JPH0512753B2 - - Google Patents

Description

[Detailed description of the invention]

<Technical Field> The present invention relates to a monocultural processing method for Japanese sentences used in applied fields of natural language processing such as question answering systems and machine translation, and in particular, the present invention relates to a monocultural processing method for Japanese sentences used in applied fields of natural language processing such as question answering systems and machine translation. This allows the surface structure of a sentence to be expressed using . <Prior art> In many conventional question answering systems, the results of syntactic analysis (or semantic analysis) are directly converted into the system's own internal structure, or syntactic analysis is performed in accordance with the database format. . For this reason, conventional systems such as those described above have problems in that the processing results lack generality, or the process for obtaining internal representations becomes unpredictable. On the other hand, in the field of natural language, a general method for expressing Japanese dependency relations concisely for machine processing has not been established, and is currently still at the research stage. <Purpose> The present invention has been made in view of the above-mentioned conventional problems, and it is possible to easily extract the internal structure of a natural language processing system that requires a deep level of internal representation of an input sentence. The purpose of the present invention is to provide a monocultural processing method for Japanese sentences that can extract simple sentence relation trees that can be easily processed by machines, and the complex dependency relations of Japanese can be easily processed by machines. <Example> The present invention will be explained in detail based on the following figures. FIG. 1 is a diagram showing a hardware configuration for executing the monocultural processing method for Japanese text according to the present invention. In the figure, 1 is the central processing unit (CPU),
2 is a main memory, 3 is a Japanese parser (Japanese syntax analysis system), 4 is a simple sentence relationship tree extraction module, 5 is a keyboard, and 6 is a CRT display device.
In this hardware configuration, sentences are input using the keyboard 5 and a simple sentence relationship tree is displayed on the CRT display device 6.
Processing is also carried out by a module of the Japanese parser 3 which outputs a parse tree and a surface case structure, and a simple sentence relation tree extraction module 4 which extracts a simple sentence relation tree from the output. FIG. 2 shows the processing flow up to finding a simple sentence relationship tree, which will be explained in order below. Morphological processing The input sentence is checked against the words stored in the dictionary of the kana-kanji conversion system and divided into word units.
By looking up the dictionary, morphological information such as part of speech and conjugation form of each word is provided. Syntactic analysis The phrase structure of a sentence is determined by applying grammatical rules based on the part-of-speech sequence and grammatical information corresponding to the input sentence given through morphological processing. Figure 3 is an example of a parsing tree (results of parsing) for the sentence ``If there is a ball in the red box, take it out and move it to the yellow box.'' The resulting parsing tree is Stored in the parse tree buffer. Case analysis Case analysis here refers to the extraction of case dominance relations on the surface level. For example, by using a case pattern dictionary of predicates for the parse tree, a surface case structure as shown in FIG. 4 is obtained and stored in a case structure buffer. Note that the surface case structure is a classification of nominal words corresponding to predicates by case particles. Simple Sentence Relationship Tree Extraction Process A simple sentence relationship tree is constructed by examining the relationships between predicates in the surface case structure from the parse tree and the surface case structure, and rearranging them in a certain order. FIG. 6 is an example of a simple sentence relation tree (see FIG. 5) stored in the simple sentence relation tree buffer. In this way, the input sentence is divided into words using the kana-kanji conversion system, processed into morphemes, and the morpheme data is input into a Japanese parser to obtain a parse tree and surface case structure.Based on these data, A simple sentence relationship tree with a binary tree structure is extracted from simple sentences and information indicating relationships between the simple sentences. Here, the basic relationships between simple sentences in the simple sentence relationship tree are considered, for example, as shown in Table 1 below.

[Table] In other words, (1) in the table is an example in which the input sentence includes an adnominal modifier. (just insert a dummy to indicate the pronoun), and each simple sentence is interpreted in terms of the relationship of "and".
In addition, (2) in the table is an example of a complex sentence that includes a conjunctive particle, a disjunctive phrase, or a conjunction. It is classified into whether it is an ``or'' relationship or an ``or'' relationship. Furthermore, (3) in the table is an example of a sentence that includes parallel phrases, and in this case, the relationship is classified as ``and'' or ``or'' depending on how the parallel phrases are connected. Next, the concept of a simple sentence relation tree will be explained based on FIG. This example is an input sentence: If there is a ball in the red box, take it out and move it to the yellow box. The output tree for / is shown. In the figure, "IF", "AND", and "RENT" indicate the relationship between simple sentences A to E. "IF" is an "if-then relationship,""AND" is an "and relationship," and "RENT ” represent the “and relationship by adjunctive modification,” and in particular, “IF” in this output tree is called the top node (or root), “AND” and “RENT” are called non-terminal nodes, and they also represent a simple sentence (single sentence). (Phrase) A to E are called terminals (or leaves). Incidentally, in this example, terminal node A has a ball in the box. /, B is / The box is red. /, the same C /takes it out/, the same D /moves it to the box/, the same E / the box is yellow. / is shown. In this way, the input sentence is parsed to obtain a parse tree, case analysis is performed to obtain the surface case structure, and based on the obtained parse tree and surface case structure, a simple sentence (single case structure) is Examine the relationship between (words and phrases),
A simple sentence relationship tree with a binary tree structure as shown in FIG. 5 is constructed using each simple sentence and relationship name.
The obtained simple sentence relation tree is then stored in a simple sentence relation tree buffer in the form shown in FIG. 6, for example. In Figure 6, the number entered in the node column indicates the item number, and in this example, the item number is 0.
The left node in the relationship name "IF" is item number 1
The relation name of "RENT" is the relation name of item number 2, that is "AND", and the left node of the relation name "RENT" is simple sentence A, and the right node of the above relation name is simple sentence B. Furthermore, the left node of the above relationship name "AND" is simple sentence C, and the right node is the relationship name "RENT" of item number 3. Similarly, the left node of the relationship name "RENT" is simple sentence D, and the right node is simple sentence C. Each indicates that it is a simple sentence E, and it is designed so that the contents of this simple sentence relation tree buffer can be expanded into a simple sentence relation tree as shown in FIG. 5 for recognition. FIG. 7 is a flowchart showing the procedure for storing data of a simple sentence relation tree into a simple sentence relation tree buffer based on the parse tree and surface case structure. However, this example assumes that each item in the case structure has information about the node number in the parse tree. First, the parse tree obtained by parsing the input sentence (see Figure 3) is
(equivalent to ) (n1), and if the first phrase encountered is a simple sentence, it is treated as a “Single” relationship (n2, n9). If it is not a simple sentence, then by examining the structure of the phrase, we determine in what relationship the items of which case structure are connected, and we use the relation name determined from the head of the simple sentence relation tree, The case structure pointer of the pragmatic phrase and the case structure pointer of the right pragmatic phrase are respectively written (n3). Next, if there is an item of case structure located further to the left of the left-hand predicate that has not been searched yet, node searches are performed from the left-hand predicate until the next predicate appears ( n4, n5,
n3). The same process is repeated for the next predicate, and if there is no predicate further to the left than the left predicate, processing moves on to the right predicate (n6, n7). The processing of the phrases on the right side is the same as the processing of the phrases on the left side described above, but when there are no more phrases that have not been searched, the process moves to the next step. In step n8, the second and third items of the simple sentence parse tree buffer that are not originally terminal nodes (in Figure 6, the number of items is 3) are replaced with pointers pointing to their own items. In this way, a simple sentence relation tree buffer as shown in FIG. 6 can be constructed. <Effects> As explained in detail above, the monocultural processing method for Japanese sentences of the present invention processes input sentences word by word using a kana-kanji conversion system, then processes the morphemes, and converts the morpheme data into Japanese. I introduced it into the parser to obtain a parse tree and surface case structure, and based on this information, I extracted a simple sentence relationship tree with a binary tree structure consisting of simple sentences and information indicating relationships between simple sentences. It can have a great effect. ○B Adnominal modification relationships, AND-OR relationships, and IF-
Various relationships contained in a sentence, such as THEN relationships, can be expressed in the same format. Furthermore, since the format is a binary tree structure (called a simple sentence relation tree) based on the relationships between simple sentences, it is possible to easily access the constituent elements and relationships of sentences. ○B Because the minimum component of a sentence is treated as a simple sentence (a predicate phrase that always includes only one pragmatic phrase, such as ``flowers are red'' or ``I look at him''),
The process of converting to internal representation has become much easier. For example, it is possible to create a desired database by making some modifications. ○C Since the simple sentence relation tree expresses the surface case structure, it does not lose its generality. In other words, it can be used for any system by simply creating the related node portion according to the purpose of the system. The conventionally used deep case structure loses its incommensurability if the deep cases differ, but the simple sentence relation tree has the advantage that it can be converted into any form because its smallest unit is a surface case structure. be.

[Brief explanation of drawings]

Figure 1 is a diagram showing the hardware configuration for implementing the monocultural processing method for Japanese text according to the present invention,
Figure 3 is a diagram showing the processing flow up to finding a simple sentence relation tree, Figure 3 is a diagram showing a parsing tree, Figure 4 is a diagram showing surface case structure, Figure 5 is a diagram showing a simple sentence relation tree,
FIG. 6 is a diagram showing a simple sentence relation tree buffer, and FIG. 7 is a flowchart showing a procedure for storing data of the simple sentence relation tree buffer into the simple sentence relation tree buffer. 1 is the central processing unit, 2 is the main memory, 3
is a Japanese parser, 4 is a simple sentence relation tree extraction module, 5 is a keyboard, and 6 is a CRT display device.

Claims (1)

[Scope of Claims] 1. Separation processing means for collating an input sentence with a dictionary in a kana-kanji conversion system and dividing it into word units; morphological processing means for providing morphological information; determining the phrase structure of a sentence based on the part of speech and the conjugated form corresponding to the input sentence determined by the morphological processing means;
a parsing means for obtaining a parse tree from the phrase structure; a case parsing means for extracting case domination relationships on the surface layer from the parse tree to obtain a surface case structure; and a simple sentence relation tree extracting means for obtaining a simple sentence relation tree from the relations between predicates in the surface case structure, wherein the relation between the predicates in the surface case structure means that the input sentence includes an adnominal modifier phrase. Each simple sentence is “and” and each simple sentence is “if”.
A Japanese sentence unit characterized by having the following relationships: "Then, then,", each simple sentence has the "and" relationship, and each simple sentence has the "or" relationship. Culture processing method.