JP2013250653A - Content summarization support device, method and program - Google Patents

Abstract

PROBLEM TO BE SOLVED: To support the generation of the summary of content.SOLUTION: A partial summary input part 1 accepts a partial summary input by an editor, and a content search part 2 searches content candidates included in the summary, and a content feature extraction part 3 extracts a feature vector relating to similarity between each of the searched content candidates and the partial summary. A ranking part 4 performs the ranking of the content candidates in an order which is beneficial for the reinforcement of the partial summary on the basis of a preliminarily learnt ranking function and the feature vector of each of the extracted content candidates.

Description

  The present invention relates to a content summary support apparatus, method, and program, and more particularly, to a content summary support apparatus, method, and program for supporting generation of a summary composed of a plurality of contents.

  While the era of social media where anyone can speak on the Internet has arrived, a large amount of information has come out, making it difficult to view and use all of it, and there is a mechanism to extract and present only useful information There is a strong demand. Among them, in recent years, the presence of editors (curators) who organize and summarize a large amount of content created by content creators and present them to consumers has attracted attention on social media. . For example, “Summary sites” such as Togetter (registered trademark) and NAVER summary (registered trademark) are widely used as a site for editors to collect information on Twitter (registered trademark), one of the representative social media. Yes. On these sites, editors collect existing content, organize it from the editor's own perspective, and deliver the resulting summary to the consumer again. As a consumer, reading content summarized through an editor can collect information more efficiently and make it easier to understand than reading content directly from the author. FIG. 1 is a diagram schematically showing this mechanism.

  In creating these summaries, the editor needs to select only the necessary content from a large amount of social media content. However, in the current “summary site”, it is necessary to perform this selection work almost manually, which puts a great deal of effort on the editor.

  Some methods have been devised so far for extracting and presenting only useful information from social media. Non-Patent Documents 1 and 2 propose a method for recommending authors of content similar to consumer preferences. Non-Patent Document 3 proposes a method of excluding unnecessary or harmful content from social media content. Non-Patent Document 4 proposes a system that automatically collects and summarizes social media content related to events.

Greene, D .; Reid, F .; Cunningham, P .; and Sheridan, G. “Supporting the curation of twitter user lists”, In NIPS Workshop on Computational Social Science and the Wisdom of Crowds, 2011. Hannon, J .; Bennett, M .; and Smyth, B. “Recommending twitter users to follow using content and collaborative filtering approaches”, In ACM Conference on Recommender Systems (RecSys10), 2010. Dan, O .; Feng, J .; and Davison, B. “Filtering microblogging messages for social TV”, In Proceedings of the World Wide Web Conference (WWW), 2011. Chakrabarti, D., and Punera, K. “Event summarization using tweets”, In Proceedings of the Fifth International AAAI Conference on Weblogs and Social Media (ICWSM), 2011.

  However, it is difficult to divert the techniques described in Non-Patent Documents 1 to 4 as a method for generating or supporting a summary reflecting the intention of the editor.

  The present invention has been made in view of the above-described facts, and an object of the present invention is to provide a content summary support apparatus, method, and program capable of supporting generation of a content summary.

  To achieve the above object, a content summary support apparatus according to the present invention selects a plurality of contents from a set of contents composed of electronic data, and generates a summary composed of the selected plurality of contents. A content summary support apparatus for supporting, comprising: a partial summary input means for accepting a partial summary that is a subset of the content that constitutes the summary input by a user; and the content based on the content included in the partial summary Content candidate search means for searching for a content candidate that is a candidate content included in the summary, and for each of the content candidates searched by the content candidate search means, the partial summary and the content candidate That extracts features related to similarity in combinations of Each of each of the content candidates extracted by the content feature extraction means, a pre-learned function for outputting a value indicating a useful degree of reinforcement of the partial summary based on the feature amount, and the content feature extraction means And ranking means for ranking the content candidates searched by the content candidate search means in an order useful for reinforcing the partial summary based on the feature amount.

  A content summarization support method according to the present invention is a content summarization support apparatus that supports generating a summary composed of a plurality of selected contents by selecting a plurality of contents from a set of contents composed of electronic data. A summary support method that accepts a partial summary, which is a subset of content constituting the summary, input by a user by a partial summary input unit, and based on content included in the partial summary by a content candidate search unit Then, a content candidate that is a candidate content included in the summary is searched from the set of contents, and the partial summarization is performed on each of the content candidates searched by the content candidate search unit by a content feature extraction unit. And the content candidate combination A pre-learned function for extracting a feature amount related to similarity in matching and outputting a value indicating a degree useful for reinforcement of a partial summary based on the feature amount by ranking means; and the content feature extraction means The content candidates searched by the content candidate search means are ranked in the order useful for reinforcing the partial summaries based on the feature quantities of the content candidates extracted by the above.

