JP2019121139A - Summarizing device, summarizing method, and summarizing program - Google Patents

Abstract

To provide a summarizing device that generates an easy-to-read summary.SOLUTION: The summarizing device 10 includes: a document analysis unit 18 that analyzes a document and generates document data; an important sentence extraction unit 20 that extracts a plurality of sentences having high importance scores as important sentences from document data analyzed by the document analysis unit 18; a compressed sentence generation unit 22 that compresses each of the extracted important sentences to generate a plurality of compressed sentences corresponding to the important sentences; and a summary generation unit 24 that selects a compressed sentence having a highest summary score from among the plurality of compressed sentences corresponding to the individual important sentences, and generates a summary.SELECTED DRAWING: Figure 1

Description

  The present invention relates to a summarizing apparatus, a summarizing method, and a summarizing program.

  Conventionally, techniques for automatically summarizing documents such as news are known.

  Patent Document 1 below discloses a document summarizing apparatus that automatically generates a summary based on a text. This device uses multiple sentences contained in the original sentence, a fused sentence obtained by fusing two of these sentences, and a divided sentence obtained by dividing the sentence as candidate sentences before compression, and compresses the candidate sentences before compression. A summary candidate sentence is being generated. Then, the document summarizing apparatus selects a summary candidate sentence satisfying a predetermined summary length from the plurality of generated summary candidate sentences and generates a summary.

  Non-Patent Document 1 below discloses an extractive summary model that extracts a part of a document, such as sentences, phrases, and words, and generates a summary based on co-occurrence of words, importance, relevance, etc. . In this extractive summary model, since the generated summary extracts the expression to be output from the original document, it can be composed of words appearing in the original document, and a grammatical summary can be created. .

  Non-Patent Document 2 below discloses a generative summarization model that generates a summarization using words that do not appear in a document by a learning model using teacher data. Since the teacher data of this generative summary model includes a summary composed of sentences of various lengths, it is not possible to easily annotate the structure information of the summary. As a result, the generated summary may be grammatically unnatural or contradictory to the document.

Unexamined-Japanese-Patent No. 2017-151863

Ramesh Nallapati, Bowen Zhow, Cicero dos Santos, Caglar Gulcehre, Bing Xiang: Abstractive Text Summarization using Sequence-to-sequence RNNs and Beyond, Proceedings of the 20th SIGNLL Conference on Computational Natural Language Learning (CoNLL), pp. 280-290, Berlin, Germany, August 7-12, 2016. Rush, A. M., Chopra, S. and Weston, J .: A Neural Attention Model for Abstractive Sentence Summarization, Proceedings of EMNLP 2015, pp. 379-389 (2015)

  In the conventional extractive summary model, for example, a news article of about 200 characters can be extracted with a grammatical structure while maintaining importance and relevance of a sentence that holds down the subject of the article. However, in this model, there is a problem that replacement of the expression, which is one of the characteristic techniques in the summary, is not properly performed, and an easy-to-read summary can not be obtained.

  On the other hand, although the generative summarization model can generate a summarization using words that do not appear in the document as described above, it is not appropriate as a grammatical structure, or there are many summaries that are inconsistent with the original text. There is a problem that it is not possible to obtain an easy summary.

  The present invention is to provide a summary device, a summary method, and a summary program that can generate an easy-to-read summary.

  According to the present invention, in a summarizing apparatus for generating a summary from a document, a plurality of sentences having high importance scores are important from a document analysis unit that analyzes documents and generates document data, and document data analyzed by the document analysis unit. The system includes: an important sentence extraction unit for extracting each sentence as a sentence; and a summary generation unit for generating a summary from the important sentences extracted by the important sentence extraction unit, wherein the summary generation unit Sentence compression to generate a compressed sentence corresponding to the important sentence, and a selection unit for selecting the compressed sentence having the highest summary score as a summary from the plurality of compressed sentences. Do.

  The summary generation unit analyzes the compressed sentence generated by the compression unit to generate compressed sentence data, and a compressed sentence analysis unit; and the compressed sentence, the compressed sentence data, and the extracted feature amount. And a score calculation unit that calculates a summary score.

  In addition, the important sentence extraction unit can extract a proper expression included in the important sentence.

