JP2005196572A - Summary making method of multiple documents - Google Patents

Abstract

<P>PROBLEM TO BE SOLVED: To provide a summary making method of a plurality of documents which can generate automatically single explanation information without exaggeration and without omission from a plurality of explanation paragraphs. <P>SOLUTION: The system of this method unites/summarizes explanation paragraphs extracted from a plurality of web pages through following processes: a simple sentence specifying processing which detects simple sentences as the smallest units of summarization processing from explanation paragraphs extracted from a plurality of web pages and divides them; a comparison processing which compares the simple sentences extracted by the specifying processing and classifies similar simple sentences corresponding to the same viewpoint into the same group; a selection processing which selects one or more typical simple sentences from the groups organized by the comparison processing; and a generation/presentation processing which displays by items the representative sentences selected from each group together with group (viewpoint) name. The system automatically generates by computer concise explanation information without exaggeration and without omission about the target terms, by integrating/summarizing the explanation paragraphs extracted from the plurality of web pages. <P>COPYRIGHT: (C)2005,JPO&NCIPI

Description

  The present invention relates to a method for creating a summary of a plurality of documents.

  In recent years, it has become a daily intellectual information activity to perform various investigations using a search engine on the World Wide Web. Using the Web like a Japanese dictionary or encyclopedia to look up unknown words and matters is a major example of such intellectual information activities. This is because the existing dictionaries and encyclopedias often do not contain new words or technical terms, but because the Web has a lot of new information and highly specialized information in circulation. It is.

  Compared to the beginning when the Web began to spread, the performance of search engines has improved, and the target information has been easily found. However, depending on the contents of the search request, there is a case where it is still unknown what to input as a keyword, or a case where it is not known how to select desired information from a huge search result. Moreover, since there is no control in the Web, it is necessary to exclude low-quality information such as typographical errors, misunderstandings, and falsehoods.

  In order to solve the above problem, the inventors of the present application search site “Cyclone” intended to use the Web encyclopedically (see Non-Patent Documents 1, 2, 11, 12, 13, and Patent Document 1 below). ) And continuously improve evaluation experiments and problems. Cyclone has a function for automatically constructing content by extracting headwords and their explanation information from a group of Web pages, and a search function for using the constructed content.

  FIG. 6 is a search result example of a conventional Cyclone when a new type of pneumonia “SARS” is input.

  In the lower half of the screen, three explanation paragraphs individually extracted (extracted) from a plurality of Web pages are presented together with the page title of the extraction source. The explanatory paragraphs are classified based on specialized fields, and are further ranked and displayed based on the likelihood as an explanation. In addition, field names, related words, and compound words for narrowing down search results are presented below boxes for entering headwords (search words).

  Therefore, the desired explanation information about the input keyword can be obtained more efficiently than existing search engines by reading the provided explanation paragraphs in turn or by narrowing down to necessary information using field names and related terms. Can do.

  However, dictionaries and encyclopedias that are generally compiled by human beings describe a concise explanation of a single headword from a multifaceted perspective. For example, in the Iwanami Information Science Dictionary (see Non-Patent Document 9 below), the essential items include inclusive definitions representing essential features, extended definitions by examples, synonyms, etc. It is described.

  On the other hand, the plurality of descriptions shown in FIG. 6 are information extracted individually from different Web pages, and are not related to each other. That is, information included in one description does not exist in the other description, or conversely, similar information is included in a plurality of descriptions. Therefore, in order to obtain explanation information from a multifaceted viewpoint, it is necessary to browse a plurality of explanation paragraphs, resulting in the waste of reading explanations of the same content many times. .

  In addition, if the user does not get the information he wants without scrolling the page or clicking on the next page to view the next page, the user may stop using the search site. Therefore, it is an effective means to summarize and present the explanatory information in order to briefly outline the first page and attract the user's interest.

