JP5441173B2 - Related information extraction apparatus, method and program thereof - Google Patents

Description

  The present invention relates to a technique for extracting information related to a plurality of specific expressions, which plays an important role in a summarization system or the like that summarizes input text.

  Conventionally, as a technique for extracting information related to a plurality of specific expressions, a dependency structure of input text, and information indicating whether or not words included in the input text are easily used as a relationship between specific expressions, There is known one that extracts relational information by machine learning using a feature as a feature (see, for example, Non-Patent Document 1).

  For example, if the text “Japan's Prime Minister Hatoyama will visit Korea next month” and a set of unique expressions “Hatoyama” and “Japan” are entered, the phrase “Japan” The information “Prime” is based on the dependency structure that is related to the phrase “Prime” and the information that the word “Prime” is easy to use as an indication of the relationship between the person and the organization. It is extracted as information indicating the relationship of the pair.

  In addition, when the relation information extraction technique described above is used, when a set of specific expressions “Hatoyama” and “Korea” is input, the information “visit” is extracted as the relation information.

Toru Hirano and two others, “Identification of relations between proper expressions using relative nounness”, Proc. Of the 15th Annual Conference of the Association for Natural Language Processing, March 2009, D5-5

  However, the technology described in Non-Patent Document 1 uses only grammatical information such as dependency structure between clauses and lexical information on whether or not the information is easy to be used as an indication of the relationship between specific expressions. For example, if machine learning is performed using large-scale text data consisting of incorrect grammar and vocabulary, the optimal relationship information may not be extracted. It was difficult to improve.

  The present invention has been made in view of the above problems, and an object of the present invention is to provide an apparatus, a method, and a program that can extract relation information between a plurality of specific expressions with high accuracy.

In order to achieve the above object, the related information extracting apparatus of the present invention is an apparatus for extracting information related to a plurality of input specific expressions, and includes a text including each of the specific expressions, and each of the specific expressions. When a specific expression set is input, the input text is morphologically analyzed , and the analysis processing unit that analyzes the specific expression set together with the dependency of the clauses constituting the input text, and obtains the analysis result by the analysis processing unit Then, at least one independent word included in the input text is extracted as a related information candidate, and it is estimated that past related information used in the past as related information of each unique expression has changed to the related information candidate over time. the change information indicating the degree to get every extracted related information candidates, select the model corresponding to the set type of the named entities based on the analysis result by the analysis unit, Based on the analysis result and the change information by the analysis processing unit each related information candidate whether the set of relationship information named entities determined by using the model, a relationship information extraction processing section for extracting relationship information ing.

Further, the related information extraction method of the present invention is a method for extracting information related to a plurality of input specific expressions using a computer in order to achieve the object, wherein the computer And a set of specific expressions composed of the specific expressions , the input text is subjected to morphological analysis , the set of specific expressions is analyzed together with the dependency of clauses constituting the input text, and the computer When the analysis result is acquired, at least one independent word included in the input text is extracted as a related information candidate, and past related information used in the past as related information of each unique expression is extracted as the related information candidate. the change information indicating the degree that is estimated to have changed with time, and obtained for each extracted related information candidate pairs to a set of types of the named entities based on the analysis result A model that was selected, the analysis results and the related information candidates based on the change information whether a set of related information of the named entities determined by using the model, and extracting the relationship information.

  Furthermore, the program of the present invention is for causing a computer to function as each means of the related information extracting apparatus.

  Furthermore, the program of the present invention is for causing a computer to execute each process of the related information extraction method.

  As a result, based on the change information indicating the degree to which the past relationship information used in the past as the relationship information of each unique expression is estimated to have changed over time to the relationship information candidate, and the analysis result, the relationship information from the relationship information candidate Therefore, it is possible to extract the relationship information in consideration of the change in the relationship between the specific expressions over time.

  According to the related information extracting apparatus, method and program of the present invention, it is possible to extract related information in consideration of changes in the relationship over time between each unique expression, so for example, “Mr. A” and “Mr. B” If the past related information of the specific expression set is “companion” and the set of related information candidates extracted from the input text is {marriage, meet, eat, drink, rival} Based on the transition of typical relationships such as “dating”, “engagement”, and “marriage” between people and people, “marriage” can be extracted as relationship information. Therefore, it is possible to extract relation information between a plurality of unique expressions with high accuracy.

