JP2009146447A - Text summarization system, text summarization method, and text summarization program - Google Patents

Abstract

<P>PROBLEM TO BE SOLVED: To achieve text summarization for generating summarization with high quality in a short time. <P>SOLUTION: A certification means 101 of a summarization configuration unit certifies a unit minimizing a clause including predicate as a summarization configuration unit by applying language knowledge information 104 for summarization configuration unit certification to each sentence of an input text, and an importance calculation means 102 of the summarization configuration unit calculates the importance of each summarization configuration unit in the text by using information such as the position of the summarization configuration unit in the text, a clue word, a title, the importance of a word and a keyword or the like as a clue to the summarization configuration unit, and successively extracts the summarization configuration unit in the order of the high importance until it reaches summarization length requested by a user on the basis of the importance to each summarization configuration unit, and a summarization result output means 103 arranges the extracted summarization configuration units in the order of the appearance in the original text, and outputs it as a summarization result. <P>COPYRIGHT: (C)2009,JPO&INPIT

Description

The present invention relates to a text summarization system, a text summarization method, and a text summarization program, and more particularly, to a text summarization system, a text summarization method, and a text summarization program that perform certification of a summary unit.

The technique described in Patent Document 1 is that “the relation analysis unit performs dependency analysis on the text in the input document, extracts binary relations between words, and the instruction input unit extracts the summary result from the user. The instruction of the word to be refined or the type of information is received from the instruction, the instruction is converted into a priority condition indicating the condition of the binary relation to be prioritized by the instruction interpreting means, and this priority condition and the extracted binary relation From the binary relation obtained in this way, the important binary relation is selected by the relation selection means, the binary relation to be supplemented to the selected important binary relation is added by the relation complementation means, "Technique for synthesizing phrase summaries by means of synthesis".

Further, the technology described in Patent Document 2 includes: “input reading means for reading input of a Japanese sentence, morphological analysis means for analyzing an input sentence using a dictionary lookup function for the input sentence and dictionary information after dictionary lookup, and syntax analysis means” In a Japanese sentence analysis method used in a Japanese processing system equipped with semantic analysis means and a memory, in Japanese, the subject indicated by the co-particle "ha" or the plural of the main case indicated by the case particle "ga" The subject / prominent shared vocabulary information holding means that pre-collects information on vocabulary that prescribes sharing between words and held in the memory, and the vocabulary information held in the subject / master shared vocabulary information holding means Technology to estimate the existence of subject / generic sharing prior to syntax analysis / semantic analysis, using themes / major sharing estimation means to infer whether or not subject / general sharing exists It is.

As a conventional method for summarizing text by a computer, there is an important sentence extraction method. This is a method of constructing a summary sentence by recognizing a sentence as a summary constituent unit in text, evaluating its importance, extracting summary constituent units with high importance, and arranging them in the order of appearance of the original text. (For example, Non-Patent Document 1).

This method is a basic method used in many summarization systems. Basically, each summary component in the text is based on information such as sentence position, clue word, title, word importance, and keyword. Is calculated, and summaries are generated by arranging the higher importance items in the order of appearance of the original text (for example, Non-Patent Documents 2 and 3).

As a method related to this, there is a text summarization method that recognizes a clause as a summary unit (for example, Non-Patent Document 4). This method is a method of generating a summary sentence by deleting a clause having low importance while preserving constituent elements of main structures such as case elements in the syntax structure of the sentence.

In addition, there is a technique related to means for analyzing the context within a sentence (for example, Non-Patent Document 5).

JP 2000-250937 A

JP 05-113994 A

H. P. Luhn, “The Automatic Creation of Literary Abstracts”, IBM Journal, 1958, p. 159-165

H. P. Edmundson, “New Methods in Automatic Extraction.”, Journal of the Association for Computing Machinery, 16 (2), 1968, p. 264-285

Okumura et al., “Research Trend on Automatic Text Summarization (instead of Preface)” Natural Language Processing, Vol. 6, no. 6, 1999, p. 1-26

Mikami et al., “Summary by Shortening in Sentences for Subtitle Generation in News Programs”, Natural Language Processing Vol. 6, no. 6, 1999, p. 65-82

Kamei et al., “Proposal of Lexical Disclosure Grammar”, IEICE Society for Language Understanding and Communication, NLC86-7,, 1986, p. 1-5

The important sentence extraction method has an advantage that a highly readable summary can be generated by a simple process because the sentence is extracted as a constituent unit of the summary. On the other hand, however, there is a limit to the method of constructing a summary with a sentence as a minimum unit in generating a short and high-quality summary. This is because when a short summary is generated, combinations of units constituting the summary are limited. Also, if the text contains long sentences, some insignificant parts cannot be removed, which also creates a limit in obtaining a good quality summary.

On the other hand, the text summarization method that uses clauses as a summary unit has a higher degree of freedom to generate summaries because it uses a smaller summary unit compared to the important sentence extraction method, which causes limitations and quality degradation in the generation of short summaries. Hateful. However, the use of clauses as a summary unit results in a change in the syntax structure in summary sentence generation, and it is difficult to generate sentences such as syntactic correctness and guarantee of semantic correctness of the generated summary sentence. Create a challenge. Obtaining accurate and detailed syntax structure information requires accurate full-fledged syntax analysis means, but such a method has a high calculation cost for summary generation.

An object of the present invention is to provide a text summarization method capable of simultaneously realizing improvement of summary quality by using an appropriate summarization unit and speeding up of processing by simplifying summary generation processing. As a result, even when the user requests a short summary, it is possible to generate a high-quality summary in a short time compared to the conventional summarization method.

The first text summarization system of the present invention reads out summary constituent unit recognition language knowledge information including information that uses a section including a sentence predicate as a summary constituent unit from a storage device, and reads the read out summary constituent unit recognition language knowledge. Extract the summary component unit certifying means that certifies the summary component unit using information, the summary component unit importance calculation method that evaluates the importance of the certified summary component unit, and the summary component unit with the higher importance. And a summary result output means for creating a summary.

The second text summarization system of the present invention reads a summary constituent unit recognition pattern including a summary constituent unit including a collation condition and a clause including a statement predicate from a storage device, and uses the read summary constituent unit recognition pattern. The summary composition unit certifying means that certifies the summary composition unit, the summary composition unit importance calculation means that evaluates the importance of the certified summary composition unit, and the summary composition unit that has a high importance is extracted and a summary is created. And a summary result output means.

The third text summarization system of the present invention reads a summary constituent unit recognition pattern including a collation condition, a summary constituent unit having a clause including a statement predicate as a unit, and an application cost from the storage device, and reads the read summary constituent unit. Summary composition unit certifying means that certifies summary composition units using patterns, summary composition unit importance calculation means for evaluating the importance of certified summary composition units, and summary structures with high importance in consideration of application costs Summarization result output means for extracting a unit and creating a summary.

A fourth text summarization system according to the present invention is the first, second, or third text summarization system, and reads a summary unit complement correction pattern including collation conditions and supplement correction information from the storage device. Using the read summary composition unit complement correction pattern, the summary composition unit recognized by the summary composition unit recognition means is supplemented / corrected based on the sentence expression, and the complement / correction summary composition unit is obtained. Summarization composition unit complementary correction means for outputting to the summary composition unit importance calculation means is provided.

A fifth text summarization system of the present invention is the first, second, or third text summarization system, and reads a summary unit complement correction pattern including collation conditions and supplement correction information from the storage device. Then, using the read summary composition unit complement correction pattern, the summary composition unit whose importance is evaluated by the importance calculation means of the summary composition unit is complemented / corrected based on the sentence expression, and complemented / corrected. It further comprises means for complementing and correcting the summary constituent unit for outputting the summarized constituent unit to the summary result output means.

The sixth text summarization system of the present invention reads a summary component unit including a collation condition, a clause including a clause including a sentence predicate, and a summary component unit recognition pattern including an extraction order constraint from a storage device, and reads the read summary component unit The summary composition unit is recognized by considering the connection relation between the summary composition units or the citation relation using the authorization pattern, and the summary composition unit certification means and the importance of the recognized summary composition unit are evaluated. Importance calculation means, extraction order constraint setting means that considers the extraction order using extraction order constraints for the summary constituent units whose importance has been evaluated, and summary constituent units that are highly important and take into consideration the extraction order And summary result output means for creating a summary.

The seventh text summarization system of the present invention reads a summary constituent unit qualification pattern including a collation condition, a summary constituent unit having a clause containing a predicate of a sentence, an application cost, and an extraction order constraint from the storage device, and reads it out. A summary component unit is certified using the summary component unit recognition pattern in consideration of the connection relationship between the summary component units or the citation relationship, and the summary component unit certification method and the summary that evaluates the importance of the recognized summary component unit Importance calculation means for composition unit, extraction order constraint that considers the extraction order using the extraction order constraint for summary composition units evaluated for importance, and extraction with high importance in consideration of application cost Summarization result output means for extracting a summary constituent unit considering the order and creating a summary.

The eighth text summarization system of the present invention reads a summary constituent unit recognition pattern including a summary constituent unit having a clause including a collation condition and a predicate of a sentence as a unit, and uses the read summary constituent unit recognition pattern. Summary composition unit recognition means for identifying summary composition unit, summary composition unit complement correction pattern including collation conditions and supplementary modification information is read from the storage device, and the summary composition unit complement modification pattern is read out and used for the summary composition unit. Complement / correction means for summarizing composition units that output after completion / correction based on sentence expressions for the units, and importance of the summarizing composition units for evaluating the importance of the summarizing composition units from the supplementary correction means for the summary composition units When the summary calculation unit and the summary unit adjacent in the summary are adjacent in the same sentence of the original document, For complementation and correction, it has expression reduction means for applying the original expression, and summary result output means for extracting a summary component unit having a high importance with respect to the summary component unit from the reduction means and creating a summary It is characterized by.

