JP2012159983A - Analogy device, analogy method and program - Google Patents

Abstract

PROBLEM TO BE SOLVED: To achieve analogy even when a concept to which a base belongs is different from a concept to which a target belongs and conditions of base and target consist of multiple words in four-term analogy.SOLUTION: A base condition partial combination generation part 21 generates all partial combinations Awhen selecting the predetermined number of words from words that constitute a base condition A, and a base result partial combination generation part 22 generates all partial combinations Bwhen selecting the predetermined number of words from words that constitute a base result B. A relation set generation part 24 extracts a word rthat associates Awith Bfrom article data, with respect to each of all combinations of Aand B. A target condition partial combination generation part 26 generates all partial combinations Cwhen selecting the predetermined number of words that constitute a target condition C, and an analogy result generation part 28 extracts a word xthat is associated with Cby the word rfrom the article data, with respect to each of all combinations of Cand each word r.

Description

  The present invention relates to an analogy device, an analogy method, and a program for performing analogy based on cases.

  As one of the basic analogy methods, four-term analogy is known. The four-term analogy is generally expressed as the following formula (1).

  A: B = C: X? ... (1)

  The above means “if A, what is B, what is C?” And X represents the analogy result. That is, A, B, C to X? It is a four-term analogy to find In the four-term analogy, a specified area (base) and a target area (target) are set as basic components. The base is existing knowledge used for analogy, and the target is an unknown problem that must be solved. That is, in the above four-term analogy example, A and B belong to the base, and C and X belong to the target. Non-Patent Document 1 proposes a system that realizes this four-term analogy algorithm. FIG. 13 is a diagram showing an outline of the algorithm.

  In the algorithm shown in FIG. 13, first, a relation extraction process for obtaining a relation set R between A and B in the base is performed. Subsequently, the relation set R obtained in the relation extraction process is transferred to the target, and a relation mapping process for obtaining X by applying the relation set R to C is performed. Hereinafter, the base A is the situation, the result B is the result of the situation, and this set of A and B is called one case.

In Non-Patent Document 1, a relation set R obtained in the relation extraction process is defined as a set of words r _i representing a similar relation as shown in the following expression (2).

R = {r _i } (i is an integer of 1 or more) (2)

In the relationship mapping process, a word x _j that is a candidate for a plurality of analogy results is obtained from C using each word r _i included in the relationship set R, and the likelihood is quantitatively determined for each of the obtained words x _j. Score (x _j ) which is a value shown in FIG. That is, the analogy result is a set composed of the word x _j and the score (x _j ) assigned to it, and is expressed by the following equation (3).

X = {x _j } (j is an integer of 1 or more) (3)

On the other hand, in Non-Patent Document 2, an analogy is performed as follows. First, a score of the TF-IDF value is assigned to each word extracted from the explanatory text corresponding to the DVD title, and the score is corrected according to the user's selection operation. Then, the recommended DVD data is selected based on the DVD data whose corrected score is composed of the upper words and the input words.
In Non-Patent Document 3, the analogy is performed as follows. First, a word vector composed of words having a higher TF-IDF value is generated from article data of the past week, and candidate words are extracted from article data of the next day of the week based on the generated word vector. A plurality of combinations of this word vector and candidate words are generated while shifting the time of article data. Then, a word vector is similarly generated from the article data of one week prior to the predicted date, and matching is performed with the word vector generated from the past article data, and candidate words corresponding to the matching word vectors are used as the prediction results. Yes.

Kazuhiro Okada, 4 others, “Analogue Method Based on Structural Mapping Theory”, Proceedings of the 35th Fuzzy Workshop, March 2010, p. 25-30 Yasuhiko Date, 7 others, “Proposal of DVD recommendation system using CFS”, Proceedings of the 35th Fuzzy Workshop, Intelligent Information Technology Fuzzy Society, March 2010, p. 33-36 Shinichiro Ito, 1 other, “Predicting Economic Trends by Language,” Proceedings of the 35th Fuzzy Workshop, March 2010, p. 37-40

  In general, in the four-term analogy, both the base situation A and the target situation C are so-called one-dimensional expressions composed of one word, and this is the same in Non-Patent Document 1. On the other hand, in Non-Patent Document 2 and Non-Patent Document 3, an analogy can be made by taking a situation represented by a plurality of words as an example, but both the target and the base must belong to the same concept. For example, in Non-Patent Document 1, an analogy can be performed even if the target belongs to the concept related to fish and the base belongs to the concept related to bird, such as “What is scale if fish? What is it if bird?”. . However, in Non-Patent Document 2, both the target and base belong to the concept of DVD, and in Non-Patent Document 3, both the target and base belong to the concept of economy.

  The present invention has been made in consideration of such circumstances, and its purpose is to provide a case in which a situation and a result in a base and a situation in a target belonging to a concept different from the concept to which the base belongs are given. In addition, an object of the present invention is to provide an analogy device, an analogy method, and a program capable of obtaining an analogy result in a target even when the base and target situations are each composed of a plurality of words.

  The present invention has been made to solve the above-mentioned problem, and a base situation partial combination for generating base situation partial combination data composed of a predetermined number of words selected from a plurality of words indicated by base situation data by different combinations. A generation unit, a base result partial combination generation unit that generates base result partial combination data including a predetermined number of words selected from a plurality of words indicated by the base result data, and the base situation partial combination data For each of the different combinations of one and one of the base result partial combination data, an association word that is a word relating the word indicated by the base situation partial combination data and the word indicated by the base result partial combination data, Relational aggregates extracted from article data stored in storage A target situation partial combination generation unit that generates target situation partial combination data composed of a predetermined number of words selected from a plurality of words indicated by the target situation data, and one of the target situation partial combination data For each different combination consisting of one and one of the association words, a word related to the word indicated by the target situation partial combination data by the association word is used as an analogy result from the article data stored in the article storage device. And an analogy result generating unit for extraction.

  In addition, the present invention is the analogy device described above, and for each of the association words extracted by the relation set generation unit, the association word obtained from the article data stored in the article storage device and the base situation part The strength of the association between the combination data and the co-occurrence of the base result partial combination data, the strength of the association between the association word and the base situation partial combination data, and the association word and the base result partial combination data Based on the strength of the relation, the relation validity score calculation part that calculates the relation validity score that quantitatively represents the validity as the word that relates the situation and the result, and the analogy result generation part extract as the analogy result For each of these words, the word obtained from the article data stored in the article storage device and the target Strength of association between the situation situation combination data and the co-occurrence of the associated word, strength of association between the word and the target situation partial combination data, and strength of association between the word and the association word, And an analogy result validity score calculation unit that calculates an analogy result validity score that quantitatively represents the validity as an analogy result based on the relation validity score calculated for the association word. Features.

  In addition, the present invention provides the analogy device described above, wherein, for each of a plurality of case data composed of base situation data and base result data, the base situation partial combination generation unit is configured to base the base situation data on the base situation data constituting the case data. Generating situation partial combination data, causing the base result partial combination generation unit to generate the base result partial combination data from the base result data constituting the case data, and causing the relation set generation unit to generate from the base situation data An association word is extracted from article data for each of different combinations of one of the generated base situation partial combination data and one of the base result partial combination data generated from the base result data, and the relation validity Relevant for each of the associated words in the sex score calculator A correctness score is calculated, and the analogy result generation unit is configured to extract an analogy result word from article data for each of the different combinations of one of the target situation partial combination data and one of the associated words, and the analogy result An analogy processing control unit that causes the result validity score calculation unit to calculate an analogy result validity score for each of the words extracted as an analogy result, and the same included in the analogy result word obtained for each case data And an analogy result accumulating unit for accumulating the analogy result validity scores calculated for the integrated identical words.

