pair storage unit 20. It may be.

  In addition, the mutual machine learning device 1 according to the present embodiment may further include an output unit (not shown) that outputs information generated in the device. The output target information may be, for example, the first relationship pair stored in the first relationship pair storage unit 26 or the second relationship pair stored in the second relationship pair storage unit 27. The information of the learning result by the first classification unit 23, the information of the learning result by the second classification unit 24, or other information may be used. The output from the output unit (not shown) may be, for example, transmitted via a communication line to a predetermined device or may be stored in a recording medium. The output unit (not shown) may or may not include a device (for example, a communication device) that performs output. The output unit (not shown) may be realized by hardware, or may be realized by software such as a driver that drives these devices.

  Moreover, although this Embodiment demonstrated the case where the 1st and 2nd classification | category part 23 and 24 classify a 1st relationship pair candidate and a 2nd relationship pair candidate after machine learning, the 1st and 1st The second classification units 23 and 24 may perform classification on the first irrelevant pair candidate and the second irrelevant pair candidate.

  In the present embodiment, the case where the mutual machine learning is performed using the first corpus and the second corpus has been described. However, the mutual machine learning apparatus according to the present embodiment using three or more corpora. It goes without saying that mutual machine learning similar to 1 may be performed. Even in such a case, if attention is paid to two corpuses among the three or more corpora, the same processing as the mutual machine learning apparatus 1 according to the present embodiment is performed.

  Moreover, although the case where the mutual machine learning device 1 is a stand-alone has been described in the above embodiment, the mutual machine learning device 1 may be a stand-alone device or a server device in a server / client system. Good. In the latter case, the output unit or the like may output information via a communication line, for example.

  In the above embodiment, each process or each function may be realized by centralized processing by a single device or a single system, or may be distributedly processed by a plurality of devices or a plurality of systems. It may be realized by doing.

  In the above embodiment, information related to processing executed by each component, for example, information received, acquired, selected, generated, transmitted, or received by each component In addition, information such as threshold values, mathematical formulas, addresses, etc. used by each component in processing is retained temporarily or over a long period of time on a recording medium (not shown) even when not explicitly stated in the above description. It may be. Further, the storage of information in the recording medium (not shown) may be performed by each component or a storage unit (not shown). Further, reading of information from the recording medium (not shown) may be performed by each component or a reading unit (not shown).

  In the above embodiment, when information used by each component, for example, information such as a threshold value, an address, and various setting values used by each component may be changed by the user Even if it is not specified in the above description, the user may be able to change the information as appropriate, or it may not be. If the information can be changed by the user, the change is realized by, for example, a not-shown receiving unit that receives a change instruction from the user and a changing unit (not shown) that changes the information in accordance with the change instruction. May be. The change instruction received by the receiving unit (not shown) may be received from an input device, information received via a communication line, or information read from a predetermined recording medium, for example. .

  In the above embodiment, when two or more components included in the mutual machine learning apparatus 1 have communication devices, input devices, etc., the two or more components have a physically single device. Or may have separate devices.

  In the above embodiment, each component may be configured by dedicated hardware, or a component that can be realized by software may be realized by executing a program. For example, each component can be realized by a program execution unit such as a CPU reading and executing a software program recorded on a recording medium such as a hard disk or a semiconductor memory. In addition, the software which implement | achieves the mutual machine learning apparatus 1 in the said embodiment is the following programs. That is, this program is different from the first method in that a plurality of first relationship pair candidates that are extracted from the first corpus by the first method and that are candidates for pairs of language expressions having a semantic relationship. A common pair that is a common pair that is extracted from the second corpus by the method 2 and is common to a plurality of second relationship pair candidates that are candidates for a pair of language expressions having a semantic relationship; A virtual common pair that is a common pair that is common to a plurality of first unrelated pair candidates and a plurality of second relationship pair candidates that are extracted from A virtual pair that is a common pair that is extracted from the second corpus and is common to a plurality of second unrelated pair candidates that are candidates for pairs of language expressions having no semantic relationship and a plurality of first relationship pair candidates A common pair storage unit that stores common pairs, and first learning data that is teacher data used in machine learning related to classification of whether or not the first relationship pair candidate has a semantic relationship is stored. A first learning data storage unit and a second learning data storage that stores second learning data that is teacher data used in machine learning related to classification of whether or not the second relationship pair candidate has a semantic relationship A computer which can access the computer and classifies whether the genuine common pair and the virtual common pair have a semantic relationship by performing machine learning using the first learning data and using the result of the machine learning Machine learning is performed using one classification unit and second learning data, and the genuine common pair and the virtual common pair have a semantic relationship using the result of the machine learning. A second classifying unit that classifies whether or not the classification result of the common pair according to the classification of the first classifying unit and at least one of the certainty and the classification result related to the common pair are second learning data And an additional unit that adds the common pair and the classification result related to the common pair to the first learning data according to at least one of the classification result and the certainty factor of the common pair by the classification of the second classification unit And the machine learning and classification by the first and second classification units and the addition of learning data by the addition unit are repeatedly executed.