  The compression unit compresses the important sentence using a neural network, and a learning model of the neural network is previously learned from a plurality of learning sentences and a compressed sentence for each of the plurality of learning sentences. Can be

  Another invention is a summary method for generating a summary from a document, comprising: a document analysis step of analyzing a document to generate document data; and a plurality of sentences having high scores from the document data analyzed by the document analysis step , And the summary generation step of generating a summary from the important sentences extracted by the important sentence extraction step, and the summary generation step includes the plurality of important sentences. Compressing each sentence to generate a compressed sentence corresponding to the important sentence, and selecting the compressed sentence with the highest score as a summary from the plurality of compressed sentences generated by the compressing step And.

  Furthermore, another invention is a program for causing a computer to function as each part of the summarizing apparatus according to any one of claims 1 to 5.

  The summary device, summary method, and summary program of the present invention provide an easy-to-read summary.

It is a figure which shows the structure of the summarizing apparatus which concerns on this embodiment. It is a figure showing composition of a learning device concerning this embodiment. It is a flowchart which shows the flow of the process which a summarizing apparatus performs. It is a flowchart which shows the flow of the process which a document analysis part performs. It is a flow figure showing the flow of processing which an important sentence extraction part performs. It is a flowchart which shows the flow of the process which a compressed sentence production | generation part performs. It is a flowchart which shows the flow of the process which a summary production | generation part performs. It is a figure which shows the hardware constitutions of the summarizing apparatus which concerns on this embodiment.

  Hereinafter, embodiments of the summarizing apparatus, the summarizing method, and the summarizing program will be described with reference to the drawings.

[Configuration of Summarizing Device]
The configuration of the summarizing apparatus according to the embodiment of the present invention will be described with reference to FIG. FIG. 1 is a diagram showing the configuration of a summary device according to the present embodiment.

  The summarizing apparatus 10 includes an input unit 12 to which information is input, an operation unit 14 that calculates information input to the input unit 12, and an output unit 16 that outputs information calculated by the operation unit 14.

  The input unit 12 receives a document to be summarized. A document contains multiple sentences, that is, text data.

  The operation unit 14 includes a document analysis unit 18, an important sentence extraction unit 20, a compressed sentence generation unit 22, and a summary generation unit 24. In addition, the calculation unit 14 has a storage unit 26 in which the document input to the input unit 12 is stored.

  The document analysis unit 18 analyzes the document stored in the storage unit 26 to generate document data. Specifically, the document analysis unit 18 performs morphological analysis on the document, and generates document data that has been separated. By morphological analysis, sentences contained in the document are divided into morphemes which are the smallest linguistic unit having meaning, and further information such as part of speech, reading, meaning, original form (only words with utilization) is added to the morpheme Ru. Then, the document analysis unit 18 performs syntactic analysis based on the document data that has already been shared, and generates parsed document data for each sentence.

  The important sentence extraction unit 20 extracts a plurality of sentences having high importance scores as important sentences from the document data analyzed by the document analysis unit 18. Specifically, the important sentence extraction unit 20 extracts feature quantities of sentence data, and assigns an important score based on the feature quantities. A model generated by a learning device 40 described later is used to assign the important score. The feature amount for calculating the importance score includes, for example, the position of the sentence in the entire document, the length of the sentence, the frequency of occurrence of the word (co-occurrence), the degree of importance of the word, the degree of relevance to the subject, the sentence and the sentence, The degree of similarity of the word, the specific expression of the word, and the information of the verb contained in the sentence are used. For example, the higher the frequency of occurrence of a word, the higher the degree of importance of the word, or the higher the degree of relevance to the subject, the higher the importance score can be set. The important sentence extraction unit 20 can also extract important sentences based on preset constraint conditions. The constraint condition is a condition other than the important score, which restricts the freedom of extraction of the important sentence, and includes the sentence length and the proper expression. If a sentence is redundant or too short, it is not suitable for summary, so the important sentence extraction unit 20 extracts an important sentence only when the sentence length is within a predetermined range. In addition, since the proper expression such as person, time, place, and matter is an essential item in the summary, the important sentence extraction unit 20 can extract the important sentence only when it includes a predetermined proper expression. . These constraints can be set arbitrarily, and by combining with extraction of important sentences based on important scores, it is possible to extract accurate important sentences that can be summarized.