  The program according to the present invention is a program for causing a computer to function as each unit of the content summary support apparatus.

  As described above, according to the content summary support apparatus, method, and program of the present invention, for each content candidate, a feature amount related to the similarity in the combination of the input partial summary and the content candidate is extracted, Based on the pre-learned function for outputting a value indicating the degree of usefulness for the summary reinforcement and the feature amount of each extracted content candidate, the content candidates are ranked in the order useful for the enhancement of the partial summary. By attaching, it is possible to obtain an effect that it is possible to support generation of a summary of content.

It is a figure which shows the relationship between the author, editor, and consumer of social media. It is a figure which shows the outline | summary of this invention. It is the schematic which shows the structure of the content summary assistance apparatus which concerns on embodiment of this invention. It is a flowchart which shows the content of the ranking function learning process routine in the content summary assistance apparatus which concerns on embodiment of this invention. It is a flowchart which shows the content of the summary assistance process routine in the content summary assistance apparatus which concerns on embodiment of this invention. It is a figure which shows an experimental result.

  Hereinafter, embodiments of the present invention will be described in detail with reference to the drawings. In the present embodiment, an example will be described in which the present invention is applied to a content summary support apparatus that supports content summarization using electronic documents provided by social services such as Twitter (registered trademark) as content. .

<Outline of the invention>
FIG. 2 outlines the concept of a method for supporting content summarization according to the present invention.

First, assume that the system has a partially-curated list that has been created by the editor so far. Hereinafter, for the sake of simplicity, this summary content being edited is referred to as a partial summary S. That is, this partial summary S is a set of contents S = {s ₁ , s ₂ ,... S _Ns } selected by the editor's hand, and is an _uncompleted summary. As an example of the partial summary, when the content is an article (tweet) of Twitter (registered trademark), a set of tweets about a specific theme selected or intended by the editor can be cited.

Next, other contents whose author a (s _i ) = a _i is the author of each content s _i (i = 1, 2,..., N _s ) included in the partial summary S, or can be viewed by the author a _i All content (timeline) is collected, content that can be viewed by the editor a ₀ is added to the content, and a candidate content set T = {t ₁ , t ₂ ,. construct t _NT }. The elements t _j (i = 1, 2,..., N _T ) of this T are ranked and presented to the editor.

  The editor selects content to be included in the summary based on this ranking, and adds the selected ~ T to the partial summary S to form a new partial summary as shown in the following equation (1). .

  Repeat the above work to complete the summary.

<Configuration of content summary support device>
As shown in FIG. 3, the content summary support apparatus 100 according to the present embodiment includes a CPU (Central Processing Unit), a RAM (Random Access Memory), and a program for executing a content summary support processing routine described later. It is comprised by the computer provided with ROM (Read Only Memory) memorize | stored, and it is comprised as shown below functionally. The content summary support apparatus 100 includes a partial summary input unit 1, a content search unit 2, a content feature extraction unit 3, a ranking unit 4, an output unit 5, a learning data input unit 6, and ranking learning data creation. Unit 7, learning content feature extraction unit 8, and ranking function learning unit 9. Social media is an example of a content provision browsing system.

  The partial summary input unit 1 receives a partial summary (a plurality of contents) being edited, which is input by an editor.

The content search unit 2 searches for content candidates that may be useful for the composition of the summary from the content database 20 that stores content sets provided by all users on social media. That is, the content is searched using each content s _i (i = 1, 2,..., N _s ) included in the input partial summary S as a query. More specific methods will depend on the characteristics of social media for which to search for content acquisition, as an example that can be used universally, it extracts key words from each content s _i included in the partial summary A method of searching related content using the word as a query can be considered. In the present embodiment, since the author a (s _i ) = a _i of the content s _i is uniquely determined, the content search unit 2 is provided from the content database 20 by the editor who has input the partial summary as the author. The contents and the contents that a _i can browse as consumers are searched, and all the searched contents are set as content candidates.