  Moreover, the following programs may be sufficient as the software which implement | achieves the mutual machine learning apparatus 1 in the said embodiment. That is, this program is different from the first method in that a plurality of first relationship pair candidates that are extracted from the first corpus by the first method and that are candidate language expression pairs having a semantic relationship. Genuine common pairs that are common pairs that are extracted from the second corpus by the second method and that are common to a plurality of second relationship pair candidates that are candidates for pairs of language expressions having semantic relations, and A plurality of first relationship pair candidates, a plurality of second relationship pair candidates, and a plurality of first irrelevant candidates that are extracted from the first corpus and have language expression pairs that do not have the semantic relationship A pair that is not a genuine common pair among a plurality of second irrelevant pair candidates that are candidates for a pair of language expressions that are extracted from the second corpus and that have no semantic relationship. A common pair storage unit that stores a virtual common pair that is a certain common pair, and first teacher data that is used in machine learning related to classification as to whether the first relation pair candidate has the semantic relationship or not. A first learning data storage unit storing learning data, and second learning data that is teacher data used in machine learning related to classification as to whether the second relationship pair candidate has the semantic relationship. A computer accessible to the stored second learning data storage unit performs machine learning using the first learning data, and uses the result of the machine learning to generate the genuine common pair and the virtual common pair. A first classifying unit that classifies whether or not has the semantic relationship, performs machine learning using the second learning data, and uses a result of the machine learning , A second classification unit that classifies whether the genuine common pair and the virtual common pair have the semantic relationship, and at least one of the classification result of the common pair and the certainty factor according to the classification of the first classification unit Accordingly, the common pair and the classification result related to the common pair are added to the second learning data, and the common pair is classified according to at least one of the classification result and the certainty factor of the common pair by the classification of the second classification unit. The pair and the classification result related to the common pair are caused to function as an addition unit that adds to the first learning data, and machine learning and classification by the first and second classification units, and addition of learning data by the addition unit It is a program that is executed repeatedly.

  In the program, functions realized by the program do not include functions that can be realized only by hardware. For example, functions that can be realized only by hardware such as a modem or an interface card in an acquisition unit that acquires information, an output unit that outputs information, and the like are not included in at least the functions realized by the program.

  Further, this program may be executed by being downloaded from a server or the like, and a program recorded on a predetermined recording medium (for example, an optical disk such as a CD-ROM, a magnetic disk, a semiconductor memory, or the like) is read out. May be executed by Further, this program may be used as a program constituting a program product.

  Further, the computer that executes this program may be singular or plural. That is, centralized processing may be performed, or distributed processing may be performed.

  FIG. 8 is a schematic diagram illustrating an example of an external appearance of a computer that executes the program and realizes the mutual machine learning device 1 according to the embodiment. The above-described embodiment can be realized by computer hardware and a computer program executed on the computer hardware.

  In FIG. 8, a computer system 900 includes a computer 901 including a CD-ROM (Compact Disk Read Only Memory) drive 905, an FD (Floppy (registered trademark) Disk) drive 906, a keyboard 902, a mouse 903, a monitor 904, and the like. Is provided.

