JP6409071B2 - Sentence sorting method and calculator - Google Patents

Description

  The present invention relates to a sentence rearranging method and a computer for rearranging a plurality of sentences.

  In recent years, information services for finding useful knowledge by analyzing a large amount of text data have been put into practical use. For example, there are efforts to automatically estimate the reputation of products using product review articles and social network texts, and to identify only the main points by analyzing news articles and automatically summarize documents. ing. The reasons for these efforts include, for example, the fact that a large amount of text data can be accessed via the Internet, information access technologies for large amounts of information such as full-text search and databases, and machine learning and The progress of automation and semi-automation of analysis that applies pattern recognition technology.

  For example, in the case of reputation analysis, positive expressions and negative expressions are maintained as a dictionary. Using the dictionary, automatic reputation analysis is realized by causing a computer to learn affirmation / denial determination patterns from a result of a person's evaluation of affirmation / denial in the past using a machine learning technique.

  In such text analysis, the user interprets the analysis result before it becomes useful. At the time of interpretation, not only the analysis result but also additional information such as the reason and basis of the judgment, and a case in accordance with the judgment can help the user's interpretation. For example, in the reputation analysis, “Product A is positively received by 70% of users. ”As well as the opinions of users who have actually made positive judgments are listed, it is easy to make use of it for product improvement, because it is understood from what viewpoints are evaluated.

  As background art of this technical field, Patent Document 1 discloses a method of analyzing user emotions from text and enumerating representative opinions thereof.

  In addition, not only enumerating a plurality of sentences but also arranging them in the order in which the meanings pass is automatic document summarization. The automatic document summarization includes a single document summary that shortens a single news article to a short number of characters, and a multi-document summary that summarizes a plurality of news articles as one short document. From the viewpoint of automatic document summarization, a task that summarizes the contents of the original news article as much as possible, or a task that generates a summary by focusing on sentences related to a given query as a query-oriented summary. There is.

  What is important in automatic document summarization is topic consistency. For this purpose, Lexical Chain and topic models are being used. For example, Lexical Chain is a model that uses WordNet to find similarities between words and connects similar words like a chain. Based on the process of using words having similar meanings, sentences including words connected by Lexical Chain are preferentially arranged.

JP 2012-256282 A

  In addition to the results of automatic text analysis, if you can gather the sentences that are the reason for the decision and the sentences that are specific examples of the decision and present them to the user in a consistent order, Interpretation of data after analysis becomes easier and reliability of interpretation is improved. For this purpose, it is necessary to appropriately rearrange a plurality of sentences obtained from a plurality of documents.

  At that time, it is difficult to arrange all sentences between sentences without taking into account the topic change. In reality, even when making a claim, the topic is changed by explaining related matters or using analogies. A connection based only on semantic similarity, such as Lexical Chain, is not enough to model the change of topic, which causes sentence rearrangement failure.

  For this reason, it is necessary to consider the connection of contents on the premise that the semantic relationship, topic similarity by topic, and topic change. Moreover, considering the practical aspects of the system, it is also necessary to cover topics on a wide range of themes.

  As described above, in the above-described conventional technology, there is a problem that it is difficult to construct a consistent claim including a change of a story to another related topic. In addition, there is a problem that it is difficult to rearrange sentences by selecting sentences comprehensively and with an appropriate abstraction level for an arbitrary theme designated by the user.

  An object of the present invention is to generate an output statement that is logically consistent.

A sentence rearranging method and a computer according to an aspect of the invention disclosed in the present application include a processor that executes a program and a memory that stores the program, and specifies a set of sentences and a kind of sentence. It is possible to access a storage device that stores a template that defines a plurality of sentence types and the order of arrangement of the plurality of sentence types, and a rule or learning model relating to a phrase, and the processor constitutes an arbitrary theme sentence A theme expression specifying process for specifying the theme expression from each sentence of a plurality of sentences in the set of sentences based on a rule or learning model relating to a word or a word that defines a theme expression that is a word related to the phrase; , The text specified by the theme expression specifying process based on a rule or learning model relating to the relationship between the theme expression and the sentence type. A sentence type determination process for determining a sentence type for each sentence including a macro expression, and the plurality of sentence types specified in the template from the plurality of sentences for which the sentence type is determined by the sentence type determination process. Select a corresponding sentence set, and execute a rearrangement process for rearranging the selected sentence set according to the rearrangement order, and an output process for outputting the rearranged sentence set by the rearrangement process It is characterized by doing.

  According to an exemplary embodiment of the present invention, a logically consistent output statement can be generated. Problems, configurations, and effects other than those described above will become apparent from the description of the following embodiments.

