JP4321336B2 - Natural language processing system, natural language processing method, and computer program - Google Patents

Description

  The present invention relates to a natural language processing system, a natural language processing method, and a computer program for mathematically handling a natural language used by humans for daily communication, and in particular, assists in proofreading or revising a natural language sentence. The present invention relates to a natural language processing system, a natural language processing method, and a computer program.

  More particularly, the present invention relates to a natural language processing system and a natural language processing system that assist in proofreading or revising word sequences and other unnatural expressions that are often found in sentences written using a language that is not a native language. A language processing system and a computer program, and in particular, a natural language processing system that supports the proofreading or revising of sentences that are unnatural from the viewpoint of a native speaker, although there are no spelling errors or grammatical errors The present invention relates to a natural language processing method and a computer program.

  Words that humans use for everyday communication, such as Japanese and English, are called “natural languages”. Many natural languages have a naturally occurring origin and have evolved with the history of mankind, people and society. Of course, people can communicate with each other by gestures and hand gestures, but natural language can realize the most natural and advanced communication.

  Natural language is inherently abstract and highly ambiguous, but it can perform computer processing by mathematically handling sentences. As a result, various applications / services related to natural language are realized by automated processing such as machine translation, dialogue system, search system, question answering system, and sentence proofing system.

  For example, when writing a sentence in a language that is not your native language, even if you pay the utmost care, avoid writing sentences that are inappropriately expressed in terms of the language as the native language. I can't. That is, it often happens that the user writes a misspelled word, commits a grammatical error, or writes a sentence that is grammatically correct but unnatural. In such a case, it is possible to correct a sentence by detecting a spelling error or grammatical error using a sentence proofreading system. Spell checkers and syntax checkers have already been put into practical use.

  Basically, grammatical errors contained in sentences can be detected by performing a parsing process. Parsing is a process of specifying a dependency relationship between words based on grammatical rules.

  For example, in addition to correct phrase structure rules, grammatical rule tables include information about error sentences that contain error phrase structures, so that sentences containing grammatical errors are parsed in the same way as correct sentences that do not contain errors. (For example, see Patent Document 1).

  In addition, it has a correction sentence generation dictionary and an information stack that stores information about errors in the learner's input sentence. It detects grammatically incorrect phrases from the learner's input sentence and corrects the erroneous phrases. Thus, the learner's input sentence can be corrected to a grammatically correct sentence (see, for example, Patent Document 2). For example, when a sentence “I want to room.” Is input, it can be correctly corrected to “I want two room”.

JP-A-2-226364 JP-A-5-204299

  As mentioned above, when writing in a language that is not your native language, even if you pay the most attention, you will write a sentence that is inappropriate when viewed from your native language. There is a problem that this is inevitable.

  Even for non-native languages, it is relatively easy for a writer who has mastered it at a certain level to avoid such grammatical errors or carelessness. Furthermore, according to the sentence proofreading system as described above, it is possible to reduce errors related to word level spelling and to avoid writing sentences that are clearly grammatically incorrect at the syntax level.

  For example, it is possible to write rules that identify non-grammatical sentences according to the criteria of whether or not they are grammatical. In fact, by accumulating such rules, a grammar checker can be realized. (For example, see Patent Document 1 and Patent Document 2).

  On the other hand, it is extremely difficult for non-native speakers to avoid writing unnatural sentences from the viewpoint of native speakers even though there are no grammatical errors. Moreover, it is difficult to identify an unnatural sentence that is grammatically correct but unnatural in a natural language processing system. This is because it is difficult to declaratively describe whether or not the native speaker feels natural as a rule. Further, even if it can be determined whether or not the sentence is natural to the native speaker, it is more difficult for the system to present a more natural sentence to the writer.

  For example, the following two English sentences (1) and (2) have almost the same meaning, and both are natural English sentences.

(1) It is unlikely that John will be selected.
(2) It is improbable that John will be selected.

  On the other hand, the following sentence (3) is a natural sentence for a speaker whose native language is English, whereas sentence (4) is an unnatural sentence.

(3) John is unlikely to be elected.
(4) John is improbable to be elected.

  It is quite possible for a writer whose English is not his native language to write the English sentence of (4) because both (1), (2) and (3) are natural sentences. However, sentence (4) is unnatural in terms of English as a native language, and it is difficult for the system to point out that (4) should be corrected to (2). .

  The present invention has been made in view of the technical problems as described above, and its main purpose is to provide an excellent natural language processing system and natural language processing capable of suitably supporting the proofreading or revising of natural language sentences. It is to provide a method and a computer program.

  A further object of the present invention is that it can favorably assist in the proofreading or revising of word sequences and other unnatural expressions that are often found in sentences written using a language other than the native language. A natural language processing system, a natural language processing method, and a computer program.

  A further object of the present invention is that it is difficult to say that it contains clear grammatical errors. Therefore, a sentence that cannot be handled by conventional technology that supports proofreading of sentences by accumulating rules that identify grammatical errors, An object of the present invention is to provide an excellent natural language processing system, natural language processing method, and computer program that can be automatically corrected into a more natural sentence.

  The present invention is a natural language processing system that supports the proofreading or revising of sentences that are unnatural from the viewpoint of a native language speaker, although there are no misspellings or grammatical errors. FIG. Is shown schematically. That is, the natural language processing system according to the present invention includes a semantic analysis unit that performs semantic analysis on a natural language sentence to be proofread, an analysis result holding unit that holds a semantic analysis result, and a plurality of natural languages based on the semantic analysis result. A generation unit that generates a sentence; a generation result holding unit that holds a generation result; and a selection unit that selects the most natural sentence from a plurality of natural language sentences held in the generation result holding unit. .

  Here, in the semantic analysis for performing the semantic analysis on the natural language sentence to be proofread and the process for generating the natural language sentence from the semantic analysis result, for example, a syntax based on Lexical Functional Grammar (LFG: Vocabulary Functional Grammar) theory A semantic analysis processing system can be applied.