The ninth text summarization system of the present invention reads a summary constituent unit qualification pattern including a summary constituent unit whose unit is a clause including a collation condition and a sentence predicate from a storage device, and uses the read summary constituent unit qualification pattern. The unit that minimizes the section including the predicate is recognized as the summary unit, and the summary sentence candidate generating means for obtaining the summary sentence candidate from the unit, and the summary unit complement correction pattern including the collation condition, supplementary correction information, and application cost Using the summary unit correction correction pattern read out from the storage device, complementing and correcting the summary sentence unit based on the sentence expression for the summary sentence candidate, and outputting the summary composition unit Summary sentence unit importance calculation means for evaluating the importance of summary sentence candidates from the unit supplementary correction means, and summary sentence candidates in each sentence based on the importance for the summary sentence candidate Searching for a summary sentence candidate that takes into account the cost for certifying the summary unit and the cost for supplementing and correcting the summary unit when determining the combination of the most likely summary sentences as a summary And a summary result output means for arranging the determined combinations of maximum likelihood summary sentence candidates in the order of appearance and outputting them as summary results.

The first text summarization method of the present invention reads summary constituent unit recognition language knowledge information including information having a section including a predicate of a sentence as a summary constituent unit from a storage device, and reads the read summary knowledge of the constituent unit recognition language knowledge. The summary composition unit qualification procedure to certify the summary composition unit using information, the summary composition unit importance calculation procedure to evaluate the importance of the recognized summary composition unit, and the summary composition unit with high importance are extracted. A summary result output procedure for creating a summary.

The second text summarization method of the present invention reads a summary constituent unit qualification pattern including a summary constituent unit whose unit is a clause including a collation condition and a sentence predicate from a storage device, and uses the read summary constituent unit qualification pattern. Summary composition unit qualification procedure, summary composition unit importance calculation procedure to evaluate the importance of the certified summary composition unit, and summary summary by extracting the most important summary composition unit And a summary result output procedure.

The third text summarization method of the present invention reads out a summary constituent unit recognition pattern including a collation condition, a summary constituent unit having a clause including a statement predicate as a unit, and an application cost from the storage device, and reads out the summary constituent unit recognition Summary composition unit qualification procedure for certifying summary composition units using patterns, summary composition unit importance calculation procedure for evaluating the importance of certified summary composition units, and high importance summary composition considering application costs A summary result output procedure for extracting a unit and creating a summary.

A fourth text summarization method according to the present invention is the first, second, or third text summarization method, and reads a summary unit complement correction pattern including collation conditions and supplement correction information from the storage device. Using the read summary composition unit complement modification pattern, the summary composition unit certified by the summary composition unit certification procedure is supplemented / modified based on the sentence expression, and the supplemented / modified summary composition unit is obtained. It has a supplementary correction procedure for the summary constituent unit that is output to the importance calculation procedure for the summary constituent unit.

A fifth text summarization method according to the present invention is the first, second, or third text summarization method, and reads a summary unit complement correction pattern including collation conditions and supplement correction information from the storage device. Then, using the read summary composition unit complement correction pattern, the summary composition unit whose importance is evaluated by the summary component unit importance calculation procedure is complemented / corrected based on the sentence expression, and complemented / corrected. A summary composition unit complementing and correcting procedure for outputting the summarized composition unit to the summary result output procedure.

The sixth text summarization method of the present invention reads out a summary constituent unit including a collation condition, a clause including a predicate of a sentence as a unit, and a summary constituent unit qualification pattern including an extraction order constraint from a storage device, and reads out the summary constituent unit The summary composition unit is evaluated by certifying the summary composition unit and evaluating the importance of the certified summary composition unit by certifying the summary composition unit in consideration of the connection relation between the summary composition units or the citation relation using the authorization pattern. Importance calculation procedure, extraction order constraint setting procedure that considers the extraction order using the extraction order constraint for the summary constituent units whose importance has been evaluated, and a summary constituent unit that is highly important and takes into consideration the extraction order And a summary result output procedure for creating a summary.

According to the seventh text summarization method of the present invention, a summary constituent unit qualification pattern including a collation condition, a summary constituent unit having a clause including a predicate of a sentence, an application cost, and an extraction order constraint is read from the storage device and read. Summarize the summary composition unit by evaluating the summary composition unit and the importance of the summary composition unit by certifying the summary composition unit by considering the connection relation between the summary composition units or the citation relation using the summary composition unit certification pattern. The importance calculation procedure of the composition unit, the setting procedure of the restriction on the extraction order that considers the extraction order using the extraction order constraint for the summary composition unit whose importance is evaluated, and the extraction with high importance in consideration of the application cost A summary result output procedure for extracting a summary constituent unit considering the order and creating a summary.

The eighth text summarization method of the present invention reads a summary constituent unit recognition pattern including a summary constituent unit having a clause including a collation condition and a predicate of a sentence as a unit, and uses the read summary constituent unit recognition pattern. Summary composition unit recognition procedure for identifying summary composition unit, summary composition unit complement correction pattern including collation condition and supplementary modification information is read from the storage device, and the summary composition unit complement modification pattern is read out and used for the summary composition unit. Complement / correction procedure for the summary component unit that outputs after completing / correcting the unit based on the sentence expression, and the importance of the summary component unit for evaluating the importance of the summary component unit from the supplementary correction procedure of the summary component unit When the summary calculation unit and the summary unit adjacent to each other in the summary are adjacent to each other in the same sentence of the original document, the sum of the summary units adjacent to the summary unit is corrected. And a reduction procedure for the expression to which the original expression is applied, and a summary result output procedure for extracting a summary component unit having a high importance with respect to the summary component unit from the reduction procedure and generating a summary. Features.

The ninth text summarization method of the present invention reads a summary constituent unit recognition pattern including a summary constituent unit having a unit including a clause including a collation condition and a statement predicate from a storage device, and uses the read summary constituent unit recognition pattern. The unit that minimizes the section containing the predicate is recognized as the summary unit, and the summary sentence unit generation procedure for obtaining the summary sentence candidate from this, and the summary unit completion correction pattern including the matching conditions, supplementary correction information, and application cost Using the read summary composition unit complement correction pattern, the summary composition unit complement correction procedure for performing the completion and correction on the summary sentence candidate based on the sentence expression, and outputting the summary composition unit Search summary sentence candidates in each sentence based on the importance calculation procedure for summary composition units that evaluate the importance of summary sentence candidates from the unit's complementary correction procedure and the importance for summary sentence candidates A procedure for searching for a summary sentence candidate that takes into account the cost for certifying the summary unit and the cost for supplementing and correcting the summary unit when determining the most likely summary sentence combination as a summary; And a summary result output procedure for arranging the determined combinations of maximum likelihood summary sentence candidates in the order of appearance and outputting them as summary results.

The first text summarization program according to the present invention reads out summary constituent unit recognition language knowledge information including information having a section including a sentence predicate as a summary constituent unit from a storage device, and reads the read out summary constituent unit recognition language knowledge. The summary composition unit qualification procedure to certify the summary composition unit using information, the summary composition unit importance calculation procedure to evaluate the importance of the recognized summary composition unit, and the summary composition unit with high importance are extracted. A summary result output procedure for creating a summary is executed by a computer.

The second text summarization program of the present invention reads a summary constituent unit qualification pattern including a summary constituent unit having a collation condition and a clause including a sentence predicate as a unit, and uses the read summary constituent unit qualification pattern. Summary composition unit qualification procedure, summary composition unit importance calculation procedure to evaluate the importance of the certified summary composition unit, and summary summary by extracting the most important summary composition unit And a summary result output procedure to be executed by a computer.

The third text summarization program according to the present invention reads a summary constituent unit recognition pattern including a collation condition, a summary constituent unit having a clause including a statement predicate as a unit, and an application cost from the storage device, and reads the read summary constituent unit. Summary composition unit qualification procedure for certifying summary composition units using patterns, summary composition unit importance calculation procedure for evaluating the importance of certified summary composition units, and high importance summary composition considering application costs A summary result output procedure for extracting a unit and creating a summary is executed by a computer.

A fourth text summarization program according to the present invention is the first, second, or third text summarization program, and reads a summary unit complement correction pattern including collation conditions and supplementary correction information from the storage device. Using the read summary composition unit complement modification pattern, the summary composition unit certified by the summary composition unit certification procedure is supplemented / modified based on the sentence expression, and the supplemented / modified summary composition unit is obtained. The computer is caused to execute a supplementary correction procedure for the summary constituent unit that is output to the importance calculation procedure for the summary constituent unit.

A fifth text summarization program according to the present invention is the first, second, or third text summarization program, and reads a summary unit complement correction pattern including collation conditions and supplementary correction information from the storage device. Then, using the read summary composition unit complement correction pattern, the summary composition unit whose importance is evaluated by the summary component unit importance calculation procedure is complemented / corrected based on the sentence expression, and complemented / corrected. The computer is caused to execute a procedure for complementing and correcting the summary composition unit for outputting the summarized composition unit to the summary result output procedure.

