JP4793932B2 - Apparatus and method for extracting sets of interrelated specific expressions - Google Patents

Description

  The present invention relates to an apparatus and method for extracting a set of specific expressions composed of a plurality of related specific expressions, which play an important role in a summarization system for summarizing input text, and the like.

  First, specific examples of interrelated specific expressions will be described below.

  For example, in the text “Prime Minister Koizumi will visit China and meet with President Hu Jintao. He plans to meet with President Bush over the United States.” The combination of proper expressions is considered between the proper expressions of “Chinese” and “America” representing place names. Here, “Koizumi” and “China” have a “visit” relationship from the statement that “Prime Minister Koizumi visits China and meets with President Hu Jintao” in the above text. In addition, “Koizumi” and “America” have a “migrating” relationship because it is interpreted that “then Prime Minister Koizumi is going to the United States to meet with President Bush”. However, there is no relationship between “Hu Jinxi” and “America” and “Bush” and “China”. Therefore, in a summarization system that summarizes input text, a search system that obtains necessary information from a large amount of text data, and the like, it is important to extract sets of specific expressions related to each other.

2. Description of the Related Art Conventionally, as this kind of specific expression pair extraction apparatus and method, one using machine learning based on word information existing between two specific expressions is known (see, for example, Non-Patent Document 1).
Kambhatla, “Combining Lexical, Syntactic, and Semantic Features with Maximum Entropy Models for Extracting Relations”, The Companion Volume to the Proceedings of 42st Annual Meeting of the Association for Computational Linguistics, Association for Computational Linguistics, July 2004, p. 178 -181

  However, the conventional unique expression set extraction device simply uses the word information existing between each unique expression as a feature, so the subject of the second sentence, that is, “Koizumi Koizumi” is omitted as in the illustrated text. If it is, the information for judging the relationship between the specific expression “Koizumi” and the specific expression “USA” is “The Prime Minister visits China and meets with President Hu Jintao. Become. In this case, what is the subject of the second sentence cannot be properly determined, so it is difficult to determine whether or not there is a relationship between each unique expression.

  The present invention has been made in view of the above problems, and an object of the present invention is to provide an apparatus and method for extracting a set of specific expressions related to each other according to each case. .

In order to achieve the above-mentioned object, the inter-specific-expression-specific group extraction apparatus of the present invention is an apparatus for extracting a specific-expression group consisting of a plurality of inter-specific expressions from an input text. When input, there is a possibility that a specific expression extraction processing unit that extracts a plurality of specific expressions included in the input text by performing morphological analysis on the input text, and an area corresponding to each of the plurality of particles as a subject. Each of the noun phrases that are included in the input text in the order of particles with high noun phrases and that are stored in the corresponding areas of the particles accompanying the noun phrases in a stack format, and each extracted by the proper expression extraction processing unit The noun phrases are read out in order from the beginning of the input text and stored in the phrase storage unit for each set of a plurality of specific expressions formed by combining the specific expressions, and correspond to the sets of specific expressions. When reading words including a unique representation of the while appearing later in the input text of the named entity, information words including the other named entity represents whether stored in ordinal number of the corresponding regions of any particle in the phrase storage unit And whether the particle corresponding to the area in which the phrase including the other proper expression is stored is a particle associated with the noun phrase that is most likely to be used as the subject among the particles corresponding to the storage area of the phrase storage unit. And information including a combination of a phrase including each of the specific expressions and a particle corresponding to an area in which the other specific expression is stored in the phrase storage unit, and including at least each acquired information The feature extraction processing unit that extracts the features, the features extracted by the feature extraction processing unit, and whether or not there is a relationship between each specific expression corresponding to a predetermined specific expression pair is determined in advance. Between the specific expressions corresponding to the set of specific expressions based on the pre-feature extracted in advance from the feature extraction processing unit using the result and the text including each specific expression corresponding to the predetermined specific expression set. And a discrimination processing unit that discriminates whether or not there is a relationship for each set of specific expressions.