  In the LFG theory, a natural language sentence is analyzed, and the semantic content of the sentence is expressed as an analysis result by a nested structure (matrix structure) of attribute-attribute value pairs called a function structure (f-structure: functional structure). f-structure is an abstraction of the semantic content of the input sentence. The greatest feature of LFG is that f-structures having the same structure can be output if the meanings expressed by sentences are the same in any language. For details of language universality of f-structure, see, for example, Dalymplle, M. et al. "Syntax and Semantics: Lexical Functional Grammar" (Academic Press (2001)).

  For example, it is possible to realize a machine translation technique in which f-structure is regarded as an intermediate language by utilizing the characteristics of f-structure universality. That is, by using f-structure as an intermediate language, a mutual machine translation system between a plurality of languages can be realized as shown in FIG. In this case, first, a natural language sentence described in the translation source language A is analyzed based on the LFG theory, and an f-structure of the language A is obtained as an analysis result. Next, the f-structure of the language A is converted into the f-structure of the translation destination language B. This conversion process can be said to be a relatively easy process in consideration of the language universality of f-structure. Finally, the obtained f-structure of language B is converted into a natural language sentence of language B.

  As for a machine translation method in which f-structure is regarded as an intermediate language, for example, “From Parallel Grammar Development towns Machine Translation” (In Proceedings of MT Supra. I want to be. Also, for conversion from f-structure to natural language sentences, for example, “LFG generation products context-free encoding in the Proceedings of the World,” by Lonald Kaplan and Jurgen Wedekind. )) Is a detailed explanation. Also, as a system that realizes conversion from natural language sentences to f-structure and vice versa, conversion from f-structure to natural language sentences, John Maxwell and Ronald Kaplan co-authored “A Method for Distinct Constitutive Constrate In Current Issues in Parsing Technology, pp. 173-190, Kluwer (1991)) and John Maxwell, Ronald Kaplan, "The interface between franss and 4". An LFG system called XLE, which is described in detail in 1993)).

  In the present invention, the universal nature of the f-structure, which does not depend on the difference in the surface description of natural language sentences, is used not for machine translation but for sentence proofreading support. That is, as shown in FIG. 2B, the sentence to be proofread is converted into f-structure by analyzing it based on the LFG theory, and then the f-structure is further converted back into a natural language sentence of the same language. To do. Since f-structure is an abstract structure, the meaning of the sentence is preserved, but the expression format is not preserved. What is called “the meaning of a sentence” here is information generally called a predicate / term structure, which is information such as what is a predicate of a sentence and what is its subject or object. Therefore, by performing the above processing, a plurality of sentences having the same semantic content but different expressions are generated.

  By performing conversion from the natural language sentence to f-structure and reverse conversion from f-structure to the natural language sentence within the same language, for example, the above-mentioned English sentence (4) is used as an input to the English sentence. (4) and English sentence (2) can be obtained as output results. What is important here is that it is extremely difficult to determine whether or not the sentence is unnatural by looking only at the English sentence (4), whereas (2) and (4) If two English sentences are given, selecting a more natural sentence from the two sentences is a relatively easy process.

  Also, by converting from natural language sentences to f-structure and reverse conversion from f-structure to natural language sentences within the same language, the meaning and content of the sentences are the same, but the expressions are different. When the sentence is generated as a candidate, for example, by comparing with a corpus, a sentence with a more natural expression can be selected from sentences having the same semantic content.

  Here, when comparing a plurality of candidate sentences with the corpus, the original sentences can be evaluated based on the result of the syntax analysis instead of directly comparing the original sentences. Here, when the evaluation is performed using the semantic analysis result, only the meaningful part of the sentence is extracted, and the difference in the surface description is discarded. On the other hand, according to the result of the syntax analysis, since the surface information is left, it is possible to appropriately determine the naturalness of the sentence expression such as the arrangement of words.

  The second aspect of the present invention is a computer program written in a computer-readable format so as to execute a process for supporting the proofreading of a sentence on a computer system. A semantic analysis step for performing semantic analysis on a language sentence; a semantic analysis result holding step for holding a semantic analysis result; and a sentence generation step for generating one or more natural language sentences from the semantic analysis result. A computer program that supports proofreading of a natural language sentence to be proofread based on a generated natural language sentence.

  The computer program according to the second aspect of the present invention defines a computer program described in a computer-readable format so as to realize predetermined processing on a computer system. In other words, by installing the computer program according to the second aspect of the present invention in the computer system, a cooperative action is exhibited on the computer system, and the natural language according to the first aspect of the present invention. The same effects as the processing system can be obtained.

According to the natural language processing system of the present invention, it is difficult to say that it contains clear grammatical errors, and therefore cannot be handled by conventional techniques that support sentence proofreading by accumulating rules that identify grammatical errors. Such a sentence can be automatically corrected to a more natural sentence, and a proofreading result that is semantically different from the original input sentence is not output.

  The natural language processing system according to the present invention enables proofreading to natural expressions that could not be realized by the prior art, and is also used for further refining the output obtained from the machine translation system. be able to.

  Other objects, features, and advantages of the present invention will become apparent from more detailed description based on embodiments of the present invention described later and the accompanying drawings.

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

  In FIG. 3, the structure of the text proofreading system 10 which concerns on one Embodiment of this invention is shown typically. As shown in the figure, the sentence proofreading system 10 includes a proofreading object input part 11, a semantic analysis part 12, a semantic analysis result holding part 13, a sentence generation part 14, a generation result holding part 15, and a corpus holding. A unit 16, a syntax analysis unit 17, and a selection unit 18 are provided. The illustrated text proofing system 10 can be configured as a dedicated hardware device, but can also be realized by introducing a text proofreading application into a general computer system such as a personal computer.

  The proofreading target sentence input unit 11 includes a user interface that receives an English sentence to be proofread from a user.