  FIG. 9 is a diagram showing an internal configuration of the computer system 900. In FIG. 9, in addition to the CD-ROM drive 905 and the FD drive 906, a computer 901 is connected to an MPU (Micro Processing Unit) 911, a ROM 912 for storing a program such as a bootup program, and the MPU 911. A RAM (Random Access Memory) 913 that temporarily stores program instructions and provides a temporary storage space, a hard disk 914 that stores application programs, system programs, and data, and an MPU 911 and a ROM 912 are interconnected. And a bus 915. The computer 901 may include a network card (not shown) that provides connection to the LAN.

  A program that causes the computer system 900 to execute the functions of the mutual machine learning apparatus 1 according to the above-described embodiment is stored in the CD-ROM 921 or the FD 922, inserted into the CD-ROM drive 905 or the FD drive 906, and the hard disk 914. May be forwarded to. Instead, the program may be transmitted to the computer 901 via a network (not shown) and stored in the hard disk 914. The program is loaded into the RAM 913 when executed. The program may be loaded directly from the CD-ROM 921, the FD 922, or the network.

  The program does not necessarily include an operating system (OS) or a third party program that causes the computer 901 to execute the functions of the mutual machine learning device 1 according to the above-described embodiment. The program may include only a part of an instruction that calls an appropriate function (module) in a controlled manner and obtains a desired result. How the computer system 900 operates is well known and will not be described in detail.

  Further, the present invention is not limited to the above-described embodiment, and various modifications are possible, and it goes without saying that these are also included in the scope of the present invention.

  As described above, according to the mutual machine learning device or the like according to the present invention, an effect that a machine learning with higher accuracy can be realized is obtained, and it is useful as a device for performing machine learning.

DESCRIPTION OF SYMBOLS 1 Mutual machine learning apparatus 11 1st corpus memory | storage part 12 2nd corpus memory | storage part 13 1st extraction part 14 2nd extraction part 15 1st relationship pair candidate memory | storage part 16 1st unrelated pair candidate memory | storage part 17 2nd Related pair candidate storage unit 18 Second unrelated pair candidate storage unit 19 Acquisition unit 20 Common pair storage unit 21 First learning data storage unit 22 Second learning data storage unit 23 First classification unit 24 Second classification unit 25 Additional unit 26 First relationship pair storage unit 27 Second relationship pair storage unit

Claims (14)