It is explanatory drawing which shows an example of an intelligent processing system. It is explanatory drawing which shows the specific example of the sentence extraction part and sentence rearrangement part which were shown in FIG. It is explanatory drawing which shows the system configuration example of a document processing system. FIG. 4 is a block diagram illustrating a hardware configuration example of the computer illustrated in FIG. 3. It is explanatory drawing which shows the example of a memory content of document DB shown in FIG. It is explanatory drawing which shows the example of the memory content of a sentence set. It is a block diagram which shows the functional structural example of a document processing system. It is explanatory drawing which shows the example of a data structure of dispute point ontology information. It is explanatory drawing which shows Example 1 of a sentence analysis result. It is explanatory drawing which shows Example 2 of a sentence analysis result. It is explanatory drawing which shows the data structure example of a template. It is explanatory drawing which shows the example of a data structure of an issue chain model. It is explanatory drawing which shows the chain condition of a dispute point. It is explanatory drawing which shows the usage example of an issue chain model. It is explanatory drawing which shows an example of the output information from an output part. It is a block diagram which shows the detailed functional structural example of a production | generation part. It is a flowchart which shows the example of a rearrangement process procedure by the computer of a document processing system.

<Explanation of terms>
First, terms used in this embodiment will be described. “Theme sentence (or simply, theme)” means a central issue, subject or topic of creation or discussion. In debates, for example, “The consumption tax should be increased. ”Is given as a theme.

  The “theme expression” is a phrase that represents a survey viewpoint of a survey that a user who uses the document processing system of this embodiment wants to perform and related phrases. The survey that the user wants to perform is given to the document processing system as a theme. For example, the theme is “Consumption tax should be increased. "Consumption tax" and "tax increase" are themes. In addition, “tax”, “value added tax”, “VAT (Value Added Tax)” and the like, which are related words of the theme expression, are also handled as theme expressions.

  “Issues” is the object, viewpoint, and focus of discussion in past documents, and “Issues” is a phrase indicating a specific example of an issue. For example, when the issue is “News”, “Economy”, “Economy”, “Product”, “Tax”, “Domestic policy”, “Diplomacy”, and the like are expressed as issue points. However, since these words have a wide range of meanings, the words of the subordinate concepts may also be used as the dispute expression. For example, in the case of “tax system”, subordinate concepts such as “national tax”, “local tax”, “direct tax”, “indirect tax”, “income tax”, and “resident tax” are also expressed as issues. The issue point and issue point expression are defined by issue point ontology information 701 (see FIG. 8) described later.

  “Positive / negative expression” is a phrase that is positive or negative about the issue. Like general reputation analysis and sentiment analysis, positive and negative expressions are defined in relation to the issue. For example, “Junk food is bad for your health. "Health" is the issue, and "bad" is a negative expression that denies the issue.

  The promotion / suppression expression is an expression that promotes or suppresses the theme expression. Promotion / suppression expressions are defined by a combination of theme expression and predicate. Specifically, in the case of promotion expression, it is a set of a certain theme expression and a predicate that promotes the theme expression, and in the case of suppression expression, it is a set of predicate that expresses the theme expression and suppression. For example, “Moderate exercise is good for your health. In the case of the sentence, “moderate exercise”, which is the theme expression, promotes “health”. “The number of passengers has decreased due to the increase in fares. In the case of the sentence “”, the theme expression “raise in fare” has the relationship that “the number of passengers” is suppressed.

  The theme, theme expression, issue point, issue expression, positive / negative expression, and promotion / suppression expression may be defined in advance by a dictionary table in the document processing system. Further, the document processing system may apply a learning model generated by machine learning to specify a theme, a theme expression, an issue point, an issue expression, an affirmation / negative expression, and a promotion / suppression expression.

<Specific examples of phrases>
Next, a specific example of the above phrase in the present embodiment will be described. For example, there are two types of phrase specifying methods: a rule-based method and a machine learning-based method. The rule-based method is a method in which a specific method is expressed as a definitive rule, and is, for example, a dictionary table storing words. Therefore, when a phrase that matches the input sentence is found in the dictionary table, the phrase is specified from the input sentence. Appropriate annotations are given to specific words in the input sentence.

  The method based on machine learning generates learning models by converting past cases into numerical vectors, and converts input sentences into numerical vectors in the same way as learning past cases. And the corresponding annotation is given to the phrase specified by the learning model in the input sentence.

  Here, a method based on machine learning will be specifically described. As a method based on machine learning, a support vector machine (SVM) or a conditional random field (CRF) can be used as a sequence labeling problem. A sequence labeling problem is a kind of formulation of a machine learning problem used in specific expression identification, morphological analysis, and the like. Specifically, it is a problem of assigning a tag corresponding to each node to data given as a sequence of nodes (sequence labeling problem). In the case of annotations on sentences, morphemes and characters are often used as nodes.

