Semantic expression generation method, semantic expression generation device, and semantic expression generation program

Abstract

To generate semantic representation data that can represent the meaning of words and the meaning of sentences in natural language data more appropriately and sufficiently than before.SOLUTION: A semantic representation generation method disclosed in the present application includes referring to a CT system table having concept information recorded therein which hierarchically and ambiguously represents the meaning of morphemes for all parts of speech to perform morphological analysis on a natural language text, thereby dividing the text into morphemes and providing concept tags to the morphemes. The method includes, through subsequent syntax analysis and context analysis, referring to an ST system table in which a semantic tag provision rule is associated with semantic tags indicating a semantic relation between a pair of text constitution elements to perform semantic analysis, thereby generating semantic representation data of the text. In the semantic representation data, the concept tags CT are provided to the morphemes, and the semantic tags ST indicating semantic information representing the semantic relation are provided to a pair of text constitution elements semantically related to each other (pair of clauses and the like).SELECTED DRAWING: Figure 13

Description

[Technical Field]

[0001] The present invention relates to a method and apparatus for generating semantic representation data necessary for acquiring knowledge from text data such as documents written in a natural language. [Background technology]

[0002] In recent years, methods have been researched and developed for computers to acquire knowledge from various text data such as documents written in natural language (hereinafter referred to as "natural language data"). Furthermore, methods have also been researched and developed for creating a knowledge base by structuring and storing the knowledge acquired in this way, and for computers to receive questions in natural language and provide answers based on the knowledge base.

[0003] To properly perform knowledge acquisition and question answering using such natural language, it is necessary to grasp the meaning of words contained in sentences hierarchically and ambiguously in semantic analysis of natural language data by computers. To address this, the concept of named entities contained in natural language data has been hierarchically defined (see, for example, Non-Patent Document 1). Furthermore, to grasp the meaning of sentences in semantic interpretation of natural language data, predicates such as verbs and adjectives are considered to be the center of meaning, and the structure of the sentence (called "predicate-argument structure") is analyzed in terms of the relationship between the predicate and arguments, which serve as complements necessary for the predicate to have meaning (see, for example, Non-Patent Document 2). [Prior art documents] [Patent documents]

[0004] [Patent Document 1] Patent Publication No. 2021-111303 [Patent Document 2] Japanese Patent Application Publication No. 6-195383 [Non-patent literature]

[0005] [Non-Patent Document 1] Koichi Takeuch, Alastair Butler, Iku Nagasaki, Takuya Okamura, Prashant Pardeshi, "Constructing Web-Accessible Semantic Role Labels and Frames for Japanese as Additions to the NPCMJ Parsed Corpus", Proceedings of the 12th Conference on Language Resources and Evaluation (LREC 2020), pp.3153-3161, Marseille, 11-16 May 2020 [Non-patent document 2] Koichi Takeuchi, Masayuki Ueno, and Nao Takeuchi, "Annotating Semantic Role Information to Japanese Balanced Corpus", Proceedings of MAPLEX 2015, 2015 Summary of the Invention [Problem to be solved by the invention]

[0006] The semantic expression data obtained by conventional semantic analysis using computers as described above does not always adequately represent the meaning of words and sentences in natural language data. As a result, the accuracy of knowledge acquisition from natural language data is insufficient, and the reusability of the acquired knowledge is also insufficient.

[0007] Therefore, it is desired to provide a method for generating semantic expression data that can express the meaning of words and sentences in natural language data more appropriately and sufficiently than conventional methods. [Means for solving the problem]

[0008] A first aspect of the present invention is a semantic expression generation method for generating semantic expression data from natural language information, the method comprising: The computer a text analysis step of receiving text data in a natural language and performing a superficial analysis including syntactic analysis on the text data to generate syntactic data representing the structure of sentences included in the text data; The computer a semantic analysis step of generating semantic expression data corresponding to the text data based on the syntactic data, The text analysis step includes: The computer records conceptual information that hierarchically and ambiguously represents the meaning of morphemes for a group of parts of speech including not only parts of speech of content words but also parts of speech of function words in the natural language. a concept tagging step of assigning a concept tag indicating concept information hierarchically expressing the meaning of each morpheme included in the text data based on the syntactic data by referring to a CT system table; The semantic analysis step includes: The computer a semantic tagging step of assigning semantic tags indicating semantic information representing a semantic relationship between a phrase or phrase chain corresponding to a predicate in the text data and another phrase or phrase chain having a dependency relationship with the predicate, based on the syntactic data; The computer and a data generation step of generating the semantic expression data based on the concept tag assigned to each morpheme included in the text data and the semantic tag assigned to a pair of a phrase or phrase chain and another phrase or phrase chain included in the text data. fruit, In the semantic tag assignment step, the computer refers to an ST system table that associates rules for determining pairs of phrases or phrase chains to which semantic tags should be assigned with other phrases or phrase chains, for each of a plurality of semantic tags that indicate a plurality of pieces of semantic information that represent the semantic relationship between a phrase or phrase chain and another phrase or phrase chain, regardless of whether the phrase or phrase chain is a phrase or phrase chain that corresponds to a predicate in the natural language, and assigns semantic tags that indicate semantic information that represent the semantic relationship to pairs of phrases or phrase chains that have a dependency relationship in the text data, even for phrases or phrase chains other than those that correspond to predicates, based on the syntactic data.

[0010] The present invention 2 The second aspect is a semantic expression device that generates semantic expression data from natural language information, comprising: a CT system storage unit that stores a CT system table that records conceptual information that hierarchically and ambiguously expresses the meaning of morphemes for a group of parts of speech, including not only parts of speech of content words in natural language but also parts of speech of function words; a text analysis unit that receives the natural language text data and performs a superficial analysis, including syntactic analysis, on the text data to generate syntactic data representing the structure of sentences included in the text data; a semantic analysis unit that generates semantic expression data corresponding to the text data based on the syntax data, the text analysis unit refers to the CT system table and assigns, to each morpheme included in the text data, a concept tag indicating conceptual information that hierarchically expresses the meaning of the morpheme based on the syntactic data; The semantic analysis unit based on the syntactic data, assigning semantic tags indicating semantic information representing the semantic relationship between a phrase or phrase chain corresponding to a predicate in the text data and another phrase or phrase chain having a dependency relationship with the predicate; The semantic expression data is generated based on the concept tag assigned to each morpheme included in the text data and the semantic tag assigned to a pair of a phrase or phrase chain and another phrase or phrase chain included in the text data. death, The meaning representation device further comprising an ST system storage unit storing an ST system table that associates, for each of a plurality of semantic tags indicating a plurality of pieces of semantic information that represent semantic relationships between a phrase or phrase chain and another phrase or phrase chain, regardless of whether the phrase or phrase chain corresponds to a predicate in the natural language, with a rule for determining a pair of the phrase or phrase chain to which the semantic tag should be assigned and the other phrase or phrase chain; The semantic analysis unit refers to the ST system table and assigns semantic tags indicating semantic information that represents the semantic relationship between phrases or phrase chains that have dependency relationships in the text data and other phrases or phrase chains, even for phrases or phrase chains other than those corresponding to predicates, based on the syntactic data. .

