JPH0517591B2 - - Google Patents

Description

[Detailed description of the invention]

(Industrial Application Field) The present invention is a natural language processing system that inputs a sentence and extracts its structure.The present invention is a natural language processing system that inputs a sentence and extracts its structure. It is about the method. (Prior art) In conventional language processing, the relationships between words, phrases, and unit sentences within a sentence and other elements are investigated.
Also, it was thought that context analysis, which analyzes what kind of author's intention can be read from the text, could only be done after all the semantic information was known. For example, various stages of machine translation in Processing of Numbers, Expressions, and Sentences (Iwanami Course Information Science 23).
A diagram showing this (p. 146 of the same book) is shown in Figure 4. As shown in this figure, the conventional system first performs word segmentation and word dictionary lookup processing called morphological analysis. Next, based on the resulting dictionary contents, individual unit sentences and unit phrases are discovered through syntactic analysis using only grammatical and syntactic information, and their mutual relationships are calculated. After that, a semantic analysis process was performed to reinterpret the structure obtained and extract the meaning, and then a context analysis was performed using the semantic information obtained. By the way, context information is useful for sentences in which there are multiple disjunctions of predicates, conjunctive particle equivalents combined with predicates, adnominals, predicates, etc. in one sentence, that is, complex sentences (complex sentences, plural pragmatic sentences). It is carried on the shoulders. When analyzing a complex sentence, if the syntax is analyzed only from grammatical and syntactic information, there are problems in that there are multiple words, phrases, and interclauses, leaving many ambiguities in the structure. Conventionally, with regard to this problem, as stated in ``Natural Language and Computer Language'' (by Yutaka Kusanagi, published by Kodansha Scientific), we have abandoned seeking solutions in natural language. Attempts were made to deal with this by setting restrictions on Japanese, such as placing restrictions on the use of vocabulary and commas, and inserting symbols to indicate starting points. (Problems to be Solved by the Invention) As described above, in the prior art, context information was considered to be information that could only be obtained after semantic processing. In addition, most of the solutions for determining the structure of a sentence from among multiple possibilities were sought outside of natural language. However, context is essentially information that humans use prior to semantic analysis to estimate the structure of a sentence that cannot be understood from grammatical and syntactic information alone. Therefore, in conventional language processing systems, where context can only be handled by using the meaning obtained after the meaning has been completely understood, the quality of context analysis cannot be guaranteed. Since it is impossible to extract completely correct meaning using current semantic analysis techniques, there is a risk that contextual interpretation based only on the results of incomplete semantic analysis will include many major errors in the analysis. . In addition, regarding the process of estimating the start of complex sentences, if we do not make full use of the context information regarding sentence structure that exists in natural languages, we will have to limit the usage of vocabulary and commas, and insert symbols that clearly indicate the structure. It is not possible to minimize the pretreatment required. The purpose of the present invention is to solve these conventional problems by analyzing the natural language expressions that characterize the context structure at the lexical analysis stage, thereby analyzing the logic in the sentence prior to syntactic and semantic analysis. The purpose of this paper is to provide a context analysis method for estimating a context structure that is consistent with the author's intentions. (Means for Solving the Problem) The context analysis method according to the present invention has a function of reading an input sentence, a function of looking up the input sentence in a dictionary, and a function of analyzing the input sentence using the dictionary information after looking up the dictionary. In a natural language analysis system, a) Vocabulary information that indicates semantically coherent sections in long sentences such as complex sentences is collected in advance, and information on the ease and difficulty of dependencies of those words and their related information are collected in advance. b) A context format retention mechanism that formalizes and retains the inclusion relations between semantically coherent intervals guided by the vocabulary of the above-mentioned context format retention mechanism; c) A context structure matching mechanism that detects candidates for a context structure, which is the structure of the semantic cohesion of the entire input sentence, by comparing it with retained lexical information, prior to syntactic analysis or semantic analysis; c) the above context structure; a context structure estimation mechanism that estimates the most probable context structure candidate from the multiple context structure candidates in the input sentence obtained by the matching mechanism, and the system estimates the logical structure and context of the input sentence prior to syntactic and semantic analysis. This is a natural language context analysis method characterized by having a context analysis unit that makes it possible to estimate the structure. (Operation) In the present invention, expressions in natural language that function to indicate the context structure and relationships between these expressions are held as lexical information in the context form holding mechanism. A context structure matching mechanism that matches this context information with the input sentence allows detection of context structure candidates in the input sentence at the stage of lexical analysis. A context structure estimation mechanism that estimates the most probable structure from among the context structure candidates obtained can estimate the context structure in the input sentence prior to syntactic analysis or semantic analysis. (Embodiment) Next, an embodiment of the present invention will be described with reference to FIGS. 1 to 3, FIG. 5, and FIG. 6. FIG. 1 is a block diagram showing one embodiment of the present invention. An input sentence read into the input sentence reading unit 1 via the communication line 10 is transmitted to the morphological analysis unit 2 via the communication line 12, and is divided into words based on the dictionary 3. Dictionary contents are assigned to each word and transmitted to the context analysis unit 4 via the communication line 26. The context analysis unit 4 includes a context format holding mechanism 5, a context structure matching mechanism 6, and a context structure estimation mechanism 7. The context structure matching mechanism 6 that receives the dictionary content string assigned to the input sentence identifies candidates for the context structure in the input sentence by referring to the vocabulary information indicating the context stored in advance in the context format storage mechanism 5. Detect. The context structure estimation mechanism 7 estimates the most probable context structure from among the context structure candidates received from the context structure matching mechanism 6 via the communication line 67, and transmits it to the syntax analysis unit 8 via the communication line 78. do. The syntactic analysis unit 8 performs syntactic analysis based on the estimated context structure, and the subsequent semantic analysis unit 9 extracts the meaning of the input sentence based on the structure of the input sentence obtained as a result of the syntactic analysis. In order to simplify the explanation, the function of the context analysis unit 4 will be explained below using Japanese as an example. The context format holding mechanism 5 comprehensively stores information about words characterizing the context as shown in FIG. Expressions that appear in natural language sentences and indicate context are registered here as vocabulary information. In the case of Japanese, such context feature words include function words such as conjunctive particles, and types and conjugations of predicates. These context characteristic words differ in their ability to create breaks in sentences depending on their type. The ``level of leading clause'' in Table 1 represents the difference in the ability of each context feature word to break a sentence. The smaller the number, the greater the force that causes the cut. A clause led by a context feature word with a small breaking power falls within a clause led by a context feature word with a large breaking power. On the other hand, a clause led by a context feature word that has a strong breaking power will not fit within a clause led by a context feature word that has a small breaking power. For example, when the conjunctive particle ``tama'' and the conjunctive particle ``node'' exist in this order in a sentence, ``tama''
The target of the clause that is led by is within the clause that is led by ``because''. Furthermore, each context feature word has certain conditions on its destination. In Figure 5, the columns ``Easy to use words'' and ``Difficult words'' indicate the difference in ease of use depending on the destination of each context characteristic word. For example, if we consider a case where there is a conjunctive particle ``kara wa'' in a sentence, followed by an adjective ``beautiful'' and a judgmental auxiliary verb ``must'', then according to Table 1, the clause ``kara wa'' leads to tends to fall into clauses led by ``must''. Next, the operation of the context analysis unit 4 will be explained in detail using an input sentence example. The following is an example of the output result of the morphological analysis unit 2 for one Japanese input sentence example, and this is an example of the input of the context analysis unit 4.