  The feature amount of the sentence data is a feature vector having the feature included in the sentence data as an element, and is a vector of feature information of a predetermined unit. The predetermined unit is a morpheme or an N-gram. A morpheme is the smallest linguistic unit that has meaning as described above. An N-gram is a language unit that occurs adjacent to N language units. The feature information of the predetermined unit includes at least one of part-of-speech, reading, meaning, and original (only words with utilization) information. The score based on a vector is a binary value (0 or 1), the number of occurrences, a value of TF-IDF (Term Frequency-Inverse Document Frequency), a value of LDA (Latent Dirichlet Allocation) or the like, and can be arbitrarily set. For example, TF-IDF is useful as an index indicating the importance of a word, because the similarity of the word can be scored. Since LDA can score the similarity of inferred topics, it is useful as an index indicating the degree of relevance of subjects or the similarity between sentences.

  In addition, the important sentence extraction unit 20 is preset to extract N lines (N is an integer of 2 or more) of important sentences from the document data. This numerical value N can be set arbitrarily. By this setting, the top N important sentences are extracted in descending order of the importance score. In addition, when two important sentences having high similarity are included in a plurality of extracted important sentences, the important sentence extraction unit 20 extracts one important sentence having a low important score from among the important sentences. It can be excluded from the target and the top N + 1 important sentences can be extracted. Thus, the important sentence extraction unit 20 can prevent the extracted plural important sentences from including sentences having duplicate contents, and can extract the important sentences that are missed even though the important score is high. it can. In the determination of the degree of similarity, the feature amounts of the important sentences of each other may be within a predetermined range, and the coincidence rate of the words included in the important sentences may be equal to or more than a predetermined value. Although the present embodiment has described the case where the number of important sentences having a high degree of similarity is two lines, the present invention is not limited to this configuration and k lines of important sentences having a high degree of similarity (k is an integer of 3 or more) ) May be more. In this case, the important sentence extraction unit 20 can exclude the k-1 important sentences having low importance scores from among the important sentences from extraction targets, and extract the top N + k-th important sentences.

  The compressed sentence generation unit 22 compresses each of the important sentences extracted by the important sentence extraction unit 20, and generates a plurality of compressed sentences corresponding to the important sentences. Specifically, the compressed sentence generation unit 22 extracts the meaning of the important sentence, generates a word from the meaning, and the compressed sentence having the same or similar meaning while the word and the syntax are different from the important sentence. Generate A neural network such as an Encoder-Decoder translation model or RNN (Recurrent Neural Network) is used for extracting the meaning of the important sentence and generating a word from the meaning, and the model is generated by a learning device 40 described later. .

  In addition, the compressed sentence generation unit 22 is preset to generate M lines (M is an integer of 2 or more) of a compressed sentence from one important sentence. This numerical value M can be set arbitrarily, and is independent of the above-mentioned numerical value N.

  The summary generation unit 24 selects a compressed sentence having the highest summary score from a plurality of compressed sentences corresponding to each important sentence generated by the compressed sentence generation unit 22, and generates a summary. Specifically, the summary generation unit 24 analyzes each compressed sentence to generate compressed sentence data, and a feature amount extracted from the compressed sentence data generated by the compressed sentence analysis unit 28. And a score calculation unit 30 that calculates a summary score based on the

  The compressed sentence analysis unit 28 performs morphological analysis on the compressed sentence to generate divided sentence data. By morphological analysis, sentences contained in the document are divided into morphemes which are the smallest linguistic unit having meaning, and further information such as part of speech, reading, meaning, original form (only words with utilization) is added to the morpheme Ru. Then, the compressed sentence analysis unit 28 performs syntactic analysis based on the compressed sentence data that has already been separated, and generates compressed sentence data that has been parsed for each compressed sentence.