The sixth text summarization program according to the present invention reads a summary constituent unit having a collation condition, a clause containing a predicate of a sentence as a unit, and a summary constituent unit recognition pattern containing an extraction order constraint from a storage device, and reads the summary constituent unit The summary composition unit is evaluated by certifying the summary composition unit and evaluating the importance of the certified summary composition unit by certifying the summary composition unit in consideration of the connection relation between the summary composition units or the citation relation using the authorization pattern. Importance calculation procedure, extraction order constraint setting procedure that considers the extraction order using the extraction order constraint for the summary constituent units whose importance has been evaluated, and a summary constituent unit that is highly important and takes into consideration the extraction order And a summary result output procedure for generating a summary by extracting the information.

The seventh text summarization program of the present invention reads a summary constituent unit recognition pattern including a collation condition, a summary constituent unit having a clause containing a predicate of the sentence, an application cost, and an extraction order constraint from the storage device, and reads Summarize the summary composition unit by evaluating the summary composition unit and the importance of the summary composition unit by certifying the summary composition unit by considering the connection relation between the summary composition units or the citation relation using the summary composition unit certification pattern. The importance calculation procedure of the composition unit, the setting procedure of the restriction on the extraction order that considers the extraction order using the extraction order constraint for the summary composition unit whose importance is evaluated, and the extraction with high importance in consideration of the application cost A summary result output procedure for extracting a summary component unit in which the order is considered and creating a summary is executed by a computer.

The eighth text summarization program of the present invention reads a summary constituent unit recognition pattern including a summary constituent unit including a collation condition and a clause including a statement predicate from a storage device, and uses the read summary constituent unit recognition pattern. Summary composition unit recognition procedure for identifying summary composition unit, summary composition unit complement correction pattern including collation condition and supplementary modification information is read from the storage device, and the summary composition unit complement modification pattern is read out and used for the summary composition unit. Complement / correction procedure for the summary component unit that outputs after completing / correcting the unit based on the sentence expression, and the importance of the summary component unit for evaluating the importance of the summary component unit from the supplementary correction procedure of the summary component unit If the summary unit adjacent to the summary calculation unit and the summary unit adjacent to each other in the same sentence of the original document are adjacent to each other, For the completion and correction of the computer, a reduction procedure for the expression to which the original expression is applied, and a summary result output procedure for extracting a summary component unit having a high importance from the summary component unit from the reduction procedure and generating a summary It is made to perform.

A ninth text summarization program of the present invention reads a summary constituent unit recognition pattern including a summary constituent unit having a section including a collation condition and a sentence predicate as a unit, and uses the read summary constituent unit recognition pattern. The unit that minimizes the section containing the predicate is recognized as the summary unit, and the summary sentence unit generation procedure for obtaining the summary sentence candidate from this, and the summary unit completion correction pattern including the matching conditions, supplementary correction information, and application cost Using the read summary composition unit complement correction pattern, the summary composition unit complement correction procedure for performing the completion and correction on the summary sentence candidate based on the sentence expression, and outputting the summary composition unit Summarization candidate for each sentence based on the importance calculation procedure for the summary component unit that evaluates the importance of the summary sentence candidate from the unit's complementary correction procedure and the importance for the summary sentence candidate Search for summary sentence candidates that are extracted in consideration of the cost for certifying the summary structural unit and the cost for complementing and correcting the summary structural unit when determining the combination of the most likely summary sentence candidates as a summary. The computer is caused to execute a procedure and a summary result output procedure in which combinations of the determined maximum likelihood summary sentence candidates are arranged in order of appearance and output as a summary result.

The first effect is that it is possible to simultaneously improve the summary quality by using an appropriate summary unit and speed up the process by simplifying the summary generation process.

The reason for this is that the unit that minimizes the clause containing the predicate for each sentence of the input text is recognized as the summary unit, and the summary is generated using this unit as the unit. This is because the short-term summary generation limit and quality degradation are less likely to occur compared to the summarization method that uses sentences as the summary unit, and the summary structure unit has a rough syntactic structure compared to the summarization method that uses sentences as the summary unit. This is because a summary can be generated by a simpler process.

The second effect is that it is possible to suppress the occurrence of expressions that are difficult to read and expressions that do not understand the meaning in the summary.

The reason is that even if the summary unit extracted from the sentence is an inappropriate expression as an independent sentence in the summary sentence, it is appropriately supplemented or corrected based on the expression of the sentence in the original document. This is to set up suppression on the order of extraction in extraction of summary units that cause misunderstanding and difficulty in reading due to the order of extraction such as citation relationships and connection relationships. This is because the continuity between sentences is prevented from being lost by reducing the portion to the original expression.

A third effect is that a higher-quality summary sentence based on a more flexible summary unit is generated.

The reason for this is that multiple summary sentence candidates are created for each sentence by performing sequential matching of the summary unit recognition pattern for each sentence of the input text for all possible combinations. This is to select the most likely candidate from among them, and at the same time, by applying a large cost to the application of the unit recognition pattern with low reliability and the application of supplementary correction, the application of the unit recognition pattern with high cost and the supplementary correction can be performed. This is because the suppression of summary sentences that are suppressed and the reliability of summary fair units and supplementary amendments are applied to improve summary quality.

It is a block diagram which shows the structure of the 1st Embodiment of this invention. It is a block diagram which shows the structure of the 2nd Embodiment of this invention. It is a block diagram which shows the structure of the 3rd Embodiment of this invention. It is a block diagram which shows the structure of the 4th Embodiment of this invention. It is a block diagram which shows the structure of the 5th Embodiment of this invention. It is a block diagram which shows the structure of the 6th Embodiment of this invention. It is a block diagram which shows the structure of the 7th Embodiment of this invention. It is a block diagram which shows the structure of the 8th Embodiment of this invention. It is a block diagram which shows the structure of the 9th Embodiment of this invention. It is a block diagram which shows the structure of the 10th Embodiment of this invention. It is explanatory drawing which shows the example of the summary structural unit recognition pattern of 2nd, 6th, 7th, 9th embodiment of this invention. It is explanatory drawing which shows the example of the summary structural unit complement correction pattern of the 6th, 9th embodiment of this invention. It is explanatory drawing which shows the example of the summary structural unit recognition pattern of the 8th Embodiment of this invention. It is explanatory drawing which shows the example of the summary structural unit recognition pattern of the 3rd, 10th embodiment of this invention. It is explanatory drawing which shows the example of the summary structural unit complement correction pattern of the 7th, 10th Embodiment of this invention. It is explanatory drawing which shows an example of the summary sentence candidate obtained in the 10th Embodiment of this invention.

Next, a first embodiment of the present invention will be described in detail with reference to the drawings.

FIG. 1 is a block diagram showing the configuration of the first exemplary embodiment of the present invention.

Referring to FIG. 1, the first embodiment of the present invention is a summary constituent unit recognition means 101, a summary constituent unit importance calculation means 102, a summary result output means 103, and a summary constituent unit recognition language. It consists of knowledge information 104.

The summary structural unit recognition language knowledge information 104 is stored in a storage device such as a memory or a hard disk, and is read and used.

Each of these means generally functions as follows.

The summary constituent unit recognition means 101 inputs text, and uses the summary constituent unit recognition language knowledge information 104 for each sentence of the input text as a summary constituent unit. Authorize. The summary component importance level calculation means 102 uses the information such as the position of the summary component unit in the text, the clue word, the title, the importance of the word, the keyword, etc. as a clue to each summary component unit. Calculate the importance of a composition unit. The summary result output means 103 sequentially extracts from the summary component units having the highest importance until the summary length requested by the user is reached based on the importance for each summary component unit. Arrange in the order of appearance in the original text and output as a summary result.

Here, as the summary constituent unit recognition means 101, for example, means for performing syntax analysis or means for performing analysis on the context within a sentence (for example, Non-Patent Document 5) can be used. Further, a rule for extracting a clause including a predicate from the result of parsing can be used as the summary unit recognition language knowledge information 104.

Next, the operation of the first exemplary embodiment of the present invention will be described in detail.

First, for each sentence of the input text, the summary unit recognition language knowledge information 104 is used to identify the unit that minimizes the clause containing the predicate as the summary unit (summary unit recognition means 101). Next, the importance of each summary component in the text is calculated using the information such as the position of the summary component in the text, clue words, title, word importance, keyword, etc. ( Summary component unit importance calculation means 102), and finally, based on the importance for each summary component unit, sequentially extract from the most important summary component unit until the summary length requested by the user is reached, The extracted summary constituent units are arranged in the order of appearance in the original text and output as a summary result (summary result output means 103).

Next, effects of the first exemplary embodiment of the present invention will be described.

In the first embodiment of the present invention, the summary unit recognition unit 101 recognizes, as a summary unit, a unit that minimizes a clause including a predicate for each sentence of the input text. Since the summary is generated, the degree of freedom in generating the summary is high, and a high-quality summary can be generated even when a short summary is generated.

Next, a second embodiment of the present invention will be described in detail with reference to the drawings.

FIG. 2 is a block diagram showing the configuration of the second exemplary embodiment of the present invention.

Referring to FIG. 2, the second embodiment of the present invention is a summary constituent unit recognition unit 201, a summary constituent unit importance calculation unit 202, a summary result output unit 203, and a summary constituent unit recognition pattern 204. It consists of and. The summary structural unit recognition pattern 204 is stored in a storage device such as a memory or a hard disk and is read and used, but will not be described below.

Each of these means generally functions as follows.

