JP2018028866A - Explanatory text creation method, explanatory text creation model learning method, and program - Google Patents

Abstract

PROBLEM TO BE SOLVED: To make it possible to learn a model for appropriately creating a text for explaining a document.SOLUTION: A document for learning division part 22 divides a document for learning that is each of documents in a set of documents for learning to form a document for learning component series that is a series of document for learning components that are components of the divided document for learning. A document component labeling part 24 provides a document for learning component label that is a label indicating the importance of the document for learning component to each of the document for learning components forming the document for learning component series, while referring to a corresponding explanatory text in a set of explanatory texts for learning. An explanatory text creation model leaning part 26 leans, as an explanatory text creation model, a model for predicting the document for learning component label series from the document for learning component series, by using the document for learning component series and a plurality of document for learning component label series that are each a series of the document for learning component labels corresponding to the document for learning component series.SELECTED DRAWING: Figure 1

Description

  The present invention relates to an explanatory sentence generation method, an explanatory sentence generation model learning method, and a program, and in particular, an explanatory sentence generation method, an explanatory sentence generation model learning method, and a program for generating a sentence that explains a given document. About.

  With the rise of social networking services (SNS) such as Twitter (R) and Facebook (R), major changes have been made in the way news is consumed. News consumption methods in the mass media era such as newspapers and television were only unilaterally receiving news sent from the news media to the mass. Basically, the news was accessible twice a day for newspapers and only during the program time on TV.

  However, with the advent of the Internet, news can be obtained not only from newspapers and TV, but also from the Internet, and in recent years, with the rise of SNS and smartphones, it has changed so that news can be easily viewed even in a little free time. A huge amount of news has overflowed on SNS, and a curation service that organizes and presents to readers has also appeared. This indicates a situation where people cannot keep up with news consumption.

  The side that sends news to get selected from a vast amount of news is not only limited to the value of the article, it is necessary to do various devices that attract the attention of many readers and lead to articles. For example, an SNS account is opened to send information, and articles can be posted on SNS.

  The title of the article has also been changed to appeal to readers. In the American viral media "Upworthy", it is said that there is a promise to write 25 headlines for one piece of content in order to add a diffuse headline. On the other hand, extreme titles without content that attracts people's attention are called “fishing titles” and prevent readers from finding the right articles. The composition of titles that appeal to readers rather than “fishing titles” relies on the editor's experience and intuition, and no specific methodology has been found.

  As a method of automatically constructing a title from an article, a method of Non-Patent Document 1 is known in which a headline of a news article is given using a sentence shortening technique. This method rewrites a sentence as short as possible while preserving 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 sentences for the purpose of summarizing the contents, and does not deal with the problem of identifying expressions that attract the interest of SNS users.

  The present invention has been made in view of the above circumstances, and an object of the present invention is to provide an explanatory note generation method and program capable of appropriately generating a sentence explaining a document.

  It is another object of the present invention to provide an explanatory sentence generation model learning method and program capable of learning a model for appropriately generating a sentence explaining a document.

  In order to achieve the above object, the explanatory note generation method according to the present invention is an output explanatory sentence that is an explanatory sentence for explaining an input document from an input document that is a given document and an explanatory sentence generation model that has been learned in advance. An explanatory text generation method in an explanatory text generation apparatus for generating an input document component that is a constituent element of an input document divided by the input document dividing unit dividing the input document into predetermined units. The input document segmentation step that constitutes the input document constituent element series that is a series and the document constituent element label estimation unit use the explanation sentence generation model, so that each input document constituent element that constitutes the input document constituent element series On the other hand, a document component label estimation step for estimating an input document component label, which is a label indicating the importance of the input document component, and an output description sentence generation unit include the input sentence Using the input document component label sequence, which is a sequence of component document sequences and input document component label corresponding to the input document component sequence, an input document component having a high importance indicated by the input document component label And an output explanation generation step for generating an output explanation sentence with emphasis.