A plurality of first relationship pair candidates that are extracted from the first corpus by the first method and that are candidates for pairs of linguistic expressions having a semantic relationship, and a second method different from the first method. Extracted from two corpora, a common pair common to a plurality of second relationship pair candidates that are candidates for a pair of language expressions having a semantic relationship, and extracted from the first corpus In addition, a plurality of first unrelated pair candidates that are candidates for pairs of language expressions that do not have the semantic relationship, a virtual common pair that is a common pair common to the plurality of second relationship pair candidates, and A common pair that is extracted from a second corpus and is common to a plurality of second unrelated pair candidates that are candidates for a pair of language expressions having no semantic relationship and the plurality of first relationship pair candidates. Ba A common pair storage unit and the Virtual common pair is stored,
A first learning data storage unit that stores first learning data that is teacher data used in machine learning related to classification of whether the first relationship pair candidate has the semantic relationship;
A first classification unit that performs machine learning using the first learning data, and classifies whether the genuine common pair and the virtual common pair have the semantic relationship using the result of the machine learning When,
A second learning data storage unit that stores second learning data that is teacher data used in machine learning related to classification of whether the second relationship pair candidate has the semantic relationship;
A second classification unit that performs machine learning using the second learning data, and classifies whether the genuine common pair and the virtual common pair have the semantic relationship using the result of the machine learning When,
Adding the common pair and the classification result related to the common pair to the second learning data according to at least one of the classification result and the certainty factor of the common pair according to the classification of the first classification unit; An additional unit that adds the common pair and the classification result related to the common pair to the first learning data according to at least one of the classification result of the common pair by the classification of the classification unit and the certainty factor,
A mutual machine learning device in which machine learning and classification by the first and second classification units and addition of learning data by the adding unit are repeatedly executed. A plurality of first relationship pair candidates that are extracted from the first corpus by the first method and that are candidates for pairs of linguistic expressions having a semantic relationship, and a second method different from the first method. Genuine common pairs extracted from two corpora and common to a plurality of second relationship pair candidates that are candidate language expression pairs having a semantic relationship, and the plurality of first relationship pairs A plurality of second irrelevant pair candidates extracted from the first corpus, a plurality of first irrelevant pair candidates that are extracted from the first corpus and are linguistic expression pair candidates that do not have the semantic relationship, Birch that is a common pair that is a pair that is not a genuine common pair among a plurality of second irrelevant pair candidates that are extracted from the corpus of 2 and that are candidate language expression pairs that do not have a semantic relationship A common pair storage unit Le common pair is stored,
A first learning data storage unit that stores first learning data that is teacher data used in machine learning related to classification of whether the first relationship pair candidate has the semantic relationship;
A first classification unit that performs machine learning using the first learning data, and classifies whether the genuine common pair and the virtual common pair have the semantic relationship using the result of the machine learning When,
A second learning data storage unit that stores second learning data that is teacher data used in machine learning related to classification of whether the second relationship pair candidate has the semantic relationship;
A second classification unit that performs machine learning using the second learning data, and classifies whether the genuine common pair and the virtual common pair have the semantic relationship using the result of the machine learning When,
Adding the common pair and the classification result related to the common pair to the second learning data according to at least one of the classification result and the certainty factor of the common pair according to the classification of the first classification unit; An additional unit that adds the common pair and the classification result related to the common pair to the first learning data according to at least one of the classification result of the common pair by the classification of the classification unit and the certainty factor,
A mutual machine learning device in which machine learning and classification by the first and second classification units and addition of learning data by the adding unit are repeatedly executed. The additional part is:
A common pair having a high certainty factor by the classification of the first classification unit and a classification result related to the common pair are added to the second learning data, and a common pair having a high certainty factor by the classification of the second classification unit; The mutual learning device according to claim 1, wherein a classification result related to the common pair is added to the first learning data. The additional part is:
A common pair having a high certainty factor by the classification of the first classification unit and having the same classification result of the first and second classification units and a classification result related to the common pair are added to the second learning data. A common pair having a high certainty factor by the classification of the second classification unit and having the same classification result of the first and second classification units and a classification result related to the common pair are added to the first learning data. The mutual machine learning device according to claim 3. The additional part is:
Adding a common pair having a high certainty factor by the classification of the first classification unit and a low certainty factor by the classification of the second classification unit and a classification result related to the common pair to the second learning data; A common pair having a high certainty factor by the classification of the second classification unit and a low certainty factor by the classification of the first classification unit and a classification result related to the common pair are added to the first learning data. The mutual machine learning device according to claim 4. A first relationship pair candidate storage unit in which the plurality of first relationship pair candidates are stored;