According to another aspect of the present invention, there is provided a method for extracting a set of related specific expressions, wherein a set of specific expressions composed of a plurality of related specific expressions is extracted from an input text using a computer. In the method, when a text is input, the computer corresponds to each of a plurality of particles, and a first step of analyzing a morphological analysis of the input text to extract a plurality of specific expressions included in the input text. With respect to a predetermined phrase storage unit having a stack format area in the order of particles associated with a noun phrase that is likely to be used as a subject, a noun phrase included in the input text is used as a corresponding area of a particle associated with the noun phrase. a second step of storing, in a plurality of named entity pairs each of which comprising a combination of the named entities extracted, the second reading of the noun phrase from beginning of a sentence of the input text in order Performs step, when reading words including a unique representation of one appearing later in the input text of each unique representation corresponding to a set of named entities, words including the other specific representation of any particle in the phrase storage unit The particle corresponding to the area in which the information indicating the number of the corresponding area stored and the other phrase including the specific expression is stored is most used as the subject among the particles corresponding to the storage area of the phrase storage unit. Information indicating whether or not a particle is associated with a highly likely noun phrase, information including a combination of a word including each of the specific expressions and a particle corresponding to the area in which the other specific expression is stored in the phrase storage unit, obtains, a third step of extracting at least including identity of each information obtained, extracted feature, the relationship between the specific representation corresponding to a set of predetermined named entities Based on the result of the determination in advance about the presence or absence and the prior feature extracted by performing the third step in advance using the text including each specific expression corresponding to the predetermined set of specific expressions, A fourth step is performed for discriminating whether or not there is a relationship between the specific expressions corresponding to the set for each set of specific expressions.

  Thus, when a phrase including one specific expression that appears later in the input text is read out from each specific expression corresponding to the set of specific expressions, information indicating the storage state of the other specific expression in the phrase storage unit is obtained. Since a feature is included and the relationship between each unique expression is determined using the feature, for example, the particle including the other specific expression is associated with each specific expression based on information such as which particle is accompanied. It is possible to determine whether or not there is a relationship.

  According to the apparatus and method for extracting a group of specific expressions related to each other of the present invention, whether or not there is a relationship between specific expressions based on information such as which particle is accompanied by a phrase including the other specific expression For example, in the text "Prime Minister Koizumi visits China and meets with President Hu Jintao. He plans to meet with President Bush over the United States." Can discriminate the relationship between each proper expression using the information that the particle is accompanied by the particle “ha”. Therefore, even when a sentence in which the phrase including the other specific expression is omitted is used, the relationship between each specific expression can be determined by judging the omitted phrase. A set of related specific expressions can be extracted.

  FIGS. 1 to 7 show a first embodiment of the present invention. FIG. 1 is a configuration diagram of a unique expression set extraction apparatus, FIG. 2 is a flowchart of a unique expression set extraction process, and FIG. 3 is a dependency. FIG. 4 is a diagram showing an overview of the analysis result by the analysis unit, FIG. 4 is a diagram showing an overview of the processing result by the base analysis result synthesis unit, FIG. 5 is a diagram showing an example of the processing result by the inter-specific expression information extraction unit, and FIG. The figure which shows the structure of a memory | storage part, FIG. 7 is a figure which shows an example of the process result by an abbreviated information extraction part.

  DESCRIPTION OF THE PREFERRED EMBODIMENTS An overview of an apparatus and method for extracting sets of related expressions according to the present invention will be described below with reference to the drawings.

  A mutual expression group extraction device (hereinafter referred to as an extraction device) according to the present invention includes a computer device mainly composed of a well-known CPU, and includes display means such as a monitor, input means such as a keyboard, It includes storage means such as a hard disk and a memory, and a communication device that can be connected to an external network (both not shown). In addition, the extraction device of the present invention is provided with a specific expression extraction processing unit 10, a feature extraction processing unit 20, a phrase storage unit 30, a discrimination processing unit 40, and a model storage unit 50.

  As shown in FIG. 1, the specific expression extraction processing unit 10 includes a morphological analysis unit 11, a specific expression extraction unit 12, a dependency analysis unit 13, a base analysis result synthesis unit 14, and a specific expression pair generation unit 15. The input text is used for morphological analysis to extract a plurality of specific expressions included in the input text.

  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 the text “Prime Minister Koizumi visits China and meets with President Hu Jintao, then plans to meet with President Bush over the United States” is entered, the processing result by morphological analysis unit 11 is , "Koizumi (noun) / Prime Minister (noun) / Ha (particle) / China (noun) / O (case particle) / Visit (verb) /, (Reading) / Hu Jintao (noun) / State (noun) / Principal ( (Noun) / to (case particle) / conversation (verb) / to (verb) /. (Phrase) / after (conjunction) / america (noun) / ni (case particle) / walking (verb) / bush (noun) / President (noun) / to (case particle) / discussion (verb) / to (case particle) / plan (verb) / do (verb) / te (connective particle) / is (verb) /. .