  The explanatory note generation model learning method according to the present invention includes a learning document set that is a set of given documents, and a learning explanatory note set that is a set of explanatory sentences explaining each document of the learning document set. An explanatory text generation model learning method in an explanatory text generation model learning device for learning an explanatory text generation model, which is a model for generating a sentence for explaining a given document, wherein the learning document dividing unit includes the learning By dividing the learning document that is each document of the document set into predetermined units, a learning document constituent element sequence that is a series of learning document constituent elements that are constituent elements of the divided learning document is configured. The learning document dividing step and the document component labeling unit are labels indicating the importance of the learning document component for each learning document component constituting the learning document component series. A document component labeling step for assigning a learning document component label with reference to a corresponding explanatory statement in the learning explanatory statement set, and an explanatory note generation model learning unit include the learning document component sequence , And a plurality of learning document component label sequences, which are sequences of learning document component labels corresponding to the learning document component sequence, to predict a learning document component label sequence from the learning document component sequence An explanatory sentence generation model learning step of learning a model for performing the explanation as the explanatory sentence generation model.

  The program according to the present invention is a program for causing a computer to execute each step of the above description sentence generation method or explanation sentence generation model learning method.

  As described above, according to the explanatory note generation method and the program of the present invention, by using the explanatory note generation model learned in advance, for each input document constituent element constituting the input document constituent element series, By appropriately estimating the input document component label and generating the output description text with emphasis on the input document component having a high importance indicated by the input document component label, a sentence explaining the document can be appropriately generated. The effect of being able to be obtained is obtained.

  According to the explanatory note generation model learning method and program of the present invention, a learning document constituent element label is assigned to each learning document constituent element constituting the learning document constituent element sequence, from the learning explanatory sentence set. A sentence for explaining a document by learning a model for predicting a learning document constituent element label series from a learning document constituent element series as an explanatory sentence generation model. The effect that a model for generating appropriately can be learned is obtained.

It is the schematic which shows the structure of the explanatory note production | generation model learning apparatus which concerns on embodiment of this invention. It is a figure which shows the example of an explanatory note production | generation model. It is a figure which shows the example of an explanatory note production | generation model. It is the schematic which shows the structure of the explanatory note production | generation apparatus which concerns on embodiment of this invention. It is a flowchart which shows the explanatory note production | generation model learning process program which concerns on embodiment of this invention. It is a flowchart which shows the explanatory note production | generation processing program 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.

[Overview]
Embodiments of the present invention provide a method for automatically constructing explanatory text that attracts the interest of SNS users from news articles. In the embodiment of the present invention, a model for selecting a sentence considered to attract the interest of an SNS user from a news article is realized by supervised learning. As the teacher information for learning this model, the news article and the most influential SNS post referring to the news article are used. As the strength of influence, for example, the number of retweets and the number of favorites in the case of Twitter (R), the number of likes and the number of shares in the case of Facebook (R) can be used.

  As described above, the embodiments of the present invention are mainly targeted for news articles. However, if a document can be used as teacher information, a document type can be used as a document type. It is not particularly limited. For example, even if the URL content included in the SNS post is not a news article, the present invention can be used similarly. In addition, the present invention can be used as a method for automatically generating an explanatory text as a review material from a book text by using a book review as teacher information instead of an SNS posting. Furthermore, the document does not need to be web content, and can be applied to all forms of documents such as magazines and advertisements, newspapers and headlines.

  In the embodiment of the present invention, (1) a learning document set which is a set of given documents, and a learning explanation set which is a set of sentences explaining each document of the learning document set are given. An explanatory sentence generation model learning method for learning an explanatory sentence generation model that is a model for generating a sentence for explaining a document, and (2) generation of an input document and an explanatory sentence that are documents given separately from the learning document An explanatory text generation method for generating an output explanatory text that is an explanatory text explaining an input document from a model, and these two methods are provided. For the purpose of generating explanatory text that attracts the interest of SNS users, the learning text that is an element of the learning text set is a web news article, and the learning text that is an element of the learning text set is a web news article. It can be interpreted as an SNS post mentioned in.