[Table] Conjunctive particles Noun Case particles Noun Case particles Amageru/Yo ni/Natsu/Ta/. /
Verb Auxiliary verb Auxiliary verb Auxiliary verb Phrase When receiving the above example sentence as input, the context structure matching mechanism 6 searches the contents of the context form holding mechanism 5, that is, FIG. 5, to find the context characteristic word in the input sentence. What is shown in FIG. 2 is a flowchart representing the operation of the context structure matching mechanism 6. As a result of checking with the context form retention mechanism 5, the conjunctive particle "tama" is first focused on as the context feature word. "Tama"
Possible destinations for clauses that lead to are: moving clauses with the conjunctive particle ``kotonori'', clauses with the assertive particle ``deru'', clauses with the verb ``protect'', clauses with the conjunctive particle ``node'', and sentence-final clauses with the conjunctive particle ``kotonori''. There are five possible clauses that contain the auxiliary verb sequence ``Yonanatsuta.'' The context structure matching mechanism 6 calculates and stores the ease of occurrence of the context feature word "tama" and its five possible destinations. Similarly, four clauses are considered as possible destinations for the clauses led by the conjunctive particle ``kotonori'' as a context characteristic word, and one clause is considered as the destination candidate for the clauses led by the conjunctive particle ``node''. Then, the ease of application to the candidate is calculated with reference to the context format storage mechanism 5, and the result is stored. Figure 6 shows the context structure matching mechanism 6 for the above example sentence.
This is the memory content of the analysis result of the input sentence, and the context structure estimation mechanism 7 along with the dictionary content string of the morphological analysis result of the input sentence.
becomes the input. FIG. 3 is a flowchart showing the operation of the context structure estimation mechanism 7. Upon receiving information on the candidate context structure for the above input sentence example from the context structure matching mechanism 6 in the form shown in FIG. 6, the context structure estimation mechanism 7
In Figure 6, the level of the leading clause and the ease of getting involved.
The most likely context structure is estimated from Since the level of the clause that the context feature word ``tama'' leads to is level 3, its destination is within the clause ``by sleeping'', which has a tendency to break at level 2. The target candidate in that clause is "by sleeping", and the ease of triggering is "normal", so as a result, the target of the clause led by "tama" is "by sleeping". It can be inferred that it is the most probable. Since the level of the clause that the context characteristic word ``by'' leads to is level 2, its destination is within the clause ``because I was able to'', which has the property of being a level 1 break. Possible targets in that clause are ``I am a carnivore,'' ``I defended myself,'' and ``I was able to do so.''
There are three types, but by comparing the ease of occurrence, it can be inferred that ``because I was able to'' is the most likely destination for a clause that is derived from ``by that''. The candidate destination for the clause that the context feature word ``node'' leads to is the sentence-final phrase ``I came to admire.'' Comparing the levels, both of them are 1, and ``easiness to apply'' is ``easy to apply,'' so as a result, ``I came to admire'' is the most likely destination for the clause that ``because'' leads to. It can be inferred that In this way, the context structure estimation unit 7 estimates the most probable syntactic structure and transmits it to the context analysis unit 8. The most probable context structure for the above input sentence example is as follows. The node located at the base of the arrow extends over the node located at the tip of the arrow, indicating that it is contained within that node. Primitive people kept a fire lit ↓ By sleeping ↓ They were able to protect themselves from carnivorous animals such as tigers and wolves ↓ They came to worship fire as a god In this way, the parser 8 The context structure can be estimated using vocabulary information before the semantic analysis unit 9 performs the analysis. (Effects of the Invention) According to the present invention, since expressions in natural language that have the function of indicating the context structure can be used as lexical information, the context structure analysis of the input sentence can be performed prior to syntactic analysis and semantic analysis. be able to.
By doing so, it is possible to streamline the syntactic analysis and semantic analysis that follow the context structure analysis, and improve the quality of the analysis. Therefore, it is possible to minimize the preprocessing of sentences and the constraints of limited languages, which are necessary to improve the quality of syntactic and semantic analysis.