1 is a configuration diagram of a related information extraction device according to an embodiment of the present invention. Relationship information extraction process flow chart Diagram showing the summary of analysis results by the dependency analysis unit The figure which shows the outline | summary of the processing result by a specific expression matching part The figure which shows the outline of the processing result by the dependency structure information acquisition section The figure which shows the outline of the processing result by the dependency structure information acquisition section The figure which shows the outline of the processing result by the dependency structure information acquisition section The figure which shows the outline of the processing result by the relation estimation information acquisition part The figure which shows an example of the data structure of the information memorize | stored in a change information storage part The figure which shows an example of the data structure of the information memorize | stored in a change information storage part The figure which shows an example of the data structure of the information memorize | stored in a change information storage part The figure which shows an example of the data structure of the information memorize | stored in a change information storage part The figure which shows an example of the data structure of the information memorize | stored in a change information storage part The figure which shows an example of the data structure of the information memorize | stored in a change information storage part The figure which shows the outline of the processing result by change information acquisition section

  FIG. 1 to FIG. 15 show an embodiment of the present invention. FIG. 1 is a configuration diagram of a related information extracting apparatus according to an embodiment of the present invention, FIG. 2 is a flowchart of related information extraction processing, and FIG. The figure which shows the outline of the analysis result by the receiving analysis part, FIG. 4 is the figure which shows the outline of the processing result by the specific expression matching part, FIG. 5 thru | or FIG. 7 is the figure which shows the outline of the processing result by the dependency structure information acquisition part, 8 is a diagram showing an overview of the processing results obtained by the relationship estimation information acquisition unit, FIGS. 9 to 14 are diagrams showing examples of the data structure of information stored in the change information storage unit, and FIG. 15 is a process performed by the change information acquisition unit. It is a figure which shows the outline | summary of a result.

  The outline of the related information extracting apparatus and method according to the present invention will be described below with reference to the drawings.

  The related information extracting apparatus of the present invention comprises a computer device mainly composed of a well-known CPU, and can be connected to display means such as a monitor, input means such as a keyboard, storage means such as a hard disk and memory, and an external network. A device (not shown) is provided. Further, the relationship information extraction apparatus of the present invention includes an analysis processing unit 10, a relationship information extraction processing unit 20, a relationship estimation information storage unit 30, a change information storage unit 40, a past relationship information storage unit 50, a model. A storage unit 60 is provided.

  As shown in FIG. 1, the analysis processing unit 10 includes a morphological analysis unit 11, a dependency analysis unit 12, and a specific expression association unit 13, and a set of specific expressions composed of two specific expressions, text, and creation of text When a date is input using an input means, the input text is analyzed by morpheme and the dependency relation of clauses constituting the input text is analyzed.

  When the morpheme analysis unit 11 acquires the input text (step S1 in FIG. 2), the input text is divided into words by performing a well-known morpheme analysis process, and parts of speech are assigned to the divided words. Output (step S2 in FIG. 2). For example, if a sentence “Junichiro Koizumi gave a speech at Tokyo Station and Ichiro Ozawa at Osaka Station” is input, the processing result by the morphological analysis unit 11 is “Junichiro Koizumi (noun) / ha (case particle). ) / Tokyo Station (noun) / de (particle) /, (reading) / Ichiro Ozawa (noun) / ha (particle) / Osaka station (noun) / de (particle) / speech (verb) / suffix (suffix) /. (Punctuation) ".

  The dependency analysis unit 12 divides the text into phrases by performing a well-known dependency analysis process on the input text that has been obtained from the morpheme analysis unit 11 and has been subjected to morpheme analysis. The dependency relationship is analyzed and output (step S3 in FIG. 2). In this case, when the illustrated input text is analyzed by the dependency analysis unit 12, information representing a dependency structure (dependency tree) as shown in FIG. 3 is output as an analysis result. Here, the phrase “Junichiro Koizumi / Ha”, the phrase “Tokyo Station / De /,”, the phrase “Ichiro Ozawa / Ha”, and the phrase “Osaka Station / De” are respectively “Speech / In the case of implementing these dependency relationships as data, for example, “(Speech. (Junichiro Koizumi) (At Tokyo Station)” (Ichiro Ozawa) (Osaka) In addition, these dependency relationships include dependency type information defined in a well-known dependency analysis technique (“D” representing a normal dependency relationship, parallel dependency). "P" representing the receiving relationship, "A" representing the receiving relationship of the same rank, or "Q" indicating the relationship of the self-related).