  The score calculation unit 30 extracts feature amounts of the compressed sentence data, and gives a summary score based on the feature amounts. A model generated by a learning device 40 described later is used for giving a summary score. The feature amount for calculating the summary score includes, for example, the genre of the document, the length of the sentence, the frequency of occurrence of the word (co-occurrence), the degree of importance of the word, the specific expression of the word, the verb included in the sentence, and the part of speech Information on relevance, character strings, word strings and part-of-speech strings are used. The relevance of the part of speech is, for example, a pair of the part of speech of the subject and the part of speech of the predicate. The summary score may be set to be higher as the master-detail relationship between the subject and the predicate becomes correct. Also, a part-of-speech sequence can be used for the summary score. Specifically, as the part-of-speech sequence pattern at the end of the sentence in the compressed sentence resembles the part-of-speech sequence pattern at the end of the sentence in the summary used for the teacher data, a high summary score is given. The part-of-speech sequence pattern at the end of the summary-specific sentence includes an excuse stop, a normal state, an ups and downs, a distinctive particle. Specific particles are "etc." For example, the important sentence is "the citrus scent such as grapefruit and lemon reduces the frustration when you want to eat sweets", and the compressed sentence generated from this important sentence is "the citrus scent is "It reduces the frustration when you want to eat sweets." "The citrus scent reduces the frustration when you want to eat sweets." "The citrus scent such as lemon eats sweets. It will reduce the frustration when it is hard. " On the other hand, the summary included in the teacher data is that "the citrus scent reduces the frustration when sweets want to eat". The importance score in this case is 0.94, 1.00, 0.94 in order from the similarity of the part-of-speech string pattern at the end of the sentence, and the part-of-speech string pattern of the sentence end is the same as that of the teacher data 2 A high summary score is given for the second compressed sentence. The assignment of the important score by this part-of-speech sequence is an example, and by combining with the other feature amount, it is possible to select a compressed sentence suitable for the summary.

  The feature amount of the compressed sentence data is a feature vector having the feature included in the compressed sentence data as an element, and is obtained by vectorizing the feature information of a predetermined unit. The predetermined units are morphemes and N-grams. A morpheme is the smallest linguistic unit that has meaning as described above. An N-gram is a language unit that occurs adjacent to N language units. The feature information of the predetermined unit includes at least one of part-of-speech, reading, meaning, and original (only words with utilization) information. The score based on a vector is a binary value (0 or 1), the number of occurrences, a value of TF-IDF (Term Frequency-Inverse Document Frequency), a value of LDA (Latent Dirichlet Allocation) or the like, and can be arbitrarily set. For example, TF-IDF is useful as an index indicating the importance of a word, because the similarity of the word can be scored.

  The summary generation unit 24 selects one compressed sentence line having the highest summary score from the compressed sentences of M lines. Since there are N lines of important sentences, the compressed sentence selected by the summary generator 24 is N lines. Then, the summary generator 24 generates a summary from the selected compressed sentences. At this time, the summary generation unit 24 arranges the selected compressed sentences in the order of the important sentences based on the relationship between the position of each important sentence in the entire document and the compressed sentences corresponding to these important sentences. , And summary generation. The summary generator 24 can also select a compressed sentence based on preset constraints. The constraint condition is a condition other than the summary score, which restricts the freedom of selection of the compressed sentence, and includes the sentence length and the specific expression. The summary generator 24 selects a compressed sentence only when the sentence length is within a predetermined range, since the sentence is redundant or too short to be suitable for a summary. In addition, since a specific expression such as a person, time, place, or matter is an essential item in the summary, the summary generation unit 24 can select a compressed sentence only when it includes a predetermined specific expression. These constraints can be arbitrarily set, and in combination with the selection of the compressed sentence based on the summary score, it is possible to select an accurate compressed sentence that can be a summary.

  The output unit 16 outputs the generated summary. The output may be a text display by a monitor, or the summary may be sent to the external environment as text data.

  With such a configuration, the important sentence extraction unit 20 can extract sentences having a high degree of importance and relevance related to the subject of the original document. In addition, the compressed sentence generation unit 22 can generate a sentence using a word that does not appear in the document. Furthermore, in the summary generation unit 24, compressed sentences whose grammatical structure is broken or not grammatically appropriate can be removed, and a summary that is not inconsistent with the document can be generated. As a result, the summary device 10 can provide an easy-to-read summary.

  Furthermore, before sentence compression, the compressed sentence generation unit 22 extracts the unique expression included in the important sentence and gives an arbitrary label. Then, the summary generation unit 24 replaces the specific expression corresponding to any label included in the selected compressed sentence. In this way, it is possible to prevent unintended conversion of words during compression by giving labels to specific expressions before sentence compression and replacing with abstract expressions corresponding to the labels after being summarized. it can.

  Although the storage unit 26 has described the configuration for storing the document input to the input unit 12, the present invention is not limited to this configuration. The storage unit 26 can also store parsed document data, important sentences, compressed sentences, summaries, important and summary scores, countering relationships between proper expressions and labels, and the like.

[Configuration of learning device]
Next, the configuration of the summarizing apparatus according to the embodiment will be described with reference to FIG. FIG. 2 is a diagram showing the configuration of a learning device according to the present embodiment.