  The semantic analysis unit 12 performs a semantic analysis process based on LFG on the English sentence received by the proofreading target sentence input unit 11 to obtain a corresponding f-structure.

  Here, semantic analysis processing based on LFG will be described. FIG. 4 schematically shows a functional configuration of the syntactic and semantic analysis processing system 200 based on the LFG grammar theory.

  The morpheme analysis unit 202 has a morpheme rule 202A and a morpheme dictionary 202B relating to a specific language such as Japanese, and performs a part-of-speech recognition process by segmenting an input sentence into morphemes that are the smallest semantic units. However, the proofreading target sentence input unit 11 can omit the morpheme analyzing unit 202 when the morphological analysis result of the English sentence to be proofread can be input. As the morphological analysis system, for example, a Japanese morphological analysis system such as “Chasen” can be applied, but the gist of the present invention is not limited to this. For example, Yuji Matsumoto, Kei Kitauchi, Tatsuo Yamashita, Yoshitaka Hirano, Hiroshi Matsuda, Kazuma Takaoka, and Masayuki Asahara, “Corporation version 2.2.1 Instruction Manual for Japanese Morphological Analysis System” (Nara Institute of Science and Technology, 2000).

  The morpheme analysis result is then input to the syntax and semantic analysis unit 203. The syntactic and semantic analysis unit 203 has dictionaries such as a grammar rule 203A and a case frame dictionary 203B. Based on the analysis of the phrase structure based on the grammar rule, the meaning of words in the sentence, and the semantic relationship between words. Analyzes the semantic structure expressing the meaning conveyed by the sentence. The case frame dictionary describes the relationship between verbs and other components in the sentence such as the subject, and the predicate and the semantic relationship between the words can be extracted.

  As a result of the syntax analysis, “c-structure (constituent structure :)” representing a phrase structure of a sentence composed of words, morphemes and the like as a tree structure is obtained, and further based on the case structure of the subject and the object. “F-structure (functional structure)” is output as a result of semantically and functionally analyzing the input sentence such as a question sentence, past tense, and polite sentence.

  c-structure represents the structure of words and phrases in a sentence in a tree structure format, and is defined by a syntax category. For example, phonological interpretation for generating a phoneme string can be performed based on c-structure. On the other hand, f-structure clearly expresses a grammatical function, and includes a grammatical function name, a semantic form, and a feature symbol. By referring to such f-structure, it is possible to obtain an understanding of the meaning of a subject, a subject, an complement, a modifier, and an adjunct. The f-structure is a set of features attached to each node of the c-structure, and is expressed in the form of an attribute-attribute value matrix. In f-structure, only the part having the meaning of the sentence, which does not depend on the difference in the surface description of the natural language sentence, is extracted, and the universal property of the language can be described. On the other hand, in c-structure, the difference in the surface description is left, and elements for judging the naturalness of the sentence expression such as the arrangement of words are included.

  For details of LFG, see, for example, R.A. M.M. Kaplan and J.H. Bresnan co-author of the paper. "Lexical-Functional Grammar: A Formal System for Grammatical Representation" (The MIT Press, Cambridge (1982) Reprinted in Formal Issues in Lexical-Functional Grammar, pp.29-130.CSLI publications, Stanford University (1995 ).) Etc.

  FIG. 5 shows the f-structure output from the semantic analysis unit 12 when the proofreading target sentence input unit 11 receives the following English sentence (5).

(5) Food quick spoils in the summer. (Food tends to rot in summer.)

  In f-structure, it is expressed in the form of an attribute-attribute value matrix, the left side in [] is the name of the feature (attribute), and the right side is the feature value (attribute value). In the example shown in FIG. 5, the main word (predicate, PRED (icate)) of English (5) is “soil”, the subject (SUBJ (ect)) is “food”, and the tense is “present (PRES (ent )) ", And further," quickly "and" in the summer "have modified" soil "as optional modifying elements (ADJUNCT). If the semantic analysis by the semantic analysis unit 12 fails, the user is notified that there is a grammatical defect in the input sentence.

  The semantic analysis result holding unit 13 holds the f-structure as a semantic analysis result based on the LFG theory obtained from the semantic analysis unit 12 in a computer constituting the sentence proofreading system 10.

  The sentence generation unit 14 obtains one or more corresponding English sentences by applying an English sentence generation process based on LFG to the f-structure held in the semantic analysis result holding unit 13. For details of the natural language sentence generation processing based on LFG, the above-mentioned Ronald Kaplan and Jurgen Wedekind's "LFG generation products context-free enumeration of the first 3 of the ensembles". Refer to).

  An example of a plurality of English sentences generated using the f-structure shown in FIG. 5 as an input will be given below. Here, it is assumed that one of the plurality of sentences generated by the sentence generation unit 14 is the same as the input English sentence (5).

(5) Food quick spoils in the summer.
(6) Food spoons quickly in the summer.
(7) Food spoons in the summer quickly.
(8) Quickly food spoils in the summer.
(9) In the summer food quick spoils.
(10) In the summer food spoils quickly.
...

  The generation result holding unit 15 holds each English sentence generated by the sentence generation unit 14 inside a computer constituting the sentence proofreading system 10.

  The corpus holding unit 16 is a language material that collects texts and utterances on a large scale or exhaustively. In the present embodiment, an English sentence written by a speaker whose native language is English is stored as a language material in a computer constituting the sentence proofreading system 10.

  The syntax analysis unit 17 performs a syntax analysis process on the English sentences held in the generation result holding unit 15 and the corpus holding unit 16. Parsing is a process of analyzing a sentence structure such as a phrase structure based on grammatical rules. Since the grammatical rule is a tree structure, by synthesizing the sentence structure as a tree structure in which individual morphemes are joined based on dependency relationships, etc. The result is expressed. A structure obtained as a result of parsing is generally called a “syntax tree”.

