JP6699031B2 - Model learning method, description evaluation method, and device - Google Patents

Description

  The present invention relates to a model learning method, an explanatory text evaluation method, and an apparatus, and more particularly to a model learning method, an explanatory text evaluation method, and an apparatus for evaluating an explanatory text for a document.

  In recent years, the rise of social networking services (hereinafter referred to as SNS) such as Twitter (registered trademark) and Facebook (registered trademark) has brought about a great change in the method of people's news consumption. In the mass media era, when news media such as newspapers and television were the main sources of news, the only way for people to consume news was to unilaterally receive news sent from the media to the mass. As a result, people were able to get in touch with the news basically twice a day in newspapers and only during news programs on TV.

  However, with the advent of the Internet, news is available not only through newspapers and television, but also through the Internet. In recent years, due to the rise of SNS and smartphones, the environment of news consumption has changed so that news can be easily contacted even in a short free time. Since a huge amount of news has been flooded on the SNS, a curation service has also appeared, which organizes news presented on the Internet and presents it to the reader. This shows that people's news consumption is not keeping up with the huge amount of news distribution.

  In order to select from a huge amount of news, the news sending side is required to take not only the value of the article, but also various mechanisms to attract a large number of readers and guide the article. For example, it is devised to open an SNS account to transmit information, and to install an article button on an SNS.

  In addition, changes have been made to appeal to readers for titles and explanations for inquiring articles on the SNS. In the US viral media “Upworthy”, it is said that there is a promise to write out 25 headline proposals for one content in order to add a diffused headline (title). On the other hand, radical titles that do not have content that attracts people's attention are called "fishing titles," and prevent readers from finding suitable articles. The issue is how to compose a title or description that appeals to the reader, rather than a “fishing title”. However, how to compose a title or description that attracts the reader's attention is left to the experience and intuition of the editor. I haven't found a general methodology.

  On the other hand, as a method of automatically constructing a title from an article, a method of adding a headline of a news article using a sentence shortening technique is known (for example, see Non-Patent Document 1). In the method of Non-Patent Document 1, a sentence is rewritten as short as possible while holding such an expression as much as possible, under the condition that an expression expected to attract the reader's attention is given in advance.

Nishikawa, Imamura, Bessho, Makino, Matsuo “Query-dependent sentence shortening and application to headline generation,” IPSJ Technical Report, 2013-NL-214, No.2, pp.1-7, 2013.

  However, the method of Non-Patent Document 1 is a method of shortening a sentence irrespective of the interest of the SNS user, and does not deal with the problem of identifying an expression that attracts the interest of the user.

  The present invention has been made to solve the above problems, and provides a model learning method and apparatus capable of learning an evaluation model for constructing an explanatory text of a document that attracts a user's interest. The purpose is to

  It is another object of the present invention to provide an explanatory text evaluation method and apparatus capable of constructing an explanatory text of a document that attracts a user's interest.

  In order to achieve the above-mentioned object, a model learning method according to a first aspect of the present invention is a model learning device that includes a feature extraction unit and a model learning unit and that learns an evaluation model for evaluating an explanatory sentence that describes a document. In the model learning method according to, the feature extraction unit, for each of the input learning explanations, a feature regarding the presence or absence of a link of an image included in the learning explanations, a feature included in the learning explanations in advance. Extracting a plurality of features including a feature regarding the presence or absence of a defined keyword, a feature regarding the structure of the learning explanation sentence, and a feature regarding a distributed expression that represents the meaning of a word included in the learning explanation sentence, and And a step of learning the evaluation model based on the plurality of features of each of the learning explanation sentences extracted by the feature extraction unit, the model learning unit executing the evaluation model. ..

  In the model learning method according to the first aspect of the invention, the step of extracting by the feature extraction unit includes the user attribute of the user who posts the learning explanation sentence for each of the input learning explanation sentences, and the learning. Based on learning posting information including a posting time of a learning explanation sentence, a learning document that is a target to be explained by the learning explanation sentence, and learning metadata including a posting time of the learning document For each of the explanation sentences, as the feature, a feature regarding whether or not the keyword included in the learning explanation sentence is included in the learning document, a distributed expression of words included in the learning explanation sentence, and the learning Features related to the similarity to the distributed expression of words included in the document, features based on the posting time of the learning explanation sentence and the posting time of the learning document, and features related to user attributes included in the learning posting information. May be further extracted.

  Further, in the model learning method according to the first aspect, when the metadata of the learning document includes a keyword related to the learning document, the feature extracting unit extracts the input learning For each of the learning explanation sentences based on the learning metadata of the learning document, a keyword relating to the learning document included in the learning metadata of the learning document is included in the learning explanation sentence as the feature. You may make it further extract the feature regarding what is done.

  An explanation sentence evaluation method according to a second aspect of the present invention is an explanation sentence evaluation method in an explanation sentence evaluation device that includes a feature extraction unit and a candidate evaluation unit and evaluates a description sentence candidate that describes a document. Part, for the input explanation sentence candidate, a feature indicating the presence or absence of a link of the image included in the explanation sentence candidate, a feature indicating the presence or absence of a predetermined keyword included in the explanation sentence candidate, the explanation sentence candidate And a plurality of features including a feature indicating a distributed representation of the meaning of words included in the explanation sentence candidates, the candidate evaluation unit, the candidate extraction unit, the feature extraction unit Based on the plurality of features extracted by the step of extracting, and an evaluation model for evaluating the pre-learned explanatory text, the step of evaluating the explanatory text candidates, including the execution. Characterize.

