JP5225219B2 - Predicate term structure analysis method, apparatus and program thereof - Google Patents

Description

  The present invention relates to a technique for analyzing a relationship between a predicate and a noun phrase included in a document composed of a plurality of sentences from the viewpoint of case.

<Description of task>
The predicate term structure analysis is a semantic analysis of a sentence described in a natural language. Specifically, a “case” corresponding to a “predicate” such as “what” in the sentence, that is, “a case” such as “who / what”, “wo case” such as “what”, “where” The part (term) corresponding to “dignity” such as “” is specified.

  For example, it is assumed that there is a document 1 “she made curry yesterday and ate it at noon” as shown in FIG. In the first sentence of this document 1, there is a predicate “create (standard form)”. In the predicate term structure analysis, it is specified that the ga case term corresponding to the predicate “create” is “she” and the wo case term is “curry”.

  However, particularly in Japanese sentences, words that should be terms (sometimes there are a plurality of words, hereinafter also referred to as noun phrases including one word) are often omitted. This is called zero pronoun. The second sentence of document 1 is an example.

  The second sentence of document 1 has “eat (standard form)” as a predicate. The term of the ga case corresponding to this predicate is “she”, the term of wo case is “curry”, and the term of d case is “lunch”. However, in the second sentence, only “noon” appears as a noun phrase that should actually be a term. In order to analyze the predicate term structure of such a sentence, the noun phrases “she” and “curry” must be supplemented from the first sentence.

<Predicate structure analysis by conventional technology>
In the prior art, analysis is performed only for a sentence in which both a predicate and its term exist in one sentence. An example of a conventional predicate term structure analyzing apparatus is shown in FIG. 2, and a flow of processing in the conventional predicate term structure analyzing apparatus is shown in FIG.

<Morphological analysis and dependency analysis>
First, the morphological analysis and dependency analysis unit 1 performs a morphological analysis for each sentence on the document input to the control unit 7 and divides it into words, specifies the part of speech of each word, and obtains a word string. Next, similarly, the morpheme analysis / dependence analysis unit 1 divides each sentence into phrases based on the morpheme analysis result (word string) to obtain a phrase string, and further, the dependency structure between phrases (which clause is A clause dependency relationship (destination clause number) is obtained by specifying which clause) (s1). A phrase is a phrase composed of one or more content words (nouns, verbs, adjectives, adverbs, etc.) and zero or more function words (particles, auxiliary verbs, etc.) in Japanese. The morpheme analysis / dependence analysis unit 1 can be configured using an existing morpheme analyzer and dependency analysis.

  At this time, for example, if the input document (object to be analyzed) is the document 1 described above, a result as shown in FIG. 4 is obtained.

<Predicate identification>
Next, the predicate identifying unit 2 identifies (extracts) all predicates of the processing target statement (s2). Specifically, based on the part-of-speech of each word of the processing target sentence, a predicate partial word string is extracted and used as a predicate. For example, “verb”, “adjective”, “sa variable noun immediately followed by verb“ do ””, and “adjective noun immediately followed by auxiliary verb“ da ”” are predicates.

  In the example of document 1, the predicate of the first sentence is “create (standard form)”, and the predicate of the second sentence is “eat”. Hereinafter, processing for the predicate “create” of the first sentence will be described.

<Candidate noun phrase extraction>
Next, the candidate noun phrase extraction unit 3 extracts all noun phrases from the processing target sentence. However, the predicate v being processed at that time in the processing target sentence is excluded. Whether it is a noun phrase is usually determined based on the part of speech. For example, if the part-of-speech at the end of a phrase content word string (content word main word) is one of a noun, a pronoun, or a noun suffix, the contents word string of the phrase is regarded as a noun phrase. Further, the special noun phrase NULL is added to the noun phrase extracted in this way, and these are used as candidate noun phrases (s3).

  This special noun phrase NULL is a noun phrase indicating that there is no case corresponding to the predicate in the term identification described later.