  The specific expression association unit 13 is for performing a specific expression association process using the input set of specific expressions and the analysis result of the dependency analysis unit 12. More specifically, when the specific expression association unit 13 acquires a set of specific expressions input using the input unit and the input text analyzed by the dependency analysis unit 12 (step S4 in FIG. 2), the input is performed. A unique expression corresponding to each unique expression is extracted from the input text, and a unique expression identifier representing the specific expression is assigned to the extracted specific expression (step S5 in FIG. 2). For example, when a specific expression “Junichiro Koizumi” is input, “Junichiro Koizumi” in the input text is expressed as “<PSN> Junichiro Koizumi </ PSN>”. Here, “PSN” is a unique expression identifier representing a person name. In this embodiment, eight types of unique expression identifiers (“PSN” representing a person name and “PSN” representing an organization name defined in a well-known unique expression extraction technique). ORG ”,“ LOC ”representing place name,“ ART ”representing artifact name,“ MNY ”representing amount,“ PNT ”representing percentage,“ TIM ”representing time, and“ DAT ”representing date) ing.

  In the present embodiment, a set of unique expressions is expressed as “Junichiro Koizumi: Tokyo Station”. In this case, the specific expression that appears first in the input text in the set of specific expressions is represented as the front specific expression on the left side of “:”, and the specific expression that appears later is displayed on the right side of “:” as the backward specific expression. Further, when a specific expression set of “Junichiro Koizumi: Tokyo Station” is input, the processing result of the specific expression association unit 13 is shown in FIG.

  Next, an outline of the relationship information extraction processing unit 20 will be described. The relationship information extraction processing unit 20 includes a relationship information candidate extraction unit 21, a dependency structure information acquisition unit 22, a relationship estimation information acquisition unit 23, a change information acquisition unit 24, a model selection unit 25, and a classifier 26. The relation information extracting unit 27 extracts relation information of a set of specific expressions based on the analysis result acquired from the analysis processing unit 10.

  When the analysis result obtained by the analysis processing unit 10 is acquired from the specific expression association unit 13, the related information candidate extraction unit 21 extracts at least one independent word included in the input text as a related information candidate (step S6 in FIG. 2). . Here, the relationship information candidate extraction unit 21, when a set of specific expressions is included in the same phrase in the input text, between the independent words adjacent to the front of the front specific expression in the specific expressions and each specific expression The existing independent words and the independent words that are adjacent to the rear of the backward unique expression are extracted as related information candidates. For example, when a set of specific expressions “Ishihara: Tokyo” is included in the same phrase, such as “<PSN> Ishihara </ PSN> <LOC> Tokyo </ LOC> governor”, An independent word “governor” adjacent to the back of the backward proper expression “Tokyo” is extracted as a candidate for related information.

  In addition, the related information candidate extraction unit 21, when the set of specific expressions in the input text is included in different clauses, the independent word adjacent to the front or rear of the front specific expression in the set of specific expressions, and the rear specific expression Independent words adjacent to the front or rear of the are extracted as related information candidates. For example, if the set of specific expressions “Joiichiro Koizumi: Liberal Democratic Party” is included in different clauses, such as “<ORG> Liberal Democratic Party </ ORG> Governors”, the backward proper expression “Liberal Democratic Party” An independent word “Governor” adjacent to the rear is extracted as a related information candidate.

  Furthermore, when the main word of the phrase that does not include the specific expression pair is an independent word, the related information candidate extraction unit 21 extracts the morphemes from the head of the phrase to the main word as related information candidates. For example, when the main sentence “Speech” is an independent word as in the phrase “Speech”, the morpheme “Speech” is extracted as a related information candidate.

  In the present embodiment, when the analysis result shown in FIG. 4 is input, “Ichiro Ozawa”, “Osaka Station”, and “Speech” are extracted as related information candidates. Further, in the present embodiment, the above three relational information candidate extraction methods are used. However, any one or two of the relational information candidate extraction methods may be used to perform the relational information candidate extraction process. .