  Here, “content that can be browsed as a consumer” will be described. An explanation will be given using Twitter (registered trademark) as an example. In the case of Twitter (registered trademark), a name that uniquely identifies each account is given in the same way as email.

  For simplicity, substitute these symbols with symbols such as A, B, C, ... Twitter (registered trademark) also has a function called timeline, in which each account designates another account such as a friend or celebrity, and the content (tweet) created by the designated account is arranged in chronological order. .

  In such a case, the timeline of account A corresponds to the content that account A can view as a consumer.

  Furthermore, it should be noted that in Twitter (registered trademark), the account A can be a creator of content (tweet) or a consumer of content (tweet) through a timeline.

  Based on this, if the partial summaries are expressed as a set of multiple contents, and the authors without duplication of those contents are A, B, and C, the timeline of A, B, and C The content creator can view the content as a consumer.

The content search unit 2 is a content candidate set T = {t ₁ , t ₂ ,..., Which is a result of searching for content provided by the editor who entered the partial summary as an author and content that a _i can browse as a consumer. .., t _NT } is passed to the content feature extraction unit 3.

  The content feature extraction unit 3 extracts a feature vector, which will be described later, for each content candidate included in the searched content candidate set, using the input partial summary.

The ranking unit 4 rearranges the content candidate set T = {t ₁ , t ₂ ,..., T _NT } output from the content search unit. That is, ranking is performed such that the useful content t _j to be added to the partial summary S is ranked higher. The framework of the machine learning method for performing this ranking is generally called ranking learning. Learning of the ranking function is performed by the learning data input unit 6, the ranking learning data creation unit 7, the learning content feature extraction unit 8, and the ranking function learning unit 9, as described below.

In addition to the partial summaries S and the content candidate sets T, the learning data input unit 6 is provided with a plurality (N _L ) of learning partial summaries S _k inputted by an input device such as a known keyboard, mouse, or storage device. = {S _{k, i} } _{i = 1} ^{Ns, k} and learning content candidate set T _k = {t _{k, j} } _{j = 1} ^{NT, k} (k = 1, 2,..., N _L ) .

The ranking learning data creation unit 7 creates learning data necessary for learning in advance functions necessary for ranking. The ranking learning data creation unit 7 determines whether each of the learning content candidates t _{k, j} (j = 1, 2,..., N _{t, k} ) is useful for reinforcing the partial summary S _k. The label lk, j to be input is received and given to the learning content candidate tk _{, j} . For example, label 5 is given when it is useful, label 1 is given when it is not useful, and labels 2, 3, and 4 are given in the order that it is not useful when it is vague. Such a labeling is used as a learning content candidate set T _L = {T ₁ , T ₂ ,... Corresponding to all the learning partial summaries S _L = {S ₁ , S _{2 ,.} is performed for T _NL}, granted to them label _{L L = {L 1, L} 2, ..., L NL}, L k = {l k, 1, l k, 2, ..., l k, _NTk } (k = 1, 2,..., N _L ) is included to create learning data (S _L , T _L , L _L ).

The learning content feature extraction unit 8 extracts a feature vector to be input to a function for performing ranking from the learning content. Specifically, each learning content candidate t _{k, j} (j = 1, 2,..., N _{T, k} ) for the learning partial summary S _k (k = 1, 2,..., N _L ). ) As feature vectors f _{k, j} (f _{k, j} ⁽²⁾ , f _{k, j} ⁽²⁾ ,..., F _{k, j} ^(NF) ) ^T , the following word similarity features and attention words Extract at least one of features and network features.

  The word similarity feature is a feature amount representing the similarity of the character face of the content, that is, the similarity related to the content. For example, it can be obtained as follows.

First, learning content feature extraction unit 8 obtains content candidate _{t k,} vector representation from _{j t k,} the _j. This vector representation is an N _W dimensional vector when there are N _W words that can appear. Each element of the vector expression is based on the frequency c (j, w) of the word w included in the contents tk _{, j} , and is at least one of the following nonlinear transformation frequency, TF-IDF, and binary weight. Is calculated by using

  The nonlinear conversion frequency is calculated according to the following equation (2).

  TF-IDF is calculated according to the following equation (3).

  However, d (w) is the frequency (sentence frequency) that the word w appears in the content.

  The binary weight is calculated according to the following equation (4).

Learning content feature extraction unit 8, the contents candidate t _k, vector representation from _j t _k, as with _j, obtaining a vector representation s _{k, i} from the content s _{k, i} included in the partial summary.

