JPH03191475A - Document summarizing system - Google Patents

Abstract

PURPOSE:To accelerate the automation of formation of a summarized document by extracting an important sentence in the meaning of dependency upon the sort of each paragraph and allowing a summarized document forming rule to correspond to an important document extracting rule in summarized document forming processing. CONSTITUTION:An unprocessed paragraph in a document is segmented. Then, the sort of the paragraph is decided based upon a heading word. Then, an important sentence is extracted from the paragraph by using an important sentence extracting rule group depending upon the sort of the selected paragraph and an important sentence extracting rule group applicable to any sorts. The extracted sentence is stored together with the paragraph number and the heading. This processing is repeated for all paragraphs in the document, and finally the paragraph numbers, the headings of respective paragraphs and the extracted sentences are displayed in the order of paragraph numbers.

Description

DETAILED DESCRIPTION OF THE INVENTION [Field of Industrial Application] The present invention relates to a document summarization method that uses a computer to generate a document summary.

[Conventional technology]

Traditionally, as a technique for summarizing the contents of documents, Iwabuchi et al.: Extraction and Processing of Semantic Information from Full-Text Information (Proceedings of the 38th National Conference of the Information Processing Society of Japan (1989)).

222) is known. This extracts important words from a document based on information such as character types and particles, determines the most important words based on the frequency of occurrence of the important words, and determines important sentences based on whether the important words and the most important words are included. It is based on that. Furthermore, when generating a summary sentence, special words such as ``above'' and ``here'' are removed from the extracted important sentences automatically.

[Problem to be solved by the invention]

In the above technology, it is determined whether each sentence is important or not, but in actual documents, certain parts have relative importance to other parts. For example, a paragraph can be considered as a grouped part. In addition to the presence of important words in a sentence, it is necessary to consider the importance of the paragraph in which the sentence is included relative to the whole.

For example, the accuracy with which a word or phrase expresses the content of an entire document does not necessarily match the frequency with which that word appears in the entire document.

Furthermore, since a word can have various syntactic and semantic roles depending on its position in a sentence, it is desirable to perform more detailed analysis than simply checking whether it appears in a sentence.

[Means to solve the problem]

The document summarization method of the present invention divides a given document into paragraph units with headings, selects a group of important sentence extraction rules corresponding to the type of paragraph for each paragraph, and The important sentences in each paragraph are extracted by referring to both types-independent important sentence extraction rules, and the set of pairs of the important sentences extracted in each paragraph and the heading of each paragraph is extracted from the given document. This is realized by determining that it is an important part of the whole and outputting it.

Further, the document summarization method of the present invention is realized by applying a summary sentence generation rule associated with each important sentence extraction rule in advance to transform, connect, and output the important sentences.

Furthermore, the document summarization method of the present invention refers to the importance set in advance for each important sentence extraction rule, selects only the important sentences that exceed a predefined boundary value, and generates a summary sentence. This will be realized as follows.

[Effect]

In the present invention, a document is divided into paragraphs, and important sentences that reflect the characteristics of the paragraphs are extracted. In this case, two types of rule groups are referred to. Importance information common to the entire document, which is within the scope of the prior art, is stored in a group of important sentence extraction rules that do not depend on paragraphs, and rules that depend on the type of paragraph are stored as separate groups of important sentence extraction rules. There is. The former is one;
The latter provides only paragraph types.

For example, suppose you have a document consisting of the following paragraphs.

1. Introduction 2. Research purpose 3. Experimental equipment 4. Experimental results 5. Discussion 6. References From the heading ``Introduction'' in the first paragraph, ``Introduction rules group'' for processing this paragraph is selected. However, this set of rules describes rules for extracting sentences related to the subject of a document. - For example, the sentence "I will talk about..." is extracted. Similarly, in the fifth paragraph, "Rules for Opinion Description Section".

In the sixth paragraph, "non-important part rule group" is selected.

The former enhances the extraction of sentences that include expressions expressing the author's assertions and opinions, while the latter describes rules that suppress the extraction of sentences as much as possible. In this way, this method makes it possible to extract important sentences that reflect the characteristics of the paragraphs included in the sentences.

Also, since each rule is detailed depending on the type of paragraph, it is possible to describe for each rule the importance of sentences to be extracted by that rule.For example, in the above example, On average, the rules in the ``opinion statement group'' are more important than the rules in the ``introduction rule group,'' and a sentence that includes a numerical expression in the fourth paragraph contains a numerical expression in another paragraph. It can be judged that it is more important than the text.

For the same reason, it is possible to describe the sentence generation rule for generating a summary sentence for each rule. If there is a rule for extracting the sentence J, this can be easily expressed as ``Since β, α is'' when generating a summary sentence.

〔Example〕

Hereinafter, the present invention will be explained with reference to the drawings.

FIG. 1 is a flow chart showing a first embodiment of the present invention. First, in paragraph cutout 11, one unprocessed paragraph in the document is cut out.Next, in paragraph type determination 12, the type of the paragraph is determined based on the words of the heading.

Then, in rule group selection 13, an appropriate important sentence extraction rule group is selected based on the type of paragraph.

In the extraction process 14, important sentences are extracted from the paragraph using two important sentence extraction rules: one that depends on the type of paragraph selected, and one that can be applied to any type of paragraph. . At this time, if rules can be written in the former group of important sentence extraction rules that suppress the application of rules in the latter group of important sentence extraction rules, it is possible to create a highly flexible extraction method. . The extracted sentence is stored in the extracted data storage 15 together with the paragraph number and heading.

The above process is repeated for all paragraphs in the document (
Step 16) Finally, in display 17, in the order of the paragraph numbers, the paragraph number 2 heading.