The summary constituent unit recognition means 201 recognizes a unit that minimizes a clause including a predicate as a summary constituent unit by using the summary constituent unit recognition pattern 204 for authorizing the summary constituent unit for each sentence of the input text. To do. The summary component importance calculation means 202 uses each of the summaries in the text as a clue to information such as the position of the summary component in the text, the clue word, the title, the importance of the word, and the keyword. Calculate the importance of a composition unit. The summary result output means 203 sequentially extracts from the summary component unit having the highest importance until the summary length requested by the user is reached, based on the importance of each summary component unit, and extracts the extracted summary component units. Arrange in the order of appearance in the original text and output as a summary result.

Next, the operation of the second exemplary embodiment of the present invention will be described in detail.

First, a unit that minimizes a clause including a predicate is recognized as a summary unit using a summary unit determination pattern 204 for certifying a summary unit for each sentence of the input text (summarization unit recognition means). 201) Next, the importance of each summary constituent unit in the text is obtained by using information such as the position of the summary constituent unit in the text, the clue word, the title, the importance of the word, and the keyword with respect to the summary constituent unit. (Summary composition unit importance calculation means 202), and finally, based on the importance for each summary composition unit, from the most important summary composition unit until the summary length requested by the user is reached The extracted summary constituent units are sequentially extracted, arranged in the order of appearance in the original text, and output as a summary result (summary result output means 203).

Next, effects of the second exemplary embodiment of the present invention will be described.

In the second embodiment of the present invention, the summary unit recognition unit 201 recognizes, as a summary unit, a unit that minimizes the clause containing the predicate for each sentence of the input text. Since a summary is generated, it is possible to generate a high-quality summary even when a short summary is generated. Since the summary constituent unit can be recognized, the summary can be generated by simpler processing.

Next, a third embodiment of the present invention will be described in detail with reference to the drawings.

FIG. 3 is a block diagram showing the configuration of the third exemplary embodiment of the present invention.

Referring to FIG. 3, the third embodiment of the present invention is a summary constituent unit recognition means 301, a summary constituent unit importance calculation means 302, a summary result output means 303, and a summary constituent unit recognition pattern 304. It consists of and. The summary structural unit recognition pattern 304 is stored in a storage device such as a memory or a hard disk and is read and used, but will not be described below.

Each of these means generally functions as follows.

The summary component unit recognition means 301 recognizes the summary component unit for each sentence of the input text, and further uses the summary component unit recognition pattern 304 in which the cost for certifying each summary component unit can be set. The unit that minimizes the section containing the part is recognized as the summary unit. The summary component unit importance calculation means 302 uses the information such as the position of the summary component unit in the text, the clue word, the title, the importance of the word, the keyword, etc. as clues in the text. Calculate the importance of each summary component. The summary result output unit 303 recognizes the summary constituent units when sequentially extracting from the summary constituent units having the highest importance until the summary length requested by the user is reached based on the importance for each summary constituent unit. The extraction is performed in consideration of the cost for the extraction, and the extracted summary constituent units are arranged in the order of appearance in the original text and output as a summary result.

Next, the operation of the third exemplary embodiment of the present invention will be described in detail.

First, by using the summary composition unit recognition pattern 304 that can recognize the summary composition unit for each sentence of the input text and set the cost for certifying the summary composition unit in each summary composition unit recognition pattern 304. Then, the unit that minimizes the section including the predicate is recognized as the summary unit (summary unit determination means 301), and then the position and clue of the summary unit in the text with respect to these summary units Using the information such as the word, title, word importance, and keyword as a clue, the importance of each summary constituent unit in the text is calculated (summary constituent unit importance calculation means 302), and finally, for each summary constituent unit Based on the importance, the summary composition unit is certified when extracting in order from the highest importance summary composition unit until the summary length requested by the user is reached. Cost of eyes was extracted in consideration arranges the extracted summarized structural units in order of appearance in the original text, and outputs the results are summarized (summarized result output unit 303).

Next, effects of the third exemplary embodiment of the present invention will be described.

In the third embodiment of the present invention, the summary unit recognition unit 201 recognizes, as a summary unit, a unit that minimizes a clause including a predicate for each sentence of the input text. Since a summary is generated, it is possible to generate a high-quality summary even when a short summary is generated. Since the summary constituent unit can be recognized, the summary can be generated by a simpler process, and the application of the unit recognition pattern with high cost can be suppressed, so that a high cost is given to a pattern with low reliability. As a result, the extraction of summary constituent units with low reliability into the summary sentence is suppressed, and an effect of generating a high-quality summary sentence can be obtained.

Next, a fourth embodiment of the present invention will be described in detail with reference to the drawings.

FIG. 4 is a block diagram showing the configuration of the fourth exemplary embodiment of the present invention.

Referring to FIG. 4, the fourth embodiment of the present invention is a summary constituent unit recognition means 401, a summary constituent unit supplementary correction means 402, a summary constituent unit importance calculation means 403, and a summary result output. It comprises means 404, summary constituent unit recognition language knowledge information 405, and summary constituent unit complementary correction language knowledge information 406. Here, for example, a means (for example, Patent Document 2) that implements a sentence analysis method for estimating the existence of a shared subject / master case can be used as the supplementary correction means 402 for the summary constituent unit. Summary unit knowledge language knowledge information 405 and summary unit supplement correction language knowledge information 406 are stored in a storage device such as a memory or a hard disk and are read and used, but will not be described below.

Each of these means generally operates as follows.

The summary constituent unit recognition unit 401 uses the summary constituent unit recognition language knowledge information 405 for each sentence of the input text to recognize the unit that minimizes the clause including the predicate as the summary constituent unit. By applying the summary component unit complement correction language knowledge information 406, the summary component unit supplementary correction means 402 is configured so that the importance level in the text can be appropriately determined only by the summary component unit information. The information shared in the sentence is supplemented to the summary constituent unit, or the summary constituent unit is corrected based on the sentence expression in the original document so that the expression becomes an appropriate expression as an independent sentence in the summary sentence.

The summary unit importance calculation means 403 calculates the position of the summary unit, the clue word, the title, and the importance of the word with respect to the summary unit in which the information is appropriately supplemented or the expression is corrected. The importance of each summary component in the text is calculated using information such as the degree and keywords. The summary result output means 404 sequentially extracts from the summary component unit having the highest importance until the summary length requested by the user is reached, based on the importance for each summary component unit, and extracts the extracted summary component units. Arrange in the order of appearance in the original text and output as a summary result.

Next, the operation of the fourth exemplary embodiment of the present invention will be described in detail.

First, using the summary component unit recognition language knowledge information 405 for each sentence of the input text, a unit that minimizes a clause including a predicate is recognized as a summary unit (summary unit recognition means 401). Next, the information shared in the original sentence is applied so that the importance in the text can be appropriately judged only by the information of the summary composition unit by applying the language knowledge information 406 for complementing the composition correction of the composition unit. Is appropriately supplemented based on the sentence representation in the original document so that the expression of the summary unit becomes an appropriate expression as an independent sentence in the summary sentence. Complementary correcting means 402), and further, the position, clue word, title, simpleness in the text of the summary constituent unit with respect to the summary constituent unit that has been appropriately supplemented or corrected for the information. The importance of each summary component in the text is calculated using information such as the importance of the keyword and keywords (summary component importance calculation means 403), and finally, based on the importance for each summary component , In order from the most important summary unit until the summary length requested by the user is reached, the extracted summary units are arranged in the order of appearance in the original text and output as a summary result (summary result output) Means 404).

Next, the effect of the 4th Embodiment of this invention is demonstrated.

In the fourth embodiment of the present invention, the summary component unit is appropriately supplemented or corrected based on the sentence expression in the original document by the summary component unit supplementary correction unit 402. Even if extracted in the summary sentence, incomplete expression does not occur as a sentence, and the importance is appropriately evaluated by the information of the supplemented summary constituent unit, so that the quality of the summary can be improved.

Next, a fifth embodiment of the present invention will be described in detail with reference to the drawings.

FIG. 5 is a block diagram showing the configuration of the fifth exemplary embodiment of the present invention.

Referring to FIG. 5, the fifth embodiment of the present invention is a summary component unit certifying unit 501, a summary component unit importance calculating unit 502, a summary unit supplementary correcting unit 503, and a summary result output. It comprises means 504, summary constituent unit recognition language knowledge information 505, and summary constituent unit complementary correction language knowledge information 506. Summary unit knowledge language knowledge information 505 and summary unit supplement correction language knowledge information 506 are stored in a storage device such as a memory or a hard disk and are read and used, but are not described below.

Each of these means generally functions as follows.

The summary constituent unit recognition unit 501 uses the summary constituent unit recognition language knowledge information 505 for each sentence of the input text to recognize the unit that minimizes the clause including the predicate as the summary constituent unit. The summary component unit importance calculation means 502 uses the information such as the position of the summary component unit in the text, the clue word, the title, the importance of the word, and the keyword as a clue to the summary component unit. Calculate the importance of each summary component. The summary composition unit supplementary correction means 503 applies the summary composition unit complement correction language knowledge information 506 so that the expression of the summary composition unit becomes an appropriate expression as an independent sentence in the summary sentence. Correct based on the expression. The summary result output means 504 sequentially extracts from the summary component units having the highest importance, sequentially until the summary length requested by the user is reached, based on the importance for each summary component unit. Arrange in the order of appearance in the original text and output as a summary result.

Next, the operation of the fifth exemplary embodiment of the present invention will be described in detail.

