JP2011085986A - Text summarization method, its device, and program - Google Patents

Abstract

A summary with less redundancy is generated when a summary is generated by selecting an important sentence from the summary target text.
A summary input text is received by a text input unit, and evaluation information including at least three elements of an evaluation target, an evaluation attribute, and an evaluation expression is extracted from the summary target text by an evaluation information extraction unit, and evaluation information is aggregated. The evaluation information extracted by the unit 4 is aggregated using the evaluation attribute similarity dictionary 3 storing the similarity between the evaluation attributes in the evaluation information, and the important sentence selection unit 5 collects the evaluation information based on the aggregated evaluation information. Then, an important sentence is selected from the summary target text, and the selected important sentence is rearranged according to a predetermined rule by the text output unit 6 and output.
[Selection] Figure 1

Description

  The present invention relates to a technique for summarizing digitized text (document).

  In recent years, a large amount of electronic text has been distributed. Therefore, the importance of so-called text summarization techniques that cause a machine to summarize texts for the purpose of quickly grasping information described in those texts is increasing.

  Currently, when letting a machine summarize text, select one or more sentences (hereinafter, important sentences) suitable for the summary from the text to be summarized (hereinafter, the text to be summarized), and rearrange them and connect them. Often done by. When selecting an important sentence, a sentence score is often defined by some method, and the sentence is often selected according to the score. As an example of defining the score of a sentence, information that is included in a sentence constituting the text is identified by extracting information in advance from the text to be summarized, and summarizing is performed according to the score of the information. Yes (Non-Patent Document 1).

  As an example of the information to be extracted, there is evaluation information that is information such as evaluation and opinion regarding a certain target. The evaluation information includes at least three elements: an evaluation object, an evaluation attribute, and an evaluation expression. For example, consider extracting an evaluation from the sentence “Gyoza in restaurant A is delicious”. At this time, the evaluation object represents an object to be evaluated, such as a store or a product, and thus becomes “A restaurant”. Since the evaluation attribute represents a specific evaluation item in the evaluation object such as the specification (properties, characteristics, etc.) of the evaluation object, a part thereof, and a viewpoint when performing the evaluation, the evaluation attribute is “gyoza”. The evaluation expression is “delicious” because it represents the evaluation itself of the evaluation attribute to be evaluated by the person who performs the evaluation (text writer).

  The evaluation information can be extracted by the method described in Patent Document 1, for example. Hereinafter, extracting the evaluation information from the text is referred to as evaluation information extraction, and the three elements of the evaluation information are separated by “,” like <evaluation object, evaluation attribute, evaluation expression>, and “<” and “>” It shall be shown in the form enclosed by].

  Another example of information to be extracted is two specific expressions and a relationship between them. For example, the sentence “Prime Minister Aso addressed in Akihabara” is a combination of three elements: the unique expressions “Prime Minister Aso” and “Akihabara” and the relationship “speech” that exists between these two specific expressions. Can be extracted. Hereinafter, such a combination of three elements is referred to as relation information.

  The relationship information can be extracted by the method described in Patent Document 2, for example. In addition, extracting the relationship information from the text is called relationship information extraction, and the three elements of the relationship information are separated by “,” as [specific expression 1, relationship, specific expression 2], and “[” It shall be shown in the form enclosed by "]".

JP 2008-140359 A (“Evaluation Information Extraction Device, Evaluation Information Extraction Method and Program”) JP 2008-225566 A ("Relational Information Extraction Apparatus and Method")

Elena Filatova et al., "A formal model for information selection in multi-sentence text extraction.", In Proceedings of the 20th international conference on Computational Linguistics (COLING), 2004.

  However, in the methods of Patent Document 1 and Patent Document 2, since it is impossible to determine whether or not the extracted evaluation information or relationship information is the same (similar), Patent Document 1 or Patent Document is simply used. When text summarization is performed by the method as described in Non-Patent Document 1 using the information extracted by the method 2, there is a possibility that a redundant summary may be generated.

  For example, consider that evaluation information is extracted from a text describing the reputation of some product or service, and the text is summarized using the result. More specifically, consider generating a single summary from multiple texts about a restaurant's reputation described by different writers.