[0011] The present invention 3 The second aspect is a semantic expression generation program for generating semantic expression data from natural language information, comprising: a text analysis step of receiving text data in a natural language and performing a superficial analysis including syntactic analysis on the text data to generate syntactic data representing the structure of sentences included in the text data; a semantic analysis step of generating semantic expression data corresponding to the text data based on the syntactic data; on the computer, The computer includes a CT system storage unit that stores a CT system table in which conceptual information is recorded that hierarchically and ambiguously expresses the meanings of morphemes for a group of parts of speech including not only parts of speech of content words but also parts of speech of function words in the natural language, the text analysis step includes a concept tagging step of assigning, to each morpheme included in the text data, a concept tag indicating concept information hierarchically expressing its meaning, based on the syntactic data, by referring to the CT system table; The semantic analysis step includes: a semantic tagging step of assigning semantic tags indicating semantic information representing a semantic relationship between a phrase or phrase chain corresponding to a predicate in the text data and another phrase or phrase chain having a dependency relationship with the predicate, based on the syntactic data; and a data generation step of generating the semantic expression data based on the concept tag assigned to each morpheme included in the text data and the semantic tag assigned to a pair of a phrase or phrase chain and another phrase or phrase chain included in the text data. fruit, the computer further includes an ST system storage unit that stores an ST system table that associates, for each of a plurality of semantic tags indicating a plurality of pieces of semantic information that represent a semantic relationship between a phrase or phrase chain and another phrase or phrase chain, regardless of whether the phrase or phrase chain corresponds to a predicate in the natural language, with a rule for determining a pair of the phrase or phrase chain to which the semantic tag should be assigned and the other phrase or phrase chain; In the semantic analysis step, by referring to the ST system table, semantic tags indicating semantic information representing the semantic relationship are assigned to pairs of phrases or phrase chains having dependency relationships in the text data and other phrases or phrase chains, in addition to phrases or phrase chains corresponding to predicates, based on the syntactic data. .

[0012] Other aspects of the present invention will be apparent from the above aspects of the present invention and the following description of the embodiments and their modifications, and therefore will not be described here. [Effects of the Invention]

[0013] According to the first aspect of the present invention, syntactic data is generated by syntactic analysis of natural language text data. A CT system table is stored with conceptual information representing the hierarchical and ambiguous meanings of morphemes for parts of speech, including not only content words but also function words. Based on the syntactic data, conceptual tags representing conceptual information representing the hierarchical meanings of morphemes are assigned to each morpheme included in the text data. Subsequently, semantic tags representing semantic information representing the semantic relationship between a phrase or phrase chain corresponding to a predicate in the text data and another phrase or phrase chain having a dependency relationship with the predicate are assigned. Based on the assignment of the conceptual tags and the semantic tag assignment, semantic expression data corresponding to the text data is generated. This allows for the generation of semantic expression data that more appropriately represents the meanings of words included in the natural language text data than in the past.

[0014] Moreover, For each of a plurality of semantic tags indicating a plurality of pieces of semantic information expressing the semantic relationship between a phrase or phrase chain and another phrase or phrase chain, regardless of whether the phrase or phrase chain corresponds to a predicate in natural language, the ST system table associates rules for determining pairs of phrases or phrase chains to which the semantic tags should be assigned with other phrases or phrase chains. By doing so, even for phrases or phrase chains other than those corresponding to predicates in the text data, semantic tags indicating semantic information expressing the semantic relationship are assigned to pairs of phrases or phrase chains and other phrases or phrase chains that have a dependency relationship in the text data. By assigning such semantic tags to the text data in addition to the concept tags, semantic expression data corresponding to the text data is generated. This allows semantic expression data to be obtained from natural language data that more appropriately and fully expresses not only the meaning of words but also the meaning of sentences than before.

[0015] The above-mentioned first aspect of the present invention 2 The above aspect and the above 3 Both of these aspects have the same effects as the first aspect of the present invention.

[0016] The effects of other aspects of the present invention are clear from the explanation of the effects of the above aspects of the present invention and the effects of the following embodiments, so a description thereof will be omitted. [Brief explanation of the drawings]

[0017] [Figure 1] 1 is a block diagram showing the configuration of a semantic expression generation device according to an embodiment of the present invention. [Figure 2] FIG. 2 is a block diagram showing the configuration of a computer that operates as the semantic expression generation device according to the embodiment. [Figure 3] FIG. 2 is a diagram for explaining a CT system table used in the embodiment. [Figure 4] FIG. 2 is a diagram for explaining a portion of the ST system table used in the above embodiment that relates to semantic tags to be assigned between phrases (pairs of phrases). [Figure 5] FIG. 10 is a diagram for explaining a portion of the ST system table used in the above embodiment that relates to semantic tags to be added between compound phrases or compound phrases. [Figure 6] FIG. 2 is a diagram for explaining a portion of the ST system table used in the above embodiment that relates to semantic tags to be assigned between sentences (sentence pairs). [Figure 7] 10 is a flowchart showing the procedure of a semantic expression generation process executed in order for a computer to operate as the semantic expression generation device according to the embodiment. [Figure 8] 10 is a flowchart showing a procedure for morphological analysis in the semantic expression generation process. [Figure 9] 10 is a flowchart showing a procedure for syntactic analysis in the semantic expression generation process. [Figure 10] 10 is a flowchart showing a procedure for context analysis in the semantic expression generation process. [Figure 11] 10 is a flowchart showing a procedure of semantic analysis in the semantic expression generation process. [Figure 12] 10A and 10B are diagrams for explaining the semantic expression process based on a specific example. [Figure 13] FIG. 2 is a diagram for explaining a first generation example of semantic expression data in the embodiment. [Figure 14] FIG. 10 is a diagram for explaining a second example of generating semantic expression data in the embodiment. [Figure 15] FIG. 10 is a diagram for explaining a third example of generating semantic expression data in the embodiment. DETAILED DESCRIPTION OF THE INVENTION

[0018] In building a knowledge base from natural language data or implementing a natural language question-answering system, in order to improve the accuracy of knowledge acquisition, it is important to generate semantic expression data that can adequately express the meanings of words and sentences in the natural language data. Hereinafter, embodiments of an apparatus and method for generating such semantic expression data will be described with reference to the drawings. The semantic expression generation apparatus according to this embodiment is typically implemented using a computer, and the semantic expression generation method according to this embodiment is typically executed using a computer. The semantic expression generation program according to this embodiment is used to cause a computer to function as the semantic expression generation apparatus.