  Specifically, in the embodiment of the present invention, the document component labeling unit assigns a label to the learning document in consideration of the description content of the learning explanatory text, and the explanatory text generation model learning unit performs the learning By using a pair of a label provided by the document and the document component labeling unit, an explanation sentence generation model for generating a sentence explaining the given document is obtained by supervised learning. By using this explanatory text generation model, the document component label estimation unit estimates a label for a newly given input document, and the output description generation unit estimates it by the document component label estimation unit. Based on the label, an output description is constructed from the input document.

[Configuration of explanatory text generation model learning device]
Hereinafter, an explanatory note generation model learning device according to an embodiment of the present invention will be described with reference to the drawings. The explanation sentence generation model learning device is configured by a computer including a CPU, a RAM, and a ROM storing a program, and is functionally configured as follows. FIG. 1 shows an outline of the configuration of the explanatory note generation model learning device 100 according to the embodiment of the present invention. As illustrated in FIG. 1, the explanatory note generation model learning device 100 according to the present embodiment includes an input unit 10 and a calculation unit 20.

  The input unit 10 accepts a learning document set that is a set of given documents and a learning explanatory text set that is a set of explanatory texts that explain each document in the learning document set.

  The computing unit 20 includes a learning document dividing unit 22, a document component labeling unit 24, an explanatory sentence generation model learning unit 26, and an explanatory sentence generation model storage unit 28.

  The learning document dividing unit 22 inputs a learning document set, and divides the learning document that is each document of the learning document set into predetermined units, thereby learning that is a component of the divided learning document. A learning document component sequence that is a sequence of document components for learning is configured, and a learning document component sequence set that is a set of the document component sequences for learning is output.

  The method for configuring the learning document constituent element series is not particularly limited, but in this embodiment, the learning document is divided into sentences and one sentence is adopted as the document constituent element.

  The document component labeling unit 24 inputs a learning document component sequence set and a learning explanation set, and uses a learning document component sequence composed of a certain learning document and a learning document that explains the learning document. For each of the learning document components that make up the learning document component series, the learning document component label, which is a label indicating the importance of the learning document component, is paired with the explanatory text. The learning document component label sequence that is a sequence of the learning document component label is configured for each pair of the learning document component sequence and the learning explanatory text Then, a learning document constituent element label series set which is a set of learning document constituent element label series is output.

  Although the method of giving the learning document component label is not particularly limited, in the present embodiment, a method of giving in consideration the content included in the learning explanatory text will be described.

  As the learning explanation sentence, a sentence determined to be appropriate as a sentence explaining the learning document is used. That is, it can be considered that the part in the learning document indicating the content included in the learning explanatory text is an important part in the learning document. Based on this idea, in this embodiment, for example, for each sentence of the learning explanation sentence, the sentence in the learning document that most closely matches the contents of the sentence, that is, the learning document constituent element is specified, and this learning A label “1” meaning “selection” is assigned to the document component. The same processing is performed for each sentence of the learning explanation sentence, and a label “−1” meaning “non-selection” is assigned to all remaining learning document constituent elements that are not assigned the label “1”. .

  Through the above procedure, a binary (1 or -1) sequence having the same number of elements as the corresponding learning document constituent element sequence can be configured for each learning document. This binary series is used as a learning document component label series.

  In the above embodiment, only two types of “selected” and “non-selected” are assumed as learning document component labels. However, when importance is set in advance in each sentence of the learning explanatory text In the embodiment, instead of the label indicating “selected” and “non-selected”, the importance is set to the label. That is, the label can take any integer value or any real value. In addition, when the learning document is a news article, since the first half of the article often contains important content, the value of the document component that is determined to be “selected” is larger in the first half of the article. It is also possible to add or multiply the importance of taking.