  In the explanation sentence evaluation method according to the second aspect of the present invention, the step of extracting by the feature extracting unit includes the user attribute of the user who posts the explanation sentence candidate for the input explanation sentence candidate and the explanation sentence candidate. Based on posting information including a posting time, a document that is a target to be explained by the explanation sentence candidate, and metadata including a posting time of the document, the keyword included in the explanation sentence candidate is the feature as the feature. Features regarding whether or not they are included in a document, features regarding the similarity between a distributed expression of words included in the explanatory sentence candidates and a distributed expression of words included in the document, and a posting time of the explanatory sentence candidates and the document The feature based on the posted time and the feature related to the user attribute included in the posted information may be further extracted.

  Further, in the explanatory text evaluation method according to the second invention, when the metadata of the document includes a keyword related to the document, the feature extracting unit extracts the metadata of the input document. Based on the above, as the feature, a feature regarding that the keyword related to the document included in the metadata of the document is included in the description sentence candidate may be further extracted.

  A model learning device according to a third aspect of the present invention is a model learning device for learning an evaluation model for evaluating a descriptive text for explaining a document, wherein the learning descriptive text for each of the input learning descriptive texts. Feature related to the presence or absence of a link of the image included in, the feature related to the presence or absence of a predetermined keyword included in the learning description, a feature related to the structure of the learning description, and of the words included in the learning description A feature extraction unit that extracts a plurality of features including a feature related to a distributed expression that represents meaning, and the evaluation model extracted based on the plurality of features for each of the learning explanation sentences extracted by the feature extraction unit. And a model learning unit for learning.

  An explanatory note evaluation device according to a fourth aspect of the present invention is an explanatory note evaluation device that evaluates explanatory note candidates that explain a document, wherein the input explanatory note candidate is a link of an image included in the explanatory note candidate. Feature indicating presence/absence, feature indicating presence/absence of a predetermined keyword included in the explanation sentence candidate, feature indicating structure of the explanation sentence candidate, and feature indicating a distributed expression regarding meaning of words included in the explanation sentence candidate Feature extraction unit that extracts a plurality of features, including, with respect to the description sentence candidates, the plurality of features extracted by the step of extracting the feature extraction unit, an evaluation model for evaluating the pre-learned description sentence And a candidate evaluation unit that evaluates the description sentence candidate based on the above.

  According to the model learning method and the apparatus of the present invention, for each of the learning explanation sentences, the feature regarding the presence or absence of the image link, the feature regarding the presence or absence of the keyword, the feature regarding the structure, and the feature regarding the distributed expression expressing the meaning of the word. By extracting multiple features including, and learning the evaluation model based on multiple features for each of the learning explanation sentences, we learn the evaluation model for constructing the explanation sentence of the document that attracts the reader's interest. The effect that can be obtained is obtained.

  According to the explanatory text evaluation method and apparatus of the present invention, the explanatory text candidates include the feature regarding the presence or absence of the link of the image, the feature regarding the presence or absence of the keyword, the feature regarding the structure, and the feature regarding the distributed expression representing the meaning of the word. An effect that a description sentence of a document that attracts a reader's interest can be configured by extracting a plurality of features and evaluating a description sentence candidate based on the plurality of features and an evaluation model learned in advance. Is obtained.

It is a block diagram which shows the structure of the model learning apparatus which concerns on the 1st Embodiment of this invention. It is a block diagram which shows the structure of the explanatory note evaluation apparatus which concerns on the 1st Embodiment of this invention. It is a flow chart which shows a model learning processing routine in a model learning device concerning a 1st embodiment of the present invention. It is a flowchart which shows the explanatory note evaluation processing routine in the explanatory note evaluation apparatus which concerns on the 1st Embodiment of this invention. It is a block diagram which shows the structure of the explanatory note evaluation apparatus which concerns on the 2nd Embodiment of this invention. It is a flowchart which shows the explanatory note evaluation processing routine in the explanatory note evaluation apparatus which concerns on the 2nd Embodiment of this invention.

  Hereinafter, embodiments of the present invention will be described in detail with reference to the drawings.

<Outline of Embodiment of the Present Invention>

  First, the outline of the embodiment of the present invention will be described. In the present embodiment, a model learning device is used to build an evaluation model for evaluating the explanatory texts that explain a document, and the explanatory text evaluation device is used to create pre-configured explanatory text candidates based on the constructed evaluation model. An example will be described in which an explanatory sentence of a news article that attracts a user's attention is formed by selecting an appropriate candidate from among them.

  For example, the model learning device constructs an evaluation model of an explanatory sentence that enables quantitative analysis of an expression that attracts a user's interest by using the learning feature extraction unit and the model learning unit. In addition, the explanation sentence evaluation device selects a suitable candidate from among the output explanation sentence candidates pre-configured by the candidate evaluation unit on the basis of the constructed evaluation model of the explanation sentence, thereby attracting the user's attention. You can configure the description of the article.