Therefore, the present invention provides a summarization technique for integrating a plurality of explanation paragraphs related to one headword and generating explanation information with no excess or deficiency. Specifically, this is processing corresponding to multiple document summarization (see Non-Patent Documents 4, 5, and 7 below). As a result, convenience can be improved even in an environment where the number of characters that can be displayed at one time is limited, such as a portable terminal.
JP 2003-085181 A Atsushi Fujii and Tetsuya Ishikawa. Unifying the World Wide Web as an encyclopedia: Extracting term descriptions from semi-structured texts. In Proceedings of the 38th Annual Meeting of the Association for Computational Linguistics, p. 488-495, 2000. Atsushi Fujii and Tetsuya Ishikawa. Organizing engineering group based on the Web and its application to question answering. In Proceedings of the 39th Annual Meeting of the Association for Computational Linguistics, pp. 196-203, 2001. Takahiro Fukushima, Hidetsugu Nanba, and Manabu Okumura. 2. Text summation challenge 2: Text summation evaluation at NTCIR workshop In Proceedings of the Third NTCIR Workshop on Research in Information Retrieval, Automatic Text Summarization and Question Answering, 2003. Inderjeet Mani. Automatic Summarization, chapter 7, pp. 169-208. John Benjamins, 2001. Dragomir R.D. Radev and Kathleen R.M. McKeown. Generating natural language summaries from multiple on-line sources. Computational Linguistics, Vol. 24, no. 3, p. 469-500, 1998. Takeshi Quartz, Hayashi Yoshihiko. Japanese compound sentence division based on connection structure analysis. IPSJ Journal, Vol. 33, no. 5, pp. 652-663, 1992. Okumura Manabu, Namba Eigo. Recent topics on automatic text summarization. Natural language processing, Vol. 9, no. 4, pp. 97-116, 2002. Masaru Ehara, Kim Jin-an. Completion of zero subject by probabilistic model. Natural language processing, Vol. 3, No. 4, pp. 67-86, 1996. Nagao Makoto. A systematic method of constructing specialized knowledge in a dictionary format. Japanese Society for Artificial Intelligence, Vol. 7, no. 2, pp. 320-328, 1992. Taku Kudo and Yuji Matsumoto. Japanese dependency analysis by applying chunking stage. IPSJ Journal, Vol. 43, no. 6, pp. 1834-1842, 2002. Satoshi Fujii, Katsunobu Ito, Tetsuya Ishikawa. Can the WWW be used as an encyclopedia? -Construction of large-scale corpus-. Information Processing Society of Japan Research Report, 2002-NL-149, pp. 7-14, 2002. Jun Fujii. Construction of an encyclopedia search site using Web information. Information Science and Technology, Vol. 53, no. 4, pp. 201-204, 2003. Jun Fujii and Tetsuya Ishikawa. Encyclopedia knowledge information extraction and organization using World Wide Web. IEICE Transactions, Vol. J85-D-II, No. 2, pp. 300-307, 2002.

  The search site “Cyclone” constructed by the inventors of the present application can efficiently present encyclopedia information related to input keywords by extracting high-quality explanatory paragraphs related to words and matters from the Web. At this time, a list of explanation paragraphs individually extracted from a plurality of Web pages is presented at present.

  However, since there is no relationship between the explanation paragraphs presented as described above, it is necessary to read a plurality of paragraphs in order to obtain sufficient explanation about the input keyword, and as a result, the same contents May have to be read many times.

  Compared to single-document summaries, research on multi-document summaries is new and no general model has been established.

  Therefore, in the present invention, in addition to the existing function of the search site “Cyclone”, a “summary” function is newly introduced.

  Here, a plurality of explanatory paragraphs are integrated (for each field) for one headword. In principle, the summary function can be applied at both offline and online stages. In the embodiment of the present invention, the summary is generated in advance offline in consideration of the response time.

  When performing an online search, if the keyword or question text entered by the user is not found in the headword, the user is required to present a headword that partially matches the character string or conceptually similar headword Natural information can be guided naturally. However, the details of this guidance are omitted because they are out of the focus of the present invention.

  SUMMARY OF THE INVENTION In view of the above, an object of the present invention is to provide a multi-document summary creation method capable of automatically generating a single piece of explanation information from a plurality of explanation paragraphs.

  Here, multiple explanation paragraphs about a certain input keyword are analyzed based on the viewpoint of explanation such as “definition” and “purpose”, and the final explanation information is obtained by selecting and integrating representative explanation sentences for each viewpoint. To construct.

  The effectiveness of the proposed method is shown by an evaluation experiment.

In order to achieve the above object, the present invention provides
[1] A single sentence specifying process that detects and divides a single sentence as a minimum unit of summary processing from explanatory information extracted from a plurality of documents, and a single sentence extracted by this specifying process are collated to correspond to the same viewpoint. A collation process for classifying similar single sentences into the same group, a selection process for selecting one or more representative single sentences from the group configured by the collation process, and representative sentences selected from each group by this selection process. Generates and presents information that is displayed in bulleted lists along with the group (viewpoint) name, integrates and summarizes the explanation information extracted from multiple documents, and automatically generates concise explanation information with no excess or deficiency for the target term It is characterized by doing.

