Sentence translator using a thesaurus and a concept-organized co- occurrence dictionary to select from a plurality of equivalent target words

• US 4,916,614 A
• Filed: 11/23/1987
• Issued: 04/10/1990
• Est. Priority Date: 11/25/1986
• Status: Expired due to Fees

• Current Assignee: Hitachi, Ltd.
• Original Assignee: Hitachi, Ltd.
• Inventors: Shimizu, Fumon, Kaji, Hiroyuki
• Primary Examiner(s): NOT, DEFINED
• Application Number:
  US07/123,791
• Time in Patent Office: 869 Days
• Field of Search: 364/419, 364/200, 364/900
• US Class Current: 704/2
• CPC Class Codes: G06F 40/247 Thesauruses; Synonyms
• CPC Class Codes: G06F 40/289 Phrasal analysis, e.g. finite state techniques or chunking
• CPC Class Codes: G06F 40/55 Rule-based translation

Abstract

In a system of performing automatic translation from natural language into another natural language based on dependence relation among concepts, a thesaurus representing implication among concepts and a co-occurrence dictionary storing co-occurrence relation between a predicate word and an argument concept are utilized to select a suitable translation word when a plurality of candidate words exist corresponding to a predicate concept. Furthermore, knowledge regarding the co-occurrence relation is obtained based on feedback information from a revise operation to translation results, thereby contents of co-occurrence relation dictionary can be more and more complete.

Claims

1. A natural language generating system, which translates a source language text into a target language, said system comprises:

a word dictionary memory storing a predicate concept represented by a source language word and at least one target language word corresponding thereto;

a co-occurrence relation dictionary memory storing co-occurrence relation between a target language word and the predicate concept;

a thesaurus memory representing implication among predicate concepts represented by source language words; and

,

a processing device referring to the co-occurrence relation dictionary memory and the thesaurus memory when a plurality of target language words exist corresponding to a predicate concept represented by a source language word to be translated in the word dictionary memory, and selecting a target language word having co-occurrence relation with an argument concept corresponding to the predicate concept represented by the source language word to be translated or with the general concept thereof.

2. A natural language generating system according to claim 1,

wherein said word dictionary memory comprises:

a source language analysis dictionary constituted by records, each including a word of a source language, a part of speech of the word, and a concept symbol representative of the concept of the word, and further including a case frame when the word is one representing a predicate concept;

said case frame including case codes representing semantic roles of arguments governed by the predicate concept, and surface case markers;

a target language generation dictionary constituted by records, each including a concept symbol, one or more words of the target language corresponding to the concept symbol, and parts of speech of the respective words, and further including a case frame when the word is one representing a predicate concept; and

,

wherein said processing device comprises means for analyzing the syntactic and semantic structure of the source language sentence by referring to words of source language, parts of speech, and case frames in source language dictionary records;

mapping source language words onto concept symbols by referring to source language words and concept symbols in source language dictionary records;

mapping concept symbols and target language words in target language dictionary records; and

generating a target language sentence by referring to target language words, parts of speech and case frames in target language dictionary records.

3. A natural language generating system according to claim 2, wherein said co-occurrence relation dictionary is constituted by co-occurrence data including a target word, a concept symbol and a case code;

and,

said processing device refers to the co-occurrence relation dictionary to examine the existence of a triplet of a predicate word of target language, and argument concept and case code.

4. A natural language generating method which translates a source language text into a target language, said method comprises:

a first step of reading a source language sentence inputted from an input device, referring to a source language analysis dictionary stored in a memory, analyzing the syntactic and semantic structure of the source language sentence and case relations, and obtaining intermediate representation comprising dependence relations among concepts;

a second step of referring to a target language generation dictionary, a co-occurrence relation dictionary and a thesaurus memory, and determining at least one target language word corresponding to each concept in the intermediate representation; and

a third step of referring to the target language generation dictionary, determining an arrangement of the target words determined at said second step based on case frames of predicate words, and outputting a translation result to an output device.

5. A natural language generating method according to claim 4, wherein said second step comprises:

a first retrieval step of retrieving co-occurrence data comprising of a candidate word corresponding to the predicate concept, an argument concept thereof and the case of the argument concept from the co-occurrence relation dictionary, when a plurality of candidate words of the target language exist corresponding to a predicate concept in the intermediate representation; and

a second retrieval step of retrieving a general concept of the argument concept from the thesaurus when retrieval fails at said first retrieval step, and retrieving co-occurrence data of the predicate concept and the general concept of the argument concept from the co-occurrence relation dictionary.

6. A natural language generating method according to claim 4, which further comprises:

a fourth step of deleting a co-occurrence relation between an old word and the concept from the co-occurrence relation dictionary and adding a co-occurrence relation between a new word and the concept to the co-occurrence relation dictionary, when the target language word for a concept is replaced with another word in post-editing.

7. A natural language generating system for translating a source language into a target language comprising:

means for inputting the source language into the system;

means for outputting the target language out of the system;

processing means for translating a source language text into a target language comprising first memory means for storing a source language analysis dictionary including relational data of a source word, a part of speech, a case frame and an associated concept;

a second memory means for storing a target language generation dictionary similarly including relational data of a target word, a part of speech, a case frame and an associated concept;

a third memory means for storing a co-occurrence relationship comprising a predicate word of target language, an argument concept, and a case thereof; and

a thesaurus memory for storing a relationship of a concept and a general concept thereof, said processing means being in communication with the memories for selected access thereto whereby a target language sentence is generated by identifying the part of speech, the concept and the case frame associated with a source word from the first memory means, identifying the predicate-argument structure of a source language sentence on the basis of the case frames of source words representing a predicate concept and parts of speech source words, identifying a plurality of candidates of target words, for a predicate concept, and the part of speech and the case frame associated with a target word from the second memory means, identifying general concepts of an argument concept from the thesaurus memory;

selecting a candidate word is a triplet comprising the candidate word, an argument concept or a general concept of an argument concept, and the case of the argument is retrieved from the third memory means; and

determining the syntactic structure of a target language sentence on the basis of the case frames of target words representing a predicate concept and the parts of speech of target words.