[0019] <1. Functional configuration of the semantic expression generation device> FIG. 1 is a block diagram showing the functional configuration of a semantic expression generation device 10 according to this embodiment. This semantic expression generation device 10 generates semantic expression data from natural language data (text data such as documents written in a natural language), and is realized by a computer executing a semantic expression generation program described below.

[0020] 1, the semantic expression generation device 10 includes a natural language analysis unit 110, a CT system table 33, and an ST system table 34. The natural language analysis unit 110 includes a morphological analysis unit 112, a syntactic analysis unit 114, a context analysis unit 116, and a semantic analysis unit 118. In this embodiment, the text data to be analyzed by the semantic expression generation device 10 is Japanese text data, and is stored in an external text data storage unit 100.

[0021] In such a semantic expression generation device 10, the natural language analysis unit 110 reads text data, which is natural language data to be analyzed, from the text data storage unit 100. In the natural language analysis unit 110, first, the morphological analysis unit 112 performs morphological analysis on the read text data (hereinafter referred to as "input text data") Din to generate data D1 in which the input text data is divided into morphemes (hereinafter referred to as "word-segmented data"). This morphological analysis also determines the parts of speech and conjugations of the morphemes included in the word-segmented data D1. Furthermore, in this morphological analysis, a concept tag (hereinafter also referred to as "CT") is assigned to each morpheme in the word-segmented data D1 by referring to the CT system table 33.

[0022] The syntactic analysis unit 114 performs syntactic analysis on the word segmentation data D1 resulting from the morphological analysis, thereby generating syntactic data D2 representing the structure (dependency structure and phrase structure) of each sentence included in the input text data Din.

[0023] The context analysis unit 116 performs context analysis on the input text data Din based on the syntactic data D2 to identify antecedents referred to by anaphors included in the input text data Din, and to identify pairs of sentences having a discourse relationship in the input text data Din, to generate context data representing the anaphoric and discourse relationships in the input text data Din, and to output context / syntax data D3 consisting of the context data and the syntactic data D2. Note that hereinafter, the morphological analysis unit 112, syntactic analysis unit 114, and context analysis unit 116 are collectively referred to as the "text analysis unit."

[0024] Based on the context and syntax data D3, the semantic analysis unit 118 assigns semantic tags (hereinafter also referred to as "ST") indicating semantic information expressing the semantic relationship between phrases or phrase chains having a dependency relationship in the input text data Din and other phrases or phrase chains (hereinafter also referred to as "phrase-phrase chain pairs") by referring to the ST system table 34, and generates semantic expression data 140 corresponding to the input text data Din based on the concept tags assigned to each morpheme included in the syntax data and the semantic tags assigned between phrases and phrase chain pairs in the syntax data. Semantic tags are also assigned between sentences having a discourse relationship, as will be described later.

[0025] 2. Hardware configuration of the semantic expression generation device 2 is a block diagram showing the configuration of a computer 20 that operates as the semantic expression generation device 10 in this embodiment by a semantic expression generation program described below, i.e., the hardware configuration of the semantic expression generation device 10 according to this embodiment. The computer 20 shown in FIG. 2 includes a CPU 21, a main memory 22, an auxiliary storage device 23, an input operation unit 24, a display device 25, a communication interface device 26, and a recording medium reader 27. The main memory 22 may be, for example, a DRAM. The auxiliary storage device 23 may be, for example, a hard disk or a solid state drive. The input operation unit 24 includes, for example, a keyboard 28 and a mouse 29. The display device 25 may be, for example, a liquid crystal display. The communication interface device 26 is an interface circuit for wired or wireless communication. The recording medium reader 27 is an interface circuit for a recording medium 30 that stores a program or the like. The recording medium 30 may be a non-transitory recording medium such as a CD-ROM, a DVD-ROM, or a USB memory.

[0026] In the computer 20 configured as described above, the auxiliary storage device 23 stores text data 32 to be analyzed, a CT system table 33, and an ST system table 34, in addition to the semantic expression generation program 31 according to this embodiment. Storing the text data 32 in the auxiliary storage device 23 realizes the text data storage unit 100 in the semantic expression generation device 10 of FIG. 1. The semantic expression generation program 31, the text data 32, the CT system table 33, and the ST system table 34 may be received from a server or another computer using the communication interface device 26, or may be read from the recording medium 30 using the recording medium reader 27, for example.

[0027] When the semantic expression generation program 31 is executed in the computer 20, the semantic expression generation program 31 is loaded into the main memory 22, and part or all of the text data 32 is loaded into the main memory 22 as input text data Din. The CPU 21 uses the main memory 22 as a working memory and executes the semantic expression generation program 31 stored in the main memory 22 to perform semantic expression generation processing on the input text data Din stored in the main memory 22. This semantic expression generation processing generates semantic expression data 140 corresponding to the input text data Din. When the CPU 21 performs the semantic expression generation processing, the computer 20 functions as the semantic expression generation device 10. Note that the configuration of the computer 20 described above is merely an example, and the semantic expression generation device 10 can be realized using various computers.

[0028] <3. CT System Table and ST System Table> In this embodiment, a CT system table and an ST system table as described below are prepared in advance and stored in the auxiliary storage device 23 as described above (FIG. 2).