  [2] In the method for creating a summary of a plurality of documents described in [1] above, when a heavy sentence is divided into a plurality of single sentences in the specifying process, the subject of the first sentence is supplemented at the head of the second and subsequent single sentences. It is characterized by doing.

  [3] In the multi-document summary creation method described in [1] above, in the collation process, in order to classify the simple sentences based on viewpoints, a fixed expression unique to each viewpoint is created in advance, and the fixed expression Are classified into corresponding groups.

  [4] The multiple document summary creation method according to [1] above, wherein the viewpoint is definition, abbreviation, illustration, purpose, synonym, book, product, advantage, defect, history, element, function. To do.

  [5] In the multiple document summarization method described in [1] above, in the collation process, a simple sentence that has not been classified is compared with a simple sentence group that has already been classified, and classified into viewpoints to which simple sentences with high similarity belong. It is characterized by.

  [6] In the method for creating a summary of a plurality of documents described in [5] above, the definition, abbreviations, examples, purposes, synonyms, books, products, advantages, disadvantages, history, elements, functions are not applied, and already A single sentence having a low similarity to any classified single sentence group is classified as “others”.

  [7] In the method for creating a summary of a plurality of documents described in [5] above, word similarity is used for calculating the similarity.

  [8] The summary creation method for a plurality of documents according to [1], wherein the selection process selects a high-quality sentence while reflecting a tendency of the entire group.

  [9] The multi-document summary creation method according to [8], wherein the reference is quantified and a representative sentence is selected based on a score obtained by combining the criteria.

[10] In the method for creating a summary of a plurality of documents described in [8] above, selecting the representative sentence,
(A) Considering the high likelihood ranking of the explanation paragraph from which the single sentence is extracted,
(B) A sentence including many words that appear in common in a single sentence included in the group is set as a candidate for a representative sentence.
(C) Of the candidates for the representative sentence, a sentence having the number of characters suitable for the purpose is selected as the representative sentence.

  [11] The multi-document summary creation method according to [8] above, wherein the criterion considers the rank of the original explanatory paragraph from which a single sentence is extracted.

  [12] The multi-document summary creation method according to [8] above, wherein the criterion is a sentence including many words that appear in common in a single sentence included in the group.

  [13] In the method for creating a summary method for a plurality of documents described in [8] above, the criterion is based on the number of explanatory characters.

  [14] In the method for creating a summary of a plurality of documents described in [5] above, the “others” is to preferentially select a single sentence that has as few overlapping words as possible from other single sentences already selected from the above viewpoints. Features.

  [15] In the method for creating a summary of a plurality of documents described in [5] above, when selecting a plurality of simple sentences from the “others”, first, the first one is selected, and a word with the already selected simple sentence is selected. Next, a simple sentence with less duplication is selected next, and various simple sentences are selected by performing this process recursively.

  [16] In the method for creating a summary of a plurality of documents described in [1], the generation / presentation process displays a representative sentence selected from each group in an itemized list together with a group (viewpoint) name.

  [17] In the multiple document summary creation method described in [16] above, the generation process replaces the sentence end expression of the single sentence extracted in the single sentence specifying process, thereby reducing the number of characters or correcting the expression to a sentence end-like expression. It is characterized by that.

  [18] In the method for creating a summary of a plurality of documents described in [16] above, in the presenting process, links are made to other explanatory texts belonging to the group and explanatory paragraphs of extraction sources so as to narrow down only to the viewpoint selected by the user. Features.

  [19] The multi-document summary creation method according to [16] above, wherein the representative sentences are presented in order from the highest score in the selection process.

  There is a great demand for a technique for extracting necessary knowledge from a vast amount of information on the World Wide Web. The present invention generates term descriptions without excess or deficiency, and has a wide range of use such as being applicable not only to search but also to creation of content, and has high commercial value.

  The purpose is to look up words and matters using the Web like a dictionary or encyclopedia, and to integrate and summarize explanations and sentences extracted from multiple Web pages, so that the target terms are concise A method for automatically generating explanatory information is proposed. According to the present invention, it is possible to efficiently obtain explanations about the target term as compared with the case of using an existing Web search engine.

  Here, the multi-document summary creation method of the present invention will be described.

  An existing method for creating a summary of a single document is disclosed in Non-Patent Document 4.