  In this embodiment, syntax analysis based on LFG grammar theory is performed, and c-structure that represents a phrase structure of a sentence composed of words, morphemes, and the like as a tree structure is handled as a syntax tree (described above). 6 to 8 show syntax trees corresponding to the above-mentioned English sentences (5) and (6) and the following English sentence (11).

(11) Raw food spoof quickly in the last summer.

  The selection unit 18 obtains the syntax tree corresponding to the English sentence held in the corpus holding unit 16 and the syntax tree corresponding to the English sentence held in the generation result holding unit 15 obtained from the syntax analysis unit 17. And a natural English sentence is selected from the English sentences held in the generation result holding means 15. Specifically, among the syntax trees corresponding to the English sentences held in the generation result holding unit 15, those having many similar sentences in the sentences held in the corpus holding unit 16 are more natural. To select as an English sentence. The similarity of the syntax trees can be calculated based on the calculation of distance between syntax trees (well-known).

  The selection unit 18 performs the selection process based on the sentence syntax analysis result, because the semantic analysis result extracts only the meaningful part of the sentence and discards the difference in the surface description. This is because, according to the syntax analysis result, these pieces of information are still left, and the naturalness of the sentence expression such as the arrangement of words can be suitably determined.

  FIG. 9 shows an English sentence selection algorithm executed in the selection unit 18 in the form of a flowchart.

Here, a set of syntax trees corresponding to the English sentences generated by the sentence generation unit 14 is X, and a set of syntax trees corresponding to the English sentences in the corpus holding unit 16 is Y. Further, among the elements of X, a syntax tree corresponding to a sentence (that is, the original sentence) input to the proofreading sentence input unit 11 is set to x _k (step S1).

The number of elements of the set X of syntax tree as m, if m is 1 (step S2), and shown as it is a user calibration target sentence corresponding to x _k (step S6).

On the other hand, when the number m of elements in the syntax tree set X is 2 or more (step S2), x of the corpus syntax tree set Y for each element x _n (1 ≦ n ≦ m) of X. _The number of elements N (x _n ) including _n is calculated (step S3).

  Here, “inclusive” of one syntax tree means that the syntax trees are similar to each other, for example, whether or not to include based on distance calculation (known) between the tree structures. be able to. A prominent example in which one syntax tree includes the other syntax tree is when the other syntax tree is completely included in one syntax tree or when the syntax trees match. For example, since the syntax tree shown in FIG. 8 has a structure that completely includes the syntax tree shown in FIG. 7 (see FIG. 10), the syntax tree shown in FIG. 8 is shown in FIG. It is determined that the syntax tree is included. However, the tense of verbs is the same even if they are different.

The number of elements N (x _n ) of the corpus syntax tree set Y calculated in step S3 basically corresponds to the number of sentences in the corpus that are similar in expression to the English sentence being processed. . That is, if the number of elements N (x _n ) is large, expressions similar to the original English sentence of the parse tree x _n are often seen in English sentences written by speakers whose native language is English, It is estimated that there is.

In step S4, N (x _n ) / N (x _k )> T _a and N (x _n )> T _b as the criteria for determining the number of elements N (x _n ), that is, the expression is similar to the generated English sentence the number of English sentences in the corpus compared by original input sentence and the threshold T _a which further determines the number itself English sentences in corpus representation is similar in certain threshold T _b.

Then, N satisfies the _{(x n) / N (x} k)> T a, and, N (x _n)> T if _b exists x _n satisfying, the x _n a sentence corresponding to N (x _n ) Are output as calibration results in descending order (step S5).

Note that T _a and T _b are thresholds set in advance, T _a is _a positive real number greater than 1, and T _b is a natural number greater than 1. If there is no x _n satisfying the above constraint, the proofreading sentence corresponding to x _k is shown to the user as it is (step S6).

  In the above algorithm, basically, a sentence set in the corpus holding unit 16 in which many sentences having a similar syntax structure exist is selected as a proofreading result. In the present embodiment, the selection is made on the object whose syntax tree is completely inclusive. For example, co-authored by Tetsuro Takahashi, Kentaro Inui, and Yuji Matsumoto, “Evaluation Method of Syntactic Similarity of Text” (Information Processing Society of Japan Research Report, 2002-NL-150, pp.163-170 (2002)). By using various syntax tree comparison methods, a more robust system can be realized.

  Alternatively, the most natural English sentence can be selected by using a machine learning method for the selection process based on the comparison of the syntax trees. In this case, learning for selecting a natural sentence from the corpus means is performed by a machine learning method, and a natural sentence is selected based on the learning result. In machine learning, rules for explaining the characteristics of data can be output by using learning data as input and using a statistical processing method.

  Examples of machine learning methods include Support Vector Machine and Maximum Entropy. For example, Support Vector Machine is one of the non-parametric pattern classifiers that performs linear discrimination using the separation hyperplane obtained as the optimal solution for learning, and learns when it is inappropriate to linearly separate learning materials. The material is nonlinearly mapped from the original pattern space to a higher-order pattern space, and a separation hyperplane is constructed in the high-dimensional space to perform linear discrimination. For details of these machine learning methods, “Machine Learning in Automated Text Categorization” by Fabrizio Sebastani (ACM Computing Surveys Vol. 34, No. 1, pp. 1-47, 2002) and the literature are cited. Please refer.

  With the above configuration, the English sentence input to the proofreading sentence input unit 11 is converted into another expression having the same semantic content and no grammatical error, and then the most natural sentence is selected from them. be able to. As a result, it is difficult to say that there is a clear grammatical error. Therefore, a sentence that cannot be handled by conventional technology that supports proofreading of sentences by accumulating rules that identify grammatical errors is more natural. Can be automatically corrected, and a calibration result that is semantically different from the original input sentence is not output.