<Configuration of Model Learning Device According to First Embodiment of the Present Invention>

  Next, the configuration of the model learning device according to the first embodiment of the present invention will be described. In the model learning apparatus according to the present embodiment, a news article is used as a learning document, and an evaluation model is learned using a user's post of SNS (eg, Twitter (registered trademark)) as a learning description. A case will be described as an example.

  As shown in FIG. 1, a model learning device 100 according to a first embodiment of the present invention includes a CPU, a RAM, and a ROM that stores a program for executing a model learning processing routine described later and various data. , Can be configured by a computer including. The model learning device 100 functionally includes an input unit 10 and a calculation unit 20 as shown in FIG.

  The input unit 10 includes a learning document (for example, a news article), a keyword related to the learning document as various related information of the learning document (for example, a publisher name, an author name, a proper noun included in the author biographies, etc.), Post learning metadata including a posting time, a plurality of learning explanatory texts (for example, postings on an SNS that mention news articles) that are explanatory texts that explain the learning document, and post each learning explanatory text. The post information for learning including the posting time of the explanatory text for learning and various user attributes of the user who posts the explanatory text for learning, and how well the learning document is expressed A learning evaluation value for each learning explanation is accepted.

  The calculation unit 20 includes a feature extraction unit 30, a model learning unit 32, a distributed expression database 40, and an evaluation model 42.

  The feature extracting unit 30 each of the learning explanation sentences based on the learning document, the learning metadata, the plurality of learning explanation sentences, and the learning contribution information of each learning explanation sentence received by the input unit 10. With respect to, a plurality of features listed in Examples 1 to 6-5 below are extracted. The feature value may be determined using the learning evaluation value.

  As an example 1, the feature extraction unit 30 extracts features related to the presence/absence of a link in an image included in the learning explanatory note. The feature expresses whether or not the description for learning includes information indicating a link to an image by a binary value.

  In addition, the feature extraction unit 30, as in Examples 2-1 to 2-3 below, includes a feature based on the posting time of the learning explanatory note and the posting time of the learning document, and the posting time of the learning explanatory note. And a feature based on.

  As Example 2-1, the feature extraction unit 30 extracts the difference between the posting time of the learning explanation sentence and the posting time of the learning document or the logarithm of the difference as the feature (Time-diff). The posting time of the learning explanatory note can be obtained from the learning posting information, and the posting time of the learning document can be obtained from the learning metadata.

  As an example 2-2, the feature extraction unit 30 sets the posting time period of the learning explanatory note as a feature (Time-period). For example, the time zone is divided into five sections from 2-5 o'clock, 6-9 o'clock, 10-15 o'clock, 16-19 o'clock, and 20-25 o'clock, and five binary values indicating which time zone they belong to. A feature expressed by a series is extracted. The time zone information is acquired from the posting time included in the learning posting information.

  The feature extraction unit 30 sets the posting day of the learning explanation text as a feature in Example 2-3 (Time-weekday). For example, the day of the week is acquired from the post information for learning, and the feature of expressing a weekday or a weekend by a binary value is extracted.

  In addition, the feature extraction unit 30 extracts a feature regarding the presence or absence of a predetermined keyword. For example, as in Example 3-1, which will be described below, a feature in which whether or not a specific keyword is included in the learning explanatory note is expressed by a binary sequence of the number of designated keywords is extracted.

  As an example 3-1, the feature extraction unit 30 sets 1 when the learning explanation includes a keyword specified in advance, sets 0 when it does not include the keyword, and expresses the feature as a binary sequence of the presence or absence of the specified keyword. To extract. In the present embodiment, some keywords related to reasons (Keyword-reason), method (Keyword-method), surprise (Keyword-surprise), special (Keyword-special), etc. are set, and the presence or absence of those keywords is set. Is a feature.

  Further, the feature extraction unit 30 may extract features related to the following keywords in addition to the case where the predetermined keywords are included as in the above Example 3-1.

  As Example 3-2, the feature extraction unit 30 extracts features related to whether or not a keyword related to the learning document included in the learning metadata of the learning document is included in the learning explanation text ( Keyword-publisher). For example, a method of counting up 1 if the learning explanation includes the author name of the learning document, and further counting up 1 if the publisher name is included can be considered. Also, if unique keywords are set in the learning document (in the case of web news articles, keywords may be set in the META tag), whether those keywords are included in the learning explanation Alternatively, a method of using the number of appearances of those keywords can be considered.

  As an example 3-3, the feature extraction unit 30 extracts a feature regarding whether or not the keyword included in the learning explanatory note is included in the learning document. For example, whether or not the general nouns and proper nouns included in the learning explanation are included in the first paragraph of the learning document is extracted by binary representation of the general nouns and proper nouns (Keyword- paragraph1).

  Further, the feature extraction unit 30 extracts features related to the structure of the learning explanatory note and the learning document, as in Examples 4-1 to 4-6 listed below.

  As Example 4-1, the feature extraction unit 30 extracts a feature that represents whether or not a parenthesis is included at the head of the learning explanation sentence by a binary value (Structure-bracket).