  Further, the present invention is the analogy device described above, wherein the relation set generation unit is a noun whose subject is the word indicated by the base situation partial combination data and a noun whose predicate is the word indicated by the base result partial combination data The verb of the predicate is extracted as the association word from the sentence of the article data, and the analogy result generation unit is configured such that the plurality of words indicated by the target situation partial combination data are subject nouns and the association word is a predicate verb. The noun of the predicate is extracted as an analogy result from the sentence of the article data.

  Further, the present invention is the analogy device described above, wherein the article storage device further stores dictionary data including words relating to a predetermined field, and the relation set generation unit is indicated by the base situation partial combination data A word included in the dictionary data is extracted as the association word from a sentence of the article data in which a word and a word indicated by the base result partial combination data co-occur, and the analogy result generation unit is configured to output the target situation partial combination A word included in the dictionary data is extracted as an analogy result from a sentence of the article data in which a word indicated by data and the associated word co-occur.

  The present invention is also an analogy method executed by an analogy estimation device, wherein the base situation partial combination generation unit includes a predetermined number of words selected from a plurality of words indicated by the base situation data, and the base situation partial combination A base situation partial combination generation process for generating data, and a base result partial combination generation unit generate base result partial combination data composed of a predetermined number of words selected by different combinations from a plurality of words indicated by the base result data The base result partial combination generation process and the relation set generation unit indicate the word indicated by the base situation partial combination data for each of the different combinations of one of the base situation partial combination data and one of the base result partial combination data. And a word related to the word indicated by the base result partial combination data The relation set generation process of extracting a certain association word from the article data stored in the article storage device, and the target situation partial combination generation unit of a predetermined number of words selected by a different combination from a plurality of words indicated by the target situation data A target situation partial combination generation process for generating target situation partial combination data consisting of the words, and an analogy result generation unit for each different combination consisting of one of the target situation partial combination data and one of the association words, An analogy result generation step of extracting a word related to the word indicated by the target situation partial combination data by the association word as an analogy result from the article data stored in the article storage device. This is an analogy method.

  Further, the present invention provides a base situation partial combination generation unit that generates base situation partial combination data including a predetermined number of words selected from a plurality of words indicated by the base situation data by using a computer used as an analogy device. A base result partial combination generation unit that generates base result partial combination data including a predetermined number of words selected from a plurality of words indicated by base result data, and one of the base situation partial combination data and the base result For each different combination consisting of one of the partial combination data, an association word, which is a word relating the word indicated by the base situation partial combination data and the word indicated by the base result partial combination data, is stored in the article storage device. Relations extracted from live article data A target situation partial combination generation unit that generates target situation partial combination data including a predetermined number of words selected from a plurality of words indicated by target situation data, and one of the target situation partial combination data and the target situation data For each different combination consisting of one of the association words, a word related to the word indicated by the target situation partial combination data by the association word is extracted from the article data stored in the article storage device as an analogy result It is a program characterized by functioning as an analogy result generation unit.

  According to the present embodiment, when a situation consisting of a situation and a result in a base and a situation in a target belonging to a concept different from the concept to which the base belongs are given, each of the situation of the base and the target is composed of a plurality of words. Even in such a case, the analogy result at the target can be obtained.

It is a figure which shows the outline | summary of the relationship extraction process of the analogy device by the 1st Embodiment of this invention. It is a figure which shows the outline | summary of the relationship mapping process of the analogy device by the same embodiment. It is a block diagram which shows the structure of the analogy device by the same embodiment. It is a figure which shows the analogy processing flow of the analogy device by the same embodiment. It is a figure which shows the analogy processing flow of the analogy device by the same embodiment. It is a figure for demonstrating the extraction process of the word r by the embodiment. It is a figure for demonstrating the extraction process of the word x by the embodiment. It is a figure for demonstrating the extraction process of the word r and the word x by the embodiment. It is a figure which shows the process outline | summary of the analogy estimation apparatus by the 2nd Embodiment of this invention. It is a block diagram which shows the structure of the analogy device by the same embodiment. It is a figure which shows the processing flow of the analogy device by the same embodiment. It is a figure which shows the processing flow of the analogy device by the same embodiment. It is a figure which shows the analogy algorithm of a prior art.

  Hereinafter, embodiments of the present invention will be described with reference to the drawings.

[First embodiment]
Generally, the situation A of the specified area (base) and the situation C of the target area (target) in the four-term analogy are so-called one-dimensional expressions composed of one word. In the present embodiment, the situation A and the situation C are extended to a multi-dimensional configuration composed of a plurality of words. If this is described in the same expression method as modus ponens shown in equation (1), the following equation (4) is obtained.

As described above, conditions A, the word _{t a1,} t a2, _..., is composed of _{t am,} status C is the word _{t c1,} t c2, _..., composed of _{t cm.} That is, the situation A and the situation C are m-dimensional word vectors having m words as elements. On the other hand, the result B in the base situation A is composed of words w _b1 , w _b2 ,..., W _bg , and the result X in the target situation C is composed of words w _x1 , w _{x2 ,.} Is done. That is, the result B and the result X are g-dimensional word vectors having g words as elements.
The analogy device of the present embodiment is based on the base situation A, which is a multidimensional vector as shown in Expression (4), the result B in the base situation A, and the target situation C, and the result in the target situation C. Processing for obtaining X is performed.

The processing outline of the analogy device of this embodiment will be described with reference to FIGS.
FIG. 1 is a diagram showing an outline of the relationship extraction process in the analogy device of this embodiment. The analogy device selects all n (n is an integer from 1 to m) words from an m-dimensional word vector that is a sequence of m (m is an integer of 2 or more) words representing the situation A. Generate a combination of Further, the analogy device selects h (h is an integer between 1 and g) words from a g-dimensional word vector that is a string of g words (g is an integer of 2 or more) representing the result B. All combinations are generated. The analog inference device searches for articles in which n words constituting the combination generated from the situation A and h words constituting the combination generated from the result B co-occur. An article includes a plurality of words, and can be regarded as a word vector composed of the plurality of words. Among the words constituting this word vector, there are also words that associate other words with each other. The analogy device extracts from the article a word r that associates n words that make up the combination generated from the situation A and h words that make up the combination generated from the result B. If the extracted words r are r ₁ , r ₂ ,..., _They can be expressed as a relation set R = {r _i } (i is an integer of 1 or more).

FIG. 2 is a diagram showing an overview of the relationship mapping process. The analog inference apparatus generates all combinations when n words are selected from an m-dimensional word vector that is a sequence of m words constituting the situation C. Analogy apparatus, by each word r _i constituting the relationship set R, extracting word associated with the combination generated from the situation C from the article, the extracted word group and analogy result X.

  FIG. 3 is a block diagram showing the configuration of the analogy device 1 according to the first embodiment of the present invention. The analogy device 1 can be realized by, for example, one or a plurality of computer devices, and is connected to an article storage device 5 that stores article data via a network. The article data is, for example, news text data, magazine text data, knowledge database contents, and the like. The article data is composed of a plurality of words, among which are words that relate other words. For example, in the case of a natural language sentence, the noun (word) of the subject and the noun (word) in the predicate are associated with the verb (word) of the predicate.

As shown in FIG. 1, the analogy device 1 includes a case storage unit 11, an input unit 12, a relationship extraction unit 13, a processing result storage unit 14, a relationship mapping unit 15, and an output unit 16.
The case storage unit 11 stores situation A data (base situation data) indicating a word vector of situation A and result B data (base result data) indicating a word vector of result B. The situation A data indicates an m-dimensional word vector having m words as elements, and the result B data indicates a g-dimensional word vector having g words as elements. The input unit 12 receives input of situation C data (target situation data) indicating a word vector of situation C using a keyboard or the like. The situation C data is an m-dimensional word vector having m words as elements. Note that the input unit 12 may receive the status C data from another computer device connected via a network, or read the status C data from a computer-readable recording medium. The processing result storage unit 14 stores processing results obtained by the units of the relationship extraction unit 13 and the relationship mapping unit 15.