[0029] FIG. 3 is a diagram illustrating a CT system table 33 used in this embodiment. The CT system table 33 records conceptual information that hierarchically and ambiguously expresses the meanings of all morphemes of parts of speech in Japanese as a natural language, i.e., morphemes of content words such as nouns, verbs, and adjectives, as well as morphemes of function words such as particles and auxiliary verbs. As shown in FIG. 3 , for example, the nouns "park" (park) and "school" are recorded as having a superordinate concept of "public facility name," and a further superordinate concept of "space." The noun "school" is also recorded as having a superordinate concept of "organization name," a further superordinate concept of "independent organization," and a still further superordinate concept of "independent entity." That is, for "school," conceptual information that hierarchically and ambiguously expresses its meaning is recorded. For example, for the particle "ni," the concepts that express its meaning are recorded as "state," "origin of action," or "cause and reason," and the superordinate concept of these concepts is recorded as "other party." That is, for the particle "ni," conceptual information that expresses its meaning hierarchically and ambiguously is recorded.

[0030] 4 to 6 are diagrams for explaining the ST system table 34 used in this embodiment. FIG. 4 is a diagram for explaining a portion of the ST system table 34 relating to semantic tags (ST) to be assigned between semantically related phrases (pairs of phrases) in Japanese text data as natural language data. FIG. 5 is a diagram for explaining a portion of the ST system table 34 relating to semantic tags to be assigned between semantically related compound phrases or compound phrases in Japanese text data. FIG. 6 is a diagram for explaining a portion of the ST system table 34 relating to semantic tags to be assigned between semantically related sentences in Japanese text data. 10 is a diagram for explaining a portion relating to a semantic tag to be assigned between and (a pair of sentences). FIG.

[0031] The ST system table associates multiple semantic tags, each representing multiple pieces of semantic information that express semantic relationships between phrases in Japanese as a natural language, with rules for determining pairs of phrases to which the semantic tags should be assigned. As shown in FIG. 4, the semantic tag assignment rules provide a method for determining which objects (pairs of phrases in this case) should be assigned the semantic tag "agt," which indicates semantic information such as "actor, subject of action with intention" as semantic information that expresses the semantic relationship between phrases. That is, according to the determination method (ST assignment rules) corresponding to the semantic tag "agt" in the ST system table 34 of FIG. 4, the semantic tag "agt" is assigned to "pairs of phrases that are in a subject-predicate relationship and whose morphemes are assigned a predetermined concept tag (predetermined CT)." Specifically, the "predetermined concept tag (predetermined CT)" here is selected based on the CT system table 33 in accordance with the semantic information that expresses the semantic relationship between the pair of phrases to which the semantic tag "agt" should be assigned. The same applies to the "predetermined CT" used to define the determination method (ST assignment rule) corresponding to other semantic tags, and an appropriate CT is selected based on the CT system table 33 according to the semantic information that represents the semantic relationship between the pair of phrases to which the semantic tag should be assigned.

[0032] 3 to 6 exemplify the contents of the CT system table 33 and the ST system table 34 for the sake of convenience in explaining this embodiment, and the contents are appropriately omitted and simplified compared to the actually used CT system table 33 and ST system table 34. Furthermore, the specific contents of the CT system table 33 and the ST system table 34 (such as the hierarchical structure of morpheme concepts and rules for assigning semantic tags) are not limited to those shown in FIGS.

[0033] <4. Semantic Representation Generation Processing> As described above, the CPU 21 of the computer 20 executes the semantic expression generation program 31, thereby performing a semantic expression generation process on natural language text data as a document to be analyzed. FIG. 7 is a flowchart showing the steps of this semantic expression generation process. This semantic expression generation process includes morphological analysis, syntactic analysis, contextual analysis, and semantic analysis. FIG. 8 is a flowchart showing the steps of morphological analysis, FIG. 9 is a flowchart showing the steps of syntactic analysis, FIG. 10 is a flowchart showing the steps of contextual analysis, and FIG. 11 is a flowchart showing the steps of semantic analysis. Note that hereinafter, morphological analysis, syntactic analysis, and contextual analysis are also collectively referred to as "text analysis." In this embodiment, the CPU 21 executes the semantic expression generation program 31, causing the computer 20 to operate as shown in FIGS. 7 to 11.

[0034] 7, first, based on an instruction by a user's input operation on the input operation unit 24, part or all of the text data 32 is read from the auxiliary storage device 23 and stored in the main memory 22 as input text data Din (step S10). The CT system table 33 and the ST system table 34 may be stored in the auxiliary storage device 23 and referred to as appropriate during the semantic expression generation process, but instead, in order to speed up the semantic expression generation process, the CT system table 33 and the ST system table 34 may be read from the auxiliary storage device 23 and stored in the main memory 22 in step S10.

[0035] Next, morphological analysis is performed on the input text data Din (step S12). As shown in Fig. 8, in the morphological analysis, morpheme boundaries in the input text data Din are identified, and the part of speech of each morpheme included in the input text data Din is determined (step S122).

[0036] Thereafter, a concept tag (CT) is assigned to each morpheme in the input text data Din by referring to the CT system table 33 (step S124). As described above, the CT system table 33 records concept information that hierarchically and ambiguously represents the meaning of morphemes used in natural language (see FIG. 3). When assigning a concept tag to each morpheme in the input text data Din, a concept tag corresponding to the morpheme is searched for in the CT system table 33, starting from the lowest concept (including examples) and moving upward, and the first concept tag determined to be a match is assigned to the morpheme. Below, the assignment of concept tags to each morpheme in the input text data Din will be described with reference to FIG. 12.

[0037] FIG. 12 is a diagram illustrating the semantic expression generation process according to this embodiment based on a specific example. Here, it is assumed that the input text data Din contains the sentence (text) shown in FIG. 12(A), i.e., "Taro went to the park." The semantic expression generation process will be described focusing on this text. In steps S122 and S124, this text is divided into seven morphemes as shown in FIG. 12(B), and a concept tag (CT) is assigned to each morpheme. In FIG. 12(B), dotted lines extending vertically indicate morpheme divisions (the same applies to FIGS. 13 and 14 below). Hereinafter, data divided into morphemes as shown in FIG. 12(B), each with a concept tag assigned and a part of speech assigned, will be referred to as "word-separated data."

[0038] Next, based on the morpheme division of the input text data Din and the assignment of part-of-speech and concept tags to each morpheme in steps S122 and S124, word-separated data D1 corresponding to the input text data Din is generated (step S126). Once the word-separated data D1 is generated, the morpheme processing (step S12) is terminated, and the process proceeds to step S14 in Figure 7 for syntactic analysis.