  The specific expression extraction unit 12 extracts a specific expression by performing a well-known specific expression extraction process on the input text obtained from the morphological analysis unit 11 and has been subjected to morpheme analysis. After assigning a kind of specific expression such as a place name, the specific expression is output (step S3 in FIG. 2). Here, when the input text that has been subjected to the morphological analysis as illustrated is input to the specific expression extraction unit 12, “Koizumi (person name)”, “China (place name)”, “Hu Jintao (person name)”, “USA (place name)” And "Bush (person name)" information is output.

  The dependency analysis unit 13 divides the text into phrases by performing a well-known dependency analysis process on the morpheme-analyzed input text acquired from the morpheme analysis unit 11, and between the plurality of divided phrases. The dependency relationship is analyzed and output (step S4 in FIG. 2). In this case, when the input text illustrated is analyzed by the dependency analysis unit 13, information representing a dependency structure (dependency tree) as shown in FIG. 3 is output as an analysis result.

  The base analysis result synthesizing unit 14 performs a process of synthesizing the information output from the specific expression extraction unit 12 and the information output from the dependency analysis unit 13 (step S5 in FIG. 2). Specifically, when the base analysis result synthesis unit 14 acquires information from the specific expression extraction unit 12 and the dependency analysis unit 13, the base analysis result synthesis unit 14 gives a tag representing the specific expression to each specific expression. For example, a tag <PSN> indicating the name of the person is added before and after “Koizumi”, “Hu Jinjing”, and “Bush”, and a tag <LOC> indicating the place name is added before and after “China” and “America”. Is done. In this case, the processing result of the analysis result combining unit 14 is shown as in FIG.

  When the specific expression pair generation unit 15 acquires the processing result of the base analysis result synthesis unit 14, the specific expression pair generation unit 15 generates a plurality of specific expression pairs by combining all the specific expressions included in the processing result (FIG. 2). Step S6). In this embodiment, the specific expression pair is composed of two specific expressions, that is, a specific expression that represents a person name and a specific expression that represents a place name, and the specific expression group is expressed as “Koizumi: America”. To do. In this case, among the specific expressions, the specific expression that appears first in the text is represented on the left side of “:” as the forward specific expression, and the specific expression that appears later is represented on the right side of “:” as the backward specific expression. From the input text shown as examples, there are six unique expressions of “Koizumi: China”, “Koizumi: America”, “China: Hu Jintao”, “Hu Jintao: America”, “China: Bush”, and “America: Bush”. A set is output.

  Next, an overview of the feature extraction processing unit 20 will be described. The feature extraction processing unit 20 includes an inter-specific expression information extraction unit 21 and an abbreviated information extraction unit 22, and each specific expression included in the combination is generated for each set of plural specific expressions generated by the specific expression pair generation unit 15. Features between them are extracted (step S7 in FIG. 2).

  When the inter-specific expression information extraction unit 21 acquires the processing result of the base analysis result synthesis unit 14 and the set of specific expressions generated by the specific expression pair generation unit 15, each corresponding to the set of specific expressions in the input text. Words that exist between proper expressions, the number of such words and the number of proper expressions that exist between each proper expression, and the words and parts of speech of the clauses that include each proper expression and the proper expressions in the same phrase Extract whether or not it exists. Here, as an example for specifically explaining the processing contents of the inter-specific expression information extraction unit 21, the processing result of the base analysis result synthesis unit 14 shown in FIG. 4 and the specific expression set “Koizumi: America” Use. First, the words existing between the unique expressions corresponding to the set of specific expressions in the input text are as shown in FIG. 5, and the number of words and the number of specific expressions existing between the specific expressions are “10”, respectively. And “2”. Also, the words and parts of speech of the phrase “Koizumi / Prime / ha” are “conversation / do /.” And “verb / verb / phrase”, respectively, and the phrase “America / ni” is related to the phrase. The word and part of speech of the phrase are “transition” and “verb”, respectively. Further, whether each unique expression exists in the same phrase is “No” because “Koizumi” and “America” are included in different phrases. Therefore, the processing result of the inter-specific expression information extraction unit 21 is shown in FIG.

  In addition, you may comprise the information extraction part 21 between specific expressions so that the distance of the shortest path | route between the clauses containing each specific expression may also be extracted. Here, the distance of the shortest path between clauses including each specific expression is represented by the distance of the shortest path of the phrase including each specific expression in the dependency tree. For example, if we look at the phrases “Koizumi / Primeer / Ha” and “China / Oh”, the phrase “Koizumi / Primeer / Ha” has the phrases “Meeting / To /.” And “Visit /,”. Since there is no route to reach “Chinese / Oh” via “”, the shortest route between each clause is “Koizumi / Prime / Ha” → “Meeting / To /” → “Visiting / , ”→“ China / O ”. Therefore, the distance of the shortest path is 3. In addition, the phrase “Koizumi / Prime / ha” and the phrase “America / de” exist in different dependency trees, so the shortest path cannot be obtained.