  In the present invention, a multi-document summary model for explanatory information is created based on the above procedure. That is, by the above-described organization process of the search site Cyclone, a plurality of explanation paragraphs obtained are input, and simple sentences (units) in the explanation paragraph are classified into groups corresponding to the “viewpoint” of the explanation. Then, final explanatory information is generated by integrating simple sentences (units) representing each group and presented to the user.

  However, the viewpoint to be prepared varies depending on the type of headword. For example, technical terms and animals and plants differ in terms of explanation. The current Cyclone contains about 600,000 headwords of various types such as technical terms, matters, names of people, animals and plants. Here, we implemented and evaluated the model for technical terms.

  Since the existing multi-document summary model is intended for newspaper articles and the like, it may be difficult to prepare a viewpoint in advance depending on the content and genre of the article. On the other hand, in the case of explanation of technical terms as in the present invention, it is possible to enumerate viewpoints to some extent by hand.

  Hereinafter, an outline of the search site Cyclone to which the present invention is applied will be described.

  FIG. 1 is a diagram showing an outline of an encyclopedia search site Cyclone according to the present invention, FIG. 2 is a system configuration diagram thereof, and FIG. 3 is a diagram showing a summary creation step of the present invention.

  In FIG. 2, 1 is a Web, 2 is a new word detecting unit, 3 is a candidate word collecting unit, 4 is a search processing unit for searching a page including a head word on Web 1, and 5 is searched using an HTML tag. An extraction processing unit that extracts descriptions of headwords from a page in units of paragraphs, 6 is an organization processing unit (functions will be described later), 7 is a summary creation unit, 8 is a headword extraction unit, 9 is encyclopedia content, 10 is a computer, 11 is a server and 12 is a browser.

  The function of the search site Cyclone to which the present invention is applied will be described with reference to FIGS. Although it is divided into offline processing for constructing encyclopedia content 9 and online processing for a user to search for content, only the offline processing according to the multi-document summary creation method of the present invention will be described here.

  Therefore, in the offline processing for constructing the encyclopedia content 9, first, “new word detection” is performed by the new word detection unit 2, and then new headword candidates are automatically collected from the Web 1 by the headword candidate collection unit 3. . Next, the search processing unit 4, the extraction processing unit 5, and the organization processing unit 6 sequentially execute “search processing”, “extraction processing”, and “organization processing” for each headword candidate to obtain the target term. Categorize descriptions by specialty. By doing this, it is possible to distinguish between explanations of multiple terms having different meanings depending on the field even with the same headword, such as “pipeline (processing / oil feeding pipe)”.

  Here, in the search process described above, a page including a headword on the Web 1 is searched. Next, in the extraction process, explanations about headwords are extracted from the retrieved page in units of paragraphs using HTML tags. Next, in the organizing process, (a) the degree of relevance to a specific field, (b) whether or not it includes a “descriptive” language expression, (c) whether or not it is an “descriptive” HTML layout, and (d) the reliability of the page The score that integrates these four measures is calculated, the paragraphs are classified into fields based on the values, and ranked based on the likelihood. Through the processing so far, a plurality of explanation paragraphs related to the headword are extracted.

  Finally, the word that characterizes the headword is acquired by the “related word extraction process” by the related word extraction unit 8. These words are used to narrow down the user's information request during online search.

  The basic principle of the related word extraction process is to detect words frequently appearing in the explanatory paragraph after the organization process of each headword. Here, a process for detecting an appropriate word and a scale for evaluating the detected word are required. Therefore, the paragraph is first subjected to morphological analysis using a Japanese morphological analyzer “tea bowl”, and a (compound) word is constructed based on the part of speech information to make a candidate for a related word. Specifically, a sequence of nouns, verb conjunctive forms, unknown words, and symbols is extracted as words. Furthermore, the appearance frequency in the paragraph and the score obtained by the organization process for the extraction source paragraph are integrated and sorted as a related word, and presented preferentially from the related word indicating the higher value. That is, words that frequently appear in a high-quality explanatory paragraph that is highly related to a specific field and that have high reliability are preferentially presented, and are effectively used when a user searches for information.

  In the present invention, based on a plurality of explanation paragraphs of the headwords obtained in this way, the summary creation unit 7 summarizes a plurality of documents, and the explanation contents 9 are integrated into one explanation paragraph so that there is no excess or deficiency. Automatically generate concise descriptive information.

Hereinafter, a procedure for summarizing a plurality of documents according to the present invention will be described in detail with reference to FIG.
(1) Specific process (step S1)
In the specific process, the minimum unit used for the summary process is detected. In the present invention, a group corresponding to the “viewpoint” is formed in the collation process described later, and therefore, a unit corresponding to the “viewpoint” needs to be extracted from the explanation paragraph.