  The learning device 40 includes an input unit 42 to which information is input, an operation unit 44 that calculates information input to the input unit 42, and an output unit 46 that outputs information calculated by the operation unit 44.

  The input unit 42 receives a plurality of learning documents to be learned. Also, the input unit 42 accepts a summary for the learning document. The summary contains multiple sentences.

  The calculation unit 44 includes a learning document analysis unit 48, a feature extraction unit 50, a model learning unit 52, and a storage unit 54.

  The learning document analysis unit 48 analyzes the learning document to generate learning document data. Specifically, the learning document analysis unit 48 performs morphological analysis on the learning document to generate the separated learning document data. By morphological analysis, sentences contained in the document are divided into morphemes which are the smallest linguistic unit having meaning, and further information such as part of speech, reading, meaning, original form (only words with utilization) is added to the morpheme Ru. Then, the learning document analysis unit 48 performs syntactic analysis based on the divided learning document data, and generates learning document data that has been syntactically analyzed for each sentence.

  The learning document is, for example, a news article of a media site online, and its text data is acquired and stored in the storage unit 54. News articles are politics, economy, sports, technology, weather, entertainment, etc., and may be one category of these, or may be multiple categories. The text for learning is not limited to the news article, but may be other text data such as the contents of a paper or an e-mail.

  The feature amount extraction unit 50 extracts feature amounts of the syntactically analyzed document data for learning. The feature amount is a feature vector whose element is the feature of document data for learning, and is obtained by vectorizing feature information of a predetermined unit included in each sentence for machine learning. The feature vector is stored in the storage unit 54 as a feature amount. The predetermined unit is a morpheme or an N-gram. A morpheme is the smallest linguistic unit with meaning. An N-gram is a language unit that occurs adjacent to N language units. The feature information of the predetermined unit includes at least one of part-of-speech, reading, meaning, and original (only words with utilization) information. The score of the feature included in the feature amount is binary (0 or 1), the number of occurrence, the value of TF-IDF (Term Frequency-Inverse Document Frequency), the value of LDA (Latent Dirichlet Allocation), or the like.

  It is preferable that the feature extraction unit 50 treats synonyms and upper / lower terms as the same feature by using a dictionary or the like, or design a feature with easy statistics such as setting a threshold to a TF-IDF value. is there. This makes it possible to reduce the dimension of the feature vector of the feature amount and to be robust against noise.

  The model learning unit 52 receives feature amounts as input, and learns a model for estimating a summary. The summary consists of multiple sentences for one learning document. The model generated by learning by the model learning unit 52 is stored in the storage unit 54.

  For model learning, teacher data and a predetermined machine learning algorithm are used. The teacher data is a learning document and a summary, and specifically, a feature amount extracted from the learning document by the feature amount extraction unit 50 and a summary corresponding to the learning document. These correspondence relationships are correct relationships because they are summarized from documents for learning through people, and are suitable as data for teachers. The feature amount is an explanatory variable for learning a model. On the other hand, the summary is the objective variable for learning the model. Therefore, the model learning unit 52 can learn whether the objective variable can be explained by the explanatory variable and can learn and generate a model that can be analyzed quantitatively by using the feature amount as the input and the summary as the output. .

  As a predetermined machine learning algorithm, linear regression, decision tree, logistic regression, neural network, k-means method, nearest neighbor method, SVM (Support vector machine) or the like can be used. A neural network is used as a model used in the compressed sentence generation unit 22 of the present embodiment. In addition, as an algorithm of machine learning, an on-line linear discriminator capable of learning one by one can be used. This algorithm is particularly useful when, after the operation of the summarizing apparatus 10, it is desired to reflect feedback information from the user in the model immediately. The online linear classifiers include perceptron, CW (Confidence Weighted Online Learning), AROW (Adaptive Regularization of Weight Vectors), and the like.

  The output unit 46 outputs the generated model. The model is used for the important sentence extraction unit 20, the compressed sentence generation unit 22, and the summary generation unit 24. The model of each of these configurations 20, 22, 24 may be the same model, or may be a model generated using a different algorithm according to each application.

[Process flow of summary generation]
Next, the flow of processing performed by the abstracting apparatus 10 will be described with reference to FIG.

  First, in step S01, the document analysis unit 18 analyzes the document and generates document data.