  As an example, suppose that there is a sentence “Restaurant A has the best taste” in a text by one writer, and a text “Restaurant A has the best seasoning” in a text by another writer. When the evaluation information is extracted from these two sentences, the evaluation information <Restaurant A, Taste, Best> is extracted from the previous sentence, and the evaluation information <Restaurant A, Seasoning, Best> is extracted from the latter sentence. The

  These pieces of evaluation information indicate that the taste (or seasoning) of the restaurant A is excellent, and have almost the same meaning. However, in order to recognize that the two pieces of evaluation information are substantially the same, it is necessary to recognize that “taste” and “seasoned” have substantially the same meaning. In other words, if the machine that handles these evaluation information cannot certify that “taste” and “seasoning” have almost the same meaning, it certifies that the above two evaluation information are almost the same. Can not.

  This is the same even when the relationship information is used. For example, when one text contains the sentence "Prime Minister Aso addressed in Akihabara" and the other text says "Prime Minister Aso made a speech in Akihabara", each of them [Prime Minister Aso, Speech, Akihabara], [Aso Even if related information such as “Prime Minister, Speech, Akihabara” is extracted, if the machine cannot recognize that “speech” and “speech” have almost the same meaning, the two related information will be almost the same. Cannot be certified.

  In turn, for summary purposes, consider extracting important sentences based on evaluation information. If you want to include as much evaluation information contained in the summary text as possible in the summary so that the reader of the summary can efficiently grasp the information, if there are multiple sentences that contain almost the same evaluation information, It is sufficient to include only one of these sentences in the summary. However, if the evaluation information that is considered to be almost the same is not treated as the same because the machine cannot recognize the identity of the evaluation attributes, two or more sentences containing the same evaluation information are summarized at the same time. May be included. Such summaries include redundancy in that the same information appears repeatedly, and are not desirable summaries for readers in terms of efficiently transmitting information.

  Therefore, an object of the present invention is to generate a summary with less redundancy by collecting (summarizing) similar evaluation information or relationship information when a summary target text is given.

  Hereinafter, the case where the evaluation information is extracted and summarized will be described as an example. However, if the evaluation attribute is read as the relationship, the present invention can be applied to the case where the relationship information is extracted and summarized.

  In order to achieve the above object, according to the present invention, text that generates at least one important sentence that is a sentence suitable for the summary from the summary target text that is the text to be summarized and generates a summary corresponding to the summary target text. A text input unit that receives summary text, and an evaluation information extraction unit that extracts evaluation information including at least three elements of an evaluation target, an evaluation attribute, and an evaluation expression from the summary target text received by the text input unit. And evaluation information aggregation that uses the evaluation attribute similarity dictionary to collect evaluation attribute similarity dictionary that stores the similarity between evaluation attributes in the evaluation information and evaluation information extracted from the text to be summarized by the evaluation information extraction unit And the summary information received by the text input unit based on the evaluation information aggregated by the evaluation information aggregation unit. A key sentence selection unit for selecting a sentence, characterized in that the key sentences selected by key sentence selection unit and a text outputting unit which outputs rearranged according to a predetermined rule.

  According to the present invention, instead of simply using the evaluation information extracted from the text to be summarized and selecting a sentence suitable for the summary, the similarity between evaluation attributes stored in the evaluation attribute similarity dictionary is used. It is possible to generate a summary with less redundancy by collecting evaluation information extracted from summary target texts and then selecting important sentences using them.

The block diagram which shows an example of embodiment of the text summarization apparatus of this invention Explanatory drawing showing an example of the input format of the text to be summarized Explanatory drawing which shows an example of evaluation information extraction processing Explanatory drawing which shows an example of an evaluation attribute similarity dictionary Explanatory drawing showing an example of summary target text Explanatory drawing which shows an example of the evaluation information list before aggregation Explanatory drawing which shows an example of the evaluation information list after aggregation Explanatory drawing which shows an example of the weight of evaluation information when not performing aggregation Explanatory drawing which shows an example of the weight of evaluation information at the time of performing aggregation Explanatory drawing which shows an example of important sentence selection processing Explanatory drawing which shows an example of the important sentence selection process at the time of not collecting evaluation information Flow chart of processing in text summarization apparatus of the present invention Flow chart showing an example of detailed processing of evaluation information aggregation

  Next, embodiments of the present invention will be described with reference to the drawings.