  When the abbreviation information extraction unit 22 acquires the processing result of the base analysis result synthesis unit 14 and the set of specific expressions generated by the specific expression pair generation unit 15, the noun phrase included in the input text is extracted from the beginning of the input text. The noun phrases read out in order are stored in the phrase storage unit 30. Further, the abbreviation information extraction unit 22 acquires information representing the storage state of the forward specific expression in the phrase storage unit 30 when reading a phrase including the backward specific expression from among the specific expressions corresponding to the set of specific expressions. Then, a feature including at least the information is extracted.

  Here, the details of the configuration of the phrase storage unit 30 and the operation of the abbreviation information extraction unit 22 will be described using an example of extracting features about a specific expression set “Koizumi: America”. As shown in FIG. 6, the phrase storage unit 30 has storage areas corresponding to a plurality of case particles, and stores noun phrases in a stack format for each storage area. Each storage area is partitioned in the order of a case particle that is most likely to be used as a subject. In this embodiment, the case particle “ha” is used as the highest case particle. In addition, the phrase storage unit 30 stores the newly stored noun phrase at the top in the storage area corresponding to the noun phrase.

  When the abbreviation information extraction unit 22 reads the noun phrase “Prime Minister Koizumi” from the input text, as shown in FIG. 6A, in the phrase storage unit 30, the memory corresponding to the case particle “ha” associated with the noun phrase is stored. The noun phrase “Prime Minister Koizumi” is stored as the first record in the area. Next, when the abbreviation information extraction unit 22 reads the noun phrase “China”, as shown in FIG. 6B, the storage area corresponding to the case particle “O” associated with the noun phrase in the phrase storage unit 30. The noun phrase "China" is memorized as the first record. Further, when the abbreviation information extraction unit 22 reads the noun phrase “President Hu Jintao” from the input text, as shown in FIG. 6C, the phrase storage unit 30 converts the noun phrase “to” into the noun phrase “to”. The noun phrase “Hu Jintao” is stored as the first record of the corresponding storage area (in this case, the area partitioned as “other”).

  Then, when the abbreviated information extraction unit 22 reads the backward proper expression “America”, whether or not the forward specific expression “Koizumi” exists in the phrase storage unit 30 and which case particle corresponds to the forward specific expression “Koizumi”. Whether it is stored in the area, what number of records in that area, whether the forward specific expression “Koizumi” is stored in the area corresponding to the highest case particle, and the forward specific expression The route to the expression is acquired as information representing the storage state of the forward unique expression “Koizumi” in the phrase storage unit 30. First, whether or not the forward unique expression “Koizumi” exists in the phrase storage unit 30 is “Yes” from FIG. Further, the forward unique expression “Koizumi” is stored in an area corresponding to the case particle “ha”. Further, the forward unique expression “Koizumi” is stored in the first record in the area corresponding to the case particle “ha”. Furthermore, since the case particle “ha” is used as the highest case particle, it is “Yes” as to whether or not the forward proper expression “Koizumi” is stored in the area corresponding to the highest case particle. It becomes. Further, the path from the backward specific expression to the forward specific expression is information combining a case particle corresponding to a storage area in which the forward specific expression is stored in the backward specific expression and a word including the forward specific expression, “America → is → Prime Minister Koizumi”. Accordingly, the processing result of the omitted information extraction unit 22 when the above example is used is shown in FIG. A tag representing a specific expression may be added to the path from the backward specific expression to the forward specific expression. In this case, the information “<LOC> America </ LOC> →→→ <PSN> Koizumi </ PSN> Prime Minister” is extracted as a feature.

  Next, an outline of the discrimination processing unit 40 will be described. The discrimination processing unit 40 includes a model selection unit 41, a classifier 42, and a specific expression pair output unit 43. Based on the features acquired from the feature extraction processing unit 20, the discrimination processing unit 40 includes each specific expression corresponding to a set of specific expressions. Whether or not there is a relationship is determined for each set of proper expressions.