  As the example 4-2, the feature extraction unit 30 includes a parenthesis at the beginning of the learning explanatory sentence and a binary feature indicating whether or not the particular keyword is included in the parenthesis, or a feature of the particular keyword. A feature expressed by a binary sequence indicating whether each is included is extracted (Structure-keyword).

  As Example 4-3, the feature extraction unit 30 extracts a feature that binary-expresses whether the first morpheme excluding the parenthesis at the beginning of the learning explanation sentence is a proper noun (Structure-propernoun).

  As Example 4-4, the feature extraction unit 30 extracts a feature that represents whether or not the second morpheme excluding the parenthesis and particle in the beginning of the learning explanation sentence is a noun by a binary value (Structure-noun).

  As Example 4-5, the feature extraction unit 30 extracts a feature represented by the edit distance between the title of the learning document and the learning explanation or the logarithm of the edit distance (Structure-title).

  As Example 4-6, the feature extraction unit 30 extracts the total number of words and the total number of characters of the learning explanation sentence, or a feature represented by logarithm thereof (Structure-length).

  Further, as Example 5, the feature extraction unit 30 extracts features related to user attributes included in the learning post information. For example, when an SNS such as Twitter (registered trademark) is targeted, the posting user attribute acquires the number of followers, the number of followers, the number of registered items in the list, the total number of posts, or the logarithm thereof from the learning posting information. Then you can make it a feature.

  In addition, the feature extraction unit 30, based on the distributed expression database 40, as in Examples 6-1 to 6-4 described below, identifies features related to the distributed expressions that represent the meanings of the words included in the learning description. Extract.

  Here, the distributed expression database 40 stores the distributed expressions of the words learned for each of the words in advance. The distributed representation of words is a framework that expresses words in fixed-dimension real-valued vectors, and the linguistic properties of words are given in this vector so that words with similar meanings become similar vectors. It is common to learn. Such a distributed expression vector of words is described in Non-Patent Document 2 (Mikolov, Sutskever, Chen, Corrado, Dean “Distributed representations of words and phrases and their compositionality,” Proc. Advances in Neural Information Processing Systems (NIPS), pp. .3111-3119, 2013.) and other methods.

  In the present embodiment, the distributed expression of each word is represented by one multidimensional vector, and the distributed expression of a set of words is represented by the sum of the multidimensional vectors corresponding to each word. Thereby, the similarity between words, the similarity between words and word sets, and the similarity between word sets and word sets can be calculated as the similarity between multidimensional vectors. The degree of similarity between multidimensional vectors is calculated using, for example, the cosine distance.

  The feature extracting unit 30 extracts the similarity of the pair of nouns having the highest similarity among the learning explanations as the feature, using the vector of the distributed expression of words as Example 6-1 (Distribute-sim). ..

  The feature extraction unit 30 uses the vector of the distributed expression of words, and as Example 6-2, all the words apart from the word set farthest from the other word set in the learning explanation sentence and the word farthest from the word set. The similarity with the set of words is extracted as a feature (Distribute-farword1). Note that the similarity when the above words are limited to nouns may be the feature.

  As an example 6-3, the feature extraction unit 30 uses the vector of the distributed expression of words, and as Example 6-3, the word farthest from the other word set in the learning explanation sentence and the most relevant word in the learning explanation sentence. The similarity with a noun pair having the highest similarity to distant words is extracted as a feature (Distribute-simfar).

  The feature extracting unit 30 uses the vector of the distributed expression of the words, as Example 6-4, for the word most distant from the other word set in the learning explanation sentence, and for the word most distant. The similarity with the word with the smallest similarity is extracted as a feature (Distribute-farwords2). Note that the similarity when the above words are limited to nouns may be the feature.

  As Example 6-5, the feature extraction unit 30 extracts features related to the degree of similarity between the distributed expression of words included in the learning explanation and the distributed expression of words included in the learning document. For example, the similarity between the union of vectors of distributed expressions of words included in the learning document and the union of vectors of distributed expressions of words included in each of the learning explanation sentences is extracted as a feature (Distribute-paragraph ).

  The model learning unit 32 is based on the plurality of features for each of the learning explanation sentences extracted by the feature extraction unit 30 and the learning evaluation value for each of the learning explanation sentences received by the input unit 10. , The evaluation model is learned and stored as the evaluation model 42. The learning evaluation value will be described as an example of the case where all learning explanations are prepared, but the learning evaluation value is not limited to this, and it is prepared for some learning explanations. Alternatively, semi-supervised learning may be performed.