  FIG. 1 shows an example of an embodiment of a text summarization device according to the present invention. The text summarization device of the present embodiment includes a text input unit 1, an evaluation information extraction unit 2, an evaluation attribute similarity dictionary 3, The evaluation information aggregation unit 4, the important sentence selection unit 5, and the text output unit 6.

<Text input part>
The text input unit 1 receives a summary target text that is directly input from a keyboard (not shown) or the like, read from a storage unit, or input from another device or the like via a communication medium (a storage device (not shown)). Etc.) FIG. 2 shows an example of the input format of the text to be summarized. Here, an example of text subjected to known dependency analysis (including well-known morphological analysis and clause analysis as a premise of dependency analysis) is shown. Yes.

  In FIG. 2, a line starting with “* (asterisk)” represents a section break and indicates information on the section. The section information includes a section number, a relation destination (the section number), a relation type, and a content part. It consists of the number of the word that becomes the main part and the number of the word that becomes the main part of the functional part. The lines that do not start with “*” indicate the information of the words that make up the phrase indicated by the line that starts with “*” immediately before, and the information of each word is written, read, and standard form (only if there is one) , Consists of parts of speech. Note that the line <EOS> indicates that the sentence ends there.

<Evaluation information extraction unit>
The evaluation information extraction unit 2 receives the summary target text received and stored in the text input unit 1, extracts the evaluation information from the summary target text, and displays a list in which all the extracted evaluation information is described. Output to 4. For extracting the evaluation information, a known technique, for example, a technique described in Patent Document 1 can be used.

  An example of processing for extracting evaluation information from the summary target text is shown in FIG.

  First, an evaluation expression dictionary storing the above-described list of evaluation expressions is prepared. In addition to the list of evaluation expressions, the evaluation expression dictionary stores information indicating whether an evaluation expression is positive, negative, or neutral. good. Such an index of positiveness or negativeness of an evaluation expression is called evaluation polarity. However, as described above, the evaluation polarity is not limited to the three types (values) of positive, negative, and neutral, and the positive maximum value is 1 and the negative maximum value is −1. You may make it represent with a numerical value.

  For example, in the evaluation expression dictionary shown in FIG. 3, since the evaluation expression “delicious” is a positive expression, its evaluation polarity is P (positive), but the evaluation expression “bad” is a negative expression. The evaluation polarity is N (negative). Further, since the evaluation expression “normal” is a neutral expression that is neither positive nor negative, its evaluation polarity is PN (meaning between P and N).

  Next, the evaluation expression included in the evaluation expression dictionary is searched from the result of dependency analysis. FIG. 3 shows the result of the dependency analysis of the sentence “Restaurant A is the most delicious!”, Which is included in the last sentence “Delicious!”. “Delicious” corresponds to the evaluation expression. When a phrase having an evaluation expression is identified from the result of dependency analysis, a phrase that is related to the phrase including the evaluation expression by “ga rating” is searched. In this case, the phrase “Musaka is” corresponds to this. Then, “Musaka” obtained by removing the case particle “ga” from this phrase is set as an evaluation attribute corresponding to the evaluation expression “delicious”. In addition, the phrase including “evaluation” is searched for the phrase including the evaluation expression. Here, the phrase “Restaurant A” corresponds to this. Then, “restaurant A” obtained by removing the case particle “ha” from this phrase is set as an evaluation target corresponding to the evaluation expression “delicious”. As a result, the evaluation information <Restaurant A, Musaka, Delicious> is extracted from the input text.

  However, when extracting the evaluation information, the evaluation expression can be simplified to the evaluation polarity (P, N, or PN) by using the evaluation polarity information included in the evaluation expression dictionary. In that case, in the example shown in FIG. 3, the evaluation information <Restaurant A, Musaka, P> is extracted.