  In the above-described embodiment, manual labeling is assumed. However, an embodiment in which labeling is automated by calculating the similarity between the learning document constituent element and the learning explanatory text is also conceivable. For the above similarity, any method can be applied as long as it is a method for calculating the similarity between sentences. The cosine distance of the word frequency vector, the cosine distance of the TF-IDF feature, the word vector (Non-patent Document 2) Methods such as a cosine distance characterized by an average vector and a cosine distance of a paragraph vector (Non-Patent Document 3) are conceivable.

(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. Internet <URL : Https://papers.nips.cc/paper/5021-distributed-representations-of-words-and-phrases-and-their-compositionality.pdf>

(Non-Patent Document 3) Le and Mikolov “Distributed representations of sentences and documents,” arXiv: 1405.4053, May 2014. Internet <URL: https://cs.stanford.edu/~quocle/paragraph_vector.pdf>

  Through the above procedure, a sequence having the same number of elements as the corresponding learning document constituent element sequence can be formed for each learning document, and this series is set as a learning document constituent label sequence.

  The explanatory note generation model learning unit 26 inputs the learning document component sequence set and the learning document component label sequence set, and from each learning document component sequence included in the learning document component sequence set, The explanatory note generation model is learned as a model for predicting the corresponding learning document component label sequence, and this explanatory note generation model is output.

  That is, in this embodiment, the purpose of the explanatory note generation model learning unit 26 is to learn a model that predicts a label given to each sentence of the learning document.

  The description sentence generation model is not particularly limited, but in this embodiment, a model using a recurrent neural network (RNN) is adopted. A schematic diagram of the model is shown in FIG.

The model adopted in this embodiment is a word sequence

Enter the label series

Prediction

Is a model that outputs Here, T is the length of the word sequence. In this model, the calculation shown in the following formula (1) is performed. However, each word

Is represented by a 1-of-K expression, that is, a binary vector in which only one of K types of words is 1 and the rest are all 0. Ε (·) is a function that embeds a 1-of-K representation of a given word in a multidimensional vector,

Is a long-short term memory (LSTM) block called a block that learns the long-term dependence of a time-series signal, and σ (•) is an arbitrary activation function.

... (1)

  The specific shape of the activation function σ (•) is not particularly limited, and for example, a sigmoid function or a tanh function can be used. Regarding the embedding function ε (•), not only a method of learning simultaneously with the entire network but also word embedding described in Non-Patent Document 2 can be used.

In the model employed in the present embodiment, the parameter obtained by learning is a linear transformation matrix.

, Bias term

, LSTM

, An internal parameter of the activation function σ (· embedding function ε). These parameters are the correct label sequence Y and the predicted label sequence

And learning to reduce the error. A cross-entropy is generally used as an error function of a label sequence, and a stochastic steepest descent method is generally used for parameter correction.

  In this embodiment, a plurality of sets of learning document constituent element sequences and learning document constituent element label sequences are given as learning data. When the learning document component is a sentence, a label is given for each sentence. Therefore, in order to apply to the explanatory sentence generation model adopted in the present embodiment, the sentence is divided into words and separately defined stop words. After removing, you need to give each word a label. As a method of giving a label to each word, a method such as giving the same thing as the sentence label to all words, or giving only the same word as the sentence label to the first word of the sentence can be considered.

  The explanatory note generation model is not limited to the above-described one. For example, a model called bidirectional LSTM described in Non-Patent Document 4 can be used. A schematic diagram of the model is shown in FIG. In this model, the calculation shown in the following formula (2) is performed.

... (2)

(Non-Patent Document 4) Graves, Mohamed, Hinton “Speech recognition with deep recurrent neural networks,” Proc. ICASSP, 2013. Internet <URL: http://www.cs.toronto.edu/~fritz/absps/RNN13. pdf>

In the embodiments so far, RNN has been used as the explanatory text generation model, but in addition to this, an embodiment using a support vector machine (SVM) is also conceivable. In this embodiment, the hidden layer vector of the learned RNN

Or

Alternatively, an SVM model that predicts a word label using the connected vector as an input of SVM is learned.