The relationship extraction unit 13 includes a base situation partial combination generation unit 21, a base result partial combination generation unit 22, a base co-occurrence article search unit 23, a relation set generation unit 24, and a relation validity score calculation unit 25.
The base situation partial combination generation unit 21 selects all the combinations when n words are selected from the m words indicated by the situation A data read from the case storage unit 11 (in the case of n = 1, for convenience of explanation, combinations) And a partial combination A _l (base situation partial combination data) which is data indicating each of these combinations is generated (1 ≦ l ≦ _m C _n , where l is an integer). That is, the partial combination _Al indicates an n-dimensional word vector having n words as elements.

The base result partial combination generation unit 22 selects all the combinations when the h words are selected from the g words indicated by the result B data read from the case storage unit 11 (for the convenience of explanation, even when h = 1) And a partial combination B _k (base result partial combination data) that is data indicating each of these combinations is generated (1 ≦ k ≦ _g C _h , k is an integer). That is, the partial combination _Bk indicates an h-dimensional word vector having h words as elements.

The base co-occurrence article retrieval unit 23 retrieves article data stored in the article storage device 5 for each of all combinations including one partial combination A _l and one partial combination B _k, and obtains a partial combination A _l. and all the words indicated by the partial combination B _k to identify the article data indicating an article co-occurring. The base co-occurrence article search unit 23 generates an article set D that is data including a set of specified article data. The relation set generation unit 24 includes, for each combination of one partial combination A _l and one partial combination B _k , n words indicated by the partial combination A _l and h numbers indicated by the partial combination B _k. A word r (association word) related to the word is extracted from each piece of article data included in the article set D, and a relation set R that is data indicating the set of extracted words r is generated. Let each extracted word r be r ₁ , r ₂ ,. The relation validity score calculation unit 25 calculates a relation validity score score (r _i ) of each word r _i (i is an integer of 1 or more) included in the relation set R generated by the relation set generation unit 24. The relation validity score score (r _i ) is a value that quantitatively represents the likelihood that the word r _i is valid as a word that associates the partial combination A ₁ and the partial combination B _k .

The relationship mapping unit 15 includes a target situation partial combination generation unit 26, a target co-occurrence article search unit 27, an analogy result generation unit 28, and an analogy result validity score calculation unit 29.
The target situation partial combination generation unit 26 generates all combinations when n words are selected from the m words indicated by the situation C data input by the input unit 12, and is data indicating each of these combinations. A certain partial combination C _f (target situation partial combination data) is generated (1 ≦ f ≦ _m C _n , f is an integer). That is, the partial combination _Cf indicates an n-dimensional word vector having n words as elements.

The target co-occurrence article search unit 27 searches the article storage device 5 for all combinations of one partial combination C _f and one word r _i included in the relation set R, and the partial combination C _f indicates Article data indicating an article in which all the words and the word r _i co-occur are specified. The target co-occurrence article search unit 27 generates an article set E that is data including a set of identified article data. The analogy result generation unit 28 extracts the word x associated with the n words indicated by the partial combination C _f by the word r _i from each article data included in the article set E, and indicates the set of extracted words x An analogy result set X that is data is generated. Let each extracted word x be x ₁ , x ₂ ,. The analogy result validity score calculation unit 29 calculates an analogy result validity score score (x _j ) of each word x _j (j is an integer of 1 or more) included in the analogy result set X generated by the analogy result generation unit 28. calculate. The analogy result validity score score (x _j ) is a value that quantitatively represents whether the word x _j is valid as an analogy result.

The output unit 16 outputs each word x _j indicated by the analogy result set X generated by the analogy result generation unit 28 and the analogy result validity score score of the word x _j calculated by the analogy result validity score calculation unit 29 ( x _j )) is displayed on the display. Alternatively, the output unit 16 may print the analogy result data by a printer or the like, write it to an information recording medium, or send it to a computer device connected via a network. This analogy result data is a set of words that are _presumed to be any of the words w _x1 , w _x2 ,..., W _xg that are elements of the g-dimensional word vector representing the result X, and the analogy result validity score. . That analogy result is represented by a fuzzy set consisting of a word x _j.

4 and 5 are diagrams showing an analogy processing flow of the analogy device 1 shown in FIG.
In our daily use language, there are about 30,000 words. From this, we extract key words representing the characteristics of each article, and analogize the situation A data and result B data generated using the extracted key words. The data is stored in the case storage unit 11 of the device 1. Generally, 20 to 50 are used as the dimension number m of the word vector of the situation A and the dimension number g of the word vector of the result B, but other dimension numbers may be used. Further, the processing result storage unit 14 stores an article set D, an article set E, a relation set R, and an analogy result set X having an initial value NULL.

In FIG. 4, the input unit 12 of the analogy device 1 receives input of situation C data (step S100). Subsequently, the base situation partial combination generation unit 21 reads the situation A data from the case storage unit 11, and from m words that are elements of the word vector indicated by the situation A data (m is an integer of 2 or more), n pieces All combinations are selected when a word (n is an integer from 1 to m) is selected. The number of combinations is _m C _n . The base situation partial combination generation unit 21 generates a partial combination A _l (1 ≦ l ≦ _m C _n , where l is an integer) indicating an n-dimensional word vector having the generated word combination as an element, and the processing result storage unit 14 (Step S105). The number n of extracted words is about 2 to 5, but other values may be used.

Subsequently, the base result partial combination generation unit 22 reads the result B data from the case storage unit 11, and from the g words that are elements of the word vector indicated by the result B data, h (h is 1 to g). All combinations when an integer word is selected are generated. The number of combinations _becomes g _{C h.} The base result partial combination generation unit 22 generates a partial combination B _k (1 ≦ k ≦ _g C _h , k is an integer) indicating an h-dimensional word vector having the generated word combination as an element, and the processing result storage unit 14 (Step S110). The number of extracted words h is about 2 to 5, but other values may be used.

Based co-occurrence article retrieval unit 23, one of the subcombination _{A l} written in the processing result storage unit 14 in step S105, consists one subcombination _{B k} written in the processing result storage unit 14 in step S110 All combinations are generated (step S115). That is, the generated combinations are A ₁ -B ₁ , A ₁ -B ₂ ,..., A ₁ -B _(gCh) , A ₂ -B ₁ , A ₂ -B ₂ , ..., A _(mCn-1) − _{B (gCh), a (mCn} ) -B 1, a (mCn) -B 2, ..., a _{a (mCn) -B (gCh)} . The base co-occurrence article search unit 23 selects a combination A ₁ -B _k that has not yet been processed in step S 125 from the combinations generated in step S 115 (step S 120).

The base co-occurrence article search unit 23 reads out the partial combination A ₁ and the partial combination B _k constituting the combination A ₁ -B _k selected in Step S120 from the processing result storage unit 14. Based co-occurrence article retrieval unit 23 searches the article data stored in the article storage device 5, and the n words indicating read subcombination A _l is a h number of words indicated by the read part combination B _k is the Identify article data indicating all included articles. The base co-occurrence article search unit 23 reads the specified article data from the article storage device 5, and adds the read article data to the article set D stored in the processing result storage unit 14 (step S125). However, the base co-occurrence article search unit 23 does not add when the extracted article data is already included in the article set D. The article data included in the relation set D is assumed to be article data d ₁ , d ₂ ,.

  The base co-occurrence article search unit 23 determines whether all the combinations generated in step S115 are the processing targets of step S125 (step S130). If it is determined that there is a combination that has not yet been processed in step S125 (step S130: NO), the base co-occurrence article search unit 23 repeats the processing from step S120. When the base co-occurrence article search unit 23 determines that all combinations are to be processed (step S130: YES), the relation set generation unit 24 executes the process of step S140.