The learning content feature extraction unit 8, the word similarity f _k, as _j ^(m), in accordance with the following equation (5), the vector representation t _k of the learning content _candidate, and _j ^(m), the corresponding learning The cosine similarity with the vector representation s _{k, i} of each content of the partial summary S _k is calculated.

The calculated word similarity f _{k, j} ^(m) is adopted as the feature amount of the learning content candidate tk _{, j} for the learning partial summary Sk. In addition to the cosine similarity, BM25 similarity frequently used in the field of information search can be used as the similarity between contents. The BM25 similarity is described in non-patent literature (Robertson, S., Zarazoga, H., and Taylor, M. 2004. Simple BM25 Extension to multiple weighted fields. Proc. Of CIKM.).

  The attention word feature is a feature amount representing the degree of similarity of information (keywords, emphasized words, etc.) important to the author. For example, it can be obtained as follows. The basic calculation method is almost the same as the word similarity, but the vector is calculated only with M words to be noticed, not all the words that can appear (M <N). The definition of the word of interest depends on the social media in which the content is searched and collected, but basically, a word specified in some form by the author of the content is used. In this embodiment, in order to collect content from Twitter (registered trademark), a word included in a phrase called a hash tag with a # (sharp) symbol added to the head of the phrase is set as a noticed word. If the word of interest does not exist in the content, the feature is zero.

Learning content feature extraction unit 8, the contents candidate t _k, the vector representation t _k of _{j, j,} and each content s _k included in the partial _sum, also vector representation from _i s _k, for _i, the content t _k, the _j Based on the frequency c (j, w) of the word of interest w included, at least one of the nonlinear transformation frequency, TF-IDF, and binary weight is calculated as each element of the vector expression. The learning content feature extraction unit 8, like the word similarity feature vector representation t _k of the learning content _candidate, and _j ^(m), the vector representation s _k for each corresponding content of the learning part summary S _{k , i} is calculated as the similarity of the word of interest, and is used as the feature amount of the learning content candidate tk _{, j} for the learning partial summary S _k .

  The network feature is a feature amount that represents the similarity related to the relationship between the author and the consumer. Social media has the feature of extracting various network features in addition to language information. For example, the following features (a) to (d) that express the similarity between contents using a social network in binary can be extracted.

(A) Is the author of the content candidate t _{k, j} and the content s _{k, i} the same?
(A) Is the content candidate t _{k, j} and the content s _{k, i} designating the same recipient?
(C) Whether the content candidate t _{k, j} and the content s _{k, i} have the same HTTP link?
(D) Is the author of the content s _{k, i} designated as the recipient of the content candidate t _{k, j} ?

  Here, the content recipient is a content consumer explicitly shown in the content by the author of the content.

Based on the above definition, a feature vector is extracted from each of the learning partial summary and the content candidate set pair (S _L , T _L ) from the learning data (S _L , T _L , L _L ), and the feature vector F _L = {F ₁ , F ₂ ,..., F _NL }, F _k = {f _{k, 1} , f _{k, 2} ,..., F _{k, NT, k} } and the label set L _L The pair (F _L , T _L ) is passed to the ranking function learning unit 9.

The ranking function learning unit 9 uses the feature vector set and label set pair (F _L , T _L ) output from the learning content feature extraction unit 8 to provide the content candidate set when the partial summary S is given. A ranking function that outputs a value for ranking each element of T and indicating a degree useful for reinforcing a partial summary is learned. Many existing learning techniques can be used to learn this function. In this embodiment, non-patent literature (Joachims, T. 2006. Training linear SVMs in linear time. In Proceedings of Knowledge Discovery and Data Mining (KDD).) Is used and a method called SVMrank is used. The ranking function obtained as a result is passed to the ranking unit 4.

  Based on the partial summary S and the content candidate set T, the content feature extraction unit 3 extracts, for each content candidate, a feature vector related to the similarity to the partial summary, like the learning content feature extraction unit 8, and Get the set F. Note that each feature vector corresponds to a content candidate.

  The ranking unit 4 uses the ranking function output from the ranking function learning unit 9 to fix the input partial summary S, and sets each element of the content candidate set T output from the content search unit 2. Ranking. Specifically, by giving each feature vector obtained by the content feature extraction unit 3 to a ranking function, a value indicating the degree to which the content is useful for reinforcing the partial summary (content ranking) is output as the function. Are obtained), and the content candidates are ranked in the order of these values.