  Through the above procedure, the explanatory note generation model learning unit 26 learns an explanatory note generation model composed of RNN or RNN and SVM, and stores this explanatory note generation model in the explanatory note generation model storage unit 28.

[Configuration of explanation generator]
Next, an explanatory note generating apparatus according to an embodiment of the present invention will be described with reference to the drawings. The explanatory note generating apparatus is configured by a computer including a CPU, a RAM, and a ROM storing a program, and is functionally configured as follows. FIG. 4 shows an outline of the configuration of the explanatory note generation apparatus 150 according to the embodiment of the present invention. As illustrated in FIG. 4, the explanatory note generation apparatus 150 according to the present embodiment includes an input unit 60, a calculation unit 70, and an output unit 90.

  The input unit 60 accepts an input document that is an explanation target of a description to be generated.

  The computing unit 70 includes an input document dividing unit 72, an explanatory sentence generation model storage unit 74, a document component label estimation unit 76, and an output explanatory sentence generation unit 78.

  The input document dividing unit 72 inputs an input document, which is a document given separately from the learning document, and divides the input document into predetermined units, so that an input document configuration which is a component of the divided input document An input document constituent element series which is a series of elements is formed, and this input document constituent element series is output.

  The input document dividing unit 72 configures an input document constituent element series by performing the same processing as the learning document dividing unit 22 described above using the input document as an input.

  The explanatory note generation model storage unit 74 stores the same explanatory note generation model as the explanatory note generation model storage unit 28.

  The document component label estimation unit 76 inputs an input document component sequence and an explanatory sentence generation model, and is an input document that is a label indicating the importance of each input document component constituting the input document component sequence The component label is estimated, and an input document component label series that is a series of the input document component labels is output.

By inputting the input document constituent element series into the explanatory note generation model, the input document constituent element label series can be obtained. For example, when bidirectional LSTM is used as the explanatory text generation model, a series of predicted labels for each word using the above equation (2).

Is obtained. In addition, when using bidirectional LSTM and SVM as the explanatory text generation model, the hidden layer vector calculated using equation (2) above

Or

Alternatively, a prediction label for each word can be obtained by inputting the connected vector to the SVM. By integrating the predicted labels for each word obtained in this way with sentences, an input document component label can be obtained. As an integration method, an average value, a maximum value, a minimum value, a mode value, and the like of predicted labels of words in a sentence can be used.

  The output description sentence generation unit 78 inputs the input document constituent element series and the input document constituent element label series, and generates an output description sentence with emphasis on the input document constituent element having a high importance indicated by the input document constituent element label. The output explanation is output by the output unit 90.

  The method for generating the output description is not particularly limited, but in the present embodiment, the following two methods will be described.

  The first method is by selecting an input document component based on the input document component label. When the input document component is a sentence, the input document component label represents an estimated value of the importance of the sentence in the input document. Therefore, a sentence is selected according to a predetermined criterion using the input document component label, and a sentence series in which the selected sentences are connected in the order of appearance in the input document is set as an output explanation sentence. As a reference here, for example, a sentence whose input document component label exceeds a predetermined threshold is selected, a predetermined number of sentences are selected in descending order of the input document component label, and input is made within a range not exceeding a predetermined number of characters. Methods such as selecting document component labels in descending order and using a selection method based on knapsack constraint optimization described in Non-Patent Document 5 are conceivable.