  There are four candidate noun phrases corresponding to the predicate “Make” of the first sentence: “Yesterday”, “She”, “Curry”, and “NULL”.

<Feature selection>
Next, for each candidate n in the candidate noun phrase, the feature selection unit 4 selects a feature from the predicate v, the candidate n, and the morpheme / dependency analysis result (analyzed sentence) to create a feature set (s4 ). As the features, for example, as shown in FIG. 5, a feature related to the predicate v (predicate related), a feature related to the candidate n (candidate related), and a feature related to the relationship between the predicate v and the candidate n (predicate / candidate relative positional relationship) can be considered. .

  As described above, regarding the predicate “make” of the first sentence, when the features of the candidate noun phrases “Yesterday”, “She”, “Curry”, and “NULL” are selected, the result is as shown in FIG.

<Term identification>
Next, by using the selected feature, the term identification unit 5 refers to the term identification model (storage unit) 6 learned in advance for each case, and the candidate n is classified into each case, that is, a case, a case, and a case. Then, the term score for the second case is calculated (s5). The term identification model 6 is learned in advance from the correct data (corpus) of the predicate term structure using the maximum posterior probability estimation method or the like. An example of the term identification model is shown in FIG.

  The term score score corresponding to the predicate v, the candidate n, and the case c is calculated using the following equation (1). This is the same as the score calculation method of the classifier using discriminative models. The higher the score, the more likely the candidate n is as the case c of the predicate v.

However, d _c (n) is a function which is 1 only when the candidate n becomes a term of case c, and 0 otherwise. X is a feature set, f _k (d _c (n) = 1, X) is a feature function, which is 1 only when the feature satisfies the condition of the argument, and 0 otherwise. If it has a real value, the feature value is returned only when the condition of the argument is satisfied. Also,

Is a weight related to the case c of the feature function f _k .

  For example, regarding the predicate v “make” and candidate n “her” in the first sentence, the weight of the feature function of each feature and the partial score

Is calculated as shown in FIG. If the sum is calculated for each case, the item score for the case is 6.568, the item score for the case is 1.967, and the item score for the case is 0.802.

The term identification unit 5 repeats the calculation of term scores for all candidates n. Then, the candidate n _max with the maximum term score is acquired for each case and used as the term of the predicate v (s6). However, if a candidate NULL is acquired, the case is “empty”.

  The term scores of all candidates related to the predicate “Make” in the first sentence are as shown in FIG. 9. As a result, the term “ga” is “she”, the term “wo” is “curry”, and the term “ni” is It becomes “empty” and the predicate term structure is completed.

  The processing described above is repeated for all sentences, and finally, the predicate term structures of all sentences are output as a list from the control unit 7 (s7).

Christopher D. Manning and Hinrich Schutze, "Foundations of Statistical Natural Language Processing", The MIT Press, 1999, pp.217-223

  As described above, in the related art, for a sentence having a predicate and its term in one sentence, the predicate term structure analysis can be performed by using a dependency relation between the predicate and the noun phrase. However, since Japanese pronouns have zero pronouns, there is a case where a term corresponding to a predicate cannot be identified by analyzing only one sentence. For example, in the second sentence of document 1, only “noon” appears as a noun phrase, so “she” is specified as the term of the ga rating corresponding to the predicate “eat” and “curry” is specified as the term of wo I can't.

  One solution is to add the noun phrase of the processed sentence (the sentence before the process target sentence) to the candidate noun phrase as well as the noun phrase of the sentence being processed in the candidate noun phrase extraction step. . However, because there is no direct grammatical dependency between predicates and candidates for different sentences, noun phrases that appear in a sentence different from the predicate are features that use grammatical features such as dependency relationships with predicates. It is very difficult to identify as a term only. For example, when the feature shown in FIG. 5 is used, since the “dependency” feature is empty for any noun phrase that appears in a sentence different from the predicate, a term cannot be identified using only this. .

  In the present invention, the above problem is solved by the following method.