  When the model selection unit 41 acquires the processing results of the inter-specific-expression information extraction unit 21 and the omitted information extraction unit 22 for each specific-expression group, the model selection unit 41 sets the specific-expression group based on the tag assigned by the base analysis result synthesis unit 14. And the type of model extracted by the classifier 42 described later is selected (step S8 in FIG. 2). For example, when a set of unique expressions “Koizumi: America” is input, the model selection unit 41 classifies the set of unique expressions into a type of “person name: place name”, and the feature acquired from the feature extraction processing unit 20 In addition, the type of the set of proper expressions is output.

  When the classifier 42 acquires the information output from the model selection unit 41, the classifier 42 selects a model from the model storage unit 50 in which a plurality of models are stored based on the type of the unique expression set selected by the model selection unit 41. Using the extracted model, it is determined whether or not there is a relationship between each of the specific expressions corresponding to the set of specific expressions (step S9 in FIG. 2).

  Here, the model uses a result determined in advance as to whether or not there is a relationship between each specific expression corresponding to a predetermined specific expression set, and text including each specific expression corresponding to the predetermined specific expression set. The information is previously generated by performing well-known machine learning using information extracted in advance from the specific expression extraction processing unit 10 and the feature extraction processing unit 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 information acquired from the abbreviated information extraction unit 22 is also used for the determination by the classifier 42, for example, it is determined whether or not there is a relationship with a specific expression set “Koizumi: America”. Can use the information that the noun phrase including the forward proper expression “Koizumi” is accompanied by the case particle “ha”.

  The classifier 42 may be configured to output a numerical value representing the degree of relationship in addition to the determination result of whether or not there is a relationship. 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 specific expression pair output unit 43 outputs the set of specific expressions determined by the classifier 42 as having a relationship between the specific expressions to the display unit (step S10 in FIG. 2). When the classifier 42 is configured to output a numerical value indicating the degree of relation, the specific expression pair output unit 43 sets the combination of specific expressions only when the degree of relation is larger than a predetermined threshold value. May be output.

  As described above, in the above-described embodiment, when text is input, a plurality of specific expressions included in the input text are extracted after morphological analysis of the input text, and then a plurality of combined specific expressions are combined. For each set of specific expressions, the noun phrases included in the input text are sequentially stored in the phrase storage unit having areas corresponding to the particles associated with the noun phrases, and among the specific expressions corresponding to the sets of specific expressions When a phrase including a backward specific expression is read, information indicating a storage state of the forward specific expression in the phrase storage unit is acquired, a feature including at least the information is extracted, and the extracted feature and the model storage unit 50 A model, that is, including a result determined in advance as to whether or not there is a relationship between each specific expression corresponding to a predetermined specific expression set, and each specific expression corresponding to the predetermined specific expression set Since each distinctive expression group determines whether or not there is a relationship between each specific expression corresponding to the specific expression pair based on the features extracted in advance using the text, the phrase including the forward specific expression For example, “Prime Minister Koizumi visits China and meets with President Hu Jintao. To determine the relationship between each proper expression using information that the noun phrase “Prime Minister Koizumi” is accompanied by the case particle “ha”. Can do. Therefore, even when using a sentence in which a phrase including a forward unique expression is omitted, it is possible to determine the relationship between each unique expression by determining the omitted word and phrase. A set of proper expressions can be extracted.

  In addition, at least information indicating whether or not a phrase including the forward unique expression is stored in the phrase storage unit 30 when a phrase including the backward specific expression is read out of each specific expression corresponding to the set of specific expressions. Since information is extracted as a feature, it is possible to determine the relationship between each unique expression by determining the omitted word even when using a sentence in which the word including the forward specific expression is omitted. A set of specific expressions related to each other can be extracted.

  Further, when a phrase including a backward specific expression is read out of each specific expression corresponding to a set of specific expressions, the phrase storage unit 30 includes at least a case particle corresponding to an area in which the phrase including the forward specific expression is stored. Since information is extracted as a feature, it is possible to determine the relationship between each unique expression by determining the omitted word even when using a sentence in which the word including the forward specific expression is omitted. A set of specific expressions related to each other can be extracted.

  Furthermore, when a phrase including a backward specific expression is read out from each specific expression corresponding to a set of specific expressions, a case particle corresponding to an area in which the phrase including the forward specific expression is stored is stored in the phrase storage unit 30. Since the information including at least information indicating whether or not the case particle corresponding to the region is the most frequently used subject particle is extracted as a feature, a sentence in which a phrase including a forward unique expression is omitted as a subject is used. Even in this case, it is possible to determine the relationship between the unique expressions by determining the omitted subject, and therefore it is possible to extract a set of specific expressions related to each other according to each case.