[0039] As shown in Fig. 9, in syntactic analysis (step S14), a dependency structure in the input text data Din, i.e., a dependency structure of phrases in the input text data Din, is determined based on the word segmentation data D1 (step S142), and a phrase structure in the input text data Din is determined (step S144). Here, the phrase structure refers to a structure that indicates how phrases are formed from morphemes contained in the input text data Din, and how compound phrases are formed from the phrases, i.e., a hierarchical structure of a sentence composed of morphemes, phrases, and compound phrases. According to these steps S142 and S144, for example, for the previously described text shown in Fig. 12(A), a dependency structure and phrase structure are determined based on the word segmentation data D1 (Fig. 12(B)), as shown in Fig. 12(C). Note that the text in Figure 12(A) does not contain any consecutive phrases, but is composed of three phrases ("Taro," "Went to the park," and "Went") as shown in Figure 12(C), and these phrases have the dependency relationships shown in Figure 12(C).

[0040] In this syntactic analysis, syntactic data D2 representing the structure (dependency structure and phrase structure) of each sentence included in the input text data Din is then generated based on the dependency structure and phrase structure obtained as described above (step S146). Once the syntactic data D2 is generated, the syntactic analysis (step S14) is terminated, and the process proceeds to step S16 in FIG. 7 to perform context analysis.

[0041] 10, in the context analysis (step S16), anaphoric analysis is performed on the input text data Din based on the syntax data D2 to identify antecedents referred to by anaphors contained in the input text data Din (step S162). Furthermore, discourse structure analysis is performed on the input text data Din based on the syntax data D2 to identify pairs of sentences having a discourse relationship in the input text data Din, i.e., pairs of sentences that are semantically related (step S164). Then, based on the results of these identifications, context data representing the anaphoric and discourse relationships in the input text data Din is generated, and the context data and the above syntax data are combined. Context and syntax data D3 are generated from data D2. When the context and syntax data D3 is generated, the context analysis (step S16) is completed, and the process proceeds to step S18 in Fig. 7 where semantic analysis is performed. Note that the use of the context and syntax data D3 obtained by such anaphora analysis and discourse structure analysis will be explained in relation to the second and third generation examples of semantic expression data described below (see Figs. 14 and 15).

[0042] 11, in the semantic analysis (step S18), by referring to the ST system table 34, semantic tags are assigned between phrases (phrase pairs), between phrase chains, and between phrase chains (phrase chain pairs) that have dependency relationships in the input text data Din based on the context and syntax data D3, and semantic tags are assigned between sentences (sentence pairs) that have anaphoric or discourse relationships in the input text data Din based on the context and syntax data D3. In the following, phrases, phrase chains, and sentences are collectively referred to as "text components," and the description will proceed assuming that pairs of text components that have dependency or discourse relationships are semantically related.

[0043] After semantic tags are assigned to pairs of text components contained in the input text data Din in step S182, semantic expression data 140 corresponding to the input text data Din is generated based on the concept tags assigned to each morpheme in the input text data Din and the semantic tags assigned to pairs of text components having a semantic relationship in the input text data Din (step S184).

[0044] For example, semantic expression data as shown in Figure 12(D) is generated for the text shown in Figure 12(A). In the semantic expression shown in Figure 12(D), phrases as text components are treated as nodes, and edges are provided between semantically related nodes (between phrases having a dependency relationship), with semantic tags (ST) indicating semantic information expressing the semantic relationship being assigned to the edges, and conceptual tags (CT) indicating conceptual information expressing the meaning of each morpheme being assigned. In the computer 20 serving as the semantic expression generation device 10 according to this embodiment, semantic expression data 140 having an appropriate data structure (a data structure suitable for computer processing) corresponding to the semantic expression shown in Figure 12(D) is generated in the main memory 22 and appropriately stored in the auxiliary storage device 23 or the like.

[0045] In the semantic representation of Figure 12(D), the two phrases "to the park" and "went" are assigned a semantic tag "lfp" indicating the semantic information of "spatial end point" that represents the semantic relationship between them (see Figure 4). In addition, the two phrases "Taro is" and "went" are assigned a semantic tag "agt" indicating the semantic information of "actor, subject of action with intention" that represents the semantic relationship between them. However, depending on the determination method (ST assignment rules) in the ST system table 34, the semantic relationship represented by the semantic information of "experienced person, ..." indicated by the semantic tag "exp" may also be determined to correspond to the dependency relationship between these phrases (see Figure 4). In this case, the two semantic tags "agt" and "exp" are assigned between "Taro is" and "went."

[0046] When the semantic expression data 140 as described above is generated in step S184, the semantic analysis (step S18) is completed. As shown in Fig. 7, the semantic expression generation process in this embodiment is completed upon completion of the semantic analysis.

[0047] In addition, with regard to the specific processing of the semantic expression generation processing (FIG. 7) in this embodiment, morphological analysis (FIG. 8), syntactic analysis (FIG. 9), and contextual analysis (FIG. 10), except for the assignment of concept tags based on the CT system table 33 (FIG. 3), any known or publicly known method may be adopted, and is not limited to the processing described above.

[0048] <5. Example of semantic expression data generation> <5.1 First generation example> Fig. 13 is a diagram for explaining a first generation example of semantic expression data in this embodiment. That is, Fig. 13 is a diagram for explaining what kind of semantic expression data is generated from the text by the semantic expression generation process (Fig. 7) when the input text data Din contains text consisting of one sentence shown in Fig. 13(A).

[0049] In this example, the text in Figure 13(A) is divided into nine morphemes as shown in Figure 13(B) by morphological analysis (step S12) in the semantic representation generation process, and a concept tag (CT) is assigned to each morpheme. Focusing on the particle morphemes "wa" and "to" in this text, as shown in Figures 3 and 13(B), the conceptual information given in the CT system table 33 for "wa" does not represent its meaning hierarchically, while the conceptual information given in the CT system table 33 for "to" represents its meaning hierarchically and ambiguously. That is, as shown in Figure 3, the concepts representing the particle "to" are recorded as "result," "comparison standard," "joint partner," "quote," or "limit," and the superordinate concept of these concepts is recorded as "partner." As shown in Figure 13(B), the particle "to" in this example is assigned a concept tag indicating the conceptual information "joint partner-partner." In the morphological analysis stage, for a morpheme with polysemic meaning, such as the particle "to," multiple conceptual tags may be assigned to the morpheme while it remains polysemic. However, if there are conflicting conceptual tags among the multiple conceptual tags, when semantic tags are assigned to pairs of text components (pairs of equal phrases) in the semantic analysis stage (S182 in Figure 11), a new conceptual tag will be assigned to the morpheme in accordance with the conceptual tags of the morphemes before and after the morpheme in the text to be analyzed.