  When the dependency structure information acquisition unit 22 acquires the analysis result by the analysis processing unit 10 from the specific expression association unit 13 and acquires the related information candidate from the related information candidate extraction unit 21, the set of specific expressions and the related information candidate Extract the smallest dependency tree that contains. In addition, the dependency structure information acquisition unit 22 includes the extracted dependency tree, the part of speech / dependency type / participant of the main part of each clause, the specific expression identifier of the set of specific expressions, the notation of the morphemes of the related information candidates, The dependency structure information is acquired by generating the tree structure information using the part of speech (step S7 in FIG. 2). For example, the tree structure information generated from the unique expression set “Junichiro Koizumi: Tokyo Station” and the related information candidate “Ichiro Ozawa” is shown in FIG. Further, the tree structure information generated from the unique expression set “Junichiro Koizumi: Tokyo Station” and the related information candidate “Osaka Station” is shown in FIG. 6 and is called “Junichiro Koizumi: Tokyo Station”. The tree structure information generated from the set of proper expressions and the related information candidate “speech” is shown in FIG. In the present embodiment, “φ” is used to express that the particle is not included in the phrase. For example, the tree structure information in FIGS. 5 to 7 indicates that the particle is not included in the phrase “Speaked”.

  When the relationship estimation information acquisition unit 23 acquires the dependency structure information of each relationship information candidate generated by the dependency structure information acquisition unit 22, the relationship estimation information indicating the degree to which the relationship information candidate is estimated to be the relationship information. Is acquired from the relationship estimation information storage unit 30 (step S8 in FIG. 2). The relationship estimation information storage unit 30 has a plurality of relationship estimation information corresponding to each of a plurality of morphemes, for example, “Ichiro Ozawa = 0.1”, “Osaka Station = 0.2”, “Speech = 1”. Each relationship estimation information is stored in advance, and is created by a human or a calculation method described later. When the relationship estimation information acquisition unit 23 acquires the dependency structure information illustrated in FIG. 7, among the relationship estimation information stored in the relationship estimation information storage unit 30, the relationship estimation corresponding to the relationship information candidate “speech” Information is searched, and relevant relationship estimation information “1” is extracted. Then, the relationship estimation information acquisition unit 23 adds the extracted relationship estimation information as a child node of the candidate node in the dependency structure information as illustrated in FIG. The relationship estimation information acquisition unit 23 performs the same process on the other relationship information candidates (“Ichiro Ozawa” and “Osaka Station”) extracted by the relationship information candidate extraction unit 21.

  As for the technique for calculating the estimated degree of related information from a large corpus, Tanaka et al., “Classification of syntactic categories of nouns based on the dispersion of semantic categories”, Transactions of Information Processing Society of Japan, vol. 40, no. 9, pp. 3387-3396, September 1999.

When the change information acquisition unit 24 acquires the creation date of the input text and the processing result of the relationship estimation information acquisition unit 23, the past relationship information used in the past as the relationship information of each unique expression changes to the relationship information candidate over time. Change information representing the degree of estimation that is estimated is acquired for each extracted relational information candidate (step S9 in FIG. 2). More specifically, the change information acquisition unit 24 is created in the past from the creation date of the input text among the relationship information of each specific expression extracted in advance from large-scale text data as a predetermined text set. acquires related information extracted from the text as the past related information R _m. Then, the change information acquisition unit 24, based on the past relationship information R _m and the information stored in the change information storage unit 40, the probability P _T (r _n | R _m ) and the rank Rank P _T ( r _n | R _m ) is acquired. Details of the probability P _T (r _n | R _m ) and the rank Rank P _T (r _n | R _m ) will be described later.

Here, information stored in the change information storage unit 40 will be described. The change information storing section 40, as shown in FIGS. 9 to 14, the two related information r _m, co-occurrence count Count (r _m, r _n) of the r _n and the relationship information r _m from the relationship information r _n A trigger model P _T (r _n | r _m ), which is a probability model representing the change over time, is stored in advance for each type of set of proper expressions. For example, FIG. 9 shows an example of the data structure in the type of the unique expression set “person name: person name”, and FIG. 10 shows an example of the data structure in the type of the set of specific expression “person name: organization name”. It is shown. 11 and 12 show the type of the unique expression set “organization name: organization name”, FIG. 13 shows the type of the specific expression set “person name: place name”, and FIG. An example of the data structure is shown for the type of the unique expression set “place name”.