  Then, in step S02, the important sentence extraction unit 20 extracts a plurality of sentences having high scores as important sentences from the document data analyzed in the document analysis step.

  Next, in step S03, the compressed sentence generation unit 22 compresses each of the important sentences extracted in the important sentence extraction step, and generates a plurality of compressed sentences corresponding to the important sentences.

  Finally, in step S04, the summary generation unit 24 selects a compressed sentence with the highest summary score from a plurality of compressed sentences corresponding to each important sentence to generate a summary, and the process ends.

  Next, the flow of processing performed by the document analysis unit 18 will be specifically described with reference to FIG.

  First, in step S11, the document analysis unit 18 performs morphological analysis on the document stored in the storage unit 24 and generates document data on which sharing has been completed.

  Then, in step S12, the document analysis unit 18 performs syntactic analysis based on the document data that has already been split, generates document data that has been syntactically analyzed for each sentence, and ends this processing.

  Next, the flow of processing performed by the important sentence extraction unit 20 will be specifically described using FIG. In this process, it is preset to extract N lines (N is an integer of 2 or more) of important sentences from document data.

  First, in step S21, the important sentence extraction unit 20 extracts feature amounts of sentence data, and assigns an important score based on the feature amounts.

  Then, in step 22, the important sentence extraction unit 20 extracts the top N important sentences in the descending order of the importance score. Here, when a plurality of important sentences to be extracted includes an important sentence having a high degree of similarity, the important sentence extraction unit 20 excludes an important sentence having a low important score from among the extraction targets, and the top N + 1 Extract the second important sentence. This process ends when the important sentences of N lines are extracted.

  Next, the flow of processing performed by the compressed sentence generation unit 22 will be specifically described using FIG. In this process, it is preset to generate M lines of compressed sentences (M is an integer of 2 or more) from one important sentence. Note that this numerical value M can be set arbitrarily, and is independent of the above-mentioned numerical value N.

  First, in step S31, before sentence compression, the compressed sentence generation unit 22 extracts unique expressions included in important sentences and gives arbitrary labels.

  Then, in step 32, the compressed sentence generation unit 22 analyzes the meaning of the important sentence, generates a word from the meaning, and generates a compressed sentence having the same meaning as the important sentence while the word and the syntax are different. This process ends.

  Next, the flow of processing performed by the summary generation unit 24 will be specifically described using FIG. 7. The summary generation unit 24 includes a compressed sentence analysis unit 28 and a score calculation unit 30, and executes this together with these configurations.

  First, in step S41, the compressed sentence analysis unit 28 analyzes each compressed sentence to generate compressed sentence data. Specifically, the compressed sentence analysis unit 28 performs morphological analysis on the compressed sentence to generate divided sentence data. Then, the compressed sentence analysis unit 28 performs syntactic analysis based on the compressed sentence data that has already been separated, and generates compressed sentence data that has been parsed for each compressed sentence.

  Then, in step S42, the score calculation unit 30 calculates a summary score based on the feature amount extracted from the compressed sentence data generated by the compressed sentence analysis unit 28. Specifically, the score calculation unit 30 extracts the feature amount of the compressed sentence data, and assigns a summary score based on the feature amount.

  Furthermore, in step S43, the summary generation unit 24 selects one compressed sentence line having the highest summary score from the compressed sentences of M lines. If there are N lines of important sentences, the summary generation unit 24 performs N times of processing and selects N lines of compressed sentences.

  Finally, in step S44, the summary generator 24 generates a summary from the selected compressed sentences. At this time, the summary generation unit 24 arranges the selected compressed sentences in the order based on the relationship between the position of each important sentence in the entire document and the compressed sentences corresponding to these important sentences, and generates a summary. It can be done.

  The summarizing device 10 only needs to have a function capable of executing the series of processes described above, and the present invention is not limited to the functional configuration shown in FIG. The above-described series of processes can be performed by hardware or software. Also, one functional block may be configured as a single piece of hardware, may be configured as a single piece of software, or may be configured as a combination of these.

  When the series of processes are executed by software, a program constituting the software is installed on a computer or the like from a network or storage medium. It should be noted that the step of writing a program can include processing performed in parallel or separately, not necessarily processing performed chronologically according to the order, as well as processing performed chronologically.

[Hardware configuration of summary device]
Next, the hardware configuration of the summarizing apparatus 10 according to the present embodiment will be described with reference to FIG.