  The method for learning the evaluation model in the model learning unit 32 is not particularly limited, but in the present embodiment, a ranking model for predicting the rank of the learning evaluation value from the extracted plurality of features is used. The model learning unit 32 uses the ranking model to learn a model for estimating the rank of the learning evaluation value corresponding to the learning feature extracted from the same learning document. Although the amount of calculation required for learning increases compared to a regression model that directly estimates the learning evaluation value, the influence of the absolute value of the learning evaluation value can be mitigated. Although various methods can be used for learning the ranking model, for example, Non-Patent Document 3 (Joachims, “Training linear SVMs in linear time,” Proc. pp.217-226, 2006.) and other methods using the support vector-based method can be considered. A regression model that directly estimates the learning evaluation value from the learning feature may be used. When the regression model is used, a model for directly estimating the learning evaluation value from the learning feature is learned. Since the learning feature is a multidimensional vector, any generally known linear/non-linear regression method can be used. For example, Non-Patent Document 4 (Chang and Lin "LIBSVM: A library for support vector machines," Support vector regression described in ACM Transactions on Intelligent Systems and Technology, Vol.2, No.27, pp.1-27, 2011.) is used.

<Structure of explanatory sentence evaluation device according to first embodiment of the present invention>

  Next, the configuration of the explanatory note evaluation device according to the embodiment of the present invention will be described. As shown in FIG. 2, the explanatory note evaluation device 200 according to the embodiment of the present invention includes a CPU, a RAM, a ROM that stores a program for executing an explanatory note evaluation processing routine, which will be described later, and various data. Can be configured with a computer including. The explanatory text evaluation device 200 is functionally provided with an input unit 210, a calculation unit 220, and an output unit 250, as shown in FIG.

  The input unit 210 describes a document to be explained, metadata including a keyword and a posting time regarding the document as various related information of the document, a plurality of explanation sentence candidates for the document, and the document. Posting information including posting time of the explanatory text candidate and various user attributes of the user who posted the explanatory text candidate, which is additional information when posting each of the explanatory text candidates, is accepted.

  The calculation unit 220 is configured to include a feature extraction unit 230, a candidate evaluation unit 232, an explanation sentence selection unit 234, a distributed expression database 240, and an evaluation model 242.

  The feature extraction unit 230 extracts a plurality of features for each of the explanation sentence candidates based on the document, the metadata, the posted information, and the plurality of explanation sentence candidates received by the input unit 210. Specifically, the same process as the feature extraction unit 30 of the model learning device 100 is performed to extract a plurality of features.

  The candidate evaluation unit 232 evaluates the explanation sentence candidates for each of the explanation sentence candidates based on the plurality of features extracted by the feature extraction unit 230 and the evaluation model 242 for evaluating the previously learned explanation sentence. To do.

  The same model as the evaluation model 42 learned by the model learning device 100 is stored in the evaluation model 242.

  The method for calculating the evaluation value in the candidate evaluation unit 232 is not particularly limited, but in the present embodiment, the evaluation value estimated when a plurality of features for each of the explanatory sentence candidates is given to the evaluation model. To use. When a ranking model is used as the evaluation model, an evaluation value serving as a ranking standard is predicted for each of the explanatory sentence candidates. A ranking obtained by ranking a plurality of explanation sentence candidates may be adopted as the evaluation value. When a regression model is used as the evaluation model, the evaluation value is directly predicted for each of the explanation sentence candidates.

  Based on the evaluation value evaluated by the candidate evaluation unit 232 for each of the explanation sentence candidates, the explanation sentence selection unit 234 selects the explanation sentence candidate that is evaluated as appropriate as the explanation sentence that explains the document and the corresponding post information. The selected set is output to the output unit 250. The method of selecting the explanation sentence candidates and the posted information is not particularly limited, but in the present embodiment, the set of the explanation sentence candidates and the posted information corresponding to the feature having the maximum evaluation value is selected. In addition, by selecting a plurality of features in descending order of evaluation value, from a plurality of explanation sentence candidates corresponding to the selected features, select a combination of explanation sentence candidates and posted information, so that the posted information does not overlap. Good.

<Operation of the model learning device according to the first embodiment of the present invention>

  Next, the operation of the model learning device 100 according to the embodiment of the present invention will be described. When the input section 10 receives a learning document, learning metadata, a plurality of learning explanation sentences, learning posting information of each learning explanation sentence, and a learning evaluation value of each learning explanation sentence, the model learning device 100 executes the model learning processing routine shown in FIG.

  First, in step S100, the learning document, the learning metadata, the plurality of learning explanatory notes, the learning posting information, and the learning evaluation value received by the input unit 10 are acquired.

  Next, in step S102, each of the learning explanation sentences is based on the learning document, the learning metadata, the plurality of learning explanation sentences, and the learning contribution information of each of the learning explanation sentences acquired in step S100. For, the plurality of features listed in Examples 1 to 6-5 above are extracted.

  In step S104, the evaluation model is learned based on the plurality of features of each of the learning explanation sentences extracted in step S102 and the learning evaluation value of each of the learning explanation sentences acquired in step S100. To do.

  In step S106, the evaluation model learned in step S104 is stored as the evaluation model 42, and the process ends.

  As described above, according to the model learning apparatus according to the first embodiment, the learning explanation text is based on the learning document, the learning metadata, the plurality of learning explanation texts, and the learning posting information. For each of the above, features related to the presence or absence of image links, features related to the presence or absence of keywords, features related to structure, features related to distributed expressions that express the meaning of words, and whether or not the keywords included in the learning explanation are included in the learning document. Features related to or, features related to the degree of similarity between the distributed expression of words included in the learning explanation and the distributed expression of words included in the learning document, the posting time of the learning explanation text, and the feature based on the posting time of the learning text , A plurality of features including features related to user attributes included in the learning post information and features related to the keywords included in the learning metadata being included in the learning explanation sentences, and By learning the evaluation model based on a plurality of features and the evaluation value for learning for each of the learning explanation text, an evaluation model for accurately constructing the explanation text of the document that attracts the reader's interest can be obtained. You can learn.