The co-occurrence number Count (r _m , r _n ) represents the number of occurrences of the relationship information r _m and r _n in each text in the large-scale text data. Specifically, the co-occurrence counting method will be described. First, a set of specific expressions having some relationship in the text and a plurality of relation information corresponding to each specific expression of the set are extracted in advance from large-scale text data. Is done. Here, the date of creation of the text is attached to each of the plurality of pieces of relation information, and each piece of relation information is arranged in the order of creation date. Then, a combination of two pieces of relation information having different creation dates among a plurality of pieces of relation information is counted as being co-occurring before and after time. For example, three pieces of relational information corresponding to a set of specific expressions “Mr. A: Mr. B” corresponding to a type of specific expression set “person name: person name” (January 10, 1998 dating, February 15, 2001) Day engagement, December 24, 2001 marriage) is extracted, the co-occurrence number Count (dating, dating, engagement), (engagement, marriage) and (engagement, marriage) related information combination (Engagement), Count (dating, marriage) and Count (engagement, marriage) are counted.

  As for the technology for extracting from a large-scale text data a set of specific expressions having some relationship in the text and the relationship information of each specific expression of the set, Non-Patent Document 1, Hirano et al. Determination of relationship between proper expressions using natural language ”, natural language processing, vol. 15, no. 4, pp. 43-58, 2008. In addition, the present inventor, by using the technique, from the 8,320,042 text created between January 1, 1991 and June 30, 2006, the group of the proper expression and the group. It has been found that 10,463,232 combinations of three relational information corresponding to each unique expression are extracted.

  Also, the change information in step S9 in FIG. 2 can be obtained by storing in advance in the past relationship information storage unit 50 a combination of the extracted unique expressions and a combination of three relationship information corresponding to each unique expression of the set. It is possible to shorten the processing time of the acquisition process.

次に、変化情報取得部２４は、取得した過去関係情報Ｒ_m と、変化情報記憶部４０に記憶されているトリガーモデルＰ_T （ｒ_n ｜ｒ_m ）と、関係推定情報取得部２３の処理結果に含まれる関係情報候補とを用いて、関係情報候補が過去関係情報Ｒ_m から経時変化したものであると推定される確率Ｐ_T （ｒ_n ｜Ｒ_m ）を、関係情報候補毎にもとめる。具体的には、確率Ｐ_T （ｒ_n ｜Ｒ_m ）は以下の式（２）でもとめられる。

例えば、過去関係情報Ｒ_m ＝｛交際、婚約｝で、関係情報候補＝｛結婚、会う、食べる、飲む、ライバル｝の場合には、関係情報候補「結婚」の確率Ｐ_T （結婚｜Ｒ_m ）は以下の式（３）でもとめられる。

ここで、図９を参照すると、Ｐ_T （結婚｜交際）は０．０４９８２７３３１であり、Ｐ_T （結婚｜婚約）は０．１５６９１７５４８であるから、Ｐ_T （結婚｜Ｒ_m ）は０．１５６９１７５４８となる。また、関係情報候補「会う」の確率Ｐ_T （会う｜Ｒ_m ）は以下の式（４）でもとめられる。

そして、変化情報取得部２４は、各関係情報候補毎にもとめられた確率Ｐ_T （ｒ_n ｜Ｒ_m ）について、値の大きい順に順位ＲａｎｋＰ_T （ｒ_n ｜Ｒ_m ）を付与する。例えば、関係情報候補「結婚」の順位ＲａｎｋＰ_T （結婚｜Ｒ_m ）には「１」が付与され、関係情報候補「会う」の順位ＲａｎｋＰ_T （会う｜Ｒ_m ）には「２」が付与される。 Next, the trigger model P _T (r _n | r _m ) determines the co-occurrence number Count (r _m , r _n ) for each type of proper expression (for example, “person name: person name”, “person name: organization name”). , “Person name: place name”). Here, the eight unique expression identifiers described above can be used for the aggregation process. In this case, the trigger model P _T (r _n | r _m ) is also obtained by the following equation (1).

Next, the change information acquisition unit 24 processes the acquired past relationship information R _m , the trigger model P _T (r _n | r _m ) stored in the change information storage unit 40, and the process of the relationship estimation information acquisition unit 23. Using the relationship information candidate included in the result, the probability P _T (r _n | R _m ) that the relationship information candidate is estimated to have changed with time from the past relationship information R _m is determined for each relationship information candidate. . Specifically, the probability P _T (r _n | R _m ) can be determined by the following equation (2).