  In the present invention, it is assumed that a simple sentence is a minimum unit corresponding to one viewpoint. Therefore, the core of the specific process is to divide the explanation paragraph into single sentences.

  Usually, many sentences and compound sentences are used as sentences, and it is still a difficult problem to divide them into simple sentences by a computer. In the present invention, each sentence in the explanation paragraph is subjected to dependency analysis using “CaboCha” (see Non-Patent Document 10), and a rule based on the sentence structure (see Non-Patent Document 6) is applied. Perform extraction.

  However, the subject is lost in the sentence behind by simply dividing it into simple sentences. For example, when the following sentence is divided into two simple sentences, “What is XML?” Must be supplemented to the beginning of the second sentence.

“XML is an abbreviation for eExtensible Markup Language and is a kind of markup language.”
When this is divided into simple sentences,
“XML is an abbreviation for eExtensible Markup Language,”
“(XML) is a kind of markup language.”
However, in any sentence, it is not always necessary to complement the subject of the first sentence to the beginning of the subsequent sentence. It is difficult to decide which elements should be complemented in what cases. For such a problem, a method of zero subject complementation in single sentence division (see Non-Patent Document 10 above) can be applied.
(2) Collation process (step S2)
It is assumed that the single sentence extracted by the specific process corresponds to a single viewpoint. In addition, similar single sentences are grouped together in the same group by collation processing. Therefore, the group configured by the matching process also corresponds to a single viewpoint.

  Currently, the following 12 types of viewpoints are prepared.

Definitions, abbreviations, examples, purposes, synonyms, books, products, benefits, drawbacks, history, elements, functions Two different methods were used in combination to classify simple sentences based on this perspective. First, a fixed expression unique to each viewpoint is created in advance, and sentences including the fixed expression are automatically classified into corresponding groups. The fixed expression here refers to a frequent expression such as “What is XX (headword)” in the viewpoint “definition” and “for example” in the viewpoint “example”. Using these expressions, initial sentence classification is performed.

  However, there is a simple sentence that does not include this standard expression prepared in advance (or a slightly different expression is used). Therefore, in the next stage, uncategorized simple sentences that could not be dealt with by the regular expression classification are compared with already classified simple sentence groups, and classified into groups to which simple sentences with high similarity belong. The word overlap was used for the similarity calculation. Therefore, the uncategorized simple sentences are classified into groups that share the most words among the groups configured by the initial classification. Specifically, normalization based on the length of the sentence is performed by calculation based on the Dice coefficient. However, not all words are used, but function words such as particles are deleted by morphological analysis.

A specific example regarding the headword “XML” is shown below.
(A) XML is an extensible markup language.

→ Definition (b) Abbreviation of eExtensible Markup Language → Abbreviation (c) Standardized by W3C in 1998 ,
→ History (d) XML is an abbreviation for Extensible Markup Language.

→ Abbreviation (e) The standardization of this XML was recommended by the W3C.

→? ? ?
In this example, the simple sentences (a) to (d) in the initial classification indicate that the underlined words and expressions correspond to the fixed expressions and are classified into the corresponding viewpoint groups. However, (e) could not be classified because it did not contain a fixed expression specific to the viewpoint group. Therefore, the similarity with the already classified simple sentences is calculated based on the word distribution. As a result, (e) was classified as “history” because it had the highest similarity with (c).

In addition, simple sentences that are not classified in any viewpoint even if the above processing is performed are classified as “others”.
(3) Selection process (step S3)
In the selection process, one or more representative sentences are selected from the group configured by the matching process. Here, it is important to select a good sentence while reflecting the trend of the entire group.

  Specifically, various criteria shown below are quantified, and a representative sentence is selected based on a score obtained by combining them.

  (A) In the Cyclone search results that are ranked and displayed based on the likelihood, it is likely that the higher-level explanation is of higher quality, so the order of the original explanatory paragraph from which the single sentence is extracted is taken into consideration.

  (B) A sentence including many words that appear in common in a single sentence included in the group is represented. As a result, minority words (noise) caused by extraction paragraphs or simple sentence extraction errors can be excluded from the final summary.

  (C) Consider the number of characters in the description. Depending on the usage environment such as a portable terminal, the number of displayed characters may be the strongest constraint. Therefore, a sentence as short as possible is selected as a representative sentence.

  Since the above three criteria have different ranges, they are combined after adjusting weights empirically.