  In the case of the above English sentence (5), the adverb “quickly” existing between the subject and the verb cannot be said to be a grammatical error, but this is what makes English a native language. Writing rare sentences is rare. For this reason, it can be expected that a more natural English sentence (6) is selected as an English sentence by the above-described method, by comparison with an English sentence written by a speaker whose native language is English in the corpus. In the case of an English sentence, it is natural that ADJUNCT representing time such as “in the summer” is located at the end of the sentence, and the English sentence (6) is also selected from this point.

  Similarly, FIG. 11 shows f-structure obtained by semantic analysis processing based on LFG in the semantic analysis unit 12 when the input English sentence is (12) below.

(12) Ann faxed the news to him.

  In this case, the sentence generation unit 14 applies an English sentence generation process based on LFG to the f-structure, and as a corresponding English sentence, together with the original English sentence (12) input, the following English sentence ( 13) is generated.

(13) Ann faxed him the news.

  In this case, it cannot be said that the English sentence (12) is a grammatically incorrect sentence, but if the indirect object is a pronoun, the sentence in which the indirect object is located immediately after the verb is natural. Therefore, it can be expected that the English sentence (13) is selected as the proofreading result by comparing the syntax tree with the English sentence written by the speaker whose native language is English in the corpus.

  FIG. 12 shows f-structure obtained by semantic analysis processing based on LFG in the semantic analysis unit 12 when the input English sentence is (14) below.

(14) John gave a hard time to Bob.

  In this case, the sentence generation unit 14 applies an English sentence generation process based on LFG to the f-structure, and as a corresponding English sentence, together with the input original English sentence (14), the following English sentence ( 15) is generated.

(15) John gave Bob a hard time.

  Again, the English sentence (14) is a grammatically correct sentence but an unnatural sentence, and by comparing the syntax tree with an English sentence written by a speaker whose native language is English in the corpus, The English sentence (15) is selected as the proofreading result.

  Further, when the above-described English sentence (4) and English sentences (16) and (17) are input, f-structure obtained by the semantic analysis processing based on the LFG in the semantic analysis unit 12 is shown in FIGS. Respectively.

(4) John is improbable to be elected.
(16) Vote for you.
(17) He is the God that is in Jerusalem.

  From the f-structure shown in FIG. 13, the English sentence (2) is displayed together with the English sentence (4), and from the f-structure shown in FIG. From the f-structure shown in FIG. 15, the English text (17) and (19) and (20) are generated by performing the English text generation processing based on the LFG for the f-structure, respectively. Then, by comparing the syntax tree with an English sentence written by a speaker whose native language is English in the corpus, English sentences (2), (18), and (19) are selected as the most natural proofreading results, respectively. Can be expected.

(2) It is impableable that John will be selected.
(18) Vote for yourself.
(19) He is the God who is in Jerusalem.
(20) He is the God who is in Jerusalem.

  In the embodiment of the present invention described above, it is assumed that the corpus holding unit 16 holds a sufficient number of English sentences for selecting the calibration result. On the other hand, when the number of English sentences held is small, a thesaurus dictionary is used to preferentially extract example sentences including similar words of words included in the English sentence to be proofread from the corpus holding unit 16. (See FIG. 16). In this case, the selection unit 18 can select a sentence after replacing a word in the sentence with a higher concept word. For example, the above selection algorithm by the selection unit 18 may be executed after replacing “John” with a mark such as “person name”.

Further, in the English sentence selection algorithm to be executed in the selector 18, the number of English sentences in the corpus representing the generated English sentence is similar compared with the case with the threshold T _a of the original input sentence, further, the expression By judging the number of English sentences in the corpus having a similar value with a predetermined threshold value T _b , the naturalness of expression for each generated English sentence is judged or evaluated (described above). Here, if the statement sentences stored in corpus holder 16 can be any sufficiently reliable, it is possible to set a threshold value of T _a and T _b to a small (close to 1) value. For example, when using the sentence proofreading support system of the present invention for proofreading an English sentence used in an English paper, only the English sentence actually used in a paper in the same field is held in the corpus holding unit 16. . Even if the number of matches in the syntax tree is small, the selection result by the selection unit 18 can be adopted as a sufficiently reliable result.

  In the above-described embodiment, the corpus holding unit 16 holds many natural language sentences with natural expressions or natural language sentences written by native speakers as language materials. Naturally, it is natural for the native speaker by selecting a sentence whose structure trees are similar (inclusive) by comparing the sentence and the sentence generated based on the semantic analysis result of the sentence to be proofread. The sentence which becomes becomes is selected. On the other hand, the corpus holding unit 16 holds a natural language sentence in which the expression of the sentence is not natural, or a natural language sentence determined not to be natural by the native speaker, and the sentence in the corpus is not similar to the structure tree. By selecting a sentence, it is possible to select a sentence that is natural for the native speaker. Or, by combining these methods, it is possible to select a sentence that is more natural as viewed from the native speaker.

  If the syntax structures are equal, the semantic analysis result (f-structure) obtained by further abstracting the syntax structures is basically the same structure. The sentence set in the corpus holding unit 16 has a f-structure equal to the f-structure held in the semantic analysis result holding unit 13 by applying the semantic analysis result in addition to the syntax analysis result in advance. The sentence may be searched from the corpus holding unit 16, and then the above-described syntax tree comparison algorithm may be performed only on the search result (see FIG. 17). As a result, the processing efficiency can be improved.

  In the above-described embodiment of the present invention, the description has been made on the assumption that the semantic analysis unit 12 uniquely determines the semantic analysis result (f-structure) for one proofreading target sentence. May have multiple semantic analysis results for one sentence. Even in such a case, it is possible to perform sentence proofing processing based on the above-described sentence generation and syntax tree comparison for each semantic analysis result.