  As the feature, the generation probability in the case c of the candidate n when the predicate v is given is used. For example, when considering the “eating” case, “eating curry” is a reasonable phrase in Japanese, but “eating yesterday” is rarely said in Japanese. In this way, in Japanese, when the predicate and case are determined, the characteristic that a noun phrase that can be taken as a term can be estimated to some extent without looking at the grammatical relationship between the two.

  ・ Similarly, the feature of whether or not it is used as a predicate term in a statement before the processing target statement is used. This takes advantage of the characteristic that noun phrases once used as terms tend to be used repeatedly as zero pronouns.

  In the candidate noun phrase extraction step, not only noun phrases in the processing target sentence but also noun phrases of sentences before the processing target sentence are candidates. However, when all the noun phrases are targeted, if the length of the document becomes long, the number of candidates increases, and the problem that the accuracy of term identification decreases. To avoid this problem, not all noun phrases in all sentences are included in the sentence up to N (N is an integer of 1 or more) previous sentences (hereinafter referred to as the immediately preceding N sentence). Only noun phrases used as terms for the predicates to be used are candidates. This makes it possible to identify terms with high accuracy while limiting the number of candidates to be small.

  According to the present invention, even if a term exists in a sentence different from the sentence in which the predicate being processed exists, it can be specified. Further, by limiting candidate noun phrases to those used as terms for the predicate of the immediately preceding N sentence, it is possible to specify the correct term from a small number of candidates.

An explanatory diagram showing an example of the input document along with its predicate term structure Configuration diagram showing an example of a conventional predicate term structure analysis apparatus Flow chart showing processing in a conventional predicate term structure analyzer Explanatory drawing showing an example of morphological analysis and dependency analysis results Explanatory diagram showing examples of features to be selected Explanatory drawing showing an example of feature set Explanatory drawing showing an example of a term identification model Explanatory drawing which shows an example of the partial score of each case of candidate n Explanatory drawing which shows an example of the item score of each case for every candidate The block diagram which shows an example of embodiment of the predicate term structure analysis apparatus of this invention Flow chart showing an example of language model construction processing Explanatory drawing showing other examples of morphological analysis and dependency analysis results Explanatory drawing showing an example of language model The flowchart which shows the process in the predicate term structure analysis apparatus of this invention Explanatory drawing showing examples of features to be added to the selection target Explanatory drawing which shows an example of the feature set by this invention Explanatory drawing which shows an example of the term identification model in this invention Explanatory drawing which shows an example of the partial score of each case of the candidate n in this invention Explanatory drawing which shows an example of the item score of each case for every candidate in this invention Explanatory drawing which shows the other example of the feature set by this invention Explanatory drawing which shows the other example of the item score of each case for every candidate in this invention

  Hereinafter, the present invention will be described in detail with reference to the illustrated embodiments.

  FIG. 10 shows an example of an embodiment of the predicate term structure analyzing apparatus of the present invention. In the figure, the same components as those of the conventional example are represented by the same reference numerals. That is, 1 is a morphological analysis / dependence analysis unit, 2 is a predicate identification unit, 11 is a term stack, 12 is a candidate noun phrase extraction unit, 13 is a language model (storage unit), 14 is a feature selection unit, and 15 is a term. Identification model (storage unit), 16 is a term identification unit, and 17 is a control unit.

  The term stack 11 stores the term of the predicate processed in the statement before the processing target statement together with the statement number.

  The candidate noun phrase extraction unit 12 extracts all noun phrases from the sentence to be processed, obtains the noun phrase used as the term of the predicate of the immediately preceding N sentence from the term stack 11, combines them, and further adds the special noun phrase NULL Is added as a candidate noun phrase.

  The language model (storage unit) 13 holds the generation probability P (n | c, v) of the noun phrase n when the predicate v and the case c are given.