<Operation of the explanatory sentence evaluation device according to the first embodiment of the present invention>

  Next, the operation of the explanatory text evaluation device 200 according to the first embodiment of the present invention will be described. When the input unit 210 receives a document, metadata that is various pieces of related information of the document, a plurality of explanation sentence candidates for the document, and posted information about each of the explanation sentence candidates, the explanation sentence evaluation device 200 displays the information in FIG. The illustrated evaluation processing routine is executed.

  First, in step S200, the document, the metadata, the plurality of explanatory sentence candidates, and the posting information received by the input unit 210 are acquired.

  Next, in step S204, a plurality of features are extracted from each of the explanation sentence candidates based on the document, the metadata, the plurality of explanation sentence candidates, and the posted information about each of the explanation sentence candidates acquired in step S200. ..

  In step S206, based on the plurality of features generated for each of the explanation sentence candidates in step S204 and the evaluation model 242 learned in advance, for each of the explanation sentence candidates, each of the features of the corresponding explanation sentence candidate is evaluated. Predict the value.

  In step S208, a set of an explanation sentence candidate and post information corresponding to the feature showing the maximum evaluation value is selected from the explanation sentence candidates predicted for each of the explanation sentence candidates in step S206.

  In step S210, the set of the explanatory sentence candidate and the post information selected in step S208 is output to the output unit 250, and the process ends.

  As described above, according to the explanation sentence evaluation device according to the first embodiment, a plurality of features are extracted for each of the explanation sentence candidates, and a plurality of features extracted for each of the explanation sentence candidates, By predicting the evaluation value for each of the explanation sentence candidates based on the evaluation model 242, the explanation sentence of the document that attracts the reader's interest can be accurately constructed.

<Configuration of model learning device according to second embodiment of the present invention>

  Next, a configuration of the model learning device according to the second exemplary embodiment of the present invention will be described. The configuration and operation of the model learning device according to the second embodiment are the same as those of the model learning device 100 according to the first embodiment, and detailed description thereof will be omitted.

<Structure of explanatory sentence evaluation device according to second embodiment of the present invention>

  Next, the configuration of the explanatory note evaluation device according to the second embodiment of the present invention will be described. It should be noted that parts having the same configurations and operations as those of the explanatory sentence evaluation device 200 according to the first embodiment are designated by the same reference numerals and description thereof will be omitted.

  The explanation sentence evaluation apparatus according to the second embodiment differs from that of the first embodiment in that each of the explanation sentence candidates is generated from a document.

  As shown in FIG. 5, the explanatory note evaluation device 400 according to the second embodiment of the present invention includes a CPU, a RAM, and a ROM that stores a program for executing the explanatory note evaluation processing routine and various data. , And can be configured by a computer including. This explanatory sentence evaluation device 400 is functionally provided with an input unit 410, a calculation unit 420, and an output unit 250, as shown in FIG.

  The input unit 410 receives a document and metadata.

  The calculation unit 420 according to the second embodiment includes a description sentence candidate generation unit 428, a feature extraction unit 430, a candidate evaluation unit 232, a description sentence selection unit 234, a distributed expression database 240, and an evaluation model 242. It is configured to include.

  Based on the document received by the input unit 410, the explanatory text candidate generation unit 428 generates a plurality of abstract text candidates that summarize the document as a plurality of explanatory text candidates. The method of generating the explanation sentence candidates is not particularly limited, but in the second embodiment, for example, a document summarization method of generating the summary sentence from a document is applied. For example, Non-Patent Document 5 (Nishikawa, Arita, Tanaka, Hirao, Makino, Matsuo “Single Document Summarization Considering Text Consistency by Discriminative Hidden Semi-Markov Model,” Proc. 492-495, 2014.) and the like can be used to select a sentence in a document so as to satisfy a predetermined character number constraint. Further, the sentence shortener shown in Non-Patent Document 1 or the like may be used in combination, and the generated summary sentence candidates may be shortened to serve as explanation sentence candidates.

  The feature extraction unit 430 extracts a plurality of features from each of the explanation sentences based on each of the document, the metadata received by the input unit 410, and the explanation sentence candidate generated by the explanation sentence candidate generation unit 428. Specifically, the same process as the feature extraction unit 30 of the model learning device 100 is performed to extract a plurality of features. In the second embodiment, only the features that can be extracted are extracted without using the posted information. However, the feature may be extracted using any predetermined posting information for each of the explanatory sentence candidates generated from the document.

  Note that the other configurations of the explanatory note evaluation device according to the second embodiment are similar to those of the explanatory note evaluation device 200 of the first embodiment, and thus detailed description thereof will be omitted.