<Evaluation attribute similarity dictionary>
The evaluation attribute similarity dictionary 3 stores the similarity between evaluation attributes in the evaluation information. The similarity between evaluation attributes indicates the degree to which evaluation attributes are semantically similar as words. An example of the evaluation attribute similarity dictionary 3 is shown in FIG.

  In FIG. 4, the similarity between each evaluation attribute is defined. For example, the similarity between the evaluation attribute “taste” and the evaluation attribute “seasoning” is 0.38. Since “taste” and “seasoned” have similar meanings as words, their similarity is high. On the other hand, the similarity between “ice cream” and “sashimi” is 0 in FIG. This reflects the fact that “ice cream” and “sashimi” are not semantically similar.

  As an example, the evaluation attribute similarity dictionary 3 uses a co-occurrence word vector having a word that co-occurs with an evaluation attribute as a vector element in advance as a corpus, for example, a domain of a text to be summarized (such as a word of mouth of a restaurant or a text of a newspaper article). It is possible to store the cosine similarity calculated using the corpus corresponding to the type).

In order to calculate the cosine similarity between the evaluation attribute a ₁ and the evaluation attribute a ₂ , for example, the following may be performed. First, using some corpus, a co-occurrence word vector whose elements are words that co-occur with the evaluation attribute a ₁ is set to V = (v ₁ , v ₂ ,..., V _n ), and is also shared with the evaluation attribute a ₂ . Let W = (w ₁ , w ₂ ,..., W _n ) be a co-occurrence word vector whose elements are words that occur. The element v _{i of the} vector V and the element w _i (i = 1, 2,... N) of the vector V are the number of times that the evaluation attribute and some word co-occur. In such a case, the cosine similarity cos (V, W) between the co-occurrence word vector V of the evaluation attribute a _{1 and} the co-occurrence word vector W of the evaluation attribute a ₂ is defined by the following equation (1).

  Other similarities between evaluation attributes include Thomas Hofmann, "Probabilistic latent semantic indexing", In Proceeding of the 22nd annual international ACM SIGIR conference on Research and development in information retrieval, pp.50-57, 1999 (references) Using the technique described in 1), the probability that the evaluation attribute belongs to some potential class can be calculated and stored. Instead of the document and word co-occurrence matrix required for the calculation, as an example, the calculation can be performed using the evaluation expression extracted from some corpus and the co-occurrence matrix of the evaluation attributes. In that case, the evaluation attribute similarity dictionary 3 stores each evaluation attribute and the probability that it belongs to each of a plurality of different potential classes.

  In addition to the above cosine similarity and the probability of belonging to some potential class, Satoru Ikehara, et al. “Japanese Vocabulary Series CD-ROM Version”, Iwanami Shoten, 1999 (reference 2) Such a structured synonym may be stored. When a dictionary in which synonyms are structured as in Reference 2 is used, as an example, if a certain evaluation attribute and another evaluation attribute belong to the same category, the similarity is 1, and if not, the similarity is Can be stored as 0 in the evaluation attribute similarity dictionary.

<Evaluation Information Aggregation Department>
The evaluation information aggregating unit 4 collects the evaluation information extracted from the summary target text by inputting the evaluation information list extracted from the summary target text by the evaluation information extraction unit 2 and the evaluation attribute similarity dictionary 3, and The aggregated list is output to the important sentence selection unit 5. Here, a case where cosine similarity between evaluation attributes is stored in the above-described evaluation attribute similarity dictionary 3 will be described.

  FIG. 5 shows an example of a summary target text, here a summary target text composed of two texts A and B described by different writers.

  When the evaluation information is extracted from the texts A and B shown in FIG. 5 by the method described in the evaluation information extraction unit 2, the evaluation information list (before aggregation) shown in FIG. 6 is obtained. From the first sentence (sentence number A1) of text A shown in FIG. 5, <restaurant A, ice cream, delicious>, and from the second sentence (sentence number A2), <restaurant A, taste, best> Information is extracted, and the first sentence (sentence number B1) of text B is <restaurant A, seasoned, best> and the second sentence (sentence number B2) is <restaurant A, sashimi, delicious> Suppose that information is extracted.