Relationship set generation unit 24, similarly to step S115, and generates all the combinations of one of one part combination _{B k} subcombination _{A l} (step S135). The relation set generation unit 24 selects the combination A ₁ -B _k that has not yet been processed in step S145 from the combinations generated in step S135 (step S140).

The relation set generation unit 24 selects the combination A _l selected from each of the articles indicated by the article data d ₁ , d ₂ ,... Included in the article set D written in the processing result storage unit 14 by the base co-occurrence article search unit 23. and n words indicated subcombinations _{a l} of -B _k, the word r relating the h pieces of words indicated by the partial combination _{B k} extracted (step S145). The relation set generation unit 24 adds the extracted word to the relation set R stored in the processing result storage unit 14. However, the relation set generation unit 24 does not add when the extracted word r is already included in the relation set R. Words r included in the relation set R are defined as words r ₁ , r ₂ ,. When the dimension numbers m and g are 20 to 50 and the extracted word numbers n and h are 2 to 5, the number of words included in the relation set R is about 20 to 50 words.

  The relation set generation unit 24 determines whether all the combinations generated in step S135 are the processing targets of step S145 (step S150). When it is determined that there is a combination that has not yet been processed in step S145 (step S150: NO), the relation set generation unit 24 repeats the processing from step S140. When the relation set generation unit 24 determines that all combinations are to be processed in step S145 (step S150: YES), the relation validity score calculation unit 25 executes the process in step S155.

The relation validity score calculation unit 25 uses the following formula (5) for each word r _i (i is an integer of 1 or more) included in the relation set R written in the processing result storage unit 14 by the relation set generation unit 24. calculating a relationship validity score score _{(r i),} the writing in the processing result storage unit 14 (step S155). Where MI (A ₁ , r _i ) is the mutual information amount between the partial combination A ₁ and the word r _i , MI (B _k , r _i ) is the mutual information amount between the partial combination B _k and the word r _i , MI (A ₁ B _k , r _i ) is a mutual information amount between the co-occurrence of the partial combination A ₁ and the partial combination B _k and the word r _i . The mutual information amount is an amount representing a measure on which two random variables depend on each other.

score (r _i ) = Σ ₁ Σ _k MI (A ₁ , r _i ) * MI (B _k , r _i ) * MI (A ₁ B _k , r _i ) (5)

As described above, the equation (5) is obtained by calculating the mutual information MI (A _l , r _i ), the mutual information MI (B _k , r _i ), and the mutual information MI (A _l B _k , r _i ). Is a value _{obtained by} cumulatively adding all the partial combinations A ₁ and all the partial combinations B _k . Accordingly, the larger the mutual information MI (A _l , r _i ), the mutual information MI (B _k , r _i ), and the mutual information MI (A _l B _k , r _i ), that is, the partial combination A _l word as strong ties r _i, the stronger the ties subcombination B _k and words r _i, also, as the strong ties subcombination a _l and subcombinations B _k of co-occurrence and word r _i, related validity score score _{(r i)} is also increased.

The relation validity score calculation unit 25 includes the mutual information MI (A _l , r _i ), the mutual information MI (B _k , r _i ), and the mutual information MI (A _l B _k , r _i ). Are calculated by the equations (6) to (8), respectively.

However, p (x) is the probability that x will appear in the article data stored in the article storage device 5, and p (x, y) is that x and y are stored in the article storage device 5 at the same time. The probability of appearing in article data. That is, p (A _l ) is the probability of article data in which all the words indicated by the partial combination A _l appear, p (B _k ) is the probability of article data in which all the words indicated by the partial combination B _k appear, p (r _i ) is the probability of the article data in which the word r _i appears. In addition, p (A ₁ B _k ) is the probability of article data in which all the words indicated by the partial combination A ₁ and all the words indicated by the partial combination B _k appear, and p (A _l , r _i ) is the partial The probability of article data in which all the words indicated by the combination A _l and the word r _i appear, p (B _k , r _i ) is the article data in which all the words indicated by the partial combination B _k and the word r _i appear. , P (A ₁ B _k , r _i ) is the probability of article data in which all words indicated by the partial combination A ₁ and all words indicated by the partial combination B _k and the word r _i appear.

The relation validity score calculation unit 25 calculates each probability as follows. The relation validity score calculation unit 25 counts the total number of articles, which is the number of article data stored in the article storage device 5. Subsequently, the relationship appropriateness score calculation unit 25, among the article data stored in the article storage device 5, subcombinations A _l number of all article data word appears indicated, subcombinations B _k are all shown P (A _l ), p (B _k ), by counting the number of article data in which the word appears, and the number of article data in which the word r _i appears, and dividing these counts by the total number of articles, respectively. p (r _i ) is calculated.

Furthermore, the relationship appropriateness score calculation unit 25, the article data and all word indicating all words and subcombination B _k indicated subcombination A _l of the article data stored in the article storage unit 5 appears The number, the number of article data in which all the words indicated by the partial combination A _l and the word r _i appear, the number of article data in which all the words indicated by the partial combination B _k and the word r _i appear, the partial combination A _l by there is all words and subcombination B _k all words and word indicating the r _i that indicates a count of article data appearing divides these count the total number of articles, respectively, p (a _l B _k ), p (A _l , r _i ), p (B _k , r _i ), p (A _l B _k , r _i ) are calculated.

Note that p (A ₁ ), p (B _k ), p (r _i ), p (A ₁ B _k ), p (A ₁ , r _i ), p used in equations (6) to (8). (B _k , r _i ), p (A _l B _k , r _i ) are the probabilities calculated based on the article data stored in the article storage device 5, but appear in the article data included in the article set D It is good also as the probability to do. In this case, the relation validity score calculation unit 25 uses the article data included in the article set D instead of the article data stored in the article storage unit 5, and p (A _l ), p ( B _k ), p (r _i ), p (A _l B _k ), p (A _l , r _i ), p (B _k , r _i ), and p (A _l B _k , r _i ) are calculated.

In FIG. 5, the target situation partial combination generation unit 26 selects all the combinations when n words are selected from m words that are elements of the word vector indicated by the situation C data input in step S100 of FIG. Is generated. The number of combinations is _m C _n . The target situation partial combination generation unit 26 generates a partial combination C _f (1 ≦ f ≦ _m C _n , where f is an integer) indicating an n-dimensional word vector having the generated word combination as an element, and the processing result storage unit 14 (Step S200).

The target co-occurrence article search unit 27 generates all combinations of one of the partial combinations C _f written in the processing result storage unit 14 and one of the words r _i included in the relation set R (Step S205). . That is, the generated combinations are C ₁ -r ₁ , C ₁ -r ₂ , ..., C ₂ -r ₁ , C ₂ -r ₂ , ..., C _mCn -r ₁ , C _mCn -r ₂ , ... . Incidentally, the target co-occurrence article retrieval unit 27, the relationship relevance score score _{(r i)} is more words _{r i} only threshold or word _{r i} of a predetermined number from those related relevance score score _{(r i)} is high Only a combination may be generated. Target cooccurrence article retrieval unit 27, among all the combinations generated in step S205, selects a combination _C f -r _i not yet been processed in step S215 (step S210).

Target cooccurrence article retrieval unit 27 reads the partial combination _{C f} and word _{r i} constituting the combination _C f -r _i selected in step S210 from the processing result storage unit 14. The target co-occurrence article search unit 27 searches for article data stored in the article storage device 5, and stores article data indicating an article including all the n words indicated by the read partial combination C _f and the word r _i. Identify. The target co-occurrence article search unit 27 reads the specified article data from the article storage device 5, and writes the read article data to the article set E stored in the processing result storage unit 14 (step S215). However, the target co-occurrence article search unit 27 does not add when the extracted article data is already included in the article set E. The article data included in the relation set E is assumed to be article data e ₁ , e ₂ ,.