First, using the summary component unit recognition language knowledge information 505 for each sentence of the input text, a unit that minimizes the clause including the predicate is recognized as a summary unit (summary unit recognition means 501). Next, with respect to these summary composition units, information such as the position of the summary composition unit in the text, clue words, title, word importance, and keywords are used as clues, and the importance of each summary composition unit in the text is determined. By calculating (summary constituent unit importance calculation means 502) and applying the summary constituent unit complement correction language knowledge information 506, the expression of the summary constituent unit becomes an appropriate expression as an independent sentence in the summary sentence. Thus, the sentence is appropriately corrected based on the sentence expression in the original document (summary composition unit supplementary correction means 503), and finally, based on the importance for each summary composition unit, The summary constituent units having the highest importance are sequentially extracted until the summary length requested by the user is reached, the extracted summary constituent units are arranged in the order of appearance in the original text, and output as summary results (summary result output means 504). ).

Next, effects of the fifth exemplary embodiment of the present invention will be described.

In the fifth embodiment of the present invention, the summary component unit is appropriately supplemented or corrected based on the expression of the sentence in the original document by the summary component complement correction unit 503. Is extracted in the summary sentence, an incomplete expression as a sentence does not occur, and the quality of the summary can be improved.

Next, a sixth embodiment of the present invention will be described in detail with reference to the drawings.

FIG. 6 is a block diagram showing the configuration of the sixth exemplary embodiment of the present invention.

Referring to FIG. 6, the sixth embodiment of the present invention is a summary constituent unit certifying means 601, a summary constituent unit supplementary correcting means 602, a summary constituent unit importance calculating means 603, and a summary result output. A means 604, a summary constituent unit recognition pattern 605, and a summary constituent unit complementary correction pattern 606 are configured. The summary constituent unit recognition pattern 605 and the summary constituent unit complementary correction pattern 606 are stored in a storage device such as a memory and a hard disk, and are read and used.

Each of these means generally functions as follows.

The summary constituent unit recognition means 601 uses the summary constituent unit recognition pattern 605 for each sentence of the input text to recognize the unit that minimizes the clause including the predicate as the summary constituent unit. The summary composition unit complement correction means 602 applies the summary composition unit complement correction pattern 606, so that the importance level in the text can be appropriately judged only by the information of the summary composition unit. The supplemented information is supplemented to the summary constituent unit, or the summary constituent unit is corrected based on the sentence expression in the original document so that the expression of the summary constituent unit becomes an appropriate expression as an independent sentence in the summary sentence. The summary component unit importance calculation means 603 is the position of the summary component unit in the text, the clue word, the title, and the importance of the word with respect to the summary component unit that has been appropriately supplemented or modified in expression. The importance of each summary component in the text is calculated using information such as the degree and keywords. The summary result output means 604 sequentially extracts the summary constituent units in order from the summary constituent unit having the highest importance until the summary length requested by the user is reached, based on the importance for each summary constituent unit. Arrange in the order of appearance in the original text and output as a summary result.

Next, the operation of the sixth exemplary embodiment of the present invention will be described in detail.

First, by using the summary constituent unit recognition pattern 605 for each sentence of the input text, the unit that minimizes the clause including the predicate is recognized as the summary constituent unit (summary constituent unit recognition means 601), and then By applying the summary composition unit complement correction pattern 606, the information shared in the original sentence is appropriately applied to the summary composition unit so that the importance level in the text can be appropriately judged only by the information of the summary composition unit. Or appropriately correcting based on the expression of the sentence in the original document so that the expression of the summary component unit becomes an appropriate expression as an independent sentence in the summary sentence (summary component unit complement correcting means 602), Furthermore, the position, clue word, title, word importance, key in the summary composition unit text in relation to the summary composition unit in which such information has been appropriately complemented or corrected. The importance of each summary component in the text is calculated using the information such as the code (summary component importance calculation means 603), and finally, the most important based on the importance for each summary component. The summary components are extracted in order from the highest summary component until the summary length requested by the user is reached, and the extracted summary components are arranged in the order of appearance in the original text and output as summary results (summary result output means 604). .

Next, effects of the sixth exemplary embodiment of the present invention will be described.

In the sixth embodiment of the present invention, the summary component unit is appropriately supplemented or corrected based on the expression of the sentence in the original document by the summary component unit supplementary correction means 602. Even if extracted in the summary sentence, incomplete expression does not occur as a sentence, and the importance is appropriately evaluated by the information of the supplemented summary constituent unit, so that the quality of the summary can be improved.

Next, a seventh embodiment of the present invention will be described in detail with reference to the drawings.

FIG. 7 is a block diagram showing a configuration of the seventh exemplary embodiment of the present invention.

Referring to FIG. 7, the seventh embodiment of the present invention is a summary constituent unit certifying means 701, a summary constituent unit complementary correcting means 702, a summary constituent unit importance calculating means 703, and a summary result output. A means 704, a summary constituent unit recognition pattern 705, and a summary constituent unit complementary correction pattern 706 are configured. The summary constituent unit recognition pattern 705 and the summary constituent unit complementary correction pattern 706 are stored in a storage device such as a memory and a hard disk, and are read and used.

Each of these means generally functions as follows.

The summary constituent unit recognition means 701 uses the summary constituent unit recognition pattern 705 for each sentence of the input text to recognize the unit that minimizes the clause including the predicate as the summary constituent unit. The summary composition unit complement correction means 702 applies the summary composition unit complement correction pattern 706 to share the original sentence so that the importance level in the text can be appropriately determined only by the summary composition unit information. Appropriately supplement the information contained in the summary unit, or modify it appropriately based on the sentence representation in the original document so that the summary unit expression is an appropriate expression as an independent sentence in the summary sentence, Further, a cost for complementing and correcting the summary constituent unit is set by each summary constituent unit complementing correction pattern 706.

The summary component unit importance calculation means 703 performs the position of the summary component unit in the text, the clue word, the title, and the importance of the word with respect to the summary component unit in which the information is appropriately supplemented or the expression is corrected. The importance of each summary component in the text is calculated using information such as the degree and keywords. The summary result output means 704 complements and corrects the summary constituent units when extracting sequentially from the summary constituent units having the highest importance until the summary length requested by the user is reached, based on the importance for each summary constituent unit. Then, the extraction is performed in consideration of the cost for performing, and the extracted summary constituent units are arranged in the order of appearance in the original text and output as a summary result.

Next, the operation of the seventh exemplary embodiment of the present invention will be described in detail.

First, by using the summary constituent unit recognition pattern 705 for each sentence of the input text, the unit that minimizes the clause including the predicate is recognized as the summary constituent unit (summary constituent unit recognition means 701), and then By applying the summary composition unit complement correction pattern 706, the information shared in the original sentence is appropriately applied to the summary composition unit so that the importance level in the text can be appropriately judged only by the information of the summary composition unit. Or appropriately correct based on the expression of the sentence in the original document so that the expression of the summary unit becomes an appropriate expression as an independent sentence in the summary sentence. To set the cost for complementing and correcting the summary constituent unit (summary constituent unit supplementary correcting means 702), and further complementing or correcting the information appropriately. For each summary component, the importance of each summary component in the text is calculated using information such as the position of the summary component in the text, clue words, title, word importance, and keywords (summary). The importance calculation means 703) of the composition unit, and finally, when extracting sequentially from the summary composition unit having the highest importance in order until the summary length requested by the user is reached based on the importance for each summary composition unit, Extraction is performed in consideration of the cost for complementing and correcting the summary constituent units, and the extracted summary constituent units are arranged in the order of appearance in the original text and output as summary results (summary result output means 704).

Next, effects of the seventh exemplary embodiment of the present invention will be described.

In the seventh embodiment of the present invention, a summary component unit is appropriately supplemented or corrected based on the sentence representation in the original document by the summary component complement correction means 702. Even if extracted in the summary sentence, incomplete expression does not occur as a sentence, and the importance is appropriately evaluated by the information of the supplemented summary constituent unit, so that the quality of the summary can be improved. In addition, since the cost for supplementary correction of the summary constituent unit can be set by each of the summary constituent unit supplemental correction patterns 706, a large cost is given to a pattern with low reliability, so that correction correction with low reliability can be performed. The summary sentence by application is suppressed, and thereby, a higher-quality summary sentence can be generated.

Next, an eighth embodiment of the present invention will be described in detail with reference to the drawings.

FIG. 8 is a block diagram showing the configuration of the eighth embodiment of the present invention.

Referring to FIG. 8, the eighth embodiment of the present invention includes a summary component unit certifying unit 801, a summary component unit importance calculating unit 802, and a constraint setting unit 803 for extracting the summary component units. , A summary result output means 804 and a summary constituent unit recognition pattern 805. The summary structural unit recognition pattern 805 is stored in a storage device such as a memory or a hard disk and is read and used, but will not be described below.

Each of these means generally functions as follows.

The summary constituent unit recognition means 801 recognizes the summary constituent unit for each sentence of the input text, and further recognizes the connection relation and the citation relation between the summary constituent units certified by the respective summary constituent unit certification pattern 805. Then, using the summary unit determination pattern 805, a unit that minimizes a clause including a predicate is recognized as a summary unit, and further, when extracting each summary unit, an extraction order that considers these relationships is used. Constraints can be set. The summary component unit importance calculation means 802 uses the information such as the position of the summary component unit in the text, the clue word, the title, the importance of the word, and the keyword as a clue to the summary component unit. Calculate the importance of each summary component. The restriction setting unit 803 regarding the extraction order of the summary constituent units sequentially extracts from the summary constituent units having the highest importance in order until the summary length requested by the user is reached based on the importance for each summary constituent unit. In addition, the extraction is performed in consideration of the restriction on the extraction order in consideration of the relationship between the respective summary constituent units. The summary result output means 804 arranges the extracted summary constituent units in the order of appearance in the original text, and outputs them as a summary result.