  In addition, when a phrase including a backward specific expression is read out from each specific expression corresponding to a set of specific expressions, the phrase including each specific expression and an area in which the front specific expression is stored in the phrase storage unit 30 are stored. Since information including at least information combined with the corresponding case particles is extracted as a feature, it is possible to determine the relationship between each proper expression by judging omitted words and phrases. A set of related specific expressions can be extracted.

  The extraction device according to the second embodiment of the present invention will be described below. The present embodiment is different from the first embodiment in that a discrimination result storage unit 60 that stores a discrimination result every time the discrimination processing unit 40 discriminates whether or not there is a relationship between each unique expression included in a set of proper expressions. When the feature corresponding to the set of specific expressions is extracted, the discrimination processing unit 40 acquires the discrimination result stored in the discrimination result storage unit 60, and the discrimination result, the extracted feature, and a predetermined feature A result determined in advance as to whether or not there is a relationship between each specific expression corresponding to a set of specific expressions, and a pre-feature extracted in advance using text including each specific expression corresponding to the predetermined specific expression set; On the basis of the above, it is configured to determine whether or not there is a relationship between each unique expression corresponding to the set of specific expressions. Since other configurations and operations are the same as those in the first embodiment, only the differences will be described with reference to FIGS.

  The feature extraction processing unit 20 according to the present embodiment includes a specific expression pair rearrangement unit 23 that rearranges a set of specific expressions generated by the specific expression pair generation unit 15 of the specific expression extraction processing unit 10.

  In addition, the discrimination processing unit 40 according to the present embodiment stores a result of the presence / absence of a relationship between each unique expression discriminated by the classifier 42 in the discrimination result storage unit 60 and also stored in the discrimination result storage unit 60. A determination result acquisition unit 44 that acquires the determination result is provided.

  With respect to the extraction apparatus of this embodiment, the operations of the feature extraction processing unit 20 and the discrimination processing unit 40 will be described using the input text exemplified in the first embodiment and the flow of FIG. Each unique expression is given a unique expression ID in the order of appearance in the input text. For example, “Koizumi”, “China”, “Hu Jintao”, “America”, and “Bush” are assigned unique expression IDs “ID1”, “ID2”, “ID3”, “ID4”, and “ID5”, respectively. Is done. Also, nothing is initially stored in the discrimination result storage unit 60.

  First, when the specific expression pairs are output from the specific expression pair generation unit 15 of the specific expression extraction processing unit 10 in the order shown in FIG. 10, the specific expression pair rearrangement unit 23 sorts the specific expression pairs. Rearrange based on (step S11). Here, in the present embodiment, the rearrangement rule “rearranged in descending order of the absolute value difference of the unique expression ID and rearrange in the order of smaller sum of the specific expression IDs when the absolute value difference is equal” is used. Yes. Further, the absolute value difference and the sum of the unique expression ID are shown in FIG. As a result, the sets of specific expressions are rearranged as shown in FIG.

  Next, the inter-specific expression information extraction unit 21 and the abbreviated information extraction unit 22 extract the first combination of unique expressions (“ID1-ID2”) as a processing target from the rearranged combination of specific expressions (step S12). Then, the feature extraction process is performed for the set of specific expressions to be processed (step S13). The contents of the feature extraction process are the same as those in the first embodiment.

  Next, the determination result acquisition unit 44 of the determination processing unit 40 determines whether or not the determination result is stored in the determination result storage unit 60 (step S14). If the determination result is stored, the determination result storage unit All the discrimination results stored in 60 are acquired (step S15). When the processing target is “ID1-ID2”, nothing is stored in the determination result storage unit 60, and thus the process proceeds to step S16.

  When the model selection unit 41 classifies the set of unique expressions (step S16), the classifier 42 determines whether or not there is a relationship between the specific expressions corresponding to the set of specific expressions (step S17), and the determination result. Is output to the specific expression pair output unit 43 (step S18). The specific expression pair output unit 43 outputs a set of specific expressions to the display means when the classifier 42 determines that there is a relationship between the specific expressions. Then, if there is a set of unique expressions that are not extracted as processing targets, the discrimination result acquisition unit 44 stores the discrimination results of the unique expressions and their relations in the discrimination result storage unit 60, and then in step S12. The process proceeds (steps S19 and S20). In this case, the next processing target follows the rearranged order.

  In step S15, all the determination results are acquired from the determination result storage unit 60. However, it is also possible to acquire only the determination results related to the set of specific expressions. Three are shown below.