  The target co-occurrence article search unit 27 determines whether all the combinations generated in step S205 are the processing targets of step S215 (step S220). If it is determined that there is a combination that has not yet been processed in step S215 (step S220: NO), the target co-occurrence article search unit 27 repeats the process from step S210. When the target co-occurrence article search unit 27 determines that all combinations are to be processed (step S220: YES), the analogy result generation unit 28 executes the process of step S225.

Analogy result generation unit 28, similarly to step S205, and generates all the combinations of one of the word _{r i} included in one relationship set R subcombination _{C f} (step S225). Analogy result generation unit 28, out of the generated combination in step S225, selects a combination _C f -r _i not yet been processed in step S235 (step S230).

The analogy result generation unit 28 selects words of combinations selected from the articles indicated by the respective article data e ₁ , e ₂ ,... Included in the article set E written by the target co-occurrence article search unit 27 in the processing result storage unit 14. A word x related to n words indicated by the partial combination C _f is extracted by r _i (step S235). In this case, the structure of the _C f _-r i -x in the target selects the _A l _-r i -B _k x to be the same as the structure of when the word _{r i} is extracted with the base. The analogy result generation unit 28 adds the extracted word x to the analogy result set X stored in the processing result storage unit 14. However, the analogy result generation unit 28 does not add when the extracted word x is already included in the analogy result set X. Words x included in the analogy result set X are respectively x ₁ , x ₂ ,.

  The analogy result generation unit 28 determines whether all combinations generated in step S225 have been processed in step S235 (step S240). If the analogy result generation unit 28 determines that there is a combination that has not yet been processed (step S240: NO), the process from step S230 is repeated. When the analogy result generation unit 28 determines that all combinations have been processed in step S235 (step S240: YES), the analogy result validity score calculation unit 29 executes the process of step S245.

The analogy result validity score calculation unit 29 uses the following equation (9) to calculate an analogy result for the word x _j (j = 1, 2,...) Indicated by the analogy result set X stored in the processing result storage unit 14. The validity score score (x _j ) is calculated and written in the processing result storage unit 14 (step S245). Where MI (C _f , x _j ) is the mutual information amount between the partial combination C _f and the word x _j , MI (r _i , x _j ) is the mutual information amount between the word r _i and the word x _j , MI (C _f r _i , x _j ) is a mutual information amount between the word x _j and the co-occurrence of the partial combination C _f and the word r _i . The analogy result validity score calculation unit 29 reads the relationship validity score score _{(r i)} from the processing result storage unit 14.

score (x _j ) = Σ _f Σ _i MI (C _f , x _j ) * MI (r _i , x _j ) * MI (C _f r _i , x _j ) (9)

As described above, the equation (9) is obtained by calculating the mutual information MI (C _f , x _j ), the mutual information MI (r _i , x _j ), the mutual information MI (C _f r _i , x _j ) and the relationship. The value obtained by multiplying the validity score score (r _i ) is a value obtained by accumulating all the partial combinations C _f and all the words r _i . Therefore, the larger the mutual information MI (C _f , x _j ), the mutual information MI (r _i , x _j ), and the mutual information MI (C _f r _i , x _j ), that is, the partial combination C _f and word as strong ties _{x j,} the more strong connection word _{r i} and the word _{x j,} the more strong connection subcombination _{C f} and word _{r i} cooccurrence and words _{x j,} also related relevance scores score As (r _i ) increases, the analogy result validity score score (x _j ) also increases.

The analogy result validity score calculation unit 29 includes the mutual information MI (C _f , x _j ), the mutual information MI (r _i , x _j ), and the mutual information MI (C _f r _i , x _j). ) Are calculated by equations (10) to (12), respectively.

However, p (C _f ) is the probability of article data in which all the words indicated by the partial combination C _f appear, and p (x _j ) is the probability of article data in which the word x _j appears. In addition, p (C _f , x _j ) is the probability of article data in which all the words indicated by the partial combination C _f and the word x _j appear, and p (r _i , x _j ) is the word r _i and the word x The probability of article data in which _j appears, p (C _f , r _i ) is the probability of article data in which all the words indicated by the partial combination C _f and the word r _i appear, p (C _f r _i , x _j ) Is the probability of article data in which all the words indicated by the partial combination C _f and the word r _i and the word x _j appear.

The analogy result validity score calculation unit 29 calculates each probability as follows. The analogy result validity score calculation unit 29 counts the total number of articles, which is the number of article data stored in the article storage device 5. Subsequently, analogy result validity score calculation unit 29, among the article data stored in the article storage device 5, the number of article data of all the words indicated by the subcombinations C _f appears, the word x _j appears the number of article data, and counts the number of article data word _{r i} appears, _p (C _f) by dividing the number of these counts by the total number of articles, respectively, p (x j), p (r i) Is calculated.

Further, the analogy result validity score calculation unit 29 calculates the number of article data in which all the words indicated by the partial combination C _f and the word x _j appear in the article data stored in the article storage unit 5 and the word r _i. the number of article data and words x _j appears as, subcombinations C _f number of article data is all words and word r _i appear indicated, all words and word r _i and word indicating the subcombination C _f By counting the number of article data in which x _j appears, and dividing these counts by the total number of articles, p (C _f , x _j ), p (r _i , x _j ), p (C _f , r _i ), p (C _f r _i , x _j ) are calculated.

Note that p (C _f ), p (x _j ), p (r _i ), p (C _f , x _j ), p (r _i , x _j ), and p (r _i , x _j ), which are used in the equations (10) to (12), p (C _f , r _i ) and p (C _f r _i , x _j ) are the probabilities calculated based on the article data stored in the article storage device 5, but the article data included in the article set E It is good also as the probability of appearing. In this case, the analogy result validity score calculation unit 29 uses the article data included in the article set E instead of the article data stored in the article storage unit 5, and p (C _f ), p Calculate (x _j ), p (r _i ), p (C _f , x _j ), p (r _i , x _j ), p (C _f , r _i ), p (C _f r _i , x _j ) To do.

Output unit 16, analogy results each word x _j which generator 28 is indicated by analogy result set X written in the processing result storage unit 14, analogy result validity of written by analogy result validity score calculation unit 29 the word x _j The analogy result data including the sex score score (x _j ) is output on the display (step S245). At this time, the output unit 16 may output only the word x _j whose analogy result validity score is greater than or equal to the threshold and its analogy result validity score score (x _j ), and in descending order of the analogy result validity score. Only a predetermined number of words x _j and their analogy result validity scores score (x _j ) may be output.

  Next, a detailed processing procedure of the word r extraction process in step S145 shown in FIG. 4 and the word x extraction process in step S235 shown in FIG. 5 will be described. In the following, two extraction processes are shown, but any of them may be used according to the case or the target of analogy. Here, for the sake of simplicity, a case will be described in which the situation A, the result B, and the situation C are three-dimensional word vectors, and the number of extracted words n and h is 2.

In this case, in step S105 of FIG. 4, partial combinations A ₁ (t _a1 , t _a2 ), A ₂ (t _a1 , t _a3 ), A ₃ from the word vector (t _a1 , t _a2 , t _a3 ) of the situation A (T _a2 , t _a3 ) is generated, and in step S110, partial combinations B ₁ (w _b1 , w _b2 ), B ₂ (w _b1 , w _b3 ), B ₃ (w _b2 , w) are generated from the word vector of the result B. _b3 ) is generated. Therefore, in step S135, the combinations A ₁ -B ₁ , A ₁ -B ₂ , A ₁ -B ₃ , A ₂ -B ₁ , A ₂ -B ₂ , A ₂ -B ₃ , A ₃ -B ₁ , A ₃ -B ₂ and A ₃ -B ₃ are generated.