  The summarizing apparatus 10 has a general computer configuration, and, for example, a central processing unit (CPU) 100, a random access memory (RAM) 102, a read only memory (ROM) 104, a storage unit 106, and a network. It has an I / F (Interface) unit 108, an input unit 110, a display unit 112, a bus 114, and the like.

  The CPU 100 is an arithmetic device that implements control and functions of the entire apparatus 10 by reading programs and data stored in the ROM 104, the storage unit 106, and the like on the RAM 102 and executing processing. The RAM 102 is a volatile memory used as a work area or the like of the CPU 100. The ROM 104 is, for example, a non-volatile memory in which a BIOS (Basic Input / Output System) executed when the summarizing apparatus 10 is started, various settings and the like are stored.

  The storage unit 106 is a storage device, such as a hard disk drive (HDD) or a solid state drive (SSD), which stores an operating system (OS) and various application programs.

  A network I / F unit 108 is a communication interface for connecting the abstract device 10 to a network and communicating with an external information terminal (not shown).

  The input unit 110 is, for example, an input device such as a keyboard and a mouse, and is used by the operator of the abstracting apparatus 10 to input various operation signals to the abstracting apparatus 10. The display unit 112 is, for example, a display device such as a display, and displays the processing result by the abstracting device 10 and the like. The input unit 110, the display unit 112, and the like may be connected to the abstracting apparatus 10 when necessary, and may be used.

  The bus 114 is commonly connected to the above-described components, and transmits, for example, address signals, data signals, various control signals, and the like.

  The summarizing apparatus 10 according to the present embodiment is not limited to the above configuration, and the hardware configuration may be realized by separate computers. Moreover, when the function as the summarizing apparatus 10 is operated by an application installed in the information terminal, part of the hardware configuration of the summarizing apparatus 10 of the present embodiment may be realized by the hardware of the information terminal.

  DESCRIPTION OF REFERENCE NUMERALS 10 summary apparatus 12 input unit 14 operation unit 16 output unit 18 document analysis unit 20 important sentence extraction unit 22 compressed sentence generation unit 24 summary generation unit 26 storage unit 28 compressed sentence analysis unit 30 score Calculation unit, 40 learning device, 42 input unit, 44 operation unit, 46 output unit, 48 learning document analysis unit, 50 feature amount extraction unit, 52 model learning unit, 54 storage unit.

Claims (6)

In a summarization device which generates a summarisation from documents:
A document analysis unit that analyzes a document and generates document data;
An important sentence extraction unit that extracts, as important sentences, a plurality of sentences having high importance scores from the document data analyzed by the document analysis unit;
A compressed sentence generation unit that generates a plurality of compressed sentences corresponding to the important sentences by respectively compressing the important sentences extracted by the important sentence extraction unit;
A summary generation unit that selects the compressed sentences with the highest summary score from the plurality of compressed sentences corresponding to the respective important sentences to generate a summary;
A summarizing device characterized by having: The summarizing apparatus according to claim 1, wherein
The summary generator
A compressed sentence analysis unit that analyzes each of the compressed sentences and generates compressed sentence data;
A score calculation unit that calculates a summary score based on the feature amount extracted from the compressed sentence data;
A summarizing device characterized by having: The summarizing apparatus according to claim 1 or 2, wherein
The compressed sentence generation unit extracts specific expressions included in the important sentence and adds an arbitrary label before the sentence compression.
The summary generator replaces the corresponding unique expression for the arbitrary label included in the selected compressed sentence.
Summary device characterized by. The summarizing device according to any one of claims 1 to 3, wherein
The compressed sentence generation unit compresses the important sentence using a neural network,
The learning model of the neural network is previously learned from a plurality of learning documents and an abstract for each of the plurality of learning documents.
Summary device characterized by. A summary method for generating a summary from documents,
A document analysis step of analyzing a document and generating document data;
An important sentence extraction step of respectively extracting, as important sentences, a plurality of sentences having high scores from the document data analyzed by the document analysis step;
A compressed sentence generation step of sentence compression of the important sentences extracted in the important sentence extraction step to generate a plurality of compressed sentences corresponding to the important sentences;
A summary generation step of selecting each of the compressed sentences having the highest summary score from the plurality of compressed sentences corresponding to each of the important sentences to generate a summary;
A summary method characterized by having: The program for functioning a computer as each part of the summarizing apparatus as described in any one of Claim 1 to 5.