  For each candidate n in the candidate noun phrase, the feature selection unit 14 uses the predicate v, the candidate n, the morpheme / dependency analysis result (analyzed sentence), the feature related to the predicate v, the feature related to the candidate n, the predicate v and the candidate n. A feature relating to the relation of, a feature whether or not it was used as a predicate term of a sentence before the sentence being processed (forward sentence), and a candidate n for each case c calculated from the language model 13 A language model score for the predicate v is selected to create a feature set.

  The term identification model (storage unit) 15 holds the weight of each case c when the feature is given.

  The term identifying unit 16 calculates the term score when the candidate n is each case using the selected feature and refers to the term identifying model 15 and acquires the candidate with the largest term score for each case. To the term of the predicate v.

  The control unit 17 controls each unit described above to analyze the predicate term structure for all the sentences in the input document, and generates a predicate term structure list including predicates in the sentence and a list of terms corresponding thereto. Output for each sentence.

  In the present embodiment, “ga”, “wo”, and “d” are identified as cases corresponding to the predicate, but other cases may be included.

  First, the construction of the language model 13 used in the present invention will be described, and then the predicate term structure analysis according to the present invention will be described.

<Language model construction>
An example of the flow of processing when constructing a language model is shown in FIG. If the generation probability P (n | c, v) of the noun phrase can be calculated, it may be constructed by another procedure.

  First, collect multiple sentences (generally tens of thousands or hundreds of thousands of sentences). This is called a plaintext corpus. For each sentence in the plaintext corpus, morphological analysis and dependency analysis are performed in the same way as the conventional <morpheme analysis / dependency analysis> to obtain a word string, phrase string, and phrase dependency relation (s11), Further, all predicates of the processing target sentence are identified (extracted) in the same manner as the <predicate identification> in the prior art (s12).

  Next, for each predicate, a clause directly related to the predicate is obtained from the clause dependency relationship (s13), and a noun phrase and a case particle are obtained from the clause (s14). The noun phrase is extracted in the same manner as described in <Candidate Noun Phrase Extraction> in the prior art, and the case particles are specified based on the part of speech and the word notation.

  At this time, if a case particle is not included in the phrase (s15: NO), nothing is done. On the other hand, when a case particle is included in the phrase (s15: YES), the noun phrase, case particle, and predicate are made into a triple, and the frequency (number of appearances) of the triple on the plaintext corpus is counted (s16). ). This is repeated for all sentences in the plaintext corpus.

  For example, suppose that there is a sentence in the plaintext corpus that “There was a person who made curry at noon”. When this is analyzed, a word string, phrase string, and phrase dependency relation as shown in FIG. 12 are obtained. When a predicate is identified from this sentence, “make” and “is” are identified. First, when the clause directly related to the predicate “Make” is taken out, “at noon” and “curry” are obtained. For “noon”, the noun phrase is “noon” and the case particle is “ni”. Therefore, the noun phrase, case particle, and predicate triplet is [create at noon]. If this is done in the same way for the phrase “curry”, the triplet [make curry] is obtained. Such a triple is created for the entire plaintext corpus and the number of appearances is counted.

  Next, the generation probability (trigram probability) of the noun phrase n when the predicate v and the case particle c are determined using the back-off smoothing method (see Non-Patent Document 1) based on the obtained triplet count. ) Is estimated (s17). When the back-off smoothing method is used, the bigram probability P (n | c) can be obtained simultaneously with the trigram probability P (n | c, v).

  Finally, the generation probability P (NULL | c) of the special noun phrase NULL for each case c is added to the bigram probability (s18), and the language model is output (s19). The generation probability of NULL is the probability that various predicates generate NULL for each case c, and when there is correct data with a predicate term structure, there is a probability that no term corresponding to the case corresponding to the predicate exists. . An example of the created language model is shown in FIG.

<Predicate term structure analysis of the present invention>
FIG. 14 shows the flow of processing in the predicate term structure analysis apparatus of FIG. 10, and the details of the configuration and operation of each part will be described below with specific examples.

  Since the object of the present invention is predicate term structure analysis in a Japanese sentence containing zero pronouns, the input is a document composed of a plurality of sentences. In this example, the description will be made using Document 1 “She made curry yesterday.