[0050] Next, by syntactic analysis (step S14), the dependency structure and phrase structure of the text of this example (FIG. 13(A)) are determined as shown in FIG. 13(C).

[0051] Then, a context analysis (step S16) is performed, followed by a semantic analysis (step S18). Through this semantic analysis, semantic tags (such as "agt", "jnt", and "pur") are assigned to pairs of semantically related text components in the text of this example (here, three pairs of phrases having a dependency relationship), and based on the concept tags assigned to each morpheme and the semantic tags assigned to pairs of semantically related text components, semantic expression data such as that shown in Figure 13(D) is generated as semantic expression data corresponding to the text of this example (Figure 13(A)).

[0052] In this semantic expression data, the semantic tag ("agt" or "agt,exp") assigned between the two phrases "Taro is" and "went" is the same as the example shown in Fig. 12. Between the two phrases "with Hanako" and "went", the semantic tag "jnt" indicating the semantic information of "joint participant" that represents the semantic relationship between them is assigned based on the determination method in the ST system table 34 of Fig. 4 (ST assignment rules for determining the semantic tag to be assigned between phrases). Between the two phrases "to support" and "went", the semantic tag "pur" indicating the semantic information of "purpose" that represents the semantic relationship between them is assigned based on the determination method in the ST system table 34 of Fig. 4.

[0053] <5.2 Second generation example> Fig. 14 is a diagram for explaining a second generation example of semantic expression data in this embodiment. That is, Fig. 14 is a diagram for explaining what kind of semantic expression data is generated from the text by the semantic expression generation process (Fig. 7) when the input text data Din contains text consisting of two sentences shown in Fig. 14(A).

[0054] In this example, the text in Figure 14(A) is subjected to morphological analysis (step S12) in the semantic expression generation process, and the first sentence ("The heating wire was hot.") and the second sentence ("The heating wire softened.") are each divided into six morphemes, as shown in Figure 14(B), and a concept tag (CT) is assigned to each morpheme.

[0055] Then, semantic analysis (step S18) is performed after syntactic analysis (step S14) and context analysis (step S16). By this semantic analysis, semantic tags ("gnr", "cap") are assigned to pairs of semantically related text components in the text of this example (Fig. 14(A)) (here, one phrase pair having a dependency relationship in the first sentence and one phrase pair having a dependency relationship in the second sentence), and based on the concept tags assigned to each morpheme and the semantic tags assigned to pairs of semantically related text components, semantic expression data such as that shown in Fig. 14(C) is generated as semantic expression data corresponding to the text of this example.

[0056] In this semantic expression data, based on the judgment method of the ST system table 34 shown in Figure 4 (ST assignment rules for determining the semantic tag to be assigned between phrases), the semantic tag "gnr" indicating the semantic information of a "general relationship" that expresses the semantic relationship between the two phrases "the heating wire" and "it was hot" in the first sentence is assigned between them, the semantic tag "cap" indicating the semantic information of "something that causes an event unintentionally" that expresses the semantic relationship between the two phrases "the heating wire" and "it softened" in the second sentence is assigned between them, and the semantic tag "eq" indicating the semantic information of "equivalent" that expresses the semantic relationship between the phrases "the heating wire" in the first sentence and "the heating wire" in the second sentence is assigned based on the context and syntax data D3 in this example.

[0057] Furthermore, in this semantic expression data, based on the context and syntax data D3 in this example, it is determined that the second sentence ("The heating wire has softened.") corresponds to the "result." Based on this determination result and the semantic tags "gnr," "cap," and "eq" that are assigned between the phrases in this example, as shown in Figure 14(C), a semantic tag indicating semantic information "cau" (cause) that represents the semantic relationship between the phrase "It was hot," which corresponds to the predicate of the first sentence, and the phrase "It softened," which corresponds to the predicate of the second sentence (the edge from "It was hot" to "It softened").

[0058] <5.3 Third Generation Example> Fig. 15 is a diagram for explaining a third example of generation of semantic expression data in this embodiment. That is, Fig. 15 is a diagram for explaining what kind of semantic expression data is generated from the text by the semantic expression generation process (Fig. 7) when the input text data Din contains text consisting of three sentences shown in Fig. 15(A).

[0059] In this example, the text in Figure 15(A) undergoes morphological analysis (step S12), syntactic analysis (step S14), and context analysis (step S16) in the semantic expression generation process, and then semantic analysis (step S18) is performed, whereby the first sentence ("I found a useful book at the bookstore."), the second sentence ("The book was red and cheap."), and the third sentence ("I bought it right away.") are divided into morphemes and a concept tag (CT) is assigned to each morpheme (see Figure 3), and semantic tags are assigned between text components that have semantic relationships in the text in this example (pairs of text components), and semantic expression data such as that shown in Figure 15(B) is generated.

[0060] In this semantic expression data, based on the judgment method of the ST system table 34 shown in FIG. 4 (ST assignment rules for determining semantic tags to be assigned between phrases), in the first sentence, the two phrases "to be helpful" and "to read a book" are assigned a "state, situation, or occasion" that expresses the semantic relationship between them. The semantic tag "sit" is assigned to indicate the semantic information of "a conjunction of two sentences," the semantic tag "obj" is assigned to indicate the semantic information of "object of a transitive verb," ​​which represents the semantic relationship between the two phrases "at the bookstore" and "found," the semantic tag "loc" is assigned to indicate the semantic information of "spatial location," which represents the semantic relationship between the two phrases "I" and "found," and the semantic tag "agt" is assigned to indicate the semantic information of "actor, subject of action with will," which represents the semantic relationship between the two phrases. In the second sentence, the two phrases "The book is" and "It's red" are tagged with the semantic tag "sit," which indicates a "state, situation, or occasion" that expresses their semantic relationship. The two phrases "The book is" and "It was cheap" are tagged with the semantic tag "sit," which indicates a "state, situation, or occasion" that expresses their semantic relationship. The two phrases "It's red" and "It was cheap" are tagged with the semantic tag "par," which indicates a "parallel relationship" that expresses their semantic relationship. In the third sentence, the two phrases "It is" and "It was bought" are tagged with the semantic tag "obj," which indicates an "object of a transitive verb" that expresses their semantic relationship. The two phrases "It was bought" and "It was bought" are tagged with the semantic tag "tim," which indicates a "temporal position" that expresses their semantic relationship. Between the phrase "book" in the first sentence and the phrase "book is" in the second sentence, the semantic tag "eq" is assigned, indicating the semantic information of "equivalent" that represents the semantic relationship between them, based on the context and syntax data D3 in this example. Between the phrase "sore" in the third sentence and the phrase "book is" in the second sentence, the semantic tag "corr" is assigned, indicating the semantic information of "anaphoric relationship" that represents the semantic relationship between them, based on the context and syntax data D3 in this example. Between the phrase "I" in the first sentence and the phrase "kat" in the third sentence, the semantic tag "agt" is assigned, indicating the semantic information of "actor, subject of action with will" that represents the semantic relationship between them, based on the context and syntax data D3 in this example.

