JPH0833895B2 - Sentence generation processing method in machine translation system - Google Patents

Description

[Detailed Description of the Invention] [Outline] In sentence generation processing using a semantic network, the sentence generation rule has a module structure, and the sentence generation rule (grammar) added to each node is specified in module units. Other modules can be called to simplify the statement generation rule description structure and facilitate maintenance.

[Industrial applications]

The present invention relates to a sentence generation processing method from a semantic network in a machine translation system, and more particularly to a generation rule configuration method used for sentence generation.

[Conventional technology]

In the machine translation system, for example, in the case of a machine translation system from Japanese to English, a Japanese sentence is first analyzed to determine the conceptual structure (semantic structure) of the sentence. This conceptual structure can be represented by a tree structure or a semantic network. In the case of a semantic network, the conceptual elements are described by nodes to which semantic symbols are added and arcs to which relation information between nodes is added. A grammar (called a sentence generation rule) necessary for sentence generation is added to each node, and the semantic network is expressed as shown in FIG. 3 (a). Circles in the figure represent nodes, letters A and B in them represent semantic symbols of nodes, arrows represent arcs, and <C> and <ST> represent arc information or arc names.

FIG. 3B shows a simplified example of the semantic network. “Buy”, “I”, and “book” are the meaning symbols of nodes, and <ST (start)>, <past>, <predicate>, <subject>, and <object> are arc information.
The arc that goes to the node is called the in-arc, and the arc that leaves the node is called the out-arc.

Next, the semantic network generated in this way is analyzed, an appropriate translated word is selected, and processing for generating a target English translation is performed.

That is, the sentence generation rule attached to each node is activated, and based on the sentence generation rule, an English word that matches the semantic symbol and its grammatical attribute are searched, and then the word, part of speech, and sentence pattern that matches the arc information are searched. Is executed to generate a sentence.

[Problems to be solved by the invention]

In the conventional machine translation system described above, when generating a sentence from the semantic network, a method of sequentially tracing the nodes and applying the added sentence generation rule to each node has been adopted.

In this case, the sentence generation rules added to each node are formed homogeneously for each node. Therefore, the sentence generation rule becomes complicated and large-scaled, and it becomes difficult for anyone other than the creator to structurally understand it, resulting in a problem that maintainability is significantly reduced.

[Means for solving problems]

An object of the present invention is to make the structure of sentence generation rules applied to sentence generation based on a semantic network clear and compact. Therefore, the sentence generation rules are modularized, and at the same time from any module to another arbitrary one. The above-mentioned module can be called so that any desired sentence generation rule system can be easily realized with a relatively simple structure.

FIG. 1 shows the basic configuration of the present invention. In the figure, 1 is semantic network data, which includes semantic symbols of nodes and arc information.

A rule analysis routine 2 executes a sentence generation rule module. That is, semantic network data 1
The word corresponding to the semantic symbol is selected for each node according to the information of the arc of, and a translated sentence that matches the semantic network is generated. This rule analysis routine can be called recursively by specifying an arc name.

Reference numeral 3 is a word dictionary, which holds a word group of translated words that can correspond to the meaning symbols of the nodes.

Reference numeral 4 denotes a word list, which is a list of candidate words searched and extracted from the word dictionary 3 by the rule analysis routine 2 using the semantic symbols of the nodes. The words in the list are arranged in an order according to the frequency of occurrence.

Reference numeral 5 is a sentence generation rule, which is modularized according to the department (category) as an element. The smallest unit that constitutes a module is called a slot. Each module has a unique sentence generation rule name.

5a is an interpretation department module, which determines the part of speech and sentence pattern from the arc information and the grammatical attribute of the word.

5b and 5b 'are syntactic division modules, which generate each element of the sentence according to the determined sentence pattern. This module is further modularized according to the type of sentence, such as sentence, gerund, and infinitive.

5c is an organization department module, which supplements the details of sentence expressions such as periods, question marks, singular and plural.

[Action]

The operation of the principle configuration of the present invention shown in FIG. 1 is performed as follows.

The rule analysis routine 2 is called from a main routine (not shown) by designating an arc name <ST> indicating a start arc.

The called rule analysis routine 2 performs the processes shown in the figure as follows.

The semantic symbol (first A) of the arc destination node of the specified arc (first <ST>) is extracted.

The word dictionary 3 is searched using the extracted semantic symbols to create a word list 4.

The word with the highest appearance frequency is selected from the word list 4.

The sentence generation rule 5 module to be executed is determined from the sentence generation rule name of the selected word.

The slots of the determined modules are read sequentially, and the contents of the slots are specified by the entry arc (
If a process related to an out arc is specified, that process is executed. On the other hand, if an in-arc / out-arc arc process other than the entry arc or a node process is specified, each process is executed.

In the case of in-arc / out-arc processing, the rule name is specified and the rule analysis routine 2 is recursively called (Call). For example, when the arc name <C> is specified, a new rule analysis routine 2 is started to analyze the arc destination node B.

In this way, the rule analysis routine can be called multiple times during execution of any rule analysis routine 2. When the processing of the called rule analysis routine is completed, control is returned to the calling rule analysis routine together with the analysis result.