(Extraction process 1): The article data is text data of a document described in a natural language. In step S145, the relation set generation unit 24 performs a morphological analysis of the document indicated by the article data d ₁ , d ₂ ,... Included in the article set D. Relationship set generation unit 24 refers to the result of the morphological analysis, n-number of words indicated by the subcombinations A _l is included in the subject, and, as a noun h number of words indicated by the partial combination B _k is a predicate from contains statements, it extracts the verb predicates relating the h pieces of words indicated by the subcombinations a _l is n words and subcombination B _k indicating a word r.
In step S235, the analogy result generation unit 28 performs morphological analysis of the document indicated by the article data e ₁ , e ₂ ,... Included in the article set E. The analogy result generation unit 28 refers to the result of the morphological analysis, and from the sentence in which the n words indicated by the subcombination C _f are included in the subject and the word r _i is included as a verb of the predicate, The noun in the part is extracted as the word x.

FIG. 6 is a diagram for explaining the extraction process of word r in step S140. In the figure, an example is shown in which a combination of partial combination A ₁ (shark, tuna) and partial combination B ₁ (fin, tail) is selected in step S135. The relation set generation unit 24 uses the word “sharks” indicated by the partial combination A _{1 as} the subject of the sentence “fish such as sharks and tuna can swim at high speed using fins and tails” indicated by the article data d ₁ . ”And“ tuna ”, and the nouns in the predicate are the words“ fin ”and“ tail ”indicated by the subcombination B ₁ , so the verb“ swim ”in the predicate of the sentence is extracted as the word r ₁ . Further, the relationship set generation unit 24, sentence indicated by the article data d ₂ "fishes such as sharks and tuna. Capable of moving a wide range with the Hireya tail" subject of indicated subcombination A ₁ Since the words “shark” and “tuna” and the nouns in the predicate are the words “fin” and “tail” indicated by the partial combination B ₁ , the verb “move” in the predicate of the sentence is referred to as the word r _2. Extract as

  In this way, by using a partial combination composed of two or more words, it is possible to extract a target relationship with higher accuracy than in the case of using one word. As the number of words included in the partial combination increases, it becomes possible to extract words indicating a more accurate relationship, but the number of sentences from which relationships are extracted decreases.

FIG. 7 is a diagram for explaining the word x extraction processing in step S235. This figure shows a case where the partial combination C ₁ (lion, wolf) is selected as _one of the combinations in step S230. The analogy result generation unit 28 includes, from any article data included in the article set E, the words “lion” and “wolf” indicated by the partial combination C ₁ are included in the subject, and the word r ₁ “swim” is described. It is assumed that a sentence including a sentence that is a part verb is not detected. On the other hand, the analogy result generation unit 28 indicates the subject of the sentence “the lion, the wolf, etc. need to move over a wide range using their feet ...” indicated by the article data e ₁ included in the article set E, as indicated by the partial combination C _1. Since the words “lion” and “wolf” and the verb of the predicate is the word r ₂ “move”, the noun “foot” in the predicate of the sentence is extracted as the word x ₁ .

  (Extraction process 2): The article data is text data of a document described in a natural language, and the article storage device 5 further stores field-specific dictionary data (corpus). The situation A, the result B, and the relation set R are terms in a specific field. Furthermore, in step S100 of FIG. 4, the input unit 12 of the analogy device 1 further receives input of information indicating a field.

In step S145 of FIG. 4, the relation set generation unit 24 identifies dictionary data corresponding to the input field information, and includes n words indicated by the partial combination A _l and h words indicated by the partial combination B _k. A word registered in the specified dictionary data is extracted as a candidate r ′ of the word r from the sentence including the word. Let the extracted candidates r ′ be candidates r ₁ ′, r ₂ ′,. The relation set generation unit 24 uses the extracted candidate r _y ′ (y = 1, 2,...) As appropriate as a word representing the relation based on the strength of association between the partial combination A ₁ and the partial combination B _k. A value indicating the property is calculated by the following equation (13). However, MI (A _l , r _y ′) is the mutual information amount between the partial combination A _l and the candidate r _y ′, and MI (B _k , r _y ′) is the difference between the partial combination B _k and the candidate r _y ′. Mutual information. Incidentally, the amount of mutual information _{MI (A l, r y '} ), the mutual information _{MI (B k, r y'} ) , by using the candidate _{r y} 'in place of the word _{r i,} equation (6), wherein Calculated in the same manner as (7).

Score (r _y ′) = MI (A ₁ , r _y ′) + MI (B _k , r _y ′) (13)

The relation set generation unit 24 selects a candidate r ′ whose calculated value is equal to or greater than a predetermined condition for determining that the relation is strong as a word r.
In step S230, the analogy result generation unit 28 is registered in the dictionary data corresponding to the input field information from the sentence including the n words indicated by the partial combination C _f and the word r _i . Extract the word as word x.

FIG. 8 is a diagram for explaining extraction processing of the word r and the word x. This figure shows an example in which the field information is a computer and the combination of the partial combination A ₁ (image, photograph) and the partial combination B ₁ (GIF, JPEG) is selected in step S140. The sentence “article or file compression such as GIF or JPEG can be used for compression of images and photos” shown in the article data d ₁ includes the words “image” and “photo” indicated by the partial combination A ₁ and the partial combination B. ₁ includes the word “GIF” and the word “JPEG”. The relation set generation unit 24 identifies the computer dictionary data from the field information, and among the words included in the sentence, “compression” and “file format” registered in the computer dictionary data are candidates r ₁ ′, r Extract as ₂ '. The relation set generation unit 24 calculates a score for the extracted candidates r ₁ ′ and r ₂ ′ using the above equation (13). As a result, the candidate r ₁ ′ “compressed” is not selected as the word r, and the candidate r ₂ 'Select “file format” as word r ₁ .

When the combination of the partial combination C ₁ (music, conversation) and the word r ₁ “file format” is selected in step S230, the analogy result generation unit 28 selects the words “music” and “conversation” indicated by the partial combination C _1. In addition, “MP3” registered in the computer dictionary data stored in the article storage device 5 is extracted from the sentence including the word r ₁ “file format” as the word x ₁ .

[Second Embodiment]
Subsequently, another embodiment of the present invention will be described.
In the first embodiment, as shown in Expression (4), there is only one pair of “A: B” as a base example, and the result X is obtained from the situation C of the target by this example. That is, in the first embodiment, the analogy device 1 performs a four-term analogy process in a multidimensional case. On the other hand, in the present embodiment, a four-dimensional analogy process of multi-dimensional multi-cases that obtains the result X from the target situation C when there are a plurality of base cases indicated by “A: B” will be considered. When this is described by the same expression method as modus ponens shown in equation (4), the following equation (14) is obtained.

  As described above, in this embodiment, there are N cases (N is an integer of 2 or more), and each case is referred to as a case (I) (I is an integer of 2 to N). Case (I) includes a base situation A (I) and a result B (I) in the base situation A (I). The analogy device of this embodiment calculates | requires the result X corresponding to the situation C using A (I): B (I).

The situation A (I) is an m-dimensional word vector having m words (m is an integer of 2 or more) words t _aI1 , t _aI2 ,..., T _aIm , and the target situation C is also the first This is the same m-dimensional word vector as in the embodiment. Further, the result B (I) is a g-dimensional word vector having g (where g is an integer of 2 or more) words w _bI1 , w _bI2 ,..., W _bIg as elements, and the target result X is also the first This is the same g-dimensional word vector as in the embodiment.

FIG. 9 is a diagram showing an outline of the operation of the analogy device according to the present embodiment.
In this embodiment, since there are many cases, it can be expressed as the following equation (15).