  The evaluation information list is extracted together with the evaluation information, the sentence number of the sentence from which the evaluation information is extracted (extraction source sentence number), and the appearance frequency of the evaluation information in the summary target text (extracted directly from the summary target text ( (Appearance frequency) before aggregation is also described.

  Looking at the evaluation attribute list shown in FIG. 6, <Restaurant A, Taste, Best> and <Restaurant A, Taste, Best> are treated as different evaluation information, but these are almost the same evaluation information. Therefore, we want to treat these as the same evaluation information rather than as separate evaluation information. This is because if these are treated as different evaluation information, a redundant summary may be generated. For this reason, the evaluation attribute similarity dictionary 3 is used to determine the identity of the evaluation attributes “taste” and “seasoning”.

  Of any two pieces of evaluation information extracted from the text to be summarized, each has evaluation attributes included in the evaluation attribute similarity dictionary 3, and the similarity between these evaluation attributes is based on a preset value (threshold). Those that are high and have the same evaluation object and evaluation expression (or evaluation polarity) are regarded as equivalent and aggregated.

  For example, when the above threshold value is 0.3, the similarity between the evaluation attributes “taste” and “seasoning” is 0.38 from FIG. 4, which is higher than the threshold value, and the evaluation target and evaluation expression are the same. , Taste, best> and <restaurant A, seasoned, best>.

  FIG. 7 shows the evaluation information list after aggregation. At the time of aggregation, it is considered that any piece of evaluation information regarded as the same has appeared for the total number of appearance frequencies of each evaluation information to be aggregated. In the example, as shown in FIG. 6, <Restaurant A, Taste, Best> and <Restaurant A, Seasoned, Best> each appear once in the summary target text. , Taste, best> appears to have appeared twice.

  Note that when different evaluation information is regarded as the same and is aggregated, there is a problem of selecting evaluation information at the aggregation destination. In the above-described example, the problem is whether the evaluation attributes “taste” and “seasoning” of each evaluation information are collected into “taste” or “seasoning”.

  For this problem, you may choose an evaluation attribute with a higher appearance frequency in the text to be summarized, an evaluation attribute with a shorter (or longer) word length, a training text, or other The appearance frequency of each evaluation attribute may be obtained using text, and the evaluation attribute of the aggregation destination may be selected based on the appearance frequency (the more frequent one). In the example given here, since “taste” has fewer characters than “seasoning”, we decided to consolidate it into “taste”.

  In this example, four pieces of evaluation information are extracted from the text to be summarized, and two of them are determined to be the same, and as a result, they are aggregated into three pieces of evaluation information. When information is extracted, a lot of evaluation information is collected. Therefore, the evaluation information included in the evaluation information list is repeatedly aggregated one by one, and the process is stopped when there is no evaluation information to be finally aggregated.

  As a result, as shown in the evaluation information list of FIG. 6, when the evaluation information extracted from the summary target text and the appearance frequency thereof are given, they are aggregated to be shown in the evaluation information list of FIG. Evaluation information and its appearance frequency are obtained.

  The processing of the evaluation information aggregation unit 4 will be summarized. The evaluation information aggregating unit 4 receives the evaluation information list extracted from the summary target text and the evaluation attribute similarity dictionary 3 as input, and selects the evaluation target and the evaluation target from among any two pieces of evaluation information included in the evaluation information list. The evaluation information (or evaluation polarity) is the same, the evaluation attributes are stored in the evaluation attribute similarity dictionary 3, and the similarity between these evaluation attributes is higher than a preset threshold value. It is considered and collected into one piece of evaluation information. The appearance frequency of evaluation information after aggregation is the sum of the appearance frequencies of two evaluation information before aggregation. For example, the evaluation attribute of the evaluation information after aggregation is short among the evaluation attributes before aggregation. The evaluation information is aggregated by repeating this until the evaluation information to be aggregated disappears from the input list of evaluation information.

<Important sentence selection part>
The important sentence selection unit 5 receives as input the list of evaluation information aggregated by the evaluation information aggregation unit 4 and the summary target text received and stored by the text input unit 1, and the important sentence suitable as a summary from the summary target text. Is output to the text output unit 6.