<Initialize>
First, the term stack 11 is emptied (s21). Then, the following processing is performed for each sentence from the top of the document input to the control unit 7.

<Morphological analysis and dependency analysis>
As in the case of the prior art, the morphological analysis and dependency analysis unit 1 performs morphological analysis for each sentence and divides it into words, specifies the part of speech of each word, and obtains a word string. Next, similarly, the morpheme analysis / dependence analysis unit 1 divides each sentence into phrases based on the morpheme analysis result to obtain a phrase string, and further specifies the dependency structure between phrases to determine the phrase dependency relation. Is obtained (s22). The morphological analysis and dependency analysis results of the document 1 are as shown in FIG.

<Predicate identification>
Next, as in the case of the prior art, the predicate identifying unit 2 identifies (extracts) all predicates of the processing target sentence (s23).

  In the example of document 1, the predicate of the first sentence is “create”, and the predicate of the second sentence is “eat”. Hereinafter, first, a process for the predicate “make” of the first sentence will be described, and a process for the predicate “eating” of the second sentence will be described later.

<Candidate noun phrase extraction>
Next, all the noun phrases are extracted from the sentence to be processed by the candidate noun phrase extracting unit 12 as in the case of the prior art. However, the predicate v being processed at that time in the processing target sentence is excluded. Whether the phrase is a noun phrase is determined based on the part of speech as in the case of the prior art. Also, the noun phrase used as a predicate term of the immediately preceding N sentence (N = 1 in this example) is extracted from the term stack 11 by the candidate noun phrase extraction unit 12 and matched. Further, the special noun phrase NULL is added to the noun phrase extracted in this way, and these are used as candidate noun phrases (s24).

  This special noun phrase NULL means that, in the term identification described later, there is no case corresponding to the predicate or there is no noun phrase in the document that should be a term corresponding to the case (this is called external reference). This is a noun phrase to show.

  Since the noun phrase for the predicate “Make” of the first sentence is only the one obtained from the sentence to be processed because the term stack 11 is empty, “Yesterday”, “She”, “Curry”, and “NULL” There will be four.

<Feature selection>
Next, the feature selection unit 14 selects a feature for each candidate n in the candidate noun phrase to create a feature set (s25). At this time, as in the case of the prior art, predicate v, candidate n, morpheme / dependency analysis result (analyzed sentence), predicate v feature, candidate n feature, predicate v and candidate n feature. In addition to selection, a feature as to whether or not it has been used as a predicate term of a sentence before the sentence being processed (forward sentence) and a feature related to the language model score are selected as shown in FIG.

  Here, regarding the feature of the language model score, referring to the language model 13, when the predicate v is given, the language model score when the candidate n is each case, that is, a case, a case, or a case. Is calculated and used. An example of the language model 13 is as shown in FIG.

  Here, if the trigram P (n | c, v) exists on the language model 13, its logarithmic probability logP (n | c, v) is used as the language model score. If not, the bigram log probability logP (n | c) and the backoff log probability log bo (c, v) are acquired from the language model 13, and the sum of the two is used as the language model score. For example, when calculating the language model score of the predicate v “Make” and candidate n “Yesterday”, there is no trigram in the language model of FIG. Therefore, log bo (ga, make) + logP (yesterday | ga) = − 0.33 + −5.00 is calculated, and −5.33 is set as the language model score.

  As described above, when the feature of each candidate noun phrase regarding the predicate “create” of the first sentence is selected, the result is as shown in FIG.

<Term identification>
Next, the term identification unit 16 refers to the term identification model 15 and uses the selected feature to calculate a term score when the candidate n is each case, that is, a case, a case, and a case. (S26). The term identification model 15 is learned from the correct data (corpus) of the predicate term structure using the maximum posterior probability estimation method or the like. An example of the term identification model in the present invention is shown in FIG.