For example, if the past relationship information R _m = {dating, engagement} and the relationship information candidate = {marriage, meet, eat, drink, rival}, the probability P _T (marriage | R _m ) Can also be stopped by the following equation (3).

Here, referring to FIG. 9, P _T (marriage | dating) is 0.049827331, and P _T (marriage | engagement) is 0.155917548, so P _T (marriage | R _m ) is 0.156397548. It becomes. Further, the probability P _T (meet | R _m ) of the related information candidate “meet” can be obtained by the following equation (4).

Then, the change information acquisition unit 24 assigns rank RankP _T (r _n | R _m ) in descending order of the probability P _T (r _n | R _m ) determined for each relationship information candidate. For example, “1” is assigned to the rank RankP _T (marriage | R _m ) of the relationship information candidate “marriage”, and “2” is assigned to the rank RankP _T (meet | R _m ) of the relationship information candidate “meet”. Is done.

  In the above formulas (2) to (4), the maximum value of each trigger model is obtained, but the minimum value and the average value of each trigger model may be obtained.

Further, when the change information acquisition unit 24 acquires the dependency structure information shown in FIG. 8, the probability (P _T (speech) that it is estimated that the relationship information candidate (“speech”) has changed over time from the past relationship information R _m. | R _m ) = 0.83) and the rank of the probability (RankP _T (speech | R _m ) = 1) are obtained as change information, and the obtained change information is modified as shown in FIG. Add as a child node of a candidate node in the information. In addition, the change information acquisition unit 24 performs the same process on other related information candidates (“Ichiro Ozawa” and “Osaka Station”).

Thus, in each text in the large-scale text data, the number Count (r _m , r _n ) in which past relationship information (here, r _m ) and relationship information candidates (here, r _n ) co-occur. Based on the probability P _T (r _n | r _m ), the change information can be acquired based on the probability.

  When the model selection unit 25 acquires the processing result of the change information acquisition unit 24, the model selection unit 25 classifies the set of proper expressions based on the specific expression identifier assigned by the specific expression association unit 13 and extracts it by a classifier 26 described later. The type of model to be selected is selected (step S10 in FIG. 2). For example, when “Joiichiro Koizumi: Tokyo Station” is input as a set of specific expressions, the model selection unit 25 classifies the set of specific expressions into a type of “person name: place name” and sets the classified specific expressions. The type of output.

  When the classifier 26 acquires the processing result of the change information acquisition unit 24 and the processing result of the model selection unit 25, a plurality of models are stored based on the type of the unique expression selected by the model selection unit 25. A model is extracted from the model storage unit 60 thus obtained. Then, the classifier 26 determines whether or not each relation information candidate becomes relation information of a set of unique expressions using the extracted model (step S11 in FIG. 2).

  Here, the model uses the analysis processing unit 10 and the result including the result determined in advance for the relationship information corresponding to the predetermined specific expression set and the text including each specific expression corresponding to the predetermined specific expression set, and It is generated in advance by performing well-known machine learning using information extracted in advance from the relationship information extraction process 20. In addition, the discrimination result for a predetermined set of specific expressions is made in advance based on human judgment. Note that each model may be configured according to the type of set of unique expressions, such as “person name: place name”, “person name: name”, or the like, or may be configured without distinguishing the types.

  In this case, since the change information given by the change information acquisition unit 24 is used for the discrimination by the classifier 26, each of the past relationship information is estimated based on the degree to which the relationship information candidate is estimated to have changed over time. It is possible to determine whether or not the relationship information candidate is related information of a set of unique expressions. In the present embodiment, it is determined that the relationship information candidate “speech” among the relationship information candidates is the relationship information of the specific expression group “Junichiro Koizumi: Tokyo Station”.

  Note that the classifier 25 may be configured to output a numerical value indicating the degree of possible relation information in addition to the determination result of whether or not the relation information is obtained. As machine learning, a well-known machine can be used. However, it is desirable to use a machine that can learn by directly inputting data of a tree structure or a graph structure.