(Non-Patent Document 5) Hirao, Suzuki, Amagasaki “Document summary as an optimization problem,” Journal of the Japanese Society for Artificial Intelligence, Vol.24, No.2, pp.223-231, 2009. Internet <URL: https://www.jstage.jst.go.jp/article/tjsai/24/2/24_2_223/_pdf>

  The second method is a method of editing the sentence selected by the first method so as to further shorten. As a sentence shortening method, for example, the methods described in Non-Patent Document 6 and Non-Patent Document 7 can be considered. In addition, as in the method described in Non-Patent Document 5, it is possible to apply a method for simultaneously selecting a sentence and shortening a sentence.

(Non-patent document 6) Hirao, Suzuki, Amagasaki “Sentence shortening as a combinatorial optimization problem by discriminative learning,” Journal of the Japanese Society for Artificial Intelligence, Vol.22, No.6A, pp.574-584, 2007. Internet <URL: https://www.jstage.jst.go.jp/article/tjsai/22/6/22_6_574/_pdf>

(Non-Patent Document 7) Hirao, Suzuki, Amagasaki “Lightweight sentence shortening method,” Proceedings of the Annual Conference of the Language Processing Society, March 2008. Internet <URL: http://www.anlp.jp/proceedings/annual_meeting/2008/pdf_dir/C3-1.pdf>

[Operation of explanation generation model learning device]
Next, the operation of the explanatory note generation model learning device according to the embodiment of the present invention will be described.

  FIG. 5 shows a flowchart showing the explanatory note generation model learning processing program according to the embodiment of the present invention. When the explanatory sentence generation model learning processing program starts, in step S100, the learning document dividing unit 22 inputs a learning document set, and divides the learning document that is each document of the learning document set into predetermined units. Thus, a learning document component sequence that is a sequence of learning document components that is a component of the divided learning document is configured, and a learning document component sequence that is a set of the learning document component sequences Output a set.

  In step S102, the document component labeling unit 24 inputs a learning document component sequence set and a learning explanation set, and stores a learning document component sequence composed of a certain learning document and the learning document. Learning document component label, which is a label indicating the importance of the learning document component for each learning document component constituting the learning document component sequence by pairing with the explanatory description for explanation Is given with reference to a pair of learning explanation texts, so that the learning document constituent element label series, which is a series of learning document constituent element labels, is converted into a learning document constituent element series and a learning explanatory text. Each pair is configured, and a learning document component label series set which is a set of the learning document component label series is output.

  In step S104, the explanatory note generation model learning unit 26 inputs the learning document component sequence set and the learning document component label sequence set, and each learning document component sequence included in the learning document component sequence set. From this, the explanatory note generation model is learned as a model for predicting the corresponding document component label sequence for learning, and the explanatory note generation model is stored in the explanatory note generation model storage unit 28, and the explanatory note generation model is stored. The learning processing program is terminated.

[Operation of the explanation generator]
Next, the operation of the explanatory note generating apparatus according to the embodiment of the present invention will be described.

  FIG. 6 shows a flowchart showing the explanatory note generation processing program according to the embodiment of the present invention. When the explanatory note generation processing program starts, in step S110, an input document is input, and the input document is divided into predetermined units, whereby an input that is a series of input document constituent elements that are constituent elements of the divided input document. A document component series is constructed, and this input document constituent series is output.

  In step S112, the document component element label estimation unit 76 inputs the input document component element sequence and the description sentence generation model, and indicates the importance level of each input document component element constituting the input document component element sequence. The input document component label is estimated, and an input document component label series which is a series of the input document component labels is output.

  In step S114, the output description sentence generation unit 78 inputs the input document constituent element series and the input document constituent element label series, and emphasizes the input document constituent element having a high importance indicated by the input document constituent element label. A sentence is generated, the output explanation is output by the output unit 90, and the explanation sentence generation processing program is terminated.

[Experimental result]
Next, in order to verify the embodiment shown so far, an experimental result of an experiment using a tweet acquired from the Twitter (R) API as an explanatory text for learning will be described. Out of tweets posted over the past year, we extracted tweets that were retweeted or favorited more than 5 times, including links to articles from specific news sites. In addition, we collected news site articles linked by these extracted tweets, and extracted titles, body text, meta information, and so on. Through the above procedure, a total of about 6,000 articles and 800,000 tweets were collected.

