JP2715875B2 - Multilingual summary generator - Google Patents

Description

Description: BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention analyzes a document described in a natural language, summarizes contents specified in advance, performs multilingual translation, and outputs the result. The present invention relates to a summary generation device. 2. Description of the Related Art Conventionally, a method of generating a summary sentence involves evaluating the importance of each sentence in a document based on an important word, a specific vocabulary, or a conjunction or end-of-sentence expression. There was a technique to compose a summary sentence. However, since these methods aim to generate summary sentences for general document contents, they cannot cope with various linguistic phenomena,
No output according to the user's intention was obtained. Information extraction methods that limit the target area by setting in advance the information that the user wants to know in detail and extract information with a certain fixed structure include, for example, “The 47th Annual Conference of IPSJ Vol. 3, page 83 ". The apparatus described here extracts information specified by a keyword relation calculation rule using a keyword and a syntax structure regarding a target area, and outputs the extracted information in a frame format. However, a format such as a frame requires a user to be familiar with the definition of each slot, and has a problem that it is difficult to understand especially when the structure of the designation information is complicated. Furthermore, although a number of machine translation devices have been proposed, there has been no device capable of translating a document including a long sentence such as a newspaper article with high quality for a general topic. For this reason, for example, in a newspaper article clipping service, an article retrieved by a keyword is provided as it is or after being summarized and translated manually. [0003] The summarization method described in the prior art aims at extracting general contents included in a document, and therefore requires a deep semantic analysis. However, since various linguistic phenomena appear in an actual document, it has been difficult to generate an accurate summary of the contents required by the user in response to all of these phenomena. In addition, in the summarization method proposed in the past, the definition of the importance of each sentence is uniquely provided in advance in the apparatus, and the contents to be summarized cannot be changed depending on what the user wants to know. A conventional information extraction method that targets only a preset area without targeting general contents can extract accurate information because the target area is specified in advance. However, since the output is given in a frame format, there is a problem that the user directly reads the output and it is difficult for a user to understand the output. Furthermore, in machine translation, there is a problem that ambiguity occurs at the syntactic level and the vocabulary level at the stage of analyzing the source language, so that the sentence cannot be analyzed if the sentence is long, and the context level beyond the sentence cannot be handled. Quality was bad. For this reason, in the case of newspaper article clipping, for example, article information obtained only by keyword search is used as a whole document or after performing work such as summarizing and translating manually. An object of the present invention is to generate a high-quality summary in multiple languages according to a user's specification for an input document. Here, since an accurate summary content can be extracted by designating a summary target in advance, an accurate summary according to the user's designation can be output. Also, since the user can change the summary target according to the use,
You can change what you summarize. Furthermore, since the information obtained by extracting the summary content is in a certain format without ambiguity, a high-quality translation result can be output. [0005] In order to solve the above-mentioned problems, the invented multilingual digest generation apparatus specifies a document input unit for receiving a document described in a natural language as an input, and designates a document to be extracted in advance. A keyword dictionary unit that stores detailed information of each keyword in the designated target and a keyword related to the summarized target, and a relationship between the keywords is identified using the keyword information stored in the keyword dictionary unit. An inter-keyword relation calculation rule storage unit that stores rules for selecting and synthesizing keywords that match the designated summarization target; a keyword information stored in the keyword dictionary unit and an inter-keyword relation calculation rule storage unit Document information that extracts the specified summary target as keywords and the relationship between keywords using the specified rules And a rule for converting a certain format consisting of the keywords output from the document information extracting unit and a relational structure between the keywords into an intermediate structure representing a sentence and, if necessary, dividing it into a plurality of intermediate structures. A sentence structure generation rule storage unit to be stored, and a sentence structure generation unit that converts the structure of the designated information output by the document information extraction unit into an intermediate structure representing a sentence using the rules stored in the sentence structure generation rule storage unit Part, a target language dictionary part indicating correspondence between each keyword and a target language, a sentence generation rule storage part storing rules for generating a natural language sentence from an intermediate structure representing a sentence, and a target language dictionary part stored in the target language dictionary part. A sentence generation unit that generates a natural language sentence from the intermediate structure output by the sentence structure generation unit using vocabulary information and rules stored in the sentence generation rule storage unit, and a natural language sentence output by the sentence generation unit Output and display It has a Yakubun output section. Next, the present invention will be described with reference to the drawings. FIG. 1 is a block diagram showing one embodiment of the first aspect of the present invention. Referring to FIG. 1, the present invention provides a document input unit 1 that receives a document described in a natural language as an input, and a keyword related to a summary target specified to be extracted in advance and a keyword specified in the specified target. Keyword dictionary 2 for storing detailed information
And a rule for calculating a relation between keywords storing rules for determining the relationship between keywords by using the keyword information stored in the keyword dictionary unit 2 and selecting and synthesizing a keyword that matches the specified summary target. Part 3