The three extracted sentences are displayed.

FIG. 2 is a flow chart showing a second embodiment of the present invention. In the figure, the steps from paragraph cutout 21 to determination process 26 are the same as in the first embodiment (see FIG. 1), except for extraction data storage 25.

In the extracted data storage 25, three items are stored: an extracted sentence and a summary sentence generation rule number.

Here, FIG. 4 shows the structure of the important sentence extraction rule group, and FIG. 5 shows the structure of the summary sentence generation rule group 40.
The summary sentence generation rule group 50 includes a sentence extraction rule content 41 and a summary sentence generation rule number 42 indicating the rule used when converting the extracted sentence into a summary sentence.The summary sentence generation rule group 50 includes a rule number 51 that is uniquely determined in the system. It consists of two items: and summary sentence generation rule content 52. The summary sentence generation number stored in the extracted data storage 25 is the one stored in this summary sentence generation rule number 42.

Next, in extracted sentence retrieval 27, one of the unprocessed extracted sentences that appears furthest in the document is selected. A summary sentence generation rule corresponding to the sentence is searched from the summary sentence generation rule group 50 based on the summary sentence generation rule number, and summary sentence generation is performed in the summary sentence generation 28. The generated sentence is displayed on display 29. The process is repeated until there are no more unprocessed extracted sentences (step 20).

FIG. 3 is a flowchart showing a third embodiment of the present invention. In the figure, the steps from paragraph cutout 31 to determination process 36 are the same as in the first embodiment (see FIG. 1), except for extraction data storage 35.

In the extracted data storage 35, two items, an extracted sentence and an extracted sentence importance level, are stored. The important sentence extraction rule group 40 in FIG. 4 includes an item of importance level 43 indicating the importance level of the extracted sentence. The extracted sentence importance stored in the extracted data storage 35 is the one stored in the importance 43.

Next, in extracted sentence retrieval 37, one of the unprocessed extracted sentences that appears furthest in the document is selected. Extracted sentence determination 38 determines whether the sentence exceeds a predetermined level of importance. If the condition is met, the display 39 outputs 0. If the condition is not met, the next extracted sentence is determined. Move to. The process is repeated until there are no more unprocessed extracted sentences (step 30).

FIG. 6 is an explanatory diagram showing the structure of the above document summarization method. In the figure, a paragraph type determination program 61 determines the type of paragraph, and an extraction processing program 62 refers to an important sentence extraction rule base 67 to extract important sentences. A group of important sentence extraction rules corresponding to various paragraphs and a group of important sentence extraction rules corresponding to various paragraphs are stored.

The extracted sentence selection program 63 selects extracted sentences that exceed the boundaries of predetermined importance, and the summary sentence generation program 64 generates a summary sentence with reference to the summary sentence generation rule base 68. A summary sentence generation rule base 68 stores a group of summary sentence generation rules.

The display program 65 displays the extracted sentence and summary sentence on the CRT 66. The overall control section 60 controls each of the programs mentioned above. The working memory 69 is used for storing extracted data 15, 25, and 35 in FIGS. 1, 2, and 3.

〔Effect of the invention〕

The important sentence extraction process of the present invention extracts important sentences in a sense that depends on the type of each paragraph and presents them in association with the heading of the paragraph when displayed, making it easy to understand the content of the entire document as well as the flow of the subject of the entire document. can be grasped.

In addition, in the summary sentence generation process, the summary sentence generation rules are associated with the important sentence extraction rules, so the summary sentence generation description can be specified in more detail, making it possible to automate summary sentence generation. promoted. Furthermore, since it is now possible to define the importance of an extracted sentence with respect to the important sentence extraction rule, a means for narrowing down the extracted sentences according to the summary level is provided.

The document summarization method of the present invention targets relatively long documents structured by paragraphs, rather than short documents with no paragraphs.Summary of a document with no paragraphs uses important sentence extraction rules that do not depend on paragraphs. This corresponds to summarizing only by In other words, this method is an extension of the summary method to documents without paragraphs.

[Brief explanation of drawings]

Fig. 1 is a flow chart showing the first embodiment of the present invention, Fig. 2 is a flow chart showing the second embodiment, Fig. 3 is a flow chart showing the third embodiment, and Fig. 4 is a flow chart showing important sentence extraction rules. FIG. 5 is an explanatory diagram showing the configuration of a group of summary sentence generation rules, and FIG. 6 is an explanatory diagram showing a configuration example. 40... Important sentence extraction rule group, 50... Summary sentence generation rule group, 60... Overall control program, 61... Paragraph type determination program, 62... Extraction processing program, 63... Extracted sentence selection program, 64...Summary sentence generation program, 65... Display program.

Claims (1)

[Claims] 1. In a document summarization method that uses a computer to generate a summary sentence of a document, a given document is divided into paragraph units with headings, and a group of important sentence extraction rules corresponding to the type of paragraph is provided. is selected for each paragraph, and the important sentences in each paragraph are extracted by referring to both the important sentence extraction rule group and the important sentence extraction rule group that does not depend on the type of paragraph, and the important sentence extracted in each paragraph is extracted. and a set of pairs of headings of each paragraph are determined to be important parts of the entire given document and are output. 2. The document summarization method according to claim 1, wherein the important sentences are transformed, concatenated, and output by applying a summary sentence generation rule that is previously associated with each important sentence extraction rule. method. 3. In the document summarization method according to claim 1 or 2, only the important sentences exceeding a predefined boundary value are selected by referring to the importance set in advance for each important sentence extraction rule, and the summarized sentences are extracted. A document summarization method characterized by generating.