  For each news article, a tweet with the largest number of retweets and favorite registrations was selected from the tweets including the link to the article, and this was set as the correct answer for the description of the news article. That is, this procedure can be regarded as a procedure for setting a learning explanation when a news article is set as a learning document.

  The explanation sentence generation model learning unit 26 and the document component label estimation unit 76 employ a method using both RNN and SVM. As an implementation of model learning, Chainer (Internet <URL: http: / /chainer.org>) and LIBSVM (Internet <URL: https://www.csie.ntu.edu.tw/~cjlin/libsvm/>) for SVM learning and evaluation, respectively, with 4-fold cross validation Model learning and evaluation were performed. The output explanation generation unit 78 only selects a sentence and does not shorten the sentence. That is, the task of this experiment is to select an important sentence from a given input document, and how accurately the correct sentence selection given by the document component labeling unit 24 can be reproduced. The standard.

  As an evaluation method of the explanation generation method in this experiment, ROC curve and area under the ROC curve (AUC) which are typical evaluation scales in the classification problem were used. The ROC curve is a graph with the misclassification rate on the horizontal axis and the correct answer rate on the vertical axis. The more the graph curve is on the upper left, the better the method. AUC is the area of the area at the lower right of the ROC, and the larger the value, the better the method. As evaluation targets, a method of randomly selecting sentences from the input document (random guess), a method of selecting sentences in order from the top of the input document (baseline), and a method according to an embodiment of the present invention (proposed) were adopted.

  The evaluation experiment results in the input document are shown in FIG. From this figure, it can be seen that the method according to the embodiment of the present invention shows the best result, and the result is superior to any evaluation object.

  As described above, according to the explanatory note generation model learning device according to the embodiment of the present invention, a learning document component label is assigned to each learning document component constituting the learning document component series. , Learning a model for predicting a learning document component label sequence from a learning document component sequence as an explanatory statement generation model, given with reference to the corresponding explanatory statement in the learning explanatory statement set Thus, it is possible to learn a model for appropriately generating a sentence explaining the document.

In addition, according to the explanatory note generation device according to the embodiment of the present invention, each input document constituting the input document constituent element series by using the explanatory note generation model learned in advance by the explanatory note generation model learning device. Estimate the input document component label for the component, and generate an output description with emphasis on the input document component with high importance indicated by the input document component label. It can be generated appropriately.
The embodiment of the present invention also relates to a document summarization method for generating an explanatory sentence that briefly explains a document composed of a plurality of sentences such as news. It is possible to automatically generate an explanatory note of news for attraction and to distribute the explanatory note on social media.

[Modification]
A program for executing each process of the explanatory note generation model learning device and the explanatory note generation device is recorded on a computer-readable recording medium, and the program recorded on the recording medium is read into a computer system and executed. By doing so, you may perform the various process mentioned above which concerns on each of an explanatory note production | generation model learning apparatus and an explanatory note production | generation apparatus. Here, the “computer system” may include an OS and hardware such as peripheral devices. Further, the “computer system” includes a homepage providing environment (or display environment) if a WWW system is used. The “computer-readable recording medium” means a flexible disk, a magneto-optical disk, a ROM, a writable nonvolatile memory such as a flash memory, a portable medium such as a CD-ROM, a hard disk built in a computer system, etc. This is a storage device.

  Further, the “computer-readable recording medium” means a volatile memory (for example, DRAM (Dynamic DRAM) in a computer system that becomes a server or a client when a program is transmitted through a network such as the Internet or a communication line such as a telephone line. Random Access Memory)), etc., which hold programs for a certain period of time. The program may be transmitted from a computer system storing the program in a storage device or the like to another computer system via a transmission medium or by a transmission wave in the transmission medium. Here, the “transmission medium” for transmitting the program refers to a medium having a function of transmitting information, such as a network (communication network) such as the Internet or a communication line (communication line) such as a telephone line. The program may be for realizing a part of the functions described above. Furthermore, what can implement | achieve the function mentioned above in combination with the program already recorded on the computer system, and what is called a difference file (difference program) may be sufficient.