  “Others” includes a large number of single sentences whose corresponding viewpoints cannot be distinguished, or noise that is not suitable as an explanatory sentence. In order to avoid redundant summarization, a simple sentence is selected preferentially from “others” in terms of general (other than “others”) with as little word overlap as possible.

When selecting a plurality of simple sentences from “Others”, first, the first one is selected, and then a single sentence with less word overlap with the already selected simple sentence is selected. Various simple sentences are selected by performing this process recursively.
(4) Generation / presentation processing (step S4)
In the generation / presentation process, representative sentences selected from each group are displayed in an itemized list together with the group (viewpoint) name. Here, the sentences are presented in order from the representative sentence with the highest score in the selection process. FIG. 4 and FIG. 5 show the explanatory paragraphs in the computer field concerning the headword “XML” and the results of summarizing them, respectively. In this example, a summary was generated that allows an overview of the headword from a multifaceted viewpoint with a small number of 397 characters.

  There is room for further improvement in the generation process. For example, it is possible to reduce the number of characters or correct the expression to the end of the sentence by replacing the end sentence expression of the single sentence extracted by the single sentence specifying process (for example, “to be standardized” in (c) above). it can. However, changing the contents of the explanation outside the scope of the mere excerpt may infringe the copyright of the page. You must be careful about the allowable range for research purposes and the restrictions when actually operating on the Web.

  In addition, as a contrivance in the presentation process, a link is made from a viewpoint name such as “definition” or “purpose” to other explanatory texts belonging to the group or an explanatory paragraph of the extraction source, and narrowed down only to the viewpoint selected by the user. There is a technique. For example, the description of “book” in FIG. 5 is extracted from a page of book sales information about XML. Using this explanation as a clue, sales information on books can be quickly acquired.

  Next, the evaluation experiment will be described.

  Since the evaluation of the summarization method largely depends on the subjectivity of the judge, it is difficult to evaluate it from the engineering viewpoint. On the other hand, there is an evaluation test collection for newspaper articles and the like (see Non-Patent Document 3 above). There is also a method for evaluating the summarization method indirectly and objectively based on the degree of performance improvement when applied to another task (such as selection of relevant documents in information retrieval) instead of evaluation of the summarization method itself.

  However, since there is no precedent for summarizing the explanation of terms used in the present invention, we have created test data on a small scale and used it for evaluation in consideration of identifying problems in the evaluation method. In addition, only the evaluation experiment on the summarization method itself was conducted.

  There are multiple options for evaluation criteria and scales. For example, there is a criterion “how close an automatically generated summary is to the explanation of an existing terminology dictionary”. However, since the search result of Cyclone is used as the input of the summarization method of the present invention, it is difficult to evaluate with this standard when there is a viewpoint that is only in the existing dictionary or not in the existing dictionary.

  Therefore, in the present invention, an evaluation experiment was performed using the following two types of scales based on the manual determination by the determiner and the summary result of the present invention. These two types of measures are in a trade-off relationship and are difficult to improve at the same time.

(1) Summarization rate How much the search result (explanation paragraph) of Cyclone was shortened?

(2) Coverage rate Of the viewpoints of explanation included in the search results (explanation paragraph) of Cyclone, how many viewpoints could be included in the summary?

  Table 1 shows the headwords used in the evaluation experiment and the experimental results.

各見出し語について、Ｃｙｃｌｏｎｅの検索結果のうちコンピュータ分野の説明段落５０件を要約処理の入力とした。５０件に限定したのは判定者による人手判定のコストを増やさないことがその理由である。

For each headword, 50 description paragraphs in the computer field among the search results of Cyclone were used as input for the summary process. The reason for limiting to 50 is that the cost of manual judgment by the judge is not increased.

  The examiner examined only the top 50 search results of Cyclone without giving a result of the summary processing according to the present invention, and gave a viewpoint to a single sentence unit. The 30 types of viewpoints requested to be given to the judge are shown below.

Definition, example, superordinate concept, subordinate concept, element, nature, attribute, function, purpose, history, present, prediction, cause, result, synonym, antonym, abbreviation, translation, indirect explanation, comparison, metaphor, product, book , Other meanings (in the case of polysemy), reading, obtaining methods, advantages, disadvantages, etymology, etc. Descriptions published in the above existing dictionary so that the judge does not give a point of view for false explanations The knowledge about the target term was given by showing etc., and one or more viewpoints were given to one simple sentence. The explanation that the target headword is an ambiguous word and is used in a meaning other than the existing terminology dictionary is given a viewpoint of “another meaning” and is effectively excluded from the evaluation.