<Operation of the explanatory sentence evaluation device according to the second embodiment of the present invention>

  Next, the operation of the explanatory note evaluation device 400 according to the second embodiment of the present invention will be described. When the input unit 410 receives a document and metadata that is various related information of the document, the explanatory note evaluation processing routine shown in FIG. 6 is executed. Regarding the operation of the explanatory sentence evaluation device of the second embodiment, the same reference numerals will be given to parts having the same operations as those of the first embodiment, and description thereof will be omitted.

  In step S400, the document and metadata received by the input unit 410 are acquired.

  In step S402, a plurality of summary sentence candidates are generated as a plurality of explanatory sentence candidates from the document acquired in step S200.

  In step S404, a plurality of features are extracted from each of the explanation sentence candidates based on the document and metadata acquired in step S400 and each of the explanation sentence candidates generated in step S402.

  Note that other operations of the explanatory note evaluation device according to the second embodiment are similar to those of the explanatory note evaluation device 200 of the first embodiment, and detailed description thereof will be omitted.

  As described above, according to the explanation sentence evaluation apparatus according to the second embodiment, each of the explanation sentence candidates is generated from the document, a plurality of features is extracted from each of the explanation sentence candidates, and the explanation sentence candidate is extracted. By accurately estimating the evaluation value for each of the explanation sentence candidates based on the plurality of features extracted for each of the above and the evaluation model 242, the explanation sentence of the document that accurately attracts the reader's interest can be configured. You can

<Experimental results>

  In order to verify the first embodiment, an experiment was performed in which a tweet acquired from Twitter(R)-API was used as an explanatory text candidate. Among tweets posted in about 3 months, extracted tweets that include links to articles on a specific news site, are retweeted or registered once or more, and have multiple tweets that include links to the same article. did. We also collected news site articles linked from these extracted tweets and extracted titles, texts, and meta information. Through the above procedure, a total of about 700 articles and about 2,800 tweets were collected.

  In addition, the total value of the number of retweets of tweets and the number of favorite registrations was adopted as the evaluation value of the learning explanation sentence candidates. Normalized discount cumulative gain (NDCG), which is one of the scales commonly used in ranking, is adopted as the evaluation method for selecting explanatory sentence candidates. This NDCG becomes 1 when the predicted ranking completely matches the correct answer, and the evaluation value is greatly reduced as the prediction of the higher ranking is missed.

  In model learning, SVMlight was used for regression model learning and SVMrank was used for ranking model learning, and a linear kernel was selected as the SVM kernel. SVM hyperparameter determination was performed with 4-fold cross validation, and SVM model learning was performed with 5-fold cross validation.

  In this experiment, 51-dimensional features among the features shown in the feature extraction unit 30 were used. The number of dimensions of each feature type is as shown in Table 1 below. In addition to the case of using all 51-dimensional features, the effectiveness of each feature is verified by comparing with each feature type and random prediction in the case of using the ranking model learning. When all features were used, regression model learning (All SVR) and ranking model (All) were compared.

  The experimental results are shown in Table 1. From Table 1, it can be seen that although all features are not always better than the results using individual feature types, each feature shows at least better results than random prediction. In addition, the ranking model was significantly better than the regression model. For each feature type, it is suggested that the presence or absence of an image (Image), the structure of the explanatory text (Structure), and the statistical information (User) of the posting user are effective for selecting the explanatory text candidates.

  The present invention is not limited to the above-described embodiments, and various modifications and applications can be made without departing from the scope of the present invention.

  For example, in the above-described embodiment, the explanation sentence evaluation device has been described by way of example in which a plurality of explanation sentence candidates are evaluated, but the present invention is not limited to this, and one explanation sentence candidate may be evaluated. You may.

  Further, in the above-described embodiment, the model learning device learns the evaluation model for evaluating the explanatory note for the news article document, and the explanatory sentence evaluation device uses the learned evaluation model to explain the explanatory text for the news article document. However, the present invention is not limited to this, and the model learning device performs the evaluation model learning to evaluate the explanatory text for the document obtained from the moving image or the sound, and the explanatory text evaluation device. By using the learned evaluation model, it is also possible to evaluate the explanatory note for the document obtained from the moving image or the sound.

  Further, in the explanation sentence evaluation device according to the above-described first embodiment, a case where each of the explanation sentence candidates is received as an input and a plurality of features is extracted using each of the received explanation sentence candidates will be described as an example. However, the present invention is not limited to this, and each of the explanatory sentence candidates corresponding to the document is acquired from the SNS, blog, etc. on the Internet, and a plurality of features are extracted using each of the acquired explanatory text candidates. You may do it.