  First, as a first method, a discrimination result of a specific expression group having the same specific expression ID as the specific expression ID corresponding to the specific expression set to be processed is acquired from the determination result storage unit 60. For example, when the set of unique expressions to be processed is “ID1-ID5”, “ID1-ID2”, “ID4-ID5”, “ID1-ID3”, “ID1-ID3”, The discrimination results corresponding to “ID3-ID5”, “ID1-ID4”, and “ID2-ID5” are acquired.

  As a second method, the discrimination result of the unique expression group having the unique expression ID existing between the unique expression IDs corresponding to the specific expression group to be processed is acquired from the discrimination result storage unit 60. For example, when the set of unique expressions to be processed is “ID1-ID5”, the unique expression IDs existing between the unique expression IDs are “ID2”, “ID3”, and “ID4”. Therefore, “ID1-ID2”, “ID2-ID3”, “ID3-ID4”, “ID4-ID5”, “ID1-ID3”, “ID2-ID4”, “ID3” among the groups of already-identified unique expressions The determination results corresponding to “−ID5”, “ID1-ID4”, and “ID2-ID5” are acquired.

  Further, as a third method, the shortest path between each specific expression corresponding to the set of specific expressions to be processed in the dependency tree is extracted, and a set of specific expressions including the specific expressions existing on the shortest path is determined. The result is acquired from the discrimination result storage unit 60. For example, when the set of specific expressions to be processed is “ID1-ID5”, the shortest path in the dependency tree of FIG. 3 is between the phrase including “Koizumi” and the phrase including “USA”. Since it does not exist, the discrimination result is not acquired.

  In addition, you may comprise so that a discrimination | determination result may be acquired combining three methods.

  In addition, when the discrimination result is acquired from the discrimination result storage unit 60, the classifier 42 determines the acquired discrimination result, the feature between each unique expression, and the model acquired from the model storage unit 50 in step S17 of the flow. Based on the above, it is determined whether or not there is a relationship between the unique expressions.

  As described above, in the above-described embodiment, each time the presence / absence of the relationship between the unique expressions corresponding to the set of specific expressions is determined, the determination result is stored in the determination result storage unit 60, and the feature corresponding to the set of specific expressions is stored. Is extracted, the determination result stored in the determination result storage unit 60 is acquired, and the presence / absence of the relationship between the determination result, the extracted feature, and each specific expression corresponding to a predetermined specific expression set is determined in advance. Based on the determined result and the feature extracted in advance using the text including each specific expression corresponding to the predetermined specific expression set, there is a relationship between each specific expression corresponding to the specific expression set. Since it is determined whether or not there is, it is possible to easily use the determination result of the set of specific expressions related to the set of specific expressions to be determined.

  In addition, the said 1st and 2nd 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 includes a program for realizing the functions shown in the configuration diagrams of FIGS. 1 and 8 or the procedures shown in the flows of FIGS. 2 and 9 via a medium or communication line in a known computer. It can also be realized by installing a program.

  Further, in the above-described embodiment, the inter-specific expression information extraction unit 21 outputs the feature shown in FIG. 5, but as shown in FIG. 13, a word other than the specific expression in the clause including each specific expression, The part of speech of the word may be included in the feature and output. In this case, with respect to the word “Japanese culture” that is not used as a specific expression, only “Japan” is extracted as a specific expression representing the place name, and it is prevented that it is determined whether there is a relationship with other specific expressions. be able to.

  Furthermore, the inter-specific expression information extraction unit 21 may output whether or not there is a noun phrase immediately after each proper expression as shown in FIG. In this case, as in the case described above, only “Japan” is extracted as a unique expression representing a place name for the word “Japanese culture”, and it is possible to prevent the presence or absence of a relationship with another specific expression from being determined. .

  Furthermore, the inter-specific expression information extraction unit 21 makes a feature whether or not another specific expression existing between the specific expressions is the same notation or a partial notation as one of the specific expressions as shown in FIG. It may be included and output. In this case, it is possible to prevent a proper expression that is a target for determining the relationship from being erroneously extracted.

1 is a configuration diagram of a unique expression set extraction apparatus according to the first embodiment of the present invention. Flow chart of specific expression group extraction processing Diagram showing the summary of analysis results by the dependency analysis unit The figure which shows the summary of the processing result by the base analysis result composition section The figure which shows an example of the processing result by the information extraction part between proper expressions The figure which shows the structure of a phrase memory | storage part The figure which shows an example of the processing result by an abbreviation information extraction part Configuration diagram of proper expression set extraction apparatus in the second embodiment of the present invention Flow diagram showing operation of feature extraction processing unit and discrimination processing unit Diagram showing processing target list The figure which shows the absolute value difference and sum of specific expression ID The figure which shows the processing target list after the rearrangement processing The figure which shows the modification of the processing result by the information extraction part between proper expressions The figure which shows the modification of the processing result by the information extraction part between proper expressions The figure which shows the modification of the processing result by the information extraction part between proper expressions

Explanation of symbols

  DESCRIPTION OF SYMBOLS 10 ... Specific expression extraction process part, 11 ... Morphological analysis part, 20 ... Feature extraction process part, 22 ... Abbreviation information extraction part, 30 ... Phrase memory | storage part, 40 ... Discrimination process part, 42 ... Classifier, 60 ... Memory | storage discrimination | determination result Department.