A first irrelevant pair candidate storage unit that stores the plurality of first irrelevant pair candidates;
A second relationship pair candidate storage unit in which the plurality of second relationship pair candidates are stored;
A second irrelevant pair candidate storage unit in which the plurality of second irrelevant pair candidates are stored;
Using the plurality of first relationship pair candidates and the plurality of second relationship pair candidates, the genuine common pair is acquired and stored in the common pair storage unit, and the plurality of first relationship pair candidates and the plurality Using the second relationship pair candidate, the plurality of first irrelevant pair candidates and the plurality of second irrelevant pair candidates, an acquisition unit that acquires the virtual common pair and accumulates it in the common pair storage unit; The mutual machine learning apparatus according to claim 1, further comprising: A first corpus storage unit in which the first corpus is stored;
A second corpus storage unit for storing the second corpus;
The plurality of first relationship pair candidates are extracted from the first corpus and stored in the first relationship pair candidate storage unit, and the plurality of first unrelated pair candidates are extracted from the first corpus. A first extraction unit that accumulates in one unrelated pair candidate storage unit;
The plurality of second relationship pair candidates are extracted from the second corpus and accumulated in the second relationship pair candidate storage unit, and the plurality of second unrelated pair candidates are extracted from the second corpus. The mutual machine learning device according to claim 6, further comprising: a second extraction unit that accumulates in the two unrelated pair candidate storage units. The first classifying unit classifies the plurality of first relationship pair candidates after repeating machine learning and classification and addition of learning data,
The mutual machine learning according to claim 6 or 7, wherein the second classifying unit classifies the plurality of second relationship pair candidates after machine learning and repeating the classification and the addition of learning data. apparatus. The first corpus is a structured corpus;
The mutual machine learning device according to claim 1, wherein the second corpus is a corpus of an unstructured natural language sentence. The mutual machine learning device according to claim 1, wherein the semantic relationship is a high-order and low-order relationship. A plurality of first relationship pair candidates that are extracted from the first corpus by the first method and that are candidates for pairs of linguistic expressions having a semantic relationship, and a second method different from the first method. Extracted from two corpora, a common pair common to a plurality of second relationship pair candidates that are candidates for a pair of language expressions having a semantic relationship, and extracted from the first corpus In addition, a plurality of first unrelated pair candidates that are candidates for pairs of language expressions that do not have the semantic relationship, a virtual common pair that is a common pair common to the plurality of second relationship pair candidates, and A common pair that is extracted from a second corpus and is common to a plurality of second unrelated pair candidates that are candidates for a pair of language expressions having no semantic relationship and the plurality of first relationship pair candidates. Ba A common pair storage unit that stores a char common pair, and first learning data that is teacher data used in machine learning related to classification of whether the first relationship pair candidate has the semantic relationship is stored. The first learning data storage unit, the first classifying unit, and the second training data used in machine learning related to the classification of whether the second relationship pair candidate has the semantic relationship or not A mutual machine learning method that is processed using a second learning data storage unit that stores learning data, a second classification unit, and an adding unit,
Whether the first classifying unit performs machine learning using the first learning data, and the genuine common pair and the virtual common pair have the semantic relationship using the result of the machine learning A first classification step to classify whether;
Whether the second classifying unit performs machine learning using the second learning data, and whether the genuine common pair and the virtual common pair have the semantic relationship using the result of the machine learning A second classification step for classifying whether;
In accordance with at least one of the classification result of the common pair and the certainty level by the classification in the first classification step, the adding unit converts the common pair and the classification result related to the common pair into the second learning data. And adding the common pair and the classification result related to the common pair to the first learning data according to at least one of the classification result and the certainty factor of the common pair based on the classification in the second classification step. An additional step,
A mutual machine learning method in which machine learning and classification in the first and second classification steps and addition of learning data in the adding step are repeatedly executed. A plurality of first relationship pair candidates that are extracted from the first corpus by the first method and that are candidates for pairs of linguistic expressions having a semantic relationship, and a second method different from the first method. Genuine common pairs extracted from two corpora and common to a plurality of second relationship pair candidates that are candidate language expression pairs having a semantic relationship, and the plurality of first relationship pairs A plurality of second irrelevant pair candidates extracted from the first corpus, a plurality of first irrelevant pair candidates that are extracted from the first corpus and are linguistic expression pair candidates that do not have the semantic relationship, Birch that is a common pair that is a pair that is not a genuine common pair among a plurality of second irrelevant pair candidates that are extracted from the corpus of 2 and that are candidate language expression pairs that do not have a semantic relationship A common pair storage unit that stores a common pair, and first learning data that is teacher data used in machine learning regarding classification of whether the first relationship pair candidate has the semantic relationship is stored A first learning data storage unit, a first classifying unit, and a second learning which is teacher data used in machine learning regarding classification of whether the second relationship pair candidate has the semantic relationship A mutual machine learning method processed using a second learning data storage unit in which data is stored, a second classification unit, and an addition unit,