  Among the viewpoints shown to the judge, there are those that Cyclone cannot output in the first place because the typical expression required in the collation processing of the present invention cannot be set in advance. Since these viewpoints are never included in the result of the summary processing by Cyclone, it has become a factor of lowering the coverage rate.

  The results and discussion will be explained.

  Table 1 shows the summary rate and coverage rate as experimental results. The number of representative sentences (N) output from each viewpoint group was changed to 1, 2, and 3. However, 5 representative sentences were always selected from “Others”. The summarization rate was calculated by equation (1). A smaller summarization rate indicates better results.

Number of characters for automatic summarization / number of characters read by user when not summarizing (1)
Here, “the number of characters read by the user when not summarizing” refers to the number of characters until all sentences corresponding to the viewpoint included in the automatic summarization according to the present invention are read when the user reads the Cyclone search results from the top. It is. In that case, it is assumed that if all the viewpoints included in the automatic summary are read, the browsing is ended even before reading the explanation paragraph 50 times. Moreover, if the viewpoint is the same, it is not necessary to be the same as the sentence included in the summary.

  The coverage rate was calculated by equation (2). The larger the coverage rate, the better the result.

The number of viewpoints included in the summary / the number of viewpoints given by the judge ... (2)
The coverage rate of the proposed summarization method was evaluated by the following two methods.

  (1) A: Only when the viewpoint given by the determiner was included as the number of viewpoints included in the summary and the viewpoint name was also correct, it was regarded as a correct answer.

  (2) B: Appropriateness of the viewpoint name is not taken into consideration (how many viewpoints given by the determiner are included as the number of viewpoints included in the summary).

  Furthermore, as a reference value for comparison with B, the coverage rate by the following simple summarization method was also calculated.

  (3) From the top of the C: Cyclone search result, extract the same number of characters as the result of the proposed summarization method (however, if the extracted point is in the sentence, it is output to the end of the sentence).

  The above C corresponds to the “read method” (simple method of extracting a certain number of characters from the beginning of the document) in the summary of a single document. Further, since the method C cannot assign a viewpoint name, whether the viewpoint name is appropriate is not considered as in the case of B.

  Consider the results in Table 1. First, it can be seen that increasing the value of N increases the summarization rate and conversely increases the coverage rate.

  Although the summarization rate varies depending on N, it is about 10%, and it can be seen that the original information can be considerably shortened. That is, in order to acquire the same amount of information, the user only has to use about one-tenth of the labor. On the other hand, the coverage of Method A was on average as low as 20%. However, in the method B that does not consider the suitability of the viewpoint name, the coverage rate is more than doubled compared to the method A.

  Comparing the method B and the method C, it was found that the coverage rate of the method of the present invention is generally high, and the method of the present invention gives more information to the user with the same number of characters. Although the difference between the two is still small, this difference may widen in the future by improving the elemental technology.

  Furthermore, the coverage rate was analyzed for each viewpoint. Here, only the case where N = 1 in the above-mentioned method B is shown below, targeting the viewpoints included in six or more headwords in the search result of Cyclone.

Definition (7/7), example (4/4), synonym (1/6), nature (3/7), function (1/7), purpose (3/6), history (4/6), Indirect explanation (5/7), comparison (1/7), book (5/6)
From this result, it is necessary to improve the coverage rate mainly from the viewpoint of being easily included in the existing dictionary such as “example” and “synonym”.

  In addition, this invention is not limited to the said Example, Based on the meaning of this invention, a various deformation | transformation is possible and these are not excluded from the scope of the present invention.

  The present invention is suitable for creating a summary of a simple term explanation that is efficient from a plurality of documents and has no excess or deficiency in a site that retrieves encyclopedia content from the Web.

It is a figure which shows the outline | summary of the encyclopedia search site Cyclone concerning this invention. It is a system configuration | structure figure of encyclopedia search site Cyclone concerning this invention. It is a summary preparation flowchart concerning this invention. It is a figure which shows the description paragraph regarding the computer field of "XML" of search site Cyclone. It is a figure which shows the result of having summarized the description paragraph regarding the computer field of "XML" of the search site Cyclone concerning this invention. It is a search result example of conventional Cyclone when new type pneumonia "SARS" is input.