This allows any node in the semantic network to
The sentence can be generated by sequentially interpreting according to the connection of arcs, performing syntactic and simplification according to the result.

〔Example〕

 The operation of the embodiment of the present invention will be described with reference to FIG.

In FIG. 2, 11 is a message board, 12 is a rule counter, 13 is a rule counter stack, 14 is a VISIT flag, 15 is a sentence generation rule, 16 is a module, 17 is a word list, and 18 is a rule analysis routine (rule interpreter). , A and B represent the semantic symbols of nodes, and C and ST represent arc information.

The message board 11 is provided for each node and is used to store various messages. The messages include:

(A) Grammar attribute of a word.

(B) Sentence generation rule name (module name) (c) Sending message sent when control is passed from another node.

(D) Return message defined when returning control to the original node.

One rule counter 12 and one rule counter stack 13 are provided for each node. Rule counter
12 is used to specify individual slots in the module of the sentence generation rule being executed by the rule parsing routine 18.

The rule counter 12 is normally incremented by 1 for each execution, and specifies a sequential slot. However, if the contents of the executed slot are macro expansions (called% types) that call modules for other sentence generation rules, the rule counter
The value of 12 is stored (pushed down) in the rule counter stack 13.

When the content of the executed slot is an instruction to return control to the original module (called R type), the content of the rule counter 12 is cleared.

One VISIT flag 14 is provided for each arc and is used to indicate whether or not the arc has been processed.

 The VISIT flag 14 is set in the following cases.

(A) When the sentence generation rule of a node that interprets the arc as an entry arc ends normally.

(B) When the VISIT process (described later) is performed by the sentence generation rule of a node that does not interpret the arc as an entry arc.

The minimum unit of the sentence generation rule 15 is a slot, and each slot has 5 fields. The module 16 is composed of a plurality of slots. The name of each module is called the sentence production rule name. The group of modules 16 is the sentence production rule 15.

The module 16 is divided into interpretation, syntactic, and organizing departments, and the syntactic department is further divided into modules such as noun phrases, sub-phrases, plain sentences, interrogative sentences, and imperative sentences. In the figure, sentences, gerunds, and indefinite poems are illustrated.

The word list 17 is a set of words having semantic symbols of arcs or nodes, and the order thereof is determined by the frequency of appearance.

The rule analysis routine 18 sequentially fetches, interprets, and executes the slots of the module 16 of the sentence generation rule name determined by the semantic symbol of the node according to the contents of the rule counter 12.

The rule analysis routine 18 can be described as a recursive algorithm as follows. The main part of this program is TXTGEN
I will call it. The subroutine (actually the central part of execution) is called VISIT.

TXTGEN: PROC; CALL VISIT (<ST>); END; As described above, the sentence generation rule starts from <ST>.

VISIT: PROC (arc name) Brings in the semantic symbol of the node beyond the arc with the specified arc name.

 Search the English dictionary by the meaning symbol.

Select the most frequent word from the word list obtained by the search.

The module to be executed is determined from the sentence generation rule name of the word.

There is a DO WHILE slot.

IF slot meets conditions (message board, arc) THEN DO; IF slot is entry arc processing THEN Performs entry arc processing ELSE IF slot is arc processing (I, O) THEN CALL VISIT (specified in slot Arc name) ELSE IF Slot is the node's process THEN Node's process END; The next slot is brought.

END; [Effect of the Invention] According to the present invention, since the sentence generation rule is modularized, the structure of the sentence generation rule is simplified, and maintenance such as correction of the sentence generation rule is facilitated.

[Brief description of drawings]

FIG. 1 is a block diagram of the principle of the present invention, FIG. 2 is a block diagram of one embodiment of the present invention, and FIGS. 3 (a) and 3 (b) are explanatory diagrams of a semantic network. In Fig. 1, 1: Semantic network data 2: Rule analysis routine 3: Word dictionary 4: Word list 5: Sentence generation rule 5a: Interpretation module 5b, 5b ′: Syntactic module 5c: Organization module

Claims (1)

[Claims] 1. A machine translation system using a semantic network composed of nodes and arcs, which is modularized according to the type of sentence generation processing department such as interpretation, syntactic, and organization, and each module has a sentence generation rule name. When given, an arbitrary module has a sentence generation rule (5) with a structure that can indicate the next arc to be processed by an arc name if necessary, and a word group that can correspond to the semantic symbol of the node. Interpret the semantic network according to the registered word dictionary (3) with sentence generation rule name added to each word and the module of sentence generation rule (5), and determine the word corresponding to the semantic symbol of each node. It also has a rule analysis routine (2) that generates a sentence by performing syntactic and simplification. The rule analysis routine (2) is started by specifying an arc name. Then, the word dictionary (3) is referred to according to the semantic symbol of the arc destination node of the arc, the corresponding word is selected, the module of the sentence generation rule name added to the word is executed, and the module being executed is the arc name. Sentence generation method in machine translation system characterized by recursively calling rule analysis routine (2) by designating another arc using