  Alternatively, it is also possible to perform sentence proofing processing based on the above-described sentence generation and syntax tree comparison after narrowing down the semantic analysis result to one (see FIG. 18). For narrowing down the semantic analysis results, for example, Support Vector Machine and other machine learning methods can be applied. Regarding narrowing down of the semantic analysis results, for example, Yoshimura, Masuichi, Okuma, Sugihara, “Selection of f-structure based on Support Vector Machine” (Information Processing Society of Japan Research Report 2003-NL-158, pp. 75-80 (2003) ) ".

  FIG. 19 shows a configuration example of a sentence proofreading system 10 according to another embodiment of the present invention. The system 10 shown in the figure is a meaning for changing the f-structure, which is the result of semantic analysis of the proofreading sentence obtained by the semantic analysis unit 12, based on user input via a user interface or other automated processing. An analysis result changing unit 19 is provided.

  The semantic analysis result changing unit 19 can delete or change the designated attribute and its attribute value in the f-structure. Thereby, it is possible to change the generated natural language sentence such as increasing or decreasing the generated English sentence according to the user's request. For example, when a sentence in which a sentence is written using a language other than the native language is targeted for proofreading and there is no confidence in the tense in the sentence, the tense attribute and the attribute value are deleted from the f-structure. As a result, the sentence generation unit 14 can generate an English sentence without being limited to the tense, and can increase the range of selection by the selection unit 18.

  For example, FIG. 20 shows an f-structure output by the semantic analysis unit 12 based on the LFG theory with the following English sentence (21) as an input.

(21) He is owning a house.

  Here, the semantic analysis result changing unit 19 presents a prepared question pattern to the user, for example, “Do you want to use progressive information when generating proofreading candidates? (Yes / No)”. , F-structure is changed. Then, when the user replies “does not use (progressive information)” (not sure whether the sentence (21) is natural about the tense) or not, the sentence from the f-structure in FIG. After deleting the information that (21) is the progressive form (that is, the part of “ASPECT PROG (ress)”), the process proceeds. As a result, a sentence such as the following English sentence (22) is generated.

(22) He owns a house.

  Actually, “own” is rarely used in the progressive form in the English sentence, and the English sentence (22) is adopted as the proofreading result.

  In each of the above-described embodiments, the sentence proofreading system generates a sentence using all information of f-structure corresponding to the input sentence. Here, since the amount of calculation increases exponentially as the number of morphemes of the sentence increases, the calculation cost may be a problem when a long sentence is targeted. As a solution to this, it is conceivable to appropriately divide the input sentence and perform the same processing as described above for each divided element. For example, the input sentence may be divided based on the morphological analysis result, and the same processing may be performed for each divided element. Alternatively, the same processing can be performed for each part of the f-structure.

  For example, after converting to a morpheme string with part-of-speech information by morpheme analysis processing, predetermined division rules are sequentially applied in order of priority, division positions are determined using the matching division rules, and the input sentence is divided Can be divided into elements before and after. By performing the semantic analysis process for each element divided in this way, a result similar to that obtained by dividing the semantic analysis result can be obtained.

  FIG. 21 shows an example of the division rule. The division rule is described as morpheme information of one morpheme or two or more continuous morphemes for determining a division position in the input sentence. In the example shown in the figure, as the division rule A with priority 1, the location where the reading mark is connected to the connection particle A is determined as the division position. Further, as a division rule B with a priority of 2, a portion where a reading point is connected to the connection particle B class is determined as a division position. Further, as a division rule C of priority 3, a place where a reading point is connected to a continuous form after use is decided as a division position. Further, as a division rule D having a priority level 4, a position where a reading mark is connected to a case particle is determined as a division position. Further, as a division rule E of priority 5, a place where a noun, an auxiliary particle and a reading point continue is determined as a division position. Further, as a division rule F having a priority level 6, a position where a reading point is connected to an adverb noun can be determined as a division position. Further, as a division rule G with a priority of 7, a place where a reading point appears is decided as a division position. In addition, as a division rule A ′ having a priority of 8, a place where a connection particle A class appears is determined as a division position. In addition, as a division rule B ′ having a priority of 9, a location where a connection particle B class appears is determined as a division position. In addition, as a division rule C ′ having a priority of 10, a portion where a continuous form after use appears is determined as a division position. In addition, as a division rule D ′ having a priority of 11, a location where a case particle appears is determined as a division position. In addition, as a division rule E ′ having a priority of 12, a part where a coordinator appears is determined as a division position. In addition, as a division rule F ′ having a priority of 13, a location where an adverb noun appears appears as a division position.

  For example, when the outermost matrix structure of f-structure has a parallel structure, that is, when the sentence is a heavy sentence, generation processing is performed for each f-structure corresponding to a single sentence. A corresponding sentence may be generated. It is also possible to delete the nested ADJUNCT (arbitrary modifier) and perform the generation process using only the main components of the sentence.

  The present invention has been described in detail above with reference to specific embodiments. However, it is obvious that those skilled in the art can make modifications and substitutions of the embodiment without departing from the gist of the present invention.

  In this specification, the semantic analysis means based on the LFG (Lexical Functional Grammer) theory has been specifically described in order to obtain the semantic analysis result for the input natural language sentence, but the gist of the present invention is not limited thereto. It is obvious that the same effect can be expected by converting the structure of a natural language sentence into an abstract semantic expression and then using a theory or technique that can reversely convert the semantic expression into a natural language sentence. As an example of such a theory other than LFG, C.I. J. et al. The case grammar proposed by Fillmore can be mentioned.

  Also, in this specification, the proofreading support system according to the present invention has been described for English sentences. However, even in languages other than English, it is possible to generate sentences from semantic analysis and semantic analysis results. Obviously, the same effect can be obtained.

  In short, the present invention has been disclosed in the form of exemplification, and the description of the present specification should not be interpreted in a limited manner. In order to determine the gist of the present invention, the claims section described at the beginning should be considered.