Claims (3)

A device that extracts a set of specific expressions composed of a plurality of related specific expressions from input text,
When a text is input, a specific expression extraction processing unit that morphologically analyzes the input text and extracts a plurality of specific expressions included in the input text;
There is an area corresponding to each of multiple particles in the order of particles associated with noun phrases that are likely to be used as subjects, and noun phrases included in the input text are stacked in the corresponding areas of particles associated with the noun phrases and phrases storage unit for storing in,
The noun phrases are read out in order from the beginning of the input text and stored in the phrase storage unit for each of a plurality of specific expression combinations formed by combining the specific expressions extracted by the specific expression extraction processing unit. When a phrase including one of the corresponding specific expressions that appears later in the input text is read, the number of the corresponding region of which particle is stored in the phrase storage unit. Is a particle associated with a noun phrase that is most likely to be used as a subject among the particles corresponding to the storage area of the phrase storage unit. Information that indicates whether or not there is a combination of a phrase containing each of the specific expressions and a particle corresponding to the area in which the other specific expression is stored in the phrase storage unit It acquires a feature extraction processing unit for extracting at least including identity of each information obtained,
Includes features extracted by the feature extraction processing unit, results determined in advance regarding the presence or absence of a relationship between each specific expression corresponding to a predetermined specific expression set, and each specific expression corresponding to the predetermined specific expression set A discrimination processing unit that discriminates, for each set of specific expressions, whether or not there is a relationship between each specific expression corresponding to the set of specific expressions, based on the prior features extracted from the feature extraction processing unit in advance using text. An apparatus for extracting sets of interrelated expressions characterized by comprising: A determination result storage unit that stores a determination result each time the determination processing unit determines whether or not there is a relationship between each unique expression included in each set of specific expressions,
The discrimination processing unit obtains the discrimination result stored in the discrimination result storage unit when the feature corresponding to the set of specific expressions is extracted by the feature extraction processing unit, and the discrimination result, the extracted feature, and a predetermined feature Extracted in advance from the feature extraction processing unit using the result determined in advance as to whether or not there is a relationship between each specific expression corresponding to the specific expression set and text including each specific expression corresponding to the predetermined specific expression set based on the has been pre-feature, set extraction device named entities interrelated according to claim 1, wherein the determining whether the relationship between the specific representation corresponding to a set of named entities . A method of extracting a set of specific expressions consisting of a plurality of related specific expressions from an input text using a computer,
A first step of extracting a plurality of specific expressions included in the input text by performing morphological analysis on the input text when the text is input;
A noun phrase included in the input text is stored in a noun phrase included in the input text with respect to a predetermined phrase storage unit having a stack format region corresponding to each of a plurality of particles in the order of particles associated with a noun phrase that is likely to be used as a subject. A second step of storing in a corresponding region of the particle associated with the phrase;
For each of a plurality of specific expression combinations obtained by combining the extracted specific expressions, the noun phrase is read in order from the beginning of the input text and the second step is performed. Among them, when reading a phrase including one specific expression that appears later in the input text, information indicating which particle in which corresponding phrase the other specific expression is stored in the phrase storage unit, and the other Indicates whether or not the particle corresponding to the area in which the phrase including the proper expression is stored is the particle associated with the noun phrase most likely to be used as the subject among the particles corresponding to the storage area of the phrase storage unit information and acquires the information other named entity in terms and phrases storage unit containing each of the specific expression combines particle corresponding to the storage area, each obtained A third step of extracting at least containing the identity of the broadcast,
Using the extracted features, the results determined in advance for the existence of a relationship between the specific expressions corresponding to the predetermined specific expressions, and the text including the specific expressions corresponding to the predetermined specific expressions And a fourth step for determining, for each set of specific expressions, whether or not there is a relationship between the specific expressions corresponding to the set of specific expressions, based on the prior features extracted by performing the third step A method for extracting sets of interrelated expressions characterized by

Images