As an example, key sentence selection unit 5, among the combinations S of any sentence s being included in the summary target text, score obtains the maximum combination S ^* using the evaluation information, the statements contained in the combination S ^* It can be formulated as an important sentence. However, the sum of the lengths (number of bytes) Length (s) of the sentence s constituting the summary must not exceed the upper limit K of the summary size (the summary size is not defined by the number of bytes, It may be based on the number of words or sentences).

  As an example, the function Score of the expression (2) can be the sum of the scores of the evaluation information a covered by the sentence combination S, and can be formulated as the following expression (3).

  Here, A is a list of evaluation information extracted and summarized from the text to be summarized. The function Include is a function as shown in the following formula (4).

  The function Include is a function that returns 1 if the sentence combination S includes some evaluation information a, and returns 0 if it does not include the evaluation information a. That is, even if S includes two or more pieces of evaluation information a, since the value returned by the function Include is only 1, the combination S of sentences having a high score in such a situation includes evaluation information having a high score. It will be covered without duplication (because additional evaluation information already included will not add to the score).

w _a is a weight of the evaluation information a, and can be weighted by a function shown in the following formula (5) as an example.

  Here, tf (a) corresponds to the number of times the evaluation information a appears in the text to be summarized, that is, the appearance frequency in FIGS. 6 and 7, and df (a) is the evaluation information in the text included in some text set. The number of texts including a. When the important sentence selection unit 5 receives the evaluation information list aggregated by the evaluation information aggregation unit 4, if some evaluation information is a result of aggregation of a plurality of evaluation information, The sum of df of evaluation information is the sum of df of evaluation information after aggregation.

  According to this equation, evaluation information that frequently appears in the text to be summarized and frequently appears in any text set is preferentially included in the summary. For example, when using evaluation information to summarize a restaurant's reputation as shown in the problem to be solved by the invention, a large text set separate from the text to be summarized, for example, a large scale The above-described df (a) can be obtained by extracting the evaluation information a from the blog text using the method shown by the evaluation information extraction unit 2. In other words, this is equivalent to selecting the evaluation information to be included in the summary by simultaneously considering the evaluation information for the evaluation target and the evaluation information for the evaluation target by many writers, which is the mainstream in the text to be summarized. To do. Of course, other weighting methods can be used.

  FIG. 8 shows an example of the weight of evaluation information when aggregation is not performed. The appearance frequency of the evaluation information list shown in FIG. 6 corresponds to tf in FIG. Let df be calculated from some text set in advance. When equation (5) is calculated using tf and df shown in FIG. 8, the weight is calculated.

  FIG. 9 shows an example of the weight of evaluation information when aggregation is performed. FIG. 9 shows the weight of evaluation information calculated based on the aggregated evaluation information list of FIG. 7, and <restaurant A, taste, best> and <restaurant A, seasoning, best> in FIG. 8 are treated as the same. Therefore, <restaurant A, taste, best> is considered to appear twice in the text to be summarized, and df is the sum of the two evaluation information df before aggregation. Therefore, the weight is also higher than before aggregation.

  The problem of selecting a sentence combination that maximizes the score while satisfying the size limit of the summary as shown in Expression (2) may require a large amount of calculation when obtaining an optimal solution. Therefore, here, as an example, a case where a combination of sentences is selected using a greedy method will be described. The greedy method is a method of selecting a sentence having the highest score when added from sentences that have not been selected yet. Of course, in addition to the greedy method, for example, Wen-tau Yih et al., "Multi-Document Summarization by Maximizing Informative Content-Words", In Proceedings of the 20th International Joint Conference on Artificial Intelligence (IJCAI), pp.1776- A search as shown in 1782, 2007 (reference document 3) can be performed to determine how to select sentences with higher scores.

  FIG. 10 shows an example of important sentence selection processing. Assume that the texts A and B constituting the text to be summarized are each composed of two sentences. Here, it is considered that two important sentences are selected from four sentences including the texts A and B. From each of the four sentences shown in FIG. 10, that is, from the sentence A1, <restaurant A, ice cream, delicious>, from the sentence A2, <restaurant A, taste, best>, and from the sentence B1, <restaurant A, seasoned, Assume that evaluation information <restaurant A, sashimi, delicious> has been extracted from sentence B2. Further, it is assumed that the extracted evaluation information is aggregated by the evaluation information aggregating unit 4 and the result weighted by the important sentence selecting unit 5 is the evaluation information list and weight shown in FIG.