Next, the operation of the eighth exemplary embodiment of the present invention will be described in detail.

First, for each sentence of the input text, a summary constituent unit is certified, and further, a connection relation and a citation relation between the summary constituent units certified by each summary constituent unit certification pattern 805 are certified, and a summary constituent unit certification pattern is obtained. Using 805, the unit that minimizes the clause containing the predicate is recognized as the summary unit, and further, when extracting each summary unit, constraints on the extraction order considering these relationships are set (summary configuration). Unit certifying means 801) Next, with respect to these summary constituent units, information such as the position of the summary constituent unit in the text, clue words, titles, importance of words, keywords, etc. are used as clues, and The importance of the summary composition unit is calculated (summary composition unit importance calculation means 802), and the weight of each summary composition unit is calculated based on the importance for each summary composition unit. When extracting in order from the highest summary component until the summary length required by the user is reached, the extraction is performed in consideration of the restriction on the extraction order considering the relationship between each summary component (summary component unit). Constraint setting means 803) regarding the extraction order, and finally, the extracted summary constituent units are arranged in the order of appearance in the original text and output as a summary result (summary result output means 804).

Next, the effect of the 8th Embodiment of this invention is demonstrated.

In the eighth embodiment of the present invention, the constraint setting means 803 regarding the extraction order of the summary constituent units is considered when extracting the respective summary constituent units, since the relationship recognized between the summary constituent units is taken into consideration. For example, the generation of a summary sentence having a different meaning from the original sentence caused by extracting only the subordinate sentence in a citation relation or a connection relation is suppressed, and an effect of generating a high-quality summary sentence is obtained.

Next, a ninth embodiment of the present invention will be described in detail with reference to the drawings.

FIG. 9 is a block diagram showing a configuration of the ninth exemplary embodiment of the present invention.

Referring to FIG. 9, the ninth embodiment of the present invention includes a summary component unit certifying unit 901, a summary component unit complementing and correcting unit 902, a summary component unit importance calculation unit 903, and an adjacent summary. The expression unit reduction unit 904, the summary result output unit 905, the summary unit recognition pattern 906, and the summary unit complement correction pattern 907 are configured. The summary constituent unit recognition pattern 906 and the summary constituent unit complementary correction pattern 907 are stored in a storage device such as a memory or a hard disk and are read and used, but will not be described one by one.

Each of these means generally operates as follows.

The summary constituent unit recognition means 901 uses the summary constituent unit recognition pattern 906 for each sentence of the input text to recognize the unit that minimizes the clause including the predicate as the summary constituent unit. The summary composition unit complement correction means 902 applies the summary composition unit complement modification pattern 907 to share the original sentence so that the importance level in the text can be appropriately determined only by the information of the summary composition unit. The supplemented information is supplemented to the summary constituent unit, or the summary constituent unit is corrected based on the sentence expression in the original document so that the expression of the summary constituent unit becomes an appropriate expression as an independent sentence in the summary sentence.

The summary component unit importance calculation means 903 is the position of the summary component unit in the text, the clue word, the title, and the importance of the word with respect to the summary component unit that has been appropriately supplemented or modified in expression. The importance of each summary component in the text is calculated using information such as the degree and keywords.

The expression reduction unit 904 in the adjacent summary component unit supposes that the summary component unit adjacent to each other in the summary is adjacent even in the same sentence of the original document. Is used. The summary result output means 905 sequentially extracts from the summary component unit having the highest importance, in order, until the summary length requested by the user is reached, based on the importance for each summary component unit. Arrange in the order of appearance in the original text and output as a summary result.

Next, the operation of the ninth exemplary embodiment of the present invention will be described in detail.

First, by using the summary constituent unit recognition pattern 906 for each sentence of the input text, the unit that minimizes the clause including the predicate is recognized as the summary constituent unit (summary constituent unit recognition means 901), and then By applying the summary composition unit complement correction pattern 907, the information shared in the original sentence is appropriately applied to the summary composition unit so that the importance level in the text can be appropriately judged only by the information of the summary composition unit. Or appropriately correct based on the expression of the sentence in the original document so that the expression of the summary composition unit becomes an appropriate expression as an independent sentence in the summary sentence (summary composition unit complement correction means 902).

Then, with respect to the summary composition unit that has been appropriately supplemented or corrected in terms of information, information such as the position of the summary composition unit in the text, clue words, title, word importance, and keywords are used as clues. , Calculate the importance of each summary component in the text (summary component importance calculation means 903), and if the summary component adjacent in the summary is adjacent in the same sentence of the original document, For the completion and correction between the summary composition units to be used, the expression before complement correction is used, and finally, the user requests in order from the most important summary composition unit based on the importance for each summary composition unit. Sequentially extracted until the summary length is reached (reduction means 904 for expression in adjacent summary constituent units), and the extracted summary constituent units are displayed in the order of appearance in the original text. Arranged to output as a summary result (summarized result output unit 905).

Next, effects of the ninth exemplary embodiment of the present invention will be described.

In the ninth embodiment of the present invention, when the summary reduction unit 904 in the adjacent summary unit is adjacent in the same sentence of the original document by the reduction means 904 in the summary, the adjacent summary unit Since the expression before completion correction is used for inter-completion and correction, the effect of mitigating the loss of continuity between sentences due to completion and correction can be obtained, and the readability of the summary can be improved.

Next, a tenth embodiment of the present invention will be described in detail with reference to the drawings.

FIG. 10 is a block diagram showing the configuration of the tenth embodiment of the present invention.

Referring to FIG. 10, the tenth embodiment of the present invention is a summary sentence candidate generation means 1001, a summary composition unit complementation correction means 1002, a summary composition unit importance calculation means 1003, and a summary sentence candidate. Search means 1004, summary result output means 1005, summary constituent unit recognition pattern 1006, and summary constituent unit complement correction pattern 1007. The summary constituent unit recognition pattern 1006 and the summary constituent unit complementary correction pattern 1007 are stored in a storage device such as a memory or a hard disk, and are read and used.

Each of these means generally operates as follows.

The summary sentence candidate generation means 1001 recognizes a unit that minimizes a clause including all possible predicates as a summary constituent unit by sequentially collating the summary constituent unit recognition pattern 1006 with respect to each sentence of the input text. Then, a summary sentence candidate is obtained. The summary composition unit complement correction means 1002 applies the summary composition unit complement correction pattern 1007 to share the original sentence so that the importance level in the text can be appropriately determined only by the information of the summary composition unit. The supplemented information is supplemented to the summary constituent unit, or the summary constituent unit is corrected based on the sentence expression in the original document so that the expression of the summary constituent unit becomes an appropriate expression as an independent sentence in the summary sentence.

The summary component unit importance calculation means 1003 is the position of the summary component unit in the text, the clue word, the title, and the importance of the word with respect to the summary component unit that has been appropriately supplemented or modified in expression. The importance of each summary component in the text is calculated using information such as the degree and keywords. The summary sentence candidate search unit 1004 searches for summary sentence candidates in each sentence in the text based on the importance of these summary sentence candidates, and determines a combination of maximum likelihood summary sentence candidates as a summary for the input text. In this case, the extraction is performed in consideration of the cost for certifying the summary unit and the cost for complementing and correcting the summary unit. The summary result output unit 1005 arranges the determined combinations of maximum likelihood summary sentence candidates in the order of appearance in the original text, and outputs the summary results.

Next, the operation of the tenth embodiment of the present invention will be described in detail.

First, a unit that minimizes a clause including all possible predicates is recognized as a summary constituent unit by sequentially collating the summary constituent unit recognition pattern 1006 for each sentence of the input text. (Summary sentence candidate generating means 1001). At this time, by applying the summary composition unit complement correction pattern 1007 to the summary composition unit in the summary sentence candidate, the importance in the text can be appropriately determined only by the information of the summary composition unit. The information shared in the original sentence is supplemented to the summary unit, or the expression of the summary unit is corrected based on the sentence representation in the original document so that it becomes an appropriate expression as an independent sentence in the summary sentence. (Summary composition unit supplementary correction means 1002).

Next, information such as position, clue word, title, word importance, keyword, etc. in the text of the summary composition unit with respect to the summary composition unit in which such information has been appropriately supplemented or corrected in expression is provided. The importance of each summary constituent unit in the text is calculated (summary constituent unit importance calculating means 1003), and the summary sentence candidate in each sentence in the text is calculated based on the importance for these summary sentence candidates. When extracting and determining the most likely summary sentence combination as a summary for the input text, it is extracted in consideration of the cost for certifying the summary unit and the cost for supplementing and correcting the summary unit (summary sentence). The candidate searching means 1004) arranges the combinations of the determined maximum likelihood summary sentence candidates in the order of appearance in the original text and outputs them as a summary result (summary result output). It means 1005).

Next, the effect of the 10th Embodiment of this invention is demonstrated.