  The relationship information extraction unit 27 extracts the relationship information candidate determined by the classifier 26 as the relationship information from among the relationship information candidates as the relationship information of the set of specific expressions, and uses the extracted relationship information as a display unit. Output (step S12 in FIG. 2). In addition, when the classifier 26 is configured to output a numerical value indicating the degree that can be related information, the related information extraction unit 27 extracts a related information candidate having the largest numerical value as related information. Alternatively, relationship information candidates whose numerical values are larger than a predetermined threshold value may be extracted as relationship information.

  In this way, the text “Junichiro Koizumi gave a speech at Tokyo Station and Ichiro Ozawa at Osaka Station”, a unique expression pair “Junichiro Koizumi: Tokyo Station”, and the text creation date are entered. Then, the relation information “speech” is extracted.

  The inventor conducted performance evaluation of the technique described in Non-Patent Document 1 described above, and found that the relevance rate representing the extraction accuracy of the optimum relationship information was 67.2%, and the reproduction rate of the relationship information in the same case. And an evaluation result of 43.9% was obtained. On the other hand, when performance evaluation was performed for this embodiment, a special effect was obtained in which the precision was improved to 70.7% and the recall was improved to 48.6%.

  As described above, in the above embodiment, at least one independent word included in the input text is extracted as a related information candidate, and past related information used in the past as related information of each unique expression is stored in the related information candidate over time. Since relation information is extracted from relation information candidates based on change information representing the degree of estimated change and analysis results, relation information is extracted in consideration of changes in the relationship over time between each specific expression Can do. Therefore, for example, the past relationship information of the unique expression set of “Mr. A” and “Mr. B” is “Dating”, and a set of relationship information candidates extracted from the input text is {marriage, meet, eat, drink , Rival}, based on the transition of typical relationships such as “dating”, “engagement”, “marriage” between people, Since it can be extracted as the relationship information, the relationship information between a plurality of unique expressions can be extracted with high accuracy.

  Further, when the analysis result obtained by the analysis processing unit 10 is acquired, at least one independent word included in the input text is extracted as a related information candidate, and a relationship estimation indicating the degree to which the related information candidate is estimated to be related information Information and the change information are acquired for each extracted related information candidate, and the related information is extracted from the related information candidate based on the analysis result, the relationship estimation information, and the change information. Since the degree estimated to be information can be used as a feature, the precision and recall can be further improved.

  In this embodiment, compared with the technique described in Non-Patent Document 1 described above, the relevance ratio representing the extraction accuracy of the optimum relationship information is improved from 67.2% to 70.7%, and the same case In particular, the reproducibility of the relational information in 4 is increased from 43.9% to 48.6%.

  Furthermore, since change information is acquired based on the number of past occurrences of past relationship information and relationship information candidates in each text in large-scale text data, it is based on the relationship between past relationship information and relationship information candidates in large-scale text data. The relationship information can be extracted, and the extraction accuracy of the optimum relationship information can be improved.

  In addition, the probability that the past relationship information has changed to the related information candidate with time is determined based on the number of past occurrences of the related information candidate and the related information candidate in each text in the large-scale text data, and the change information is determined based on the probability. Since it is acquired, it is possible to extract the relationship information based on the probability that the past relationship information has changed to the relationship information candidates, and the extraction accuracy of the optimum relationship information can be further improved.

  In addition, the said embodiment is only a specific example of this invention, and this invention is not limited only to the said embodiment. For example, the present invention installs a program for realizing the functions shown in the configuration diagram of FIG. 1 or a program having the procedure shown in the flow of FIG. 2 via a recording medium or a communication line in a known computer. This is also possible.

  In addition, the related information candidate extracting unit 21 includes a predetermined particle included at the end of one of the consecutive clauses that appears first in the input text, and a predetermined verb included in the other of the consecutive clauses. When included in the beginning of a phrase, a morpheme sequence including one phrase and a morpheme from the beginning of the other phrase to the main word may be extracted as a related information candidate. For example, the relationship information candidate extraction unit 21 stores a functional verb “effect” and a particle “wo” corresponding to the functional verb. When the phrase “influence” and the phrase “influence” appear consecutively in the input text, the relationship information candidate extraction unit 21 selects one of the phrases “influence” and the other “influence”. The morpheme from the beginning of the clause to the main word, that is, the morpheme string “influence” consisting of “influence” is extracted as a candidate of related information.