A (I): B (I) = C: X (I)? (I is an integer not less than 2 and not more than N) (15)

  As shown in the figure, the analogy device of the present embodiment performs a relationship extraction process similar to that of the analogy device 1 of the first embodiment for each case (I), so that the partial combination of the situation A (I) A relation set R (I) that is a set of words indicating a relation with the partial combination of the result B (I) is generated. The analogy device of this embodiment generates an analogy result set X (I) by performing the relationship mapping process for each relation set R (I) in the same manner as the analogy device 1 of the first embodiment, and the analogy result is valid. Calculate the sex score. The analogy device of this embodiment integrates analogy result sets X (1) to X (N), and deletes duplicates when there are duplicate words x. The analogy result validity score after duplicate deletion is a value obtained by summing up the analogy result validity scores for the overlapping word x.

  FIG. 10 is a block diagram showing the configuration of the analogy device 1a according to the second embodiment of the present invention. In this figure, the same parts as those in the analogy device 1 according to the first embodiment shown in FIG. The analogy device 1a shown in FIG. 10 differs from the analogy device 1 of the first embodiment shown in FIG. 3 in that it includes a case storage unit 11a instead of the case storage unit 11, an analogy processing control unit 17 and an analogy result. It is a point provided with the integrating | accumulating part 18. FIG.

  The case storage unit 11a includes situation A (I) data (base situation data) indicating a word vector of situation A (I) and result B (I) data (base result data) indicating a word vector of result B (I). Case (I) consisting of The analogy processing control unit 17 causes the relationship extraction unit 13 and the relationship mapping unit 15 to perform the same processing as in the first embodiment on the cases (1) to (N) stored in the case storage unit 11a. Instruct. The analogy result accumulation unit 18 integrates analogy results X (I), which is data indicating a set of analogy result words x obtained for each case (I).

11 and 12 are diagrams showing a processing flow of the analogy device 1a shown in FIG.
The case storage unit 11a of the analogy device 1a stores a case (I) composed of situation A (I) data and result B (I) data (I is an integer of 2 or more and N or less). The situation A (I) data represents an m-dimensional word vector having m (m is an integer of 2 or more) words t _aI1 , t _aI2 ,..., T _aIm (m is an integer of 2 or more). The result B (I) data indicates a g-dimensional word vector having g (where g is an integer of 2 or more) words w _bI1 , w _bI2 ,..., W _bIg as elements. Further, the processing result storage unit 14 stores an article set D, an article set E, a relation set R (I), and an analogy result set X (I) with an initial value NULL.

In FIG. 11, the input unit 12 of the analogy device 1a receives input of status C data (step S300), similarly to step S100 of the first embodiment shown in FIG. The analogy process control unit 17 selects a case (I) that has not yet been processed from cases (I) stored in the case storage unit 11a (step S305). The analogy process control unit 17 instructs the relationship extraction unit 13 to execute the process for the selected case (I). As a result, the relationship extraction unit 13 reads the situation A (I) data and the result B (I) data of the case (I) selected by the analogy processing control unit 17 from the case storage unit 11, and in the first embodiment By using the situation A data and the result B data in place of the situation A data, the same processing as steps S105 to S155 in the first embodiment shown in FIG. 4 is executed (steps S310 to S360). Thus, the relationship between the set R for case (I), Relationship relevance scores score of each word _{r i} in the relation set R _{(r i)} is obtained. The relation set R obtained by using the case (I) is set as a relation set R (I), the words r ₁ , r ₂ ,... Included in the relation set R are set as words r _I1 , r _I2,. Let the relation validity score score (r _i ) calculated for _i be the relation validity score score (r _Ii ). These data are written in the processing result storage unit 14 in association with the identification information of the case (I).

  The analogy processing control unit 17 determines whether there is a case (I) that is not yet processed in the case (I) stored in the case storage unit 11a. When it is determined that there is a case (I) that has not yet been processed (step S365: NO), the analogy process control unit 17 repeats the process from step S305. If the analogy processing control unit 17 determines that all cases (I) are to be processed (step S365: YES), the relationship mapping unit 15 is activated.

In FIG. 12, the target situation partial combination generation unit 26, as in step S200 of the first embodiment shown in FIG. 5, uses a partial combination C _f (1 ≦ f ≦ _m C _n , f is an integer) and is written in the processing result storage unit 14 (step S300).

The analogy processing control unit 17 identifies a case (I) that has not yet been processed in the relationship mapping unit 15 (step S305). The analogy process control unit 17 instructs the relationship mapping unit 15 to execute the process for the identified case (I). Accordingly, the relationship mapping unit 15 associates the identification information of the case (I) selected by the analogy processing control unit 17 with the relationship set R (I) and the relationship validity score stored in the processing result storage unit 14. The score (r _Ii ) is read out and used in place of the relation set R and the relation validity score score (r _i ) in the first embodiment, and the processes of steps S210 to S245 in the first embodiment shown in FIG. 5 are performed. Execute (Steps S415 to S440). Thereby, the analogy result set X and the analogy result validity score score (x _j ) of each word x _j included in the analogy result set X are obtained. This analogy result set X obtained by using the relation set R (I) corresponding to the case (I) and the relation validity score score (r _Ij ) is set as an analogy result set X (I) and included in the analogy result set X The words x ₁ , x ₂ ,... _Are the words x _I1 , x _I2 ,..., Respectively, and the analogy result validity score score (x _j ) calculated for the word x _j is the relation validity score score _I (x _Ij ). And These data are written in the processing result storage unit 14 in association with the identification information of the case (I).

  The analogy processing control unit 17 determines whether there is a case (I) that is not yet processed in the relationship mapping unit 15. If it is determined that there is a case (I) that has not yet been processed (step S455: NO), the analogy process control unit 17 repeats the process from step S405. When the analogy processing control unit 17 determines that all cases (I) are to be processed (step S455: YES), the analogy result accumulation unit 18 is activated.

  The analogy result accumulation unit 18 generates an integrated analogy result set X that is data obtained by integrating the analogy result sets X (1) to (N) read from the processing result storage unit 14 as shown in the following equation (16). To do.

_{X = Σ I X (I)} (I = 1~N) ... (16)

The analogy result accumulating unit 18 eliminates duplication by leaving only one word if there is duplication in the words _xIj included in the analogy result sets X (1) to (N), and sets a set of words _{xIj from} which duplication has been eliminated. An integrated analogy result set X, which is data to be shown, is generated. It is assumed that words included in the integrated analogy result set X are x ₁ , x ₂ ,.

Subsequently, analogy result integration unit 18, by the following equation (17), and calculates the inferred result relevance score of each word x _j included in the integrated analogy result set X. However, score _I (x _j ) is the analogy result validity score of the word x _j obtained for the case (I).

score (x _j ) = Σ _I score _I (x _j ) (17)

That analogy result relevance score of a word _{x j} included in the integrated analogy result set X is the analogy results Relevance Score score _I value obtained by integrating _{(x Ij)} word _{x Ij} integrated in the word _{x j} . The analogy result accumulating unit 18 writes the word x _j included in the integrated analogy result set X and the analogy result validity score score (x _j ) of the word x _{j in} the processing result storage unit 14 (step S460).

The output unit 16 includes a analogy result accumulation unit 18 is a processing result each word _{x j} storage unit 14 in the integrated analogy result set X written by the show and the word _{x j} analogical results relevance score score _{(x j)} in step S445 The analogy result data consisting of is output on the display (step S465). At this time, the output unit 16 may output only the word x _j having an analogy result validity score equal to or higher than the threshold, or may output only a predetermined number of words x _{j in} descending order of the analogy result validity score. Good.

In the embodiment described above, the article storage device 5 is an external device connected to the analogy device 1 of the first embodiment and the analogy device 1a of the second embodiment via a network. The apparatus 1a may be configured to include the article storage device 5 therein.
In the above-described embodiment, the article sets D and E are set of article data. However, the article data identification information such as a URL (Universal Resource Locator) where the article data is stored and the data name of the article data is used. It may be the data shown. In this case, the analogy device 1 and the analogy device 1a refer to the article data in the article storage device 5 specified by the identification information of the article data included in the article sets D and E, and execute the above-described processing.
In the above-described embodiment, the situation A, situation A (1) to situation A (N) and situation C have the same number of word vectors. However, situation A, situation A (1) to situation A The number of dimensions of (N) and the number of dimensions of the word vector of situation C may be different. Further, the number of dimensions of the word vectors in the situations A (1) to A (N) may not be the same. Further, the number of dimensions of the word vectors of the results B (1) to B (N) may not be the same.