  If the sentence including the evaluation information with the highest weight as described above is selected as the important sentence according to the evaluation information and its weight shown in FIG. 10, the sentence A2 including the evaluation information <restaurant A, taste, best> is selected first. Will be. Since it is assumed here that the number of sentences to be selected is two, the sentence to be selected is sentence B2 including the next highest evaluation information <restaurant A, sashimi, delicious>. These two sentences are important sentences. Selected as.

  FIG. 11 shows an example of important sentence selection processing when evaluation information is not collected. When the evaluation information is not collected, the sentence A2 is first selected as an important sentence from the evaluation information list and the weight shown in FIG. Another sentence selected is sentence B1 including evaluation information <restaurant A, seasoning, best> having the next highest weight. As a result, sentence A2 and sentence B1 are selected as important sentences. However, as shown in FIG. 11, they describe almost the same thing and are more verbose than the summary generated when the evaluation information is aggregated. It has become.

<Text output part>
The text output unit 6 rearranges the important sentences selected by the important sentence selection unit 5 in accordance with a predetermined rule and outputs them as a single text. Sorting may be based on, for example, sentence numbers, based on the time the summarized text was written or posted, or as a known example, Mirella Lapata, "Probabilistic Text Structuring: Experiments A method of rearranging sentences described in “With Sentence Ordering”, In Proceedings of the 41st Meeting of the Association for Computational Linguistics, pp.545-552, 2003 (reference 4) may be used. If there is only one important sentence selected by the important sentence selector 5, the one sentence is output as a single text as it is.

<Text Summarization Method of the Present Invention>
FIG. 12 shows the flow of processing in the text summarization apparatus of the present invention described above.

  The text input unit 1 receives and stores the summary target text (s1), and the evaluation information extraction unit 2 extracts the evaluation information from the summary target text (s2). The evaluation information aggregating unit 4 aggregates the evaluation information extracted from the summary target text using the evaluation attribute similarity dictionary 3 (s3), and the important sentence selection unit 5 based on the aggregated evaluation information from the summary target text. An important sentence is selected (s4). The text output unit 6 rearranges and outputs the selected important sentences according to a predetermined rule (s5).

<Details of evaluation information aggregation processing>
FIG. 13 shows an example of the detailed processing flow of the evaluation information aggregation described above.

  That is, the evaluation information aggregating unit 4 extracts a combination of two pieces of evaluation information from the list describing the evaluation information extracted from the summary target text by the evaluation information extraction unit 2 (s11), and evaluates between the evaluation attributes of the two pieces of evaluation information. Are read from the evaluation attribute similarity dictionary 3 (s12). At this time, if there is no similarity between the corresponding evaluation attributes (s13: No), the process returns to s11 to extract another combination of two pieces of evaluation information. If there is a similarity between the corresponding evaluation attributes (s13: Yes), it is checked whether it is higher than a preset threshold value. If it is lower (s14: No), the process returns to s11 as described above, and another 2 A combination of two pieces of evaluation information is extracted.

  On the other hand, when the degree of similarity between the corresponding evaluation attributes is higher than a preset threshold value (s14: Yes), the evaluation objects and the evaluation expressions of the two evaluation information are further compared and collated (s15). As a result of the comparison and collation, if one or both of the evaluation objects or the evaluation expressions do not match (s16: No), the process returns to s11 as described above, and another combination of two pieces of evaluation information is taken out.

  Further, if both the evaluation objects or the evaluation expressions are the same as a result of the comparison and collation (s16: Yes), the two evaluation information are regarded as equivalent and are aggregated into one evaluation information. The list is updated with information (s17).

  The above processing is repeated for all combinations of evaluation information in the list including evaluation information after aggregation (s18).

  Note that the order of the processes of s12 to s14 and the processes of s15 to s16 may be interchanged.