  The set of ranked content candidates is output to the user by the output unit 5.

<Operation of content summary support device>
Next, the operation of the content summary support apparatus 100 according to this embodiment will be described. First, when a plurality of pairs of learning partial summaries and content candidate sets prepared in advance are input to the content summary support apparatus 100, the ranking function learning processing routine shown in FIG. Executed.

  First, a plurality of pairs of learning partial summaries and learning content candidate sets input are received in step S101, and each content included in the learning content candidate set received in step S101 is received in step S102. For each candidate, an input of a label indicating a degree useful for reinforcing the partial summary is received.

  In step S103, learning data composed of a plurality of pairs of learning partial summaries and learning content candidate sets received in step S101 and labels of each content candidate received in step S102 is created. To do.

  In the next step S104, a feature vector is extracted for each of the learning content candidates included in the learning data created in step S103. In step S105, based on the feature vector for each of the learning content candidates extracted in step S104 and the label of each of the learning content candidates included in the learning data created in step S103. Then, the ranking function is learned, and the ranking function learning processing routine is terminated.

  When the partial summary being edited is input by the user (editor), the content summary support apparatus 100 executes a summary support processing routine shown in FIG.

  First, in step S111, the input partial summary is received, and in step S112, content related to the content included in the partial summary is searched from the content database 20 using the partial summary received in step S111. The set of content items is set as a content candidate set.

  In step S113, a feature vector is extracted for each content candidate included in the content candidate set searched in step S112, using the partial summary received in step S111.

  In the next step S114, using the learned ranking function, based on the feature vector for each content candidate extracted in step S114, a value indicating a useful degree is obtained, and the content candidate sets are ranked. .

  In step S115, the content candidate sets ranked in step S114 are rearranged and output by the output unit 5.

  The user adds content to the partial summary with reference to the ranked content candidate set output by the output unit 5, and generates a content summary. Further, the partial summary to which the content is added may be further input to the content summary support apparatus 100, and the above summary support processing routine may be repeatedly executed.

<Example>
Examples are shown below. As the basis of learning data, 5,000 (completed) summaries collected from Togetter (registered trademark) and 1.3 million articles (content) collected from Twitter (registered trademark) were collected. Randomly selected from the contents included in each summary, a partial summary for learning was created that included 12 contents on average. In addition, a candidate content set for learning was created by mixing all the contents that were not selected and mixing randomly selected contents from many contents collected from Twitter (registered trademark). This learning candidate content set includes 192 contents on average. About 30% of the content is included in the actual summary, and each content is given a label (l = 5) indicating that it is useful, and the rest, that is, Twitter (registered trademark) The contents (randomly selected from) were given a label (l = 1) indicating that they were not useful.

  By collecting 2000 (completed) summaries from Togetter (registered trademark) and 580,000 content from Twitter (registered trademark) as test data, and performing the same work as learning data, partial summaries and content A candidate set was constructed. In the experiment, a feature vector is extracted from the content candidate set by the method described in the above embodiment, is ranked using a ranking function, and a higher content candidate is selected according to the ranking. did. Then, we evaluated how correctly the content actually included in the summary was selected.

  As an evaluation scale, Mean Average Precision (MAP) is used in this experiment. FIG. 6 shows the experimental results. When the method described in the above embodiment was used, the highest evaluation value was shown as MAP = 0.857 (see SVMrank in FIG. 6). On the other hand, when only word similarity is used without learning, MAP = 0.825 (see Word-TFIDF in FIG. 6), and when randomly selected without using the ranking unit, MAP = 0.53 (FIG. 6). (See Random). As a result, the effectiveness of the method described in the above embodiment was experimentally shown.

  As described above, according to the content summary support apparatus according to the present embodiment, for each of the content candidates, a feature vector relating to similarity to the input partial summary is extracted, and the degree useful for reinforcing the partial summary By ranking the content candidates in the order useful for augmenting the partial summaries based on a pre-learned ranking function for outputting a value indicating and a feature vector extracted for each of the content candidates, It can assist in generating a summary of the content.

  In addition, when the editor inputs a partial summary (partial summary being edited) that is naturally created in the process of creating the summary, the content summarization support apparatus is adapted to the content ( The editing operation can be supported by recommending and presenting content that is likely to be included in the summary.

  Also, by ranking, contents useful for summarization can be quickly found, and editing work can be made more efficient.