  Here, in the sequence labeling problem, there are two learning phases: a learning phase in which a learning model is generated from a manually provided training document, and an application phase in which a tag is added to unknown data using the generated learning model.

  First, the learning phase will be described. In the learning phase, the computer manually identifies certain features (eg, themes, theme expressions, issues, issues, positive / negative expressions, and promotion / suppression expressions) as training documents and adds annotations that indicate the specified features. Prepare training documents. In the following, issues are cited as an example of features.

  Next, the computer converts the character string of the sentence in the training document into a series of numerical vectors. Specifically, the computer first performs a morphological analysis on the sentence in the training document, and sets each morpheme as a series node. Next, the computer lists the surface character strings and parts of speech of each morpheme. Assume that there are a total of Vs types of surface character strings and a total of Vp types of parts of speech in the training document. The computer sequentially adds each surface character string and each part of speech to the dictionary table as a dictionary key, and assigns natural numbers one by one so as not to overlap each other in the order of registration in the dictionary table.

  The computer uses the dictionary table to convert the surface character string and part of speech of each node into numerical values. Then, each node can be expressed as a numerical vector in which the dimension corresponding to the surface character string and the part of speech is set to 1, and the other dimensions are set to 0. For this numerical vector, a tag indicating whether or not it is an issue is paired. If the node corresponds to the issue, an issue tag is assigned. If the node does not correspond to the issue, an O (other) tag is assigned.

  As another tag assignment method, when a compound word issue is assumed, the issue tag is distinguished such as B-issue to the node that is the start position of the issue and I-issue to the node that is the intermediate position. There is also a way to do it. As a variant, there is a method of further distinguishing the end position as E-issue and the issue of a single node as U-issue. The computer inputs the combination of the numerical vector sequence and the tag to the SVM or CRF, and generates a learning model.

  When applying a learning model to an input sentence, the computer applies the learning model by converting the input sentence to a numeric vector sequence in the same way as the sentence in the training document, and annotates the corresponding phrase of the input sentence. .

  As described above, when a phrase is specified in the following embodiment, either a rule-based method or a machine learning-based method is applied. In the above-described machine learning, the issue has been described by taking the issue as an example, so that the tag notation is “issue”, but in the case of other features, the tag notation is in accordance with the feature. For example, the theme expression tag is the theme tag, the positive expression is the positive tag, the negative expression is the negative tag, the promotion expression is the promote tag, and the suppression expression is the suppress tag.

<Example of intelligent processing system>
FIG. 1 is an explanatory diagram illustrating an example of an intelligent processing system. The intelligent processing system 1 has a generation system 100 and a data management system 101. In the intelligent processing system 1, the generation system 100 executes intelligent processing in cooperation with the data management system 101.

  First, an overview of the intelligent processing system 1 of the present invention will be described. The intelligent processing system 1 of the present invention includes a generation system 100 in which nine modules are combined, and a data management system 101.

  When the agenda is input, the generation system 100 outputs a descriptive sentence that describes an opinion on the agenda. The data management system 101 stores data processed in advance and is accessible from the generation system 100. In the following, the generation system 100 will be described first, and then the data management system 101 will be described.

  In the generation system 100, the nine modules 102 to 110 are executed in order. These modules 102 to 110 are configured by independent algorithms.

  In the generation system 100, the input unit 102 first receives an agenda input from a user. The generation system 100 may further accept an input regarding whether to generate a positive opinion for the agenda item or to generate a negative opinion. In this way, by clarifying the position of the sentence to be generated, the intelligent processing system 1 can be used in a discussion such as a debate.

  Next, the agenda analysis unit 103 analyzes the agenda, and the generation system 100 determines the polarity of the agenda and the keyword used for the search. Subsequently, the search unit 104 searches for an article using the keyword and the dispute word indicating the issue in the debate. For example, if the agenda is “Casino should be banned”, the keyword is “casino” which is a noun phrase. Further, the generation system 100 can determine whether to use a positive or negative issue word for “Casino” by determining the polarity. The issue word here is, for example, a word that expresses a value that becomes an issue point in the discussion. Further, the “promotional word” here means “an event that promotes the issue”. On the other hand, the “suppression word” here refers to “an event that suppresses the issue”.

  When the generation system 100 wants to output an affirmative opinion on the above agenda, the search unit 104 selects “casino” as the keyword and “suppression word” that suppresses the casino as the issue word. Perform a search. Here, since the agenda is negative with respect to “Casino”, the processing using “suppression word” as the issue word is performed. In the search, an article that states whether or not “casino” is right or wrong can be searched by searching for a question word such as “casino” and a keyword. Only keywords extracted from the agenda include many articles that do not need to be considered in the debate, such as advertised articles about "Casino" or blog articles that describe only impressions about "Casino". It is included and cannot be searched properly.