  In this case, since a morpheme sequence consisting of a combination of a predetermined particle and a functional verb can be extracted as related information candidates, the number of related information candidates can be increased, and the extraction accuracy of optimum related information can be improved. Can do.

  DESCRIPTION OF SYMBOLS 10 ... Analysis process part, 11 ... Morphological analysis part, 12 ... Dependency analysis part, 20 ... Relation information extraction process part, 21 ... Relation information candidate extraction part, 23 ... Relation estimation information acquisition part, 24 ... Change information acquisition part, 27: Relationship information extraction unit, 40: Change information storage unit, 50: Past relationship information storage unit

Claims (10)

An apparatus for extracting information related to a plurality of input proper expressions,
When a text including each unique expression and a set of specific expressions including the specific expressions are input , the input text is subjected to morphological analysis, and the set of specific expressions is analyzed together with the dependency of the clauses constituting the input text. An analysis processing unit to
When the analysis result by the analysis processing unit is acquired, at least one independent word included in the input text is extracted as a related information candidate, and past related information used in the past as related information of each specific expression is extracted as the related information candidate Change information indicating the degree of change estimated to have changed over time is obtained for each extracted relational information candidate, and a model corresponding to the type of the specific expression set is selected and analyzed based on the analysis result by the analysis processing unit. A relational information extraction processing unit that determines whether each relational information candidate is relational information of the set of specific expressions based on the analysis result and change information by the processing unit , and extracts the relational information. The relationship information extraction apparatus characterized by this. When the relation information extraction processing unit acquires an analysis result by the analysis processing unit, the relation information extraction processing unit extracts at least one independent word included in the input text as a related information candidate, and the related information candidate is estimated to be related information. The relationship estimation information indicating the degree and the change information are acquired for each extracted relationship information candidate, and a model corresponding to the type of the specific expression set is selected based on the analysis result by the analysis processing unit, and the analysis processing unit The relationship information is extracted by determining whether each relationship information candidate is the relationship information of the set of specific expressions based on the analysis result, relationship estimation information, and change information by using the model. The related information extracting apparatus according to 1. 3. The relationship according to claim 1, wherein the relationship information extraction processing unit acquires change information based on the number of past relationship information and relationship information candidates that co-occur in each text in a predetermined text set. Information extraction device. The related information extraction processing unit determines the probability that the past related information has changed to the related information candidate with time based on the number of past related information and related information candidates that co-occurred in each text in the predetermined text set, Change information is acquired based on probability. The relation information extraction device according to claim 3 characterized by things. A method of extracting information related to a plurality of input unique expressions using a computer,
When a text including each specific expression and a set of specific expressions composed of the specific expressions are input, the computer performs morphological analysis on the input text, and includes the dependency of the phrase constituting the input text and the specific expression. Analyzing the tuple ,
When the computer obtains the analysis result, the computer extracts at least one independent word included in the input text as a related information candidate, and past related information used in the past as related information of each specific expression is the related information. Change information indicating the degree of change estimated to have been changed over time for each candidate is obtained for each extracted relational information candidate, and a model corresponding to the type of the specific expression set is selected based on the analysis result. A relation information extraction method characterized by discriminating, using the model, whether each relation information candidate is relation information of the set of specific expressions based on information, and extracting relation information. When the computer acquires the analysis result, the computer extracts at least one independent word included in the input text as a related information candidate, and relationship estimation information representing a degree that the related information candidate is estimated to be related information; The change information is acquired for each extracted relationship information candidate , a model corresponding to the type of the specific expression set is selected based on the analysis result, and each relationship based on the analysis result, the relationship estimation information, and the change information is selected. 6. The relation information extraction method according to claim 5, wherein the information candidate is discriminated using the model to determine whether or not the information candidate is relation information of the set of the specific expressions . The relation information extraction method according to claim 5, wherein the computer acquires change information based on a number of past relation information and relation information candidates that co-occur in each text in a predetermined text set. The computer determines the probability that the past relationship information has changed to the related information candidate over time based on the number of past relationship information and the related information candidate that co-occurred in each text in the predetermined text set, and based on the probability Change information is acquired. The relationship information extraction method of Claim 7 characterized by the above-mentioned.   The program for functioning a computer as each means of the related information extraction apparatus in any one of Claims 1 thru | or 4.   A program for causing a computer to execute each process of the related information extraction method according to any one of claims 5 to 8.

Images