  In addition, since a wide range of content candidates are collected, it is possible to secure a wide viewpoint that makes the partial summary being edited more valuable while reducing the time and effort of referring to all the content.

  Note that the present invention is not limited to the above-described embodiment, and various modifications and applications are possible without departing from the gist of the present invention.

  For example, the content may be electronic data other than an electronic document. For example, electronic data such as video data and image data may be used as the content.

  The present invention can also be realized by installing a program on a known computer via a medium or a communication line.

  Further, the above-described content summary support apparatus has a computer system inside, but the “computer system” includes a homepage providing environment (or display environment) if a WWW system is used. .

  In the present specification, the embodiment has been described in which the program is installed in advance. However, the program can be provided by being stored in a computer-readable recording medium.

DESCRIPTION OF SYMBOLS 1 Partial summary input part 2 Content search part 3 Content feature extraction part 4 Ranking part 6 Learning data input part 7 Learning data creation part 8 Learning content feature extraction part 9 Ranking function learning part 20 Content database 100 Content summary support apparatus

Claims (7)

A content summary support apparatus that supports a plurality of contents selected from a set of contents composed of electronic data and generates a summary composed of the selected plurality of contents,
Partial summary input means for accepting a partial summary that is a subset of the content that constitutes the summary input by a user;
Content candidate search means for searching for content candidates that are candidates for content included in the summary from the set of content based on content included in the partial summary;
Content feature extraction means for extracting a feature quantity related to similarity in the combination of the partial summary and the content candidate for each of the content candidates searched by the content candidate search means;
Based on a pre-learned function for outputting a value indicating a degree useful for reinforcement of a partial summary based on the feature amount, and the feature amount of each of the content candidates extracted by the content feature extraction unit. Ranking means for ranking the content candidates searched by the content candidate search means in an order useful for reinforcing the partial summary;
A content summary support device including: The set of contents is provided from all users in a content providing and browsing system in which each user provides contents as an author and browses contents that can be browsed among the contents provided by other users as an author. A collection of content,
2. The content summary according to claim 1, wherein the content candidate search unit searches, as the content candidate, content that can be browsed by an author of content included in the partial summary and content that can be browsed by a user who has input the partial summary. Support device. The learning included in learning data, which is prepared in advance and includes a partial digest for learning, a set of content candidates for learning, and a label with a value indicating the useful degree for each of the content candidates. Learning content feature extracting means for extracting the feature amount for the combination of the learning partial summary and the learning content candidate for each of the content candidates for learning;
The function is learned based on the feature amount of each of the learning content candidates extracted by the learning content feature extracting unit and the label of each of the learning content candidates included in the learning data. A ranking function learning means,
The content summary support apparatus according to claim 1 or 2, further comprising: A content summarization support method in a content summarization support device that supports generating a summary composed of a plurality of selected contents by selecting a plurality of contents from a set of contents composed of electronic data,
Accepting a partial summary, which is a subset of the content constituting the summary, input by the user by the partial summary input means;
Based on the content included in the partial summary, a content candidate search unit searches for a content candidate that is a candidate content included in the summary from the set of content,
For each of the content candidates searched by the content candidate search means, a feature amount relating to the similarity in the combination of the partial summary and the content candidate is extracted by the content feature extraction means,
A function learned in advance for outputting a value indicating a degree useful for reinforcement of a partial summary based on the feature amount by the ranking unit, and the content candidates extracted by the content feature extraction unit A content summary support method that ranks the content candidates searched by the content candidate search means in an order useful for reinforcing the partial summary based on feature amounts. The set of contents is provided from all users in a content providing and browsing system in which each user provides contents as an author and browses contents that can be browsed among the contents provided by other users as an author. A collection of content,
5. The content summary according to claim 4, wherein the content candidate search means searches for content that can be browsed by an author of the content included in the partial summary and content that can be browsed by a user who has input the partial summary as the content candidate. Support method. The learning content feature extraction unit includes a partial summary for learning, a set of content candidates for learning, and a label with a value indicating the useful degree for each of the content candidates. For each of the learning content candidates included in the learning data, extract the feature amount for the combination of the learning partial summary and the learning content candidate,
By the ranking function learning means, the feature amount of each of the learning content candidates extracted by the learning content feature extraction means and the label of each of the learning content candidates included in the learning data The content summarizing support method according to claim 4, further comprising learning the function based on:   The program for functioning a computer as each means which comprises the content summary assistance apparatus of any one of Claims 1-3.