[0061] Furthermore, in this semantic representation data, based on the context and syntax data D3 in this example, the third sentence ("I bought it right away.") is determined to be a "result." Based on this determination result and the semantic tags assigned between the phrases in this example, as shown in Figure 15(B), a semantic tag indicating the semantic information "rea" (reason) that expresses the semantic relationship is assigned between the phrase "I found it," which corresponds to the predicate of the first sentence, and the phrase "I bought it," which corresponds to the predicate of the third sentence. Similarly, semantic tags indicating the semantic information "rea" (reason) that express the semantic relationship are assigned between the phrase "it was helpful" in the first sentence and the phrase "I bought it" in the third sentence, between the phrase "in red" in the second sentence and the phrase "I bought it" in the third sentence, and between the phrase "it was cheap" in the second sentence and the phrase "I bought it" in the third sentence.

[0062] <6. Effects> According to the present embodiment, in the morphological analysis (FIGS. 7 and 8) of the input text data Din, which is natural language data, the CT system table 33 (FIG. 3) is referenced, which stores conceptual information that hierarchically and ambiguously expresses the meanings of all morphemes of parts of speech. Concept tags (CT) are assigned not only to the morphemes of content words such as nouns and verbs, but also to the morphemes of function words such as particles (see FIGS. 12(B), 13(B), and 14(B)). In the semantic analysis (FIGS. 7 and 11), semantic expression data 140 is generated based on the assignment of such conceptual tags to the morphemes (FIGS. 12(D), 13(D), 14(C), etc.). This allows for the generation of semantic expression data that more appropriately expresses the meanings of words contained in natural language data than in the past.

[0063] According to this embodiment, the input text data Din, which is natural language data, is subjected to morphological analysis (FIG. 8) including the assignment of concept tags to morphemes as described above, followed by syntactic analysis (FIG. 9) and contextual analysis (FIG. 10), and then semantic analysis (FIG. 11). In this semantic analysis, by referring to the ST system table 34, semantic tags are assigned to each pair of semantically related text components (phrases, compound phrases, or sentences) in the input text data Din. The semantic expression data 140 is generated based on the assignment of conceptual tags to morphemes and semantic tags to pairs of such text components (see, for example, Figures 12(D), 13(D), and 14(C)). The ST system table 34 referenced here, as shown in Figure 4, provides ST assignment rules (methods for determining which pairs of text components should be assigned the semantic tags) for each of a plurality of semantic tags indicating a plurality of pieces of semantic information that represent the semantic relationship between pairs of semantically related text components (phrases, compound phrases, or sentences) in natural language, regardless of whether the text components are phrases such as predicates. Therefore, according to this embodiment, semantic expression data that adequately and sufficiently expresses not only the meaning of words but also the meaning of sentences or paragraphs from natural language data can be obtained compared to conventional methods. Furthermore, the ST assignment rules provided in the ST system table 34 include rules that define pairs of text components to which semantic tags should be assigned using the conceptual tags (CT) of the morphemes included in the pairs. According to such ST assignment rules, semantic tags to be assigned to pairs of text components can be determined more accurately.

[0064] Furthermore, if the semantic expression data generated by this embodiment is used in knowledge acquisition from natural language data or in a natural language question-answering system, the accuracy of knowledge acquisition and the reusability of the acquired knowledge can be improved.

[0065] <7. Variations> The present invention is not limited to the above-described embodiment, and various modifications can be made without departing from the scope of the present invention.

[0066] For example, in the above embodiment, the input text data Din for generating semantic expression data is Japanese text data. However, semantic expression data can be generated from input text data Din that is text data in another natural language, for example, English text data, by a semantic expression generation device or a semantic expression generation method configured in the same manner as the semantic expression generation device or the semantic expression generation method according to the above embodiment.

[0067] In the CT system table 33 used in the above embodiment, concepts representing the meanings of morphemes in natural language are hierarchically classified as shown in Fig. 3. However, the hierarchical classification shown in Fig. 3 is merely exemplary, and concepts representing the meanings of morphemes in natural language may be hierarchically classified in other ways. Furthermore, the ST attachment rules associated with each semantic tag in the ST system table 34 used in the above embodiment may be defined in a form different from the ST attachment rules shown in Figs. 4 to 6. Furthermore, in the ST system table 34 used in the above embodiment, an ST attachment rule is assigned to a pair of semantically related text components (phrases, compound phrases, or sentences) in natural language, regardless of whether the text components are phrases corresponding to predicates. However, instead of this, an ST attachment rule may be assigned to a pair of semantically related text components in natural language only when one of the pair is a phrase corresponding to a predicate, as in the case of conventional deep case assignment. [Explanation of symbols]

[0068] 10...Semantic expression generation device 20...computer 21...CPU 22...Main memory 23…Auxiliary storage device 30...Recording media 31 ...Semantic expression generation program 32...Text data 33...CT system table 34...ST system table 100...Text data storage unit 112...Morphological analysis section 114...Syntax analysis unit 116…Context analysis section 118...Semantic analysis section 140...Semantic expression data