FIG. 1 is a diagram schematically showing a functional configuration of a natural language processing system according to the present invention. FIG. 2 is a diagram for explaining a system that uses the language universal nature of f-structure, which does not depend on the difference in surface description of natural language sentences, for machine translation and sentence proofreading support. FIG. 3 is a diagram schematically showing the configuration of the text composition system 10 according to an embodiment of the present invention. FIG. 4 is a diagram schematically showing a functional configuration of the syntax and semantic analysis processing system 200 based on the LFG grammar theory. FIG. 5 is a diagram illustrating a configuration example of f-structure output by the semantic analysis unit 12. FIG. 6 is a diagram illustrating a configuration example of a syntax tree. FIG. 7 is a diagram illustrating a configuration example of a syntax tree. FIG. 8 is a diagram illustrating a configuration example of a syntax tree. FIG. 9 is a flowchart showing an English sentence selection algorithm executed in the selection unit 18. FIG. 10 is a diagram for explaining that the syntax tree shown in FIG. 8 has a structure that completely includes the syntax tree shown in FIG. FIG. 11 is a diagram illustrating a configuration example of f-structure output by the semantic analysis unit 12. FIG. 12 is a diagram illustrating a configuration example of f-structure output by the semantic analysis unit 12. FIG. 13 is a diagram illustrating a configuration example of f-structure output from the semantic analysis unit 12. FIG. 14 is a diagram illustrating a configuration example of f-structure output by the semantic analysis unit 12. FIG. 15 is a diagram illustrating a configuration example of f-structure output by the semantic analysis unit 12. FIG. 16 is a diagram illustrating a configuration example of the sentence proofreading system 10 including a thesaurus dictionary. FIG. 17 is a diagram illustrating a configuration example of the sentence proofreading system 10 that performs sentence proofreading using the syntax analysis result having the same structure of the semantic analysis results in the corpus. FIG. 18 is a diagram illustrating a configuration example of the sentence proofreading system 10 that performs proofreading after narrowing down the semantic analysis result to one when a plurality of semantic analysis results are obtained for one proofreading target input sentence. It is. FIG. 19 is a diagram illustrating a configuration example of the sentence proofreading system 10 that performs proofreading by changing the semantic analysis result of the proofreading sentence. FIG. 20 is a diagram illustrating a configuration example of f-structure output from the semantic analysis unit 12. FIG. 21 is a diagram illustrating an example of a division rule that determines a division position of an input sentence.

Explanation of symbols

DESCRIPTION OF SYMBOLS 10 ... Text structure system 11 ... Proofread sentence input part 12 ... Semantic analysis part 13 ... Semantic analysis result holding part 14 ... Text generation part 15 ... Generation result holding part 16 ... Corpus holding part 17 ... Syntax analysis part 18 ... Selection part

Claims (18)