<Supplementary items>
The description so far has described the case where each text generates a summary from a plurality of texts including at least one sentence, that is, the case where the text to be summarized is composed of a plurality of texts each including at least one sentence. Even if the summary is generated from one text including at least two sentences, that is, even if the text to be summarized consists of one text including at least two sentences, The present invention is applicable (unless it is denied that there is a possibility that these are extracted in duplicate).

  The present invention also installs a program for realizing the functions shown in the configuration diagram of FIG. 1 or a program having the procedure shown in the flowchart of FIG. 12 via a medium or communication line in a known computer. Is also feasible.

  1: text input unit, 2: evaluation information extraction unit, 3: evaluation attribute similarity dictionary, 4: evaluation information aggregation unit, 5: important sentence selection unit, 6: text output unit.

Claims (7)

A text summarization apparatus that selects at least one important sentence that is a sentence suitable for a summary from a summary target text that is a text to be summarized and generates a summary corresponding to the summary target text,
A text input part that accepts the text to be summarized;
An evaluation information extraction unit that extracts evaluation information including at least three elements of an evaluation target, an evaluation attribute, and an evaluation expression from the summary target text received by the text input unit;
Evaluation attribute similarity dictionary storing similarity between evaluation attributes in evaluation information;
An evaluation information aggregating unit for aggregating evaluation information extracted from the summary target text by the evaluation information extraction unit using an evaluation attribute similarity dictionary;
Based on the evaluation information aggregated by the evaluation information aggregation unit, an important sentence selection unit that selects an important sentence from the summary target text received by the text input unit,
A text summarization apparatus comprising: a text output unit that outputs the important sentences selected by the important sentence selection unit according to a predetermined rule. A combination of two pieces of evaluation information is taken out from the list describing the evaluation information, and the similarity between the two evaluation attributes of the two pieces of evaluation information is read from the evaluation attribute similarity dictionary, which is higher than a preset threshold value and the 2 If both evaluation objects of evaluation information and evaluation expressions are compared and collated, if both are the same, the two evaluation information are regarded as equivalent and aggregated into one evaluation information. The text summarization apparatus according to claim 1, further comprising an evaluation information aggregating unit that repeatedly performs all combinations of evaluation information in the list. Among the combinations S of arbitrary sentences s included in the text to be summarized, a combination S ^* having the maximum score using the evaluation information is obtained, and an important sentence selection unit that uses the sentence included in the combination S ^* as an important sentence is provided. The text summarization apparatus according to claim 1, wherein the text summarization apparatus is a text summarization apparatus. A text summarization method for selecting at least one important sentence which is a sentence suitable for summarization from text to be summarized, which is a text to be summarized, and generating a summary corresponding to the text to be summarized,
A step in which the text input unit accepts the text to be summarized;
An evaluation information extraction unit that extracts evaluation information including at least three elements of an evaluation object, an evaluation attribute, and an evaluation expression from the accepted summary object text;
An evaluation information aggregating unit that aggregates the evaluation information extracted from the summary target text using an evaluation attribute similarity dictionary storing similarity between evaluation attributes in the evaluation information;
An important sentence selecting unit selecting an important sentence from the summary target text based on the aggregated evaluation information;
A text summarizing method, comprising: a text output unit rearranging and outputting the selected important sentences according to a predetermined rule. Evaluation information aggregation step
Extracting a combination of two pieces of evaluation information from a list describing the evaluation information;
Reading the similarity between the evaluation attributes of the two evaluation information from the evaluation attribute similarity dictionary;
If this is higher than a preset threshold value, the step of comparing and collating the evaluation objects and evaluation expressions of the two evaluation information,
If both are the same, the two pieces of evaluation information are regarded as equivalent and aggregated into one piece of evaluation information;
The text summarizing method according to claim 4, further comprising a step of repeatedly performing this for all combinations of the evaluation information in the list including the evaluation information after aggregation. The important sentence selection step is
Of the combinations S of arbitrary sentences s included in the text to be summarized, obtaining a combination S ^* having the maximum score using evaluation information;
The text summarizing method according to claim 4, further comprising a step of setting a sentence included in the combination S ^* as an important sentence.   The program for functioning a computer as each means of the text summarization apparatus in any one of Claims 1 thru | or 3.

Images