  The analogy device 1 and analogy device 1a described above have a computer system inside. Then, the relationship extraction unit 13, the relationship mapping unit 15 and the output unit 16 of the analogy device 1, and the relationship extraction unit 13, the relationship mapping unit 15, the output unit 16, the analogy process control unit 17 and the analogy result integration unit of the analogy device 1a The operation process 18 is stored in a computer-readable recording medium in the form of a program, and the above-described processing is performed by the computer system reading and executing this program. The computer system here includes a CPU, various memories, an OS, and hardware such as peripheral devices.

Further, the “computer system” includes a homepage providing environment (or display environment) if a WWW system is used.
The “computer-readable recording medium” refers to a storage device such as a flexible medium, a magneto-optical disk, a portable medium such as a ROM and a CD-ROM, and a hard disk incorporated in a computer system. Furthermore, the “computer-readable recording medium” dynamically holds a program for a short time like a communication line when transmitting a program via a network such as the Internet or a communication line such as a telephone line. In this case, a volatile memory in a computer system serving as a server or a client in that case, and a program that holds a program for a certain period of time are also included. The program may be a program for realizing a part of the above-described functions, or may be a program that can realize the above-described functions in combination with a program already recorded in a computer system.

DESCRIPTION OF SYMBOLS 1, 1a ... Analogical inference apparatus 11, 11a ... Case memory | storage part 12 ... Input part 13 ... Relation extraction part 14 ... Processing result memory | storage part 15 ... Relation mapping part 16 ... Output part 17 ... Analogical-inference process control part 18 ... Analogical-inference-results integration part 21 ... base situation partial combination generation part 22 ... base result partial combination generation part 23 ... base co-occurrence article search part 24 ... relation set generation part 25 ... validity score calculation part 26 ... target situation partial combination generation part 27 ... target co-occurrence article Search unit 28 ... analogy result generation unit 29 ... analogy result validity score calculation unit 5 ... article storage device

Claims (7)

A base situation partial combination generation unit that generates base situation partial combination data composed of a predetermined number of words selected from a plurality of words indicated by the base situation data, and
A base result partial combination generation unit that generates base result partial combination data composed of a predetermined number of words selected from a plurality of words indicated by the base result data;
The word indicated by the base situation partial combination data and the word indicated by the base result partial combination data are associated with each different combination of one of the base situation partial combination data and one of the base result partial combination data. A relation set generation unit that extracts association words that are words from article data stored in the article storage device;
A target situation partial combination generation unit that generates target situation partial combination data composed of a predetermined number of words selected from a plurality of words indicated by the target situation data;
For each different combination of one of the target situation partial combination data and one of the association words, a word related to the word indicated by the target situation partial combination data by the association word is stored in the article storage device. An analogy result generation unit that extracts as analogy results from the posted article data,
An analogy device comprising: For each of the association words extracted by the relation set generation unit, co-occurrence of the association word obtained from the article data stored in the article storage device, the base situation partial combination data, and the base result partial combination data The relation between the situation and the result based on the strength of the relation, the strength of the relation between the related word and the base situation partial combination data, and the strength of the relation between the related word and the base result partial combination data A relation validity score calculation unit for calculating a relation validity score that quantitatively represents validity as a word to perform,
For each of the words extracted as an analogy result by the analogy result generation unit, the word obtained from the article data stored in the article storage device, the co-occurrence of the target situation partial combination data, and the association word Strength of association, strength of association between the word and the target situation partial combination data, strength of association between the word and the association word, and the relation validity score calculated for the association word An analogy result validity score calculation unit for calculating an analogy result validity score that quantitatively represents the validity as an analogy result,
The analogy device according to claim 1, further comprising: For each of a plurality of case data composed of base situation data and base result data, the base situation partial combination generation unit generates base situation partial combination data from the base situation data constituting the case data, and the base result partial combination The generation unit generates the base result partial combination data from the base result data constituting the case data, and the relation set generation unit has one of the base situation partial combination data generated from the base situation data and , For each different combination consisting of one of the base result partial combination data generated from the base result data, the association word is extracted from the article data, and the relation validity score calculation unit makes the relation validity for each of the association words Generate sex score and generate analogy result In addition, the analogy result word is extracted from the article data for each different combination of one of the target situation partial combination data and one of the association words, and the analogy result validity score calculation unit extracts the analogy result as an analogy result An analogy processing control unit for calculating an analogy result validity score for each of the words
An analogy result accumulating unit that integrates the same words included in the analogy result words obtained for each case data and accumulates the analogy result validity scores calculated for the integrated same words Prepare
The analogy device according to claim 2, wherein: The relation set generation unit obtains a predicate verb from the sentence of the article data in which the word indicated by the base situation partial combination data is a subject noun and the word indicated by the base result partial combination data is a predicate noun. Extract as related words,
The analogy result generation unit extracts a predicate noun as an analogy result from a sentence of the article data in which a plurality of words indicated by the target situation partial combination data are subject nouns and the association word is a predicate verb ,
The analogy device according to any one of claims 1 to 3, characterized in that: The article storage device further stores dictionary data including words related to a predetermined field,
The relation set generation unit converts a word included in the dictionary data from the sentence of the article data in which the word indicated by the base situation partial combination data and the word indicated by the base result partial combination data co-occurs. Extract as
The analogy result generation unit extracts a word included in the dictionary data as an analogy result from a sentence of the article data in which the word indicated by the target situation partial combination data and the association word co-occur.
The analogy device according to any one of claims 1 to 3, characterized in that: An analogy method performed by an analogy device,
A base situation partial combination generation process in which a base situation partial combination generation unit generates base situation partial combination data including a predetermined number of words selected from a plurality of words indicated by the base situation data by different combinations;
A base result partial combination generation process in which a base result partial combination generation unit generates base result partial combination data composed of a predetermined number of words selected from a plurality of words indicated by the base result data;
The relation set generation unit includes a word indicated by the base situation partial combination data and the base result partial combination data for each different combination of one of the base situation partial combination data and one of the base result partial combination data. A relation set generation process for extracting an association word, which is a word relating to a word to be shown, from article data stored in the article storage device;
A target situation partial combination generation unit that generates target situation partial combination data including a predetermined number of words selected by different combinations from a plurality of words indicated by the target situation data; and
The analogy result generation unit, for each different combination consisting of one of the target situation partial combination data and one of the association words, a word related to the word indicated by the target situation partial combination data by the association word, An analogy result generation process for extracting as analogy results from the article data stored in the article storage device;
An analogy method characterized by comprising: A computer used as an analogy device
A base situation partial combination generation unit for generating base situation partial combination data composed of a predetermined number of words selected from different words indicated by the base situation data;
A base result partial combination generation unit that generates base result partial combination data composed of a predetermined number of words selected from a plurality of words indicated by the base result data.
The word indicated by the base situation partial combination data and the word indicated by the base result partial combination data are associated with each different combination of one of the base situation partial combination data and one of the base result partial combination data. A relation set generation unit that extracts association words, which are words, from article data stored in an article storage device;
A target situation partial combination generation unit that generates target situation partial combination data composed of a predetermined number of words selected from a plurality of words indicated by the target situation data;
For each different combination of one of the target situation partial combination data and one of the association words, a word related to the word indicated by the target situation partial combination data by the association word is stored in the article storage device. An analogy result generator that extracts as analogy results from the article data
A program characterized by functioning as

Images