Document information for extracting a specified summary target as a relationship between keywords using the keyword information stored in the keyword dictionary unit 2 and the rules stored in the inter-keyword relationship calculation rule storage unit 3 Extraction unit 4
And a rule for converting a certain format consisting of the keywords and the relational structure between the keywords output by the document information extracting unit 4 into an intermediate structure representing a sentence and, if necessary, dividing it into a plurality of intermediate structures. A sentence structure generation rule storage unit 5 and a sentence structure for converting the structure of the designated information output by the document information extraction unit 4 into an intermediate structure representing a sentence by using the rules stored in the sentence structure generation rule storage unit 5 A generation unit 6, a target language dictionary unit 7 indicating correspondence between each keyword and a target language, a sentence generation rule storage unit 8 storing rules for generating a natural language sentence from an intermediate structure representing a sentence, and the target language dictionary unit A sentence generation unit 9 for generating a natural language sentence from the intermediate structure output by the sentence structure generation unit 6 using the vocabulary information stored in the sentence generation rule 7 and the rules stored in the sentence generation rule storage unit 8; Natural language sentence output by the generator 9 Output, and a summary output unit 10 for displaying. Next, the operation of the embodiment of the present invention will be described with reference to FIG. As one embodiment of the present invention, a user knows the contents of semiconductor manufacturing technology such as "what kind of company develops, manufactures, sells, or uses what kind of semiconductor manufacturing technology". Let's consider the case. It is also assumed that the document information extraction unit 4 is provided with an output format including, for example, keywords and relationships between keywords, as shown in FIG. FIG. 2 shows that the contents related to the semiconductor manufacturing technology exist in the document, and that the summary content consists of a developer called NEC and an etching technology. Here, it is shown that NEC exists in Tokyo. The etching technology is classified into a plasma etching technology, which is a technology corresponding to a DRAM for 64 megabits. Here, for example, consider the input document shown in FIG. The document input from the document input unit 1 is passed to a document information extracting unit 4. The keyword dictionary unit 2 stores keywords related to a target area specified by the user. For example, for semiconductor manufacturing technology, a keyword indicating a company such as “NEC”, a keyword indicating a semiconductor manufacturing technology such as “plasma etching”, A keyword such as "development" indicating the relationship between the company and the semiconductor manufacturing technology and detailed information on each are stored. The keyword relationship calculation rule storage unit 3 stores a method of combining keywords by using a relationship between keywords stored in the keyword dictionary unit 2, a co-occurrence relationship, a syntax structure, and the like, and detailed information attached to the keyword. Is stored, which describes a method for calculating the relationship between keywords, which describes a method of synthesizing. The document information extraction unit 4 analyzes the input document by using the keyword information stored in the keyword dictionary unit 2 and the keyword relation calculation rule stored in the keyword relation calculation rule storage unit 3, and summarizes the contents to be summarized. Is extracted and output in the specified format. For example, in the case of the document shown in FIG. 3, first, the keyword “NEC” stored in the keyword dictionary unit 2
Recognize words such as "plasma etching" and "development" as keywords. Further, the rules stored in the keyword relation calculation rule storage unit 3 are applied. For example, the document information extraction unit 4 outputs the structure shown in FIG. 2 by recognizing the relationship between the keywords using the syntax structure and synthesizing the detailed information of the related keywords. The sentence structure generation unit 6 receives the data output from the document information extraction unit 4, converts the received data into an intermediate structure indicating the structure of the sentence by applying the rules stored in the sentence structure generation rule storage unit 5. . For example, the sentence structure generation unit 6 receiving the data shown in FIG. 2 applies the sentence structure generation rules described based on the knowledge about the structure output from the document information extraction unit 4 and converts the conversion result shown in FIG. Is output. In FIG. 4, according to the format of the extraction target of "what kind of company develops, manufactures, sells, or uses what kind of semiconductor manufacturing technology", "NEC" focuses on the word "development". Is converted to a tree structure whose main purpose is "plasma etching technology". When the structure output from the document information extraction unit 4 is complicated, a sentence division rule corresponding to a preset output from the document information extraction unit 4 is described and stored in the sentence structure generation rule storage unit 6. It is also possible to divide into multiple tree structures. The sentence generation unit 9 receives the output of the sentence structure generation unit 6 and generates a sentence using the bilingual translation stored in the target language dictionary unit 7 and the grammar rules stored in the sentence generation rule storage unit 8. For example, when the target language dictionary unit 7 and the sentence generation rule storage unit 8 are configured with the target language being English, the output of FIG. 5 is obtained for the input of the tree structure of FIG. Summary sentence output unit 1
0 indicates the sentence output by the sentence generation unit 9. According to the present invention, by setting a target area and an intermediate structure, which is an output format, of information serving as a view point of a summary, a summary sentence can be generated accurately according to a user's intention. . In addition, since the target area can be changed, it is possible to provide the user with a plurality of summary viewpoints. Furthermore, the summary can be extracted by partial analysis centering on the keyword without analyzing the entire sentence, and a certain intermediate structure format without ambiguity can be obtained. That is, in the translation processing, since the translation is performed directly from the intermediate structure, a high-quality multilingual summary sentence can be output.