In the tenth embodiment of the present invention, the summary sentence candidate generating means 1001 and the summary sentence candidate searching means 1004 recognize a unit that minimizes a clause including all possible predicates as a summary constituent unit. Since a plurality of summary sentence candidates are created from now on, it is possible to realize the selection of summary sentences from more flexible summary constituent units. At this time, when searching from the most likely summary sentence candidate, the cost for certifying the summary unit and the cost for supplementing and correcting the summary unit are also considered. The generation of unknown expressions can be suppressed at the same time, and the effect of generating a higher-quality summary based on a flexible summary unit can be obtained.

Next, examples of the first to third embodiments of the present invention will be described in detail with reference to the drawings.

FIG. 11 is an explanatory diagram showing an example of the summary constituent unit certification pattern 204, the summary constituent unit certification pattern 605, the summary constituent unit certification pattern 705, and the summary constituent unit certification pattern 906.

FIG. 14 is an explanatory diagram showing an example of the summary constituent unit recognition pattern 304 and the summary constituent unit recognition pattern 1006.

An example of a specific implementation of the summary component unit certifying language knowledge information 104 in the first embodiment is the summary component unit qualification pattern 204 in the second embodiment, and therefore, here, the second embodiment. Will be described.

First, the summary constituent unit recognition means 201 uses the summary constituent unit certification pattern 204 for authorizing the summary constituent unit for each sentence of the input text in the summary constituent unit certification pattern 204 to include a clause including a predicate. The smallest unit is recognized as the summary unit. At this time, there is a sentence in the text that says, “Traditionally, there has been a tendency to cut off students mechanically, but in the last few years there has been a moment to take care of students.” Assume that the unit recognition pattern 204 is as shown in FIG. Each pattern is composed of a “collation condition” that describes whether or not to apply a pattern to a sentence or summary unit, and a “summary unit” that describes the form of the summary unit obtained by application. Yes.

“S” in the collation condition represents an object to which the pattern is applied, and “P1” and “P2” represent a portion of S that is collated with an expression defined by “P1 =” or “P2 =”. The “use” is matched with a verb, an adjective, or an adjective verb, and “*” is matched with an arbitrary character string having a length of 0 or more. As a result, the pattern 1 functions to collate with a sentence in which two simple sentences are connected by the conjunction “ga”, for example, and to recognize each single sentence as a summary constituent unit. Therefore, the sentence in the text conforms to the matching condition of the pattern 1 in FIG. 11, and as a result, “similar to the summary structural units S1 and S2,“ there has been a tendency to mechanically cut off the students by points. And "The momentum to take care of the students has come out in the last few years."

Next, in the summary component unit importance calculation means 202, the position of the summary component unit in the text, clue word, title, importance of the word, and keyword for all the summary component units including these two units. The importance of each summary composition unit in the text is calculated using information such as the above, and finally, the user requests in order from the most important summary composition unit based on the importance for each summary composition unit. The summary result output unit 203 arranges the extracted summary constituent units in the order of appearance in the original text and outputs the summary results.

In the third embodiment, the configuration is the same as that of the second embodiment, except that a summary configuration unit recognition pattern 304 that can set a cost for certifying each summary configuration unit is used. This will be described in detail in an example of the tenth embodiment.

Next, examples of the fourth to seventh and ninth embodiments of the present invention will be described in detail with reference to the drawings.

FIG. 12 is an explanatory diagram illustrating an example of the summary constituent unit complement correction pattern 606 and the summary constituent unit complement correction pattern 907.

Summary component unit recognition language knowledge information 405 and summary component unit complement correction language knowledge information 406 in the fourth embodiment, or summary component unit recognition language knowledge information 505 in the fifth embodiment, and An example of concrete implementation of the summary constituent unit complement correction language knowledge information 506 is the summary constituent unit recognition pattern 605 and the summary constituent unit supplement correction pattern 606 in the sixth embodiment. The embodiment will be described.

First, the summary constituent unit recognition means 601 uses the summary constituent unit recognition pattern 605 for each sentence of the input text to recognize the unit that minimizes the clause including the predicate as the summary constituent unit. At this time, there is a sentence in the text that “a movie fan has stopped going to the theater and can only use a rental video”, and a summary unit recognition pattern 605 is shown in FIG. Suppose that

Then, the sentence in the text conforms to the matching condition of pattern 2 in FIG. 11, and as a result, “movie fans will not go to the movie theater” corresponding to the summary structural units S1 and S2. “Rental video can be used” is obtained as a summary unit.

Next, the summary composition unit complement correction means 602 applies the summary composition unit complement correction pattern 606 so that the importance level in the text can be appropriately determined only by the summary composition unit information. The information shared in the sentence is supplemented to the summary constituent unit, or the summary constituent unit is corrected based on the sentence expression in the original document so that the expression becomes an appropriate expression as an independent sentence in the summary sentence.

At this time, it is assumed that the summary structural unit complement correction pattern 606 is as shown in FIG. Each pattern has a “matching condition” that describes whether or not the pattern is applied to the sentence or summary unit, and a “completion” that describes the form of the summary unit that is obtained as a result of supplementary correction by application. It is composed of “Fix”. “S” in the collation condition represents a sentence unit represented by a column of summary constituent units to which the pattern is applied. “S1”, “Sk”, and “Sn” are “S1 =”, “ It represents the summary constituent unit in S that matches the formula defined by “Sk =” and “Sn =”. Here, since S1 meets the matching condition of pattern 2 in FIG. 12, complementary correction is applied. Here, since the tense of S2 is in the past, “movie fans have gone to the movie theater” is obtained as a result of the complementary correction of S1. Further, since S2 meets the matching condition of pattern 1, complementary correction is similarly applied.

Here, P is collated with the subject “movie fan” of S1, and as a result of the complementary correction of S2, “movie fans can now use rental videos” is obtained. Further, the importance calculation means 603 of the summary constituent unit performs the position, clue word, title, word in the text of the summary constituent unit with respect to the summary constituent unit in which these information is appropriately complemented or corrected in expression. The importance of each summary component in the text is calculated using information such as importance and keywords, and finally, the most important summary component is ordered in order based on the importance for each summary component. The summary results output means 604 arranges the extracted summary constituent units in the order of appearance in the original text and outputs the summary results until the summary length requested by the user is reached.

Further, in the fifth embodiment, the configuration is almost the same as that of the sixth embodiment, but the summary component unit supplementary correction unit 602 is implemented before the summary unit importance calculation unit 603. Is different. For this reason, the result of complementary correction affects the importance calculation in the fourth and sixth embodiments, but is different in that it does not affect the fifth embodiment.

In the seventh embodiment, the configuration is the same as that of the sixth embodiment, except that the summary unit complement correction pattern 706 that can set the cost for supplementary correction of each summary unit is used. Different. This will be described in detail in an example of the tenth embodiment.

The ninth embodiment has substantially the same configuration as that of the sixth embodiment, except that expression reduction means 904 is added in adjacent summary constituent units. In the sixth embodiment and the ninth embodiment, in contrast to the example sentence “movie fans are no longer going to the movie theaters, they can now use rental videos”. If you get two summarization units, “movie fans are now able to do rental videos,” both of which are extracted during the summarization, the ninth implementation In the case of the above, when the summary unit adjacent in the summary is adjacent in the same sentence of the original document, since the original expression is used for complementation and correction between the adjacent summary units, It ’s different from the fact that it ’s no longer necessary to go to the pavilion and it ’s done with rental videos. ”

Next, an example of the eighth embodiment of the present invention will be described in detail with reference to the drawings.

FIG. 13 is an explanatory diagram illustrating an example of the summary constituent unit recognition pattern 805.

The summary component unit recognition means 801 first recognizes the summary component unit for each sentence of the input text, and further recognizes the connection relationship and the citation relationship between the recognized summary component units. Using the certified pattern 805, a unit that minimizes a clause including a predicate is recognized as a summary constituent unit, and further, when extracting each summary constituent unit, a restriction on the extraction order considering these relationships is set.

For example, there is a sentence in the text that says “If a gene related to lifespan is elucidated, a drug for longevity and longevity will be realized”, and the summary structural unit recognition pattern 805 is as shown in FIG. Suppose there was. The pattern includes “extraction order constraint” in addition to “collation condition” and “summary constituent unit”. Pattern 3 is checked against the example sentence, and as a result, “the genes related to lifespan” corresponding to the summary building blocks S1 and S2 are elucidated. As obtained. The expression “S2> S1” in the extraction order constraint in pattern 3 expresses the constraint that when extracting S1 and S2 obtained as summary constituent units into a summary, S1 is always extracted with S2. Yes.

The summary component importance calculation unit 802 calculates the importance of each summary component in the text, and the constraint setting unit 803 regarding the summary component extraction order is based on the importance for each summary component. When extracting in order from the most important summary unit until the summary length requested by the user is reached, the extraction is performed taking into account the constraints on the extraction order considering the relationship between each summary unit. The summary result output means 804 arranges the extracted summary constituent units in the order of appearance in the original text and outputs the summary results.

Next, an example of the tenth embodiment of the present invention will be described in detail with reference to the drawings.

FIG. 15 is an explanatory diagram illustrating an example of the summary constituent unit complement correction pattern 706 and the summary constituent unit complement correction pattern 1007.

FIG. 16 is an explanatory diagram illustrating an example of summary sentence candidates.

For example, in the text, “It is difficult to obtain an understanding of the Socialist Party regarding the consumption tax rate increase, and the prime minister said,“ I have to make an effort. " "Exists. In the summary sentence candidate generation means 1001, first, all the possible combinations are made by sequentially performing the collation of the summary constituent unit recognition pattern 1006, thereby creating all the summary constituent units and obtaining the summary sentence candidates therefrom. .