Explanation of symbols

1 Web
2 New word detection unit 3 Headword candidate collection unit 4 Search processing unit 5 Extraction processing unit 6 Organization processing unit 7 Summary creation unit 8 Headword extraction unit 9 Encyclopedia content 10 Computer 11 Server 12 Browser

Claims (19)

(A) a single sentence specifying process for detecting and dividing a single sentence as a minimum unit of the summary process from the description information extracted from a plurality of documents;
(B) a collation process for collating simple sentences extracted by the specific process and classifying similar single sentences corresponding to the same viewpoint into the same group;
(C) a selection process for selecting one or more representative single sentences from the group configured by the matching process;
(D) A generation / presentation process of displaying a representative sentence selected from each group in the selection process together with a group (viewpoint) name in an itemized form,
(E) A method for creating a summary of a plurality of documents, wherein the explanation information extracted from a plurality of documents is integrated / summarized, and concise explanation information about the target term is automatically generated by a computer.   2. The method for creating a summary of a plurality of documents according to claim 1, wherein when the heavy sentence is divided into a plurality of single sentences in the specifying process, the subject of the first sentence is supplemented at the head of the second and subsequent single sentences. Multi-document summary creation method.   The summary creation method for a plurality of documents according to claim 1, wherein in the collation process, in order to classify the single sentence based on a viewpoint, a fixed expression unique to each viewpoint is generated in advance, and a sentence including the fixed expression is included. A method for creating a summary of a plurality of documents, which is classified into corresponding groups.   2. The multiple document summarization method according to claim 1, wherein said viewpoint is definition, abbreviation, illustration, purpose, synonym, book, product, advantage, defect, history, element, function. Summarization method.   2. The multi-document summary creation method according to claim 1, wherein, in the collation process, a simple sentence that has not been classified is compared with a single sentence group that has already been classified, and is classified into a viewpoint to which a single sentence having a high degree of similarity belongs. How to create a summary of multiple documents.   6. The multiple document summarization method according to claim 5, wherein the definition, abbreviation, example, purpose, synonym, book, product, advantage, defect, history, element, function, and any of those already classified A method for creating a summary of a plurality of documents, wherein simple sentences having a low similarity to a single sentence group are classified as “others”.   6. The multi-document summary creation method according to claim 5, wherein word similarity is used in the similarity calculation.   2. The multi-document summary creation method according to claim 1, wherein in the selection process, a high-quality sentence is selected while reflecting a tendency of the entire group.   9. The multi-document summary creation method according to claim 8, wherein the criteria are quantified and a representative sentence is selected based on a score obtained by combining the criteria. 9. The multiple document summary creation method according to claim 8, wherein the representative sentence is selected.
(A) Considering the high likelihood ranking of the explanation paragraph from which the single sentence is extracted,
(B) A sentence including many words that appear in common in a single sentence included in the group is set as a candidate for a representative sentence.
(C) A method for creating a summary of a plurality of documents, wherein a sentence having a character number suitable for a purpose is selected as a representative sentence from the representative sentence candidates.   9. The multi-document summary creation method according to claim 8, wherein the criteria consider the rank of the original explanatory paragraph from which a single sentence is extracted.   9. The multi-document summary creation method according to claim 8, wherein the reference is a sentence including many words appearing in common in a single sentence included in the group.   9. The method for creating a summary of a plurality of documents according to claim 8, wherein the criterion is based on the number of explanatory characters.   6. The multiple document summary creation method according to claim 5, wherein a plurality of simple sentences are selected preferentially from said "others" with as few overlapping words as possible from other single sentences already selected from said viewpoint. How to summarize documents.   6. In the method for creating a summary of a plurality of documents according to claim 5, when selecting a plurality of simple sentences from the “others”, first, the first one is selected, and there is little overlap of words with the already selected simple sentences. A method of creating a summary of a plurality of documents, wherein a single sentence is selected next, and various simple sentences are selected by performing this process recursively.   The summary creation method for a plurality of documents according to claim 1, wherein the generation / presentation processing displays a representative sentence selected from each group in an itemized list together with a group (viewpoint) name. Method.   17. The multi-document summary creation method according to claim 16, wherein the generation process replaces a sentence end expression of a single sentence extracted by a single sentence specifying process, thereby reducing the number of characters or correcting the expression to a sentence end-like expression. To create multiple document summaries.   17. The method for creating a summary of a plurality of documents according to claim 16, wherein in the presenting process, links are made to other explanatory texts belonging to the group and explanatory paragraphs of the extraction source so as to narrow down only to the viewpoint selected by the user. How to summarize documents.   17. The method for creating a summary of a plurality of documents according to claim 16, wherein the summaries are presented in order from a representative sentence having a high score in the selection process.

Images