10, 210, 410 Input unit 20, 220, 420 Operation unit 30, 230, 430 Feature extraction unit 32 Model learning unit 40, evaluation model 40, 240 Distributed expression database 42, 242 Evaluation model 100 Model learning device 200 Explanatory sentence evaluation device 232 candidate evaluation unit 234 explanatory sentence selection unit 250 output unit 428 explanatory sentence candidate generation unit

Claims (9)

A model learning method in a model learning device for learning an evaluation model for evaluating whether or not a description sentence for explaining a document is a configuration that appeals to a reader, including a feature extraction unit and a model learning unit,
The feature extraction unit, for each of the input learning explanation sentences, a feature regarding the presence or absence of a link of an image included in the learning explanation sentence, a feature regarding the presence or absence of a predetermined keyword included in the learning explanation sentence A step of extracting a plurality of features including a feature relating to the structure of the learning explanation sentence and a feature relating to a distributed expression expressing the meaning of words included in the learning explanation sentence,
The model learning unit learns the evaluation model based on the plurality of features of each of the learning explanation sentences extracted by the feature extraction unit. The model learning unit learns the evaluation model based on the plurality of features and the evaluation value for learning so that the evaluation value increases as the degree to which the explanation sentence appeals to the reader increases. The model learning method described. The feature extraction unit extracts the learning posting information including the user attribute of the user who posts the learning explanatory text for each of the input learning explanatory text and the posting time of the learning explanatory text, and , For each of the learning explanation sentences based on the learning document that is the target of the learning explanation sentence and the learning metadata including the posting time of the learning document, Regarding the feature regarding whether or not the keyword included in the explanation sentence is included in the learning document, and the similarity between the distributed expression of the words included in the learning explanation sentence and the distributed expression of the words included in the learning document Features, features based on the posting time of the learning explanation sentence and the posting time of the learning document, features relating to user attributes included in the learning posting information, and whether the learning explanation sentence includes parentheses at the beginning The feature expressed by a binary value and the feature expressed by a binary value indicating whether or not a parenthesis is included at the beginning of the learning explanation and the particular keyword is included in the parenthesis are further extracted. The model learning method described. When the metadata of the learning document includes a keyword related to the learning document,
The feature extraction unit extracts the learning metadata of the learning document based on the input learning metadata of the learning document, and includes the learning metadata of the learning document as the features. The model learning method according to claim 3, further comprising extracting a feature regarding that the keyword related to the learning document included in the learning explanation sentence is included. A description sentence evaluation method in a description sentence evaluation device for evaluating a description sentence candidate for explaining a document, comprising a feature extraction unit and a candidate evaluation unit,
The feature extraction unit, for the input explanation sentence candidate, a feature indicating the presence or absence of a link of an image included in the explanation sentence candidate, a feature indicating the presence or absence of a predetermined keyword included in the explanation sentence candidate, the A step of extracting a plurality of features including a feature indicating the structure of the explanation sentence candidates and a feature indicating a distributed expression regarding the meaning of the words included in the explanation sentence candidates;
The candidate evaluation unit, for the explanation sentence candidate, an evaluation for evaluating whether or not the plurality of features extracted by the step of extracting by the feature extraction unit, and a pre-learned explanation sentence appeals to the reader A model, and evaluating the explanatory sentence candidates based on
Evaluation method including the description. The step of extracting by the feature extraction unit is the posting information including the user attribute of the user who posts the explanation sentence candidate for the input explanation sentence candidate and the posting time of the explanation sentence candidate, and the explanation sentence candidate is explained. Based on the document that is the target and the metadata including the posting time of the document, as the feature, a feature regarding whether the keyword included in the description sentence candidate is included in the document, and the description sentence candidate Related to the degree of similarity between the distributed expression of the words included in and the distributed expression of the words included in the document, a feature based on the posting time of the explanatory sentence candidate and the posting time of the document, and a user included in the posted information Attribute features, a binary representation of whether or not a parenthesis is included at the beginning of the learning explanation sentence, and whether a parenthesis is included at the beginning of the learning explanation sentence and a specific keyword is included in the parenthesis. The explanatory sentence evaluation method according to claim 5, further comprising extracting a feature expressing whether it is binary . When the metadata of the document includes a keyword related to the document,
The step of extracting the feature extraction unit, based on the metadata of the input document,
7. The explanation sentence evaluation method according to claim 6, further comprising extracting, as the feature, a feature regarding that a keyword related to the document included in the metadata of the document is included in the explanation sentence candidate. A model learning device for learning an evaluation model for evaluating whether or not a description explaining a document is a structure appealing to a reader ,
For each of the input learning explanations, the feature regarding the presence or absence of the link of the image included in the learning explanation, the feature regarding the presence or absence of a predetermined keyword included in the learning explanation, the learning explanation A feature extraction unit that extracts a plurality of features including a feature related to the structure of, and a feature related to a distributed expression that represents the meaning of words included in the learning explanation sentence
A model learning unit, which is extracted by the feature extracting unit, learns the evaluation model based on the plurality of features for each of the learning explanation sentences,
Model learning device including. An explanatory sentence evaluation device for evaluating explanatory sentence candidates for explaining a document,
For the input explanation sentence candidates, a feature indicating the presence or absence of a link of an image included in the explanation sentence candidates, a feature indicating the presence or absence of a predetermined keyword included in the explanation sentence candidates, and a structure of the explanation sentence candidates. A feature extraction unit that extracts a plurality of features including a feature indicating the feature and a distributed expression related to the meaning of words included in the description sentence candidate,
For the explanation sentence candidates, based on the plurality of features extracted by the step of extracting the feature extraction unit, and an evaluation model for evaluating whether the pre-learned explanation sentence is a configuration to appeal to the reader, A candidate evaluation unit for evaluating the explanatory sentence candidates,
An explanation sentence evaluation device including.

Images