Whether the first classifying unit performs machine learning using the first learning data, and the genuine common pair and the virtual common pair have the semantic relationship using the result of the machine learning A first classification step to classify whether;
Whether the second classifying unit performs machine learning using the second learning data, and whether the genuine common pair and the virtual common pair have the semantic relationship using the result of the machine learning A second classification step for classifying whether;
In accordance with at least one of the classification result of the common pair and the certainty level by the classification in the first classification step, the adding unit converts the common pair and the classification result related to the common pair into the second learning data. And adding the common pair and the classification result related to the common pair to the first learning data according to at least one of the classification result and the certainty factor of the common pair based on the classification in the second classification step. An additional step,
A mutual machine learning method in which machine learning and classification in the first and second classification steps and addition of learning data in the adding step are repeatedly executed. A plurality of first relationship pair candidates that are extracted from the first corpus by the first method and that are candidates for pairs of linguistic expressions having a semantic relationship, and a second method different from the first method. Extracted from two corpora, a common pair common to a plurality of second relationship pair candidates that are candidates for a pair of language expressions having a semantic relationship, and extracted from the first corpus In addition, a plurality of first unrelated pair candidates that are candidates for pairs of language expressions that do not have the semantic relationship, a virtual common pair that is a common pair common to the plurality of second relationship pair candidates, and A common pair that is extracted from a second corpus and is common to a plurality of second unrelated pair candidates that are candidates for a pair of language expressions having no semantic relationship and the plurality of first relationship pair candidates. Ba A common pair storage unit that stores a char common pair, and first learning data that is teacher data used in machine learning related to classification of whether the first relationship pair candidate has the semantic relationship is stored. A first learning data storage unit, and second learning data that is teacher data used in machine learning related to classification of whether the second relationship pair candidate has the semantic relationship is stored. A computer that can access the learning data storage unit
A first classification unit that performs machine learning using the first learning data, and classifies whether the genuine common pair and the virtual common pair have the semantic relationship using the result of the machine learning ,
A second classification unit that performs machine learning using the second learning data, and classifies whether the genuine common pair and the virtual common pair have the semantic relationship using the result of the machine learning ,
Adding the common pair and the classification result related to the common pair to the second learning data according to at least one of the classification result and the certainty factor of the common pair according to the classification of the first classification unit; In accordance with at least one of the classification result and the certainty of the common pair according to the classification of the classification unit, the common pair and the classification result related to the common pair are functioned as an additional unit that adds to the first learning data,
A program in which machine learning and classification by the first and second classification units and addition of learning data by the addition unit are repeatedly executed. A plurality of first relationship pair candidates that are extracted from the first corpus by the first method and that are candidates for pairs of linguistic expressions having a semantic relationship, and a second method different from the first method. Genuine common pairs extracted from two corpora and common to a plurality of second relationship pair candidates that are candidate language expression pairs having a semantic relationship, and the plurality of first relationship pairs A plurality of second irrelevant pair candidates extracted from the first corpus, a plurality of first irrelevant pair candidates that are extracted from the first corpus and are linguistic expression pair candidates that do not have the semantic relationship, Birch that is a common pair that is a pair that is not a genuine common pair among a plurality of second irrelevant pair candidates that are extracted from the corpus of 2 and that are candidate language expression pairs that do not have a semantic relationship A common pair storage unit that stores a common pair, and first learning data that is teacher data used in machine learning regarding classification of whether the first relationship pair candidate has the semantic relationship is stored A first learning data storage unit that stores second learning data that is teacher data used in machine learning related to classification as to whether or not the second relationship pair candidate has the semantic relationship. A computer accessible to the learning data storage unit of
A first classification unit that performs machine learning using the first learning data, and classifies whether the genuine common pair and the virtual common pair have the semantic relationship using the result of the machine learning ,
A second classification unit that performs machine learning using the second learning data, and classifies whether the genuine common pair and the virtual common pair have the semantic relationship using the result of the machine learning ,
Adding the common pair and the classification result related to the common pair to the second learning data according to at least one of the classification result and the certainty factor of the common pair according to the classification of the first classification unit; In accordance with at least one of the classification result and the certainty of the common pair according to the classification of the classification unit, the common pair and the classification result related to the common pair are functioned as an additional unit that adds to the first learning data,
A program in which machine learning and classification by the first and second classification units and addition of learning data by the addition unit are repeatedly executed.

Images