  The embodiment of the present invention has been described in detail with reference to the drawings. However, the specific configuration is not limited to this embodiment, and includes designs and the like that do not depart from the gist of the present invention.

10, 60 Input unit 20, 70 Arithmetic unit 22 Document dividing unit for learning 24 Document component labeling unit 26 Explanation sentence generation model learning unit 28 Explanation sentence generation model storage part 72 Input document division part 74 Explanation sentence generation model storage part 76 Document Component element label estimation unit 78 Output explanation generation unit 90 Output unit 100 Explanation sentence generation model learning device 150 Explanation sentence generation device

Claims (8)

An explanatory text generation method in an explanatory text generation apparatus that generates an output explanatory text that is an explanatory text explaining an input document from an input text that is a given document and an explanatory text generation model that has been learned in advance.
An input document dividing step in which an input document dividing unit forms an input document constituent element series that is a series of input document constituent elements that are constituent elements of the divided input document by dividing the input document into predetermined units. When,
An input document in which the document component label estimation unit is a label indicating the importance of the input document component with respect to each input document component constituting the input document component series by using the description sentence generation model A document component label estimation step for estimating a component label;
The output description sentence generation unit uses the input document component label sequence that is a sequence of the input document component sequence and the input document component label corresponding to the input document component sequence to indicate the input document component label. An output description generation step for generating an output description with an emphasis on input document components having high importance,
An explanatory note generation method comprising: 2. The explanatory note generation method according to claim 1, wherein in the input document dividing step, a given input document is divided into sentence units, and one sentence is used as a unit of constituent elements. In the document component label estimation step, a pre-learned neural network for inputting an input document component sequence and predicting a document component label sequence is used as the description sentence generation model. The explanatory note generation method according to claim 1 or claim 2. In the output description sentence generation step, the output description sentence is selected by focusing on an input document constituent element having a high importance indicated by the input document constituent element label and editing the selected input document constituent element. The explanatory note generation method according to any one of claims 1 to 3, wherein: To generate a sentence that explains a given document from a learning document set that is a set of given documents and a learning explanation set that is a set of explanatory sentences that explain each document in the learning document set An explanatory text generation model learning method in an explanatory text generation model learning device for learning an explanatory text generation model that is a model of
The learning document dividing unit divides the learning document, which is each document of the learning document set, into predetermined units, and is a sequence of learning document constituent elements that are constituent elements of the divided learning documents. A learning document dividing step that constitutes a learning document component sequence; and
The document component labeling unit, for each learning document component constituting the learning document component sequence, a learning document component label that is a label indicating the importance of the learning document component, A document component labeling step to be given while referring to the corresponding explanatory text in the learning explanatory text set;
The explanation sentence generation model learning unit learns by using a plurality of learning document component label sequences that are sequences of the learning document component sequence and learning document component labels corresponding to the learning document component sequence. An explanatory sentence generation model learning step of learning, as the explanatory sentence generation model, a model for predicting a learning document constituent element label series from the original document constituent element series;
An explanatory note generation model learning method characterized by comprising: 6. The explanatory text generation model learning method according to claim 5, wherein in the learning document dividing step, the learning document is divided into sentence units, and one sentence is used as a unit of constituent elements. In the explanatory sentence generation model learning step, as the explanatory sentence generation model, a neural network for inputting a learning document component sequence and predicting a learning document component label sequence is learned. The explanatory note generation model learning method according to claim 5 or claim 6.   In order for a computer to execute each step of the explanatory note generation method according to any one of claims 1 to 4 or the explanatory note generation model learning method according to any one of claims 5 to 7. Program.