BRIEF DESCRIPTION OF THE DRAWINGS FIG. 1 is a diagram for explaining a block diagram which is an embodiment of claim 1 of the present invention. FIG. 2 is a diagram illustrating an embodiment of the present invention. FIG. 3 is a diagram illustrating an embodiment of the present invention. FIG. 4 is a diagram illustrating an embodiment of the present invention. FIG. 5 is a diagram illustrating an embodiment of the present invention. [Description of codes] 1 Document input unit 2 Keyword dictionary unit 3 Inter-keyword relation calculation rule storage unit 4 Document information extraction unit 5 Document structure generation rule storage unit 6 Sentence structure generation unit 7 Target language dictionary unit 8 Sentence generation rule storage unit 9 Sentence generation unit 10 Summary sentence output unit

Claims (1)

(57) [Claims] [Claim 1] A document input unit that receives a document described in a natural language as an input, and a keyword related to a summary target specified to be extracted in advance and a specified target A keyword dictionary unit for storing detailed information of each keyword, and a relationship between the keywords is identified using the keyword information stored in the keyword dictionary unit, and selection and synthesis of keywords that match the specified summary target are performed. A storage unit for calculating a relation between keywords storing rules to be performed, and a keyword for summarizing the designated summarization target by using the keyword information stored in the keyword dictionary unit and the rules stored in the storage unit for calculating the relation between keywords. A document information extraction unit for extracting a relationship between keywords and keywords; and a keyword and keyword output by the document information extraction unit. A sentence structure generation rule storage unit that stores a rule that converts a certain format consisting of relational structures between documents into an intermediate structure representing a sentence and, if necessary, stores a rule for dividing the sentence structure into a plurality of intermediate structures; A sentence structure generating unit that converts the structure of the designated information output by the document information extracting unit into an intermediate structure representing a sentence by using rules stored in a unit, and a target language dictionary unit that indicates a correspondence between each keyword and a target language A sentence generation rule storage unit storing rules for generating a natural language sentence from an intermediate structure representing a sentence, using vocabulary information stored in the target language dictionary unit and rules stored in the sentence generation rule storage unit A sentence generation unit that generates a natural language sentence from the intermediate structure output by the sentence structure generation unit; and a summary sentence output unit that outputs and displays the natural language sentence output by the sentence generation unit. Multilingual Summary Generator