  Subsequently, in the issue system 100, the issue determination unit 105 classifies the output articles and determines an issue to be used when generating an opinion. Next, in the generation system 100, the sentence extraction unit 106 extracts a sentence describing the issue from the output article. Subsequently, the generation system 100 generates a sentence by rearranging the extracted sentences in the sentence rearrangement unit 107. Subsequently, in the generation system 100, the evaluation unit 108 evaluates the generated sentence. Subsequently, in the paraphrase unit 109, the generation system 100 inserts an appropriate conjunction and deletes unnecessary expressions. Next, the generation system 100 outputs the sentence with the highest evaluation as an essay statement that expresses the opinion at the output unit 110.

  The data management system 101 includes an interface / structure unit 111 and four databases (databases) 112 to 115.

  The interface / structure unit 111 provides an access unit for data managed in the DBs 112 to 115. The interface / structuring unit 111 is an interface that provides an access means to the text data DB 112, the text annotation data DB 113, the search index DB 114, and the dispute point ontology DB 115, and is implemented by a technique such as REST.

  The text data DB 112 is text data such as news articles. The text annotation data DB 113 is data assigned to the text data DB 112. The search index DB 114 is an index for enabling the text data DB 112 and the annotation data DB 113 to be searched. The issue ontology DB 115 is a database in which issues that are frequently discussed in debates and related words are linked. In addition, the issue point ontology DB 115 includes issue point ontology information 701 (see FIG. 8).

  In the following, first, the data management system 101 will be described first, and then each part of the generation system 100 will be described.

  Data stored in the text data DB 112 is text data such as news articles. The generation system 100 extracts an appropriate sentence for composing an opinion sentence from the text data, and generates an argument sentence by arranging the extracted sentences. Therefore, the text data DB 112 is a data source of sentences that constitute the output statement. The text data DB 112 is constructed by crawling English and Japanese news articles from the Internet. Each data is managed by assigning doc_id as a unique identifier, for example.

  The text annotation data DB 113 is a DB in which data assigned to the text data DB 112 is accumulated. The text annotation data DB 113 stores annotations such as positive, negative, promote, and suppress described above.

  The search index DB 114 is index data for enabling the text data DB 112 and the text annotation data DB 113 to be searched. As index data for keyword search, a vector value of a characteristic word statistic (for example, a calculation result by TF-IDF) of each text data for similarity search is stored as an index for similarity search. Alternatively, the search index 114 can be automatically generated by inputting the text data 112 and the text annotation data 113 to the Solr index generation API using software such as Solr.

  Further, a vector value of a calculation result by TF-IDF representing a statistic of a characteristic word of each text data for similar search may be stored as an index for similar search.

  The issue ontology DB 115 is a database in which issues that are frequently discussed in debates and related words are linked. As an example of the issue ontology here, the value of an issue that is often discussed in debates, an expression that promotes the value, and an expression that suppresses the value are associated. For example, there is health as an example of the value of an issue. This means that at times debates let you fight as an issue in terms of whether the value of health increases or decreases. Here, for example, there are excise, doctor, organ donation, and medicine as expressions that promote the value of health. Examples of expressions that suppress the value of health include junk food, tabacco, alchohol, and smoke. The issue ontology is manually created with reference to past debates, for example. In addition, the issue point ontology DB 115 holds issue point ontology information 701 (see FIG. 8).

<Specific Examples of Sentence Extracting Unit 106 and Sentence Rearranging Unit 107>
FIG. 2 is an explanatory diagram showing a specific example of the sentence extraction unit 106 and the sentence rearrangement unit 107 shown in FIG. The sentence extraction unit 106 identifies the sentences 200 a to 200 d related to the dispute recognized by the dispute determination unit 105 from the text data 13 a to 13 d that are related articles searched by the search unit 104. In the sentences 200a to 200d, underlined phrases such as “health” and “taxation” are issues recognized by the issue determination unit 105, and a bolded phrase such as “junk food” is a theme expression.

  The sentence rearrangement unit 107 generates the statement statement 201 by rearranging the plurality of sentences 200 a to 200 d specified by the sentence extraction unit 106. The sentence rearrangement unit 107 identifies sentence types (claims, reasons, and examples) of the sentences 200a to 200d. In this example, it is assumed that the sentence 200a is identified as a case, the sentence 200b is identified as a sentence type, the sentence 200c is asserted, and the sentence 200d is identified as a reason. Since rearrangement in the order of assertion → reason → examples results in a consistent sentence, the sentence rearrangement unit 107 rