JPS62221775A - Natural language processing system - Google Patents

Abstract

PURPOSE:To facilitate the extension of a grammar rule and at the same time to reduce the description amount of the rule, by defining a language model with an object of a class level and describing the grammar rule within the class in division into rule sets. CONSTITUTION:A word model 14 expresses the attribute relation and the high- order/low-order relation between objects in terms of the subject field related to the language meaning. A grammar rule is described in division into rule sets within a class where a language model 15 formed in a hierarchical structure by an object of a class level has a high-order/low-order rank relation. A modeling system 16 controls the objects forming a class. When an input language like a Japanese sentence, etc. is inputted through an input part 11, a word division processing part 12 refers to a word dictionary 13 to divide the input sentence into words. A sentence structure/meaning analysis part 17 uses the model 15 to interpret the sentence structure and produces a sentence structure analysis tree for input sentences. Then the part 17 makes use of the model 14 based on the sentence structure/analysis tree to interpret the meaning and produces a speech act of the input sentence.

Description

[Detailed Description of the Invention] [Summary] A natural language processing method that performs syntactic and semantic analysis of natural language using a world model and a language model with a hierarchical structure, in which the language model is defined by class-level objects. By dividing the grammar rules into rule sets within that class, it is easy to expand the rules, and
These rules are written in classes that have a superordinate relationship, and the inheritance of the rules is used to reduce the amount of rules written.

[Industrial application field]

The present invention relates to a natural language processing method that analyzes input natural language such as Japanese sentences so that it can be understood by a computer.In particular, the present invention relates to a natural language processing method that can improve the extensibility of grammatical rules without reducing the amount of description. It is related to the method.

[Conventional technology]

In order to create an easy-to-use database search system or machine translation system, computers must be able to interpret the natural language input correctly as intended by the user. In order for a computer to understand natural language, the computer generates a correct syntax tree, and for each word, it grasps the correct semantic concept corresponding to the target field expressed by the input sentence.

The above syntax tree is difficult to solve solely from a grammatical point of view, and in order to obtain correct modification relationships, a world model showing semantic relationships regarding target objects is used. The world model expresses the attribute relationships and superior/subordinate relationships between objects 1 and each other in fields related to the meaning of language.As a modeling system using this world model, for example, The number etc. are known.

When performing syntactic analysis of natural language, a language model is created that describes the grammatical rules, and based on this language model, the grammatical function of each word that makes up the sentence is analyzed. ing.

Conventionally, as a method of writing grammar rules according to the language of the input sentence for such syntactic analysis, there is a method of writing the rules in a single bucket (j('?)).

[Problem that the invention seeks to solve]

With the rule writing method, there is a problem that it is difficult to modify and expand according to the language of the input sentence, and with the packetization method (the rules are divided, the in/return of the rules is Because this is not possible, there was a problem in that the same rules had to be written over and over again.

The present invention has been made to solve the above problems,
The purpose of this invention is to provide a natural language processing method that facilitates rule modularization and rule expansion, and reduces the amount of rule description.

[T stage for solving problems]

FIG. 1 shows a basic configuration diagram of the natural language processing method of the present invention.

In the figure, 11 is an input unit for inputting natural language;
is a word division processing unit that divides an input sentence into words according to information in a word dictionary I3, and the word dictionary 13 stores information such as parts of speech and objects represented in the world model 14 corresponding to each word. ing. In addition, the above world model 1
4 expresses attribute relationships and superior/subordinate relationships between objects in a target field related to the meaning of language.

Reference numeral 15 denotes a language model configured in a hierarchical structure using class-level objects, in which grammar rules are divided into rule sets and described in classes that are in a superior-subordinate relationship. 16 is a modeling system that manages objects that make up a class; 17 is a syntactic/semantic analysis that uses the world model 14 and the language model 15 to generate a parse tree for an input sentence, and generates a speech act for the input sentence; Department.

[Effect]

In the present invention, when an input language sentence such as a Japanese sentence is input by the input unit 11, the word division processing unit 12 refers to the word dictionary 13 and divides the input sentence into words. syntax·
The semantic analysis unit 17 generates a parse tree of the input sentence by interpreting the syntax using the language model 15 having a hierarchical structure, and also creates a world model 1 based on this parse tree.
4 to interpret the meaning and generate speech acts of the input sentence.

In the case of the present invention, in the class object of the language model,
Grammar rules are classified into two words, one sentence, and other grammatical attributes, and these are written in classes that have a superior-subordinate relationship. This makes it possible to inherit rules, and this reduces the amount of rules written. Decrease. Furthermore, since grammar rules are divided and described as rule sets within a class, modularization and expansion of rules becomes easy.

〔Example〕

FIG. 2 is an example of a system configuration to which the present invention is applied, FIG. 3 is an example of a language model related to the present invention, FIG. 4 is an example of a world model also related to the present invention, and FIGS. 5 and 6 is an example of the ruleset description for the object “verb sentence class”, No. 7
The figure shows another example of a language model relevant to the present invention.

FIG. 2 shows an example in which the natural language processing method of the present invention is applied to a database search system.

In this Figure 2, 20 is a dictionary for kana-kanji conversion;
1 is an input/output device such as a keyboard/display;
This corresponds to the input section 11 in the figure. Reference numeral 22 denotes an interactor that performs interaction between the user and the system, and 23 is a morphological analysis unit corresponding to the word division processing unit 12 shown in FIG. 1, which refers to the word dictionary 13 and divides the input sentence into words. , the attribute described in the dictionary class in the word dictionary 13 is added to each word in the input sentence.

The syntactic/semantic analysis section 17 is a main part of the present invention, and generates a syntactic analysis tree of an input sentence and generates a speech act of the input sentence, as in the case shown in FIG. When the input sentence is the second or subsequent sentence and it is determined that it is related to the previous sentence, the context processing unit 24 executes the syntactic/semantic analysis unit 17.
Modify the speech act generated by the speech act generated from the input of the preamble. This modified speech act is the output of the context processing section 24. Further, if no relationship with the preamble is found, the output of the syntax/semantic analysis section 17 becomes the output of the context processing section 24 as is.

The paraphrase generation unit 25 executes processing to show the user the interpretation of the input sentence of the system and obtain confirmation from the user, and generates a Japanese sentence from the speech act that is the output of the context processing unit 24 to explain the meaning to the user. Ask for confirmation.

The command generation unit 26 generates a command for database search from the speech act that is the output of the context processing unit 24. The data search unit 27 sends a command to the database management system 29 and searches the database 30 . The response generation unit 28 displays the search results of the database 30, and indicates the number of data searched initially. Furthermore, when the user requests display, the search data is displayed sequentially. Modeling System 162 World Model 1
4 and language model 15 are similar to those shown in FIG. 31 is a model editor that edits information regarding the modeling system including the word dictionary 13, the world model 14, and the language model 15.

FIG. 3 shows details of the language model 15, and explains how language models are classified and how rules are written and operated in the language model.

In Figure 3, the language model (LM) classes are roughly divided into three classes, with the "unit" class at the top, and the "word" class, "clause" class, and "sentence" class below it. ing.

The above "word" class is r? ii particle" class, "verb" class, 1-noun" class, and 1-pause" class, and the "auxiliary" class is further classified into the "particle" class and the "auxiliary verb" class. Also, the "verb" class is "
The verbs are classified into three types: the "arc verb" class, the "class verb" class, and the "general verb j class."
Class verb is defined as "class". For example, for a sales model, "sell", "ship", etc. Verbs corresponding to the arcs of the world model (those connected by broken lines in Figure 3) are defined as the "1-arc verb" class. This is, for example, the verb ``There is one in the sentence ``Factory j in Nakaatsu''. This verb corresponds to the arc connecting the factory location and the factory on the world model. Put verbs that do not correspond to the world model into the "general verb" class (e.g., display.

I want to see it, etc.) The "noun" class has a "class noun" class below it, and the "class noun j class has "
It is divided into three classes: the "interrogative word" class (for example, it, how much, etc.), the "4 compound noun class" (for example, 500 million Å or more, north of Tokaido, etc.), and the "value noun" class. The "value noun" class can be a value when instantiated in the world model, and has a Value attribute in the dictionary. For example, [I sold 200 bottles of beer at ○○ shop. ”, when the store name “○○ya” is instantiated in the “store name” class, “○○ya” is written as a value in that class.

``The period j class is a class corresponding to the period and comma of the input sentence.

The "1 clause" class is divided into three classes: a "1 noun clause" class, a "1 punctuated clause" class, and a "verb clause" class.

Here, the ``particle'' class is the ``1 noun'' class or the ``1 noun'' class.
It is combined with the ``Verb'' class to generate the ``Noun Clause'' class and the ``Verb Clause'' class. It is also possible to combine nodes to generate new nodes.

The "verb clause" class is classified into one of three types: "arc verb" class, "1 class verb" class, and "general verb" class, similar to the "1 verb" class of the "word" class above, and furthermore, the verb In the conjugated form, "arc verb final form"
It is subdivided into the ``general verb final form'' class, and the ``class verb final form'' class. For example, "produced" in "500 bottles of beer were produced" is the final form of the class verb. Further, the "verb clause" class has the "final form" and "adjunctive form" classes below it, and the "final form" class is related to each of the lower arc, general, and class conjugative final form classes. For example, ``sell'' corresponds to the adnominal form of the verb clause.

The "Sentence" class is classified into the "1 Noun Sentence" class and the "Verb Sentence j class. The "1 Noun Sentence" class is classified into the "Ha Noun Sentence" class.
The class is divided into three types: the ``noun sentence'' class, the ``noun sentence'' class, and the ``desu noun sentence'' class.The ``verb sentence'' class is classified into arc,
The verbs are divided into three verb classes: general and class, and are further classified according to the inflected form of the verb as shown in the figure.

In this case, you can also insert a noun clause in the middle to tie it together.

The language model configured in a hierarchical structure in this way is defined by object 1- at the class level, and the grammar rules of the class object width are divided into rule sets and described. These classes then have a super-sub relationship, and the ruleset is triggered on a message run,
Furthermore, the class object has other methods and data for language analysis.

For example, in the "verb sentence" class, the first rule is to generate a sentence when the previous buffer L1 is a noun, the second is the rule to generate a sentence when the previous buffer L1 is nil, and the third is to end the sentence generation. Three types of rules are described as a ruleset.

FIG. 5 shows a description example of a ruleset corresponding to an object of the "verb sentence" class. "@5uper J is a slot, [\value J is a facet, "sentence Δ
is Basoyou. The slot "'@" indicates that it is a system-defined slot name, and the facet "\" indicates that it is a system-defined facet name.

The slot SLI in the first line indicates that the upper class is "sentence". SL2 indicates control information and “
'doall' instructs that all rules that satisfy the conditions should be executed.SL3 is
Shows a list of local variables used in rule cents. This “@sea:case-frame” is a local variable. SL4 indicates an argument list when calling the ruleset. SL5 indicates by "t" information that when the trace information is hidden and the ruleset is applied, write the before and after states of the rule, that is, the states before and after execution. SL6 indicates a variable list to be output before executing a rule when applying trace information. SL7 is similar to SL3 and shows a list of variables used in the rulesent. SL8 indicates information to write the specified variable when the execution of the rule is completed when the trace information output command is issued. SL9 is an example of writing a rule, and rule ■ is, as shown in Figure 6 (al), if the left buffer Ll position is a noun clause, write it below the verb sentence at the current C position of the right buffer. This is a rule that says to attach.Rule
), if there is nothing in the left buffer L1, the rule is to shift the current position C of the right buffer one position to the left. Further, rule (2) is a rule for ending a verb sentence, as shown in FIG. 6(C), and is a rule for outputting the syntax tree when it is completed.

Since the rules are written in parts in this way, the rules are modularized and can be easily expanded.

In addition, there are also classes such as ``ha noun sentence'' class 5, ``ka noun sentence'' class, and ``ka noun sentence'' class.
Rules common to the "Noun Sentence" class are described in the "Noun Sentence" class, which is a higher class. Therefore, the rule set for the "noun sentence" class includes a rule called "rule noun sentence," which functions in common to all of the above classes. That is, the rule is that when a noun sentence is completed, the result is output. This rule does not exist in the ``ha noun sentence'' class, ``ka noun sentence'' class, and ``desu noun sentence'' class, but by using the inheritance of the rule, ” class and is invoked. Therefore, the amount of rules to be written will be reduced.

That is, according to the above embodiment of the present invention, a language model is defined by class-level objects, and grammar rules are divided into rule sets and described within the class, thereby making it easy to modularize and expand rules. , and the amount of rules written can be reduced by using the inheritance of the rules.

Furthermore, Figure 4 shows an example of a world model for checking the consistency of the meaning of an input sentence.A solid line indicates that there is an attribute relationship between each object, and a broken line indicates that there is a superior-lower relationship between objects. represents.

FIG. 7 shows another embodiment of the language model according to the present invention. The difference from Fig. 3 is that the language model includes the "connective category" class, the "adjunctive connective" class, the "no" class, the "kaj class, and the "to" class (=J plus, one adjunct Above the "conjunctive form" class is the "conjunction category" class, below it are the "adjunctive form" class (both sentences and clauses), and the "no" class and "ka" under the "1 noun clause" class.
Class 3 This is the point where the class is connected.

In this example, the common connection rule is set to the "adjunctive connection" class in the "noj class,""to" class, and "ka" class, which are subordinate to the sentence adnominal class, the clause adnominal class, and the noun clause class. Described. For this reason, until now each class (2
The common connection rules that were written for each adjunct class or noun clause class have been removed, and the inheritance of rules written only for the upper class of these classes allows execution of the lower ones, and the rule description quantity can be reduced.

For example, in ``My house is'' or ``The house being sold'', ``House is'' is a noun clause, and the previous word is used as a noun clause.
Make the same connection. This common rule is written in the upper class.

Therefore, the embodiment shown in FIG. 7 has the effect of further modularizing the rules and reducing the amount of rules written.

〔Effect of the invention〕

As explained above, according to the present invention, the language model is structured hierarchically using class-level objects, and the grammar rules are divided into rule sets within the class, so that the rules can be expanded as the rules become modular. This has the effect of making it easier and reducing the amount of rules to be written by using the inheritance of the rules.

[Brief explanation of drawings]

FIG. 1 is a basic configuration diagram of the present invention, FIG. 2 is a block diagram showing an example of a database search system configuration to which the present invention is applied, FIG. 3 is a diagram showing a language model related to the present invention, and FIG. Figure 5 is a diagram showing a world model related to the present invention, Figure 5 is a diagram showing a description example of rule cents in the case of the object "verb sentence class", Figure 6 is a diagram also explaining the rule set, Figure 7 is a diagram showing the book FIG. 7 is a diagram showing another example of the language model of the invention. In the figure, 11 is an input unit, 12 is a word segmentation processing unit, 13 is a word dictionary, 14υ11゛ world model, 15 is a language model, 16 is a modeling system, and 17 is a syntactic/semantic analysis unit. Patent applicant Fujitsu Ltd. August Makusan Uori - Hirozu Figure 1 (: l-0 (J

Claims (1)

[Claims] A word division processing unit (12) that divides an input sentence in natural language into words according to a word dictionary performs semantic analysis using a world model (15) that expresses attributes and superior-lower relations between objects in fields related to the meaning of language. At the same time, it is equipped with a syntactic/semantic analysis unit (17) that performs syntactic analysis using a language model, defines the language model using class-level objects in a superior-lower relationship, and divides and describes the grammar rules within the class into rule sets. A language processing method characterized in that the inheritance of rules is made possible by using the superior-subordinate relationship between these classes.