Claims

1. A semantic expression generation method for generating semantic expression data from natural language information, comprising: a text analysis step in which a computer receives text data in a natural language and performs a superficial analysis, including syntactic analysis, on the text data to generate syntactic data representing the structure of sentences included in the text data; a semantic analysis step in which the computer generates semantic expression data corresponding to the text data based on the syntactic data, the text analysis step includes a concept tagging step in which the computer refers to a CT system table in which conceptual information is recorded that hierarchically and ambiguously represents the meaning of morphemes for a group of parts of speech including not only parts of speech of content words but also parts of speech of function words in the natural language, and assigns, to each morpheme included in the text data, a concept tag that indicates conceptual information that hierarchically represents its meaning, based on the syntactic data; The semantic analysis step includes: a semantic tagging step in which the computer assigns semantic tags indicating semantic information representing a semantic relationship between a phrase or phrase chain corresponding to a predicate in the text data and another phrase or phrase chain having a dependency relationship with the predicate, based on the syntax data; a data generation step in which the computer generates the semantic expression data based on the concept tag assigned to each morpheme included in the text data and the semantic tag assigned to a pair of a phrase or phrase chain and another phrase or phrase chain included in the text data, In the semantic tag assignment step, the computer refers to an ST system table that associates rules for determining pairs of phrases or phrase chains to which semantic tags should be assigned with other phrases or phrase chains, for each of a plurality of semantic tags that indicate a plurality of pieces of semantic information that represent the semantic relationship between a phrase or phrase chain and other phrases or phrase chains, regardless of whether the phrase or phrase chain is a phrase or phrase chain that corresponds to a predicate in the natural language, and assigns semantic tags that indicate semantic information that represent the semantic relationship to pairs of phrases or phrase chains that have a dependency relationship in the text data, even for phrases or phrase chains other than those that correspond to predicates, based on the syntactic data.

2. In the ST system table, a plurality of semantic tags each indicating a plurality of pieces of semantic information representing a semantic relationship between sentences in the natural language are associated with a rule for determining a pair of sentences to which the semantic tag should be assigned; the text analysis step further includes a context analysis step in which the computer performs a context analysis on the text data based on the syntactic data to generate context data representing anaphoric relations and discourse relations in the text data; 2. The semantic representation generation method according to claim 1, wherein in the semantic tagging step, the computer refers to the ST system table to assign semantic tags indicating semantic information that represents the semantic relationship to pairs of sentences in the text data that have an anaphoric or discourse relationship based on the context data.

3. 2. The semantic representation generation method of claim 1, wherein in the semantic tag assignment step, if there are multiple semantic tags to be assigned to a pair in the ST system table consisting of a phrase or phrase chain in the syntax data and another phrase or phrase chain, the computer assigns the multiple semantic tags to the pair.

4. 2. The semantic representation generation method according to claim 1, wherein in the ST system table, for at least some of the plurality of semantic tags, rules for determining pairs of phrases or phrase chains to which semantic tags should be assigned and other phrases or phrase chains are defined using conceptual tags of morphemes contained in the phrases or phrase chains that make up the pair.

5. A semantic expression device that generates semantic expression data from natural language information, a CT system storage unit storing a CT system table in which conceptual information is recorded that hierarchically and ambiguously expresses the meaning of morphemes for a group of parts of speech including not only parts of speech of content words but also parts of speech of function words in natural language; a text analysis unit that receives the natural language text data and performs a superficial analysis, including syntactic analysis, on the text data to generate syntactic data representing the structure of sentences included in the text data; a semantic analysis unit that generates semantic expression data corresponding to the text data based on the syntax data, the text analysis unit refers to the CT system table and assigns, to each morpheme included in the text data, a concept tag indicating concept information that hierarchically expresses the meaning of the morpheme, based on the syntactic data; The semantic analysis unit based on the syntactic data, assigning semantic tags indicating semantic information representing the semantic relationship between a phrase or phrase chain corresponding to a predicate in the text data and another phrase or phrase chain having a dependency relationship with the predicate; generating the semantic expression data based on the concept tags assigned to each morpheme included in the text data and the semantic tags assigned to pairs of phrases or phrase chains and other phrases or phrase chains included in the text data; The meaning representation device an ST system storage unit that stores an ST system table that associates, for each of a plurality of semantic tags indicating a plurality of pieces of semantic information that represent semantic relationships between a phrase or phrase chain and another phrase or phrase chain, regardless of whether the phrase or phrase chain corresponds to a predicate in the natural language, with a rule for determining a pair of the phrase or phrase chain to which the semantic tag should be assigned and the other phrase or phrase chain; The semantic analysis unit refers to the ST system table and, based on the syntactic data, assigns semantic tags indicating semantic information that represents the semantic relationship between phrases or phrase chains that have a dependency relationship in the text data and other phrases or phrase chains, even for phrases or phrase chains other than those that correspond to predicates.

6. A semantic expression generation program for generating semantic expression data from natural language information, a text analysis step of receiving text data in a natural language and performing a superficial analysis including syntactic analysis on the text data to generate syntactic data representing the structure of sentences included in the text data; a semantic analysis step of generating semantic expression data corresponding to the text data based on the syntactic data; on the computer, the computer includes a CT system storage unit storing a CT system table in which conceptual information is recorded that hierarchically and ambiguously expresses the meanings of morphemes for a group of parts of speech including not only parts of speech of content words but also parts of speech of function words in the natural language; the text analysis step includes a concept tagging step of assigning, to each morpheme included in the text data, a concept tag indicating concept information hierarchically expressing its meaning, based on the syntactic data, by referring to the CT system table; The semantic analysis step includes: a semantic tagging step of assigning semantic tags indicating semantic information representing a semantic relationship between a phrase or phrase chain corresponding to a predicate in the text data and another phrase or phrase chain having a dependency relationship with the predicate, based on the syntactic data; a data generation step of generating the semantic expression data based on the concept tag assigned to each morpheme included in the text data and the semantic tag assigned to a pair of a phrase or phrase chain and another phrase or phrase chain included in the text data, the computer further includes an ST system storage unit that stores an ST system table that associates, for each of a plurality of semantic tags indicating a plurality of pieces of semantic information that represent a semantic relationship between a phrase or phrase chain and another phrase or phrase chain, regardless of whether the phrase or phrase chain corresponds to a predicate in the natural language, with a rule for determining a pair of the phrase or phrase chain to which the semantic tag should be assigned and the other phrase or phrase chain; In the semantic analysis step, by referring to the ST system table, a semantic expression generation program assigns semantic tags indicating semantic information that represents the semantic relationship between phrases or phrase chains that have a dependency relationship in the text data and other phrases or phrase chains, based on the syntactic data, even for phrases or phrase chains that do not correspond to predicates.

Images