  The difference from the term identification model in the prior art is that the feature weight related to “Used (Used: 0, Used: 1)” and the feature weight related to “Language Model Score (LM Ga, LM wo, LM Ni)” have been added. That is. Since these are optimized in the learning process, the values of other feature weights are also different.

  Based on these feature weights, the term score score is calculated using the above-described equation (1).

  For example, regarding the predicate v “make” and candidate n “her” in the first sentence, the weight of the feature function of each feature and the partial score

Is calculated as shown in FIG. If the sum is calculated for each case, the item score for the case is 6.031, the item score for the case is 0.391, and the item score for the case is 0.345.

In term identification, the calculation of term scores is repeated for all candidates n. Then, the candidate n _max with the maximum term score is acquired for each case and used as the term of the predicate v (s27). However, if a candidate NULL is acquired, the case is “empty”.

  Since the term scores of all candidates related to the predicate “Make” in the first sentence are as shown in FIG. 19, as a result, the term “ga” is “she”, the term “wo” is “curry”, and the term “d” is It becomes “empty” and the predicate term structure is completed.

  The noun phrases “she” and “curry” recognized as terms are recorded in the term stack 11 together with the sentence number 1 (s28).

<< Process after the second sentence >>
After the second sentence, first, the above-described <morpheme / dependency analysis> and <predicate identification> are performed.

  The morphological analysis and dependency analysis results of the document 1 are as shown in FIG. The second sentence of document 1 has “eat” as a predicate. Hereinafter, the processing of <candidate noun phrase extraction>, <feature selection>, and <term identification> will be described with a focus on the predicate “eat” of the second sentence.

  In <candidate noun phrase extraction>, “noon” is extracted as a noun phrase in the second sentence. Further, from the term stack 11, noun phrases used as terms in the predicate of the immediately preceding N sentence (N = 1) are extracted. The two sentences used in the predicate of the first sentence are “she” and “curry”. Since NULL is also added, there are four candidate noun phrases, “noon”, “her”, “curry”, and “NULL”.

  In <Feature Selection>, features are selected for the above four candidates n using the language model 13 by the same method as described above. Then, a feature set as shown in FIG. 20 is obtained.

  Subsequently, in <term identification>, the term identification model 15 is referred to, and the term score for each case of the candidate n is calculated using the selected feature. In the case of the above feature, the term score of each candidate n is as shown in FIG.

  As a result, for the predicate “eat” in the second sentence, “she” is specified as the term of the ga case, “curry” as the term of the wo case, “noon” as the term of the second case, and the predicate term structure is completed. . In addition, the noun phrases “she”, “curry”, and “noon” that are recognized as terms are recorded in the term stack 11 together with the sentence number 2.

  Since there are no more sentences in the input document, the predicate term structures of all the sentences are finally output as a list from the control unit 17 (s29).

<Other embodiments>
In the above-described embodiment, the noun phrase used as the term of the predicate of the immediately preceding N sentence is added from the term stack, but all the nouns stored in the term stack are not limited to this. A phrase may be used as a target. Further, only the feature related to the language model score may be used as the feature to be added, and the feature as to whether or not it has been used as a predicate term of a statement before the processing target statement may be omitted.

  Further, the present invention installs a program for realizing the functions shown in the configuration diagram of FIG. 10 or a program having the procedure shown in the flowchart of FIG. 14 via a medium or a communication line in a known computer. Is also feasible.

  1: morphological analysis / dependence analysis unit, 2: predicate identification unit, 11: term stack, 12: candidate noun phrase extraction unit, 13: language model (storage unit), 14: feature selection unit, 15: term identification model (Storage unit), 16: term identification unit, 17: control unit.