[Brief explanation of the drawing]

FIG. 1 is a block diagram showing one embodiment of the present invention. FIG. 2 is a flowchart showing the operation of the context structure matching mechanism 6 shown in FIG. FIG. 3 is a flowchart showing the operation of the context structure estimation mechanism 7 shown in FIG. 1, FIG. 4 is a block diagram showing the conventional context analysis method, and FIG. 6 are diagrams showing an example of the analysis results of the context structure matching mechanism. In the figure, 1... input sentence reading section, 2...
Morphological analysis section, 3... Dictionary, 4... Context analysis section,
5...Context form holding mechanism, 6...Context structure matching mechanism, 7...Context structure estimation mechanism, 8...Syntax analysis section, 9...Semantic analysis section.

Claims (1)

[Scope of Claims] 1. A natural language analysis system having a function of reading an input sentence, a function of looking up the input sentence in a dictionary, and a function of analyzing the input sentence using the dictionary information after looking up the dictionary, which includes: a) complex sentences, etc. Vocabulary information that indicates semantically coherent sections within a long sentence is collected in advance, and information on the ease and difficulty of dependence of those vocabulary and the semantically coherent sections guided by those vocabulary are collected in advance. b) a context format retention mechanism that formalizes and retains the inclusion relationship between the two; b) collating the dictionary content string obtained as a result of morphological analysis of the input sentence with the lexical information retained in the context format retention mechanism described above; c) A context structure matching mechanism that detects candidates for a context structure, which is the structure of the semantic cohesion of the entire input sentence, before syntactic analysis or semantic analysis; a context structure estimation mechanism that estimates the most probable context structure candidate from a plurality of context structure candidates; A natural language context analysis method characterized by having an analysis section.