Of the summary constituent unit recognition patterns 1006 shown in FIG. 14, the patterns 2 and 5 can be collated first. The summary sentence candidates 1 and 2 shown in FIG. 16 are obtained from the summary structural units S1 and S2 obtained by the pattern 2, respectively. Candidate 1 in the figure indicates that the supplementary correction unit 1002 for the summary unit applies the expression completion correction pattern 2.

Next, a summary sentence candidate obtained by applying pattern 5 is created. Since the summary unit S1 or S2 cannot be extracted independently for the pattern 5 due to the extraction order constraint, these cannot be summary candidates, but the pattern 4 can be further applied to S2. If one of the two obtained by applying pattern 4 to S2 is accompanied by S1 of pattern 5, it satisfies the extraction order constraint and can be a summary sentence candidate.

Candidate 3 shown in FIG. 16 is obtained by combining S1 of pattern 5 and S1 of pattern 4, and candidate 4 is obtained by combining S1 of pattern 5 and S2 of pattern 4. The remaining independent summary sentence candidates are summary sentence candidates obtained by applying pattern 2 to candidate 3 and candidate 4, and from S2 obtained by applying pattern 2 to each candidate, candidates 5 and 6 Get. The sum of the cost of the unit recognition pattern applied to each summary sentence candidate and the total value of the cost of the complementary correction pattern applied to each summary sentence candidate are calculated by the summary component candidate importance level calculation means 1003. Are used for candidate selection in the summary sentence candidate search means 1004.

The summary sentence candidate searching means 1004 selects and combines one of the following: outputting an input sentence for each sentence, outputting nothing, or extracting any summary sentence candidate in the summary Then, a summary candidate suitable for the summary length requested by the user is created. The summary result output means 1005 has a summary score (summary score = total importance value−α × total unit accreditation cost−β × sum of supplementary correction costs) among the summary candidates thus obtained. Is output as a summary result. Here, α and β represent parameters (positive real numbers).

Increasing α suppresses the application of the unit recognition pattern to sentences, so the tendency to use the sentences in the original text as a summary sentence becomes stronger, and increasing β increases the costly complementary correction pattern. Since the appearance of the summary sentence candidate in the summary is suppressed, the tendency that the summary sentence candidate subjected to the correction with low reliability does not appear is increased by giving a large cost to the correction with low reliability. The formula for calculating the summary score shown here is an example, and other formulas may be used.

Next, an eleventh embodiment of the present invention will be described with reference to the drawings.

The eleventh embodiment of the present invention is a method in which each means shown in FIGS.

Next, a twelfth embodiment of the present invention will be described with reference to the drawings.

The twelfth embodiment of the present invention is a program that causes a computer to execute the procedures of the eleventh embodiment of the present invention.

101 Summary constituent unit recognition means 102 Summary constituent unit importance calculation means 103 Summary result output means 104 Summary constituent unit recognition language knowledge information 201 Summary constituent unit recognition means 202 Summary constituent unit importance calculation means 203 Summary result output Means 204 Summarized composition unit recognition pattern 301 Summary composition unit recognition means 302 Summary composition unit importance calculation means 303 Summary result output means 304 Summary composition unit recognition pattern 401 Summary composition unit recognition means 402 Summary composition unit complement correction means 403 Summary constituent unit importance calculation means 404 Summary result output means 405 Summary constituent unit recognition language knowledge information 406 Summary constituent unit complement correction language knowledge information 501 Summary constituent unit recognition means 502 Summary constituent unit importance calculation means 503 Summary Complementary correction means 504 for the composition unit Summary result output means 505 Summary component unit linguistic knowledge information 506 Summary component unit supplemental correction language knowledge information 601 Summary component unit recognition unit 602 Summary unit complement correction unit 603 Summary unit importance calculation unit 604 Summary result output unit 605 Summary Composition unit authorization pattern 606 Summary composition unit complement correction pattern 701 Summary composition unit recognition means 702 Summary composition unit complement correction means 703 Summary composition unit importance calculation means 704 Summary result output means 705 Summary composition unit authorization pattern 706 Summary composition unit Complementary correction pattern 801 Summary composition unit recognition means 802 Summary composition unit importance calculation means 803 Summary composition unit extraction order restriction setting means 804 Summary result output means 805 Summary composition unit authorization pattern 901 Summary composition unit authorization means 902 Summarizing unit supplementary correction means 9 3 Summary component unit importance calculation unit 904 Expression reduction unit 905 Summary result output unit 906 Summary component unit recognition pattern 907 Summary component unit complement correction pattern 1001 Summary sentence candidate generation unit 1002 Summary component unit generation unit 1002 Complement correction means 1003 Summary component unit importance calculation means 1004 Summary sentence candidate search means 1005 Summary result output means 1006 Summary composition unit recognition pattern 1007 Summary composition unit complement correction pattern

Claims (6)

A collation condition that is a pattern that collates with the entire input sentence, and that includes an extraction unit that is a pattern that collates with a clause that includes a statement predicate in the pattern, and includes the extraction unit Restrictions on the extraction order when extracting summary composition units from sentences containing connection relations and citation relations using the summary composition unit patterns and the summary composition unit patterns that correspond to the forms of multiple summary composition units were defined. A summary unit recognition pattern storage unit for storing extraction order constraints;
When a sentence is input and the input sentence and the collation condition are collated and matched, a part (corresponding part) corresponding to the extraction part of the collation condition is extracted from the input sentence, and the form of the summary constituent unit A recognizing unit for summarizing constituent units to obtain a plurality of summarizing constituent units by replacing the extraction unit of
The importance calculation means for the summary constituent unit for calculating the importance of the summary constituent unit obtained by the authorization means;
Search means for combining the plurality of summary constituent units based on the extraction order constraint to create a plurality of summary candidates having a predetermined summary length;
Among the summary candidates created by the search means, summary result output means for outputting the sum of the importance of the summary constituent units as a summary result;
A text summarization system characterized by comprising: The summary constituent unit certification pattern stored in the summary constituent unit certification pattern storage unit further includes an application cost of the summary constituent unit certification pattern,
The importance calculation means further adds a value of reliability corresponding to the application cost of the summary composition unit certification pattern used for the certification of the summary composition unit to the importance value of the summary composition unit. Calculate the unit score,
The summary result output means outputs, as a summary result, a sum of the scores of the summary constituent units that is the maximum among the summary candidates created by the search means.
The text summarization system according to claim 1. A collation condition that is a pattern that collates with the entire input sentence, and that includes an extraction unit that is a pattern that collates with a clause that includes a statement predicate in the pattern, and includes the extraction unit Restrictions on the extraction order when extracting summary composition units from sentences containing connection relations and citation relations using the summary composition unit patterns and the summary composition unit patterns that correspond to the forms of multiple summary composition units were defined. Read out the extraction order constraints from the summary unit recognition pattern storage unit,
When a sentence is input and the input sentence and the collation condition are collated and matched, a part (corresponding part) corresponding to the extraction part of the collation condition is extracted from the input sentence, and the form of the summary constituent unit Substituting the corresponding extraction part with the corresponding part to obtain a plurality of summary constituent units,
Performing importance calculation processing for calculating the importance of the plurality of summary constituent units;
Combining the plurality of summary units based on the extraction order constraint, creating a plurality of candidates for a summary of a predetermined summary length;
A summary result output process is performed to output a summary result having a maximum sum of importance of summary constituent units among the plurality of created summary candidates.
A text summarization method characterized by that. The summary constituent unit certification pattern stored in the summary constituent unit certification pattern storage unit further includes an application cost of the summary constituent unit certification pattern,
In the importance calculation process, a summary configuration obtained by adding a reliability value corresponding to the application cost of the summary configuration unit recognition pattern used for the certification of the summary configuration unit to the importance value of the summary configuration unit Calculate the unit score,
In the summary result output process, among the created summary candidates, the sum of the scores of the summary constituent units is output as a summary result.
The text summarization method according to claim 3. A collation condition that is a pattern that collates with the entire input sentence, and that includes an extraction unit that is a pattern that collates with a clause that includes a statement predicate in the pattern, and includes the extraction unit Restrictions on the extraction order when extracting summary composition units from sentences containing connection relations and citation relations using the summary composition unit patterns and the summary composition unit patterns that correspond to the forms of multiple summary composition units were defined. Read out the extraction order constraints from the summary unit recognition pattern storage unit,
When a sentence is input and the input sentence and the collation condition are collated and matched, a part (corresponding part) corresponding to the extraction part of the collation condition is extracted from the input sentence, and the form of the summary constituent unit And replacing the extraction unit with the corresponding part to obtain a plurality of summary constituent units,
A summary component unit importance calculation step for calculating a summary component unit importance obtained by the authorization step;
A search step of combining a plurality of summary units based on the extraction order constraint to create a plurality of summary candidates having a predetermined summary length;
A summary result output step for outputting, as a summary result, a summary sum of importance of summarization units among the summary candidates created by the search step;
A text summarization program for causing a computer to execute. The summary constituent unit certification pattern stored in the summary constituent unit certification pattern storage unit further includes an application cost of the summary constituent unit certification pattern,
The importance calculation step further includes a summary configuration in which a reliability value corresponding to the application cost of the summary configuration unit certification pattern used for the certification of the summary configuration unit is added to the importance value of the summary configuration unit. Calculate the unit score,
The summary result output step outputs, as a summary result, the sum of the scores of the summary constituent units among the summary candidates created by the search step,
The text summarization program according to claim 5, wherein:

Images