Claims (7)

A method for identifying a term that is a noun phrase corresponding to a case corresponding to a predicate extracted from each sentence in an input document through morphological analysis and dependency analysis,
The candidate noun phrase extraction unit extracts all noun phrases from the processing target sentence, obtains the noun phrase from the term stack for storing the predicate terms processed in the sentence before the processing target sentence, and combines both, Adding a special noun phrase NULL to make it a candidate noun phrase;
The feature selection unit selects, for each candidate n in the candidate noun phrase, a feature related to the predicate, a feature related to the candidate n, a feature related to the relationship between the predicate v and the candidate n, and a noun when the predicate v and the case c are given. If a language model holding the generation probability of phrase n is used, and a trigram probability exists in the language model, a value obtained from the trigram probability is used as a feature related to the language model score, and the trigram probability is the language model. A value obtained from the bigram probability and the backoff probability if it does not exist as a feature related to the language model score, and creating a feature set;
The term identifying unit calculates the term score when the candidate n is each case by using the selected feature and referring to the term identification model that holds the weight of each case c when the feature is given, Obtaining a candidate with the highest term score for each case and making it a term of the predicate v.
Predicate term structure analysis method characterized by the above. For each candidate n in the candidate noun phrase, the feature selection unit selects a feature related to the predicate, a feature related to the candidate n, and a feature related to the relationship between the predicate v and the candidate n, and as a predicate term of a sentence before the processing target sentence A language model that retains the feature of whether it was used and the generation probability of a noun phrase n when the predicate v and case c are given, and if the trigram probability exists in the language model, the trigram A step of creating a feature set using a value obtained from the probability as a feature related to the language model score, and if a trigram probability does not exist in the language model, a value obtained from the bigram probability and the backoff probability as a feature related to the language model score The predicate term structure analysis method according to claim 1, comprising: The candidate noun phrase extraction unit extracts all noun phrases from the processing target sentence and is used as a predicate term of the immediately preceding N sentence from the term stack that stores predicate terms processed in a sentence before the processing target sentence. 3. The predicate term structure analysis method according to claim 1, further comprising a step of obtaining a noun phrase, combining both, and adding a special noun phrase NULL to make a noun phrase as a candidate noun phrase. A device for identifying a term that is a noun phrase corresponding to a case corresponding to a predicate extracted from each sentence in an input document through morphological analysis and dependency analysis,
A term stack that stores the predicate terms processed in the statement before the processing target statement;
Extracting all noun phrases from the sentence to be processed, obtaining a noun phrase from the term stack, combining both, and adding a special noun phrase NULL to make a candidate noun phrase extracting part,
A language model that retains the generation probability of a noun phrase n given a predicate v and a case c;
For each candidate n in the noun phrase, a feature related to the predicate, a feature related to the candidate n, and a feature related to the relationship between the predicate v and the candidate n are selected, and a trigram probability exists in the language model. If the trigram probability does not exist in the language model, the value obtained from the bigram probability and the backoff probability is assumed as the feature related to the language model score. A feature selector for creating a set;
A term identification model that retains the weight of each case c when given a feature;
Using the selected feature, referring to the term identification model, the term score when the candidate n is each case is calculated, and the term with the largest term score is obtained for each case to obtain the term of the predicate v With an identification unit,
Predicate term structure analysis device characterized by the above. For each candidate n in the candidate noun phrase, a feature related to the predicate, a feature related to the candidate n, and a feature related to the relationship between the predicate v and the candidate n are selected, and whether or not the predicate is used as a term of a predicate before the processing target sentence If a trigram probability exists in the language model using the feature and the language model, the value obtained from the trigram probability is set as a feature related to the language model score, and the trigram probability must exist in the language model. 5. The predicate term structure analysis apparatus according to claim 4, further comprising: a feature selection unit that creates a feature set by using a value obtained from a bigram probability and a backoff probability as a feature related to a language model score . Extract all noun phrases from the sentence to be processed, obtain the noun phrase used as the predicate term of the previous N sentence from the term stack, combine both, and add the special noun phrase NULL to make it a candidate noun phrase The predicate term structure analysis apparatus according to claim 4 or 5, further comprising a candidate noun phrase extraction unit.   The program for functioning a computer as each means of the apparatus in any one of Claims 4 thru | or 6.

Images