A natural language processing system that supports proofreading of sentences,
Semantic analysis means for performing semantic analysis on natural language sentences to be proofread;
Semantic analysis result holding means for holding a semantic analysis result by the semantic analysis means;
Sentence generation means for generating one or more natural language sentences from the semantic analysis result by the semantic analysis means;
Generation result holding means for holding a generation result by the sentence generation means;
A corpus holding means for holding a large number of natural language sentences;
A parsing means for parsing a natural language sentence;
The natural language sentence held in the generation result holding means and the natural language sentence held in the corpus holding means are respectively compared with the result of the syntax analysis by the syntax analysis means, and held in the generation result holding means. A selecting means for selecting, as a natural sentence, a sentence having a lot of sentences having similar parsing results among the natural language sentences held in the corpus holding means among the two or more natural language sentences;
Comprising
The semantic analysis means performs a semantic analysis based on a lexical functional grammar (LFG) theory, and converts a semantic content of a natural language sentence to be proofread into a functional structure including a nested structure of attribute-attribute value pairs ( functional
output as structure (f-structure)),
The sentence generation unit generates a natural language sentence from a semantic analysis result described in a functional structure obtained by a semantic analysis process based on the vocabulary functional grammar theory in the semantic analysis unit;
Natural language processing system characterized by that. The parsing means outputs the structure of a word or phrase in a sentence as a syntax tree that is expressed in a tree structure format,
The selection means calculates a similarity based on a distance calculation between syntax trees of the natural language sentences held in the generation result holding means and the natural language sentences held in the corpus holding means, Select a natural sentence that has a lot of sentences with similar parsing results among the natural language sentences held in the corpus holding means.
The natural language processing system according to claim 1. It also has a thesaurus dictionary,
The selection means replaces a word in a sentence to be proofread with a higher concept word based on the thesaurus dictionary, and then selects a sentence.
The natural language processing system according to claim 1. Means for obtaining a semantic analysis result for the natural language sentence held in the corpus holding means;
The selecting means targets a sentence having a semantic analysis result equal to a semantic analysis result for a sentence to be proofread held in the semantic analysis result holding means among sentences in the corpus holding means. Select sentence based on comparison,
The natural language processing system according to claim 2. The semantic analysis means outputs a relationship between a feature name (attribute) and a feature value (attribute value) included in a natural language sentence to be analyzed as a semantic analysis result,
Semantic analysis result changing means for changing or deleting the name of the predetermined feature and the value of the feature included in the semantic analysis result held in the semantic analysis result holding means,
The natural language processing system according to claim 1. Semantic analysis result dividing means for dividing the semantic analysis result held in the semantic analysis result holding means into two or more, or dividing the sentence to be proofread and performing semantic analysis to obtain two or more semantic analysis results. Prepared,
The sentence generation means performs a sentence generation process for each divided semantic analysis result.
The natural language processing system according to claim 1. Semantic analysis means, semantic analysis result holding means, sentence generation means, generation result holding means, corpus holding means, syntax analysis means, and sentence correction on a natural language processing system constructed using a computer functioning as a selection means A natural language processing method to support,
The semantic analysis means for performing a semantic analysis on a natural language sentence to be proofread;
The semantic analysis result holding means holds a semantic analysis result holding step of holding a semantic analysis result by the semantic analysis means;
A sentence generating step said sentence generating means, for generating one or more natural language text from the semantic analysis result by the semantic analysis means,
Wherein the generation result holding unit, a generation result holding step of holding the product result by the sentence generating means,
And corpus holding step the corpus holding means for holding a number of natural language text,
The parsing means for parsing a natural language sentence;
The selection unit compares the natural language sentence held in the generation result holding unit and the natural language sentence held in the corpus holding unit by the parsing result by the syntax analysis unit, and the generation result A selection step of selecting, as a natural sentence, a sentence having a lot of sentences having similar parsing results among natural language sentences held in the corpus holding means among two or more natural language sentences held in the holding means. When,
Have
In the semantic analysis step, semantic analysis based on Lexical Functional Grammar (LFG) theory is performed, and the semantic content of a natural language sentence to be proofread is converted into a functional structure consisting of a nested structure of attribute-attribute value pairs ( functional
output as structure (f-structure)),
In the sentence generation step, a natural language sentence is generated from a semantic analysis result described in a functional structure obtained by a semantic analysis process based on the vocabulary functional grammar theory in the semantic analysis step.
A natural language processing method characterized by that. In the parsing step, the structure of the words and phrases in the sentence is output as a syntax tree expressed in a tree structure format,
In the selection step, a similarity is calculated based on a distance calculation between syntax trees of the natural language sentence held in the generation result holding unit and the natural language sentence held in the corpus holding unit, Select a natural sentence that has a lot of sentences with similar parsing results among the natural language sentences held in the corpus holding means.
The natural language processing method according to claim 7. In the selection step, based on the thesaurus dictionary, the word in the sentence to be proofread is replaced with a higher-order concept word, and then the sentence is selected.
The natural language processing method according to claim 7. Further comprising the step of obtaining a semantic analysis result for a natural language sentence as a corpus;
In the selection step, a sentence is selected based on a comparison of syntax trees for a target whose semantic analysis result is equal to a semantic analysis result for the sentence to be proofread in the corpus.
The natural language processing method according to claim 8. In the semantic analysis step, a relationship between a feature name (attribute) and a feature value (attribute value) included in a natural language sentence to be analyzed is output as a semantic analysis result,
The computer also functions as a semantic analysis result changing means,
The semantic analysis result changing means further includes a semantic analysis result changing step for changing the semantic analysis result obtained by the semantic analysis step.
The natural language processing method according to claim 7. The computer also functions as a semantic analysis result dividing means,
The semantic analysis result dividing unit divides the semantic analysis result in the semantic analysis step into two or more, or divides a sentence to be proofread and performs semantic analysis to obtain two or more semantic analysis results. Further comprising
In the sentence generation step, a sentence generation process is performed for each divided semantic analysis result.
The natural language processing method according to claim 7. The computer is described in a computer-readable format so that the processing for supporting the proofreading of the sentence is executed on the computer. The computer is a semantic analysis unit, a semantic analysis result holding unit, a sentence generation unit, a generation result holding unit, and a generation result holding unit. , A computer program for functioning as a corpus holding means, a syntax analysis means, and a selection means ,
The semantic analysis means, lexical functional grammar for a natural language sentence as a calibration target (Lexical Functional
Semantic analysis based on Grammar (LFG) theory is performed, and the semantic content of the natural language sentence is converted into a functional structure consisting of nested structure of attribute-attribute value pairs.
(F-structure)) ,
The semantic analysis result holding unit holds the semantic analysis result by the semantic analysis means,
It said sentence generating means generates the vocabulary features one or more natural language sentences from the result of the semantic analysis is described with the resulting functional structures by semantic analysis processing based on grammar theory in the semantic analysis means,
The product result holding means holds the generated result of the sentence generating means,
The corpus holding unit holds a number of natural language text,
The syntax analysis means, parses the natural language sentence,
The selection means compares the natural language sentence held in the generation result holding means and the result of syntactic analysis of the natural language sentence held in the corpus holding means by the syntax analysis means, and the generation result Of the two or more natural language sentences held in the holding means, the natural language sentence held in the corpus holding means is selected as a natural sentence in which many sentences having similar syntax analysis results exist.
A computer program characterized by being a thing . The parsing means outputs the structure of a word or phrase in a sentence as a syntax tree that is expressed in a tree structure format,
The selection means calculates a similarity based on a distance calculation between syntax trees of the natural language sentences held in the generation result holding means and the natural language sentences held in the corpus holding means, Select a natural sentence that has a lot of sentences with similar parsing results among the natural language sentences held in the corpus holding means.
The computer program according to claim 13. The selection means selects a sentence after replacing a word in a sentence to be proofread with a higher-order concept word based on a thesaurus dictionary.
The computer program according to claim 13. The computer program of claim 14, the computer program further the computer, but the order is made to function as means for acquiring the semantic analysis result for natural language text which is retained in the corpus holding means Yes ,
The selecting means targets a sentence having a semantic analysis result equal to a semantic analysis result for a sentence to be proofread held in the semantic analysis result holding means among sentences in the corpus holding means. Select sentence based on comparison,
A computer program characterized by being a thing. The semantic analysis means outputs a relationship between a feature name (attribute) and a feature value (attribute value) included in a natural language sentence to be analyzed as a semantic analysis result,
Causing the computer to further function a semantic analysis result changing unit for changing or deleting a name and a feature value of a predetermined feature included in the semantic analysis result held in the semantic analysis result holding unit;
The computer program according to claim 13. 14. The computer program according to claim 13, wherein the computer program further causes the computer to function as a semantic analysis result dividing unit .
The semantic analysis result dividing means divides the semantic analysis result held in the semantic analysis result holding means into two or more, or divides a sentence to be proofread and performs semantic analysis to obtain two or more semantic analysis results. obtained,
The sentence generation means performs a sentence generation process for each divided semantic analysis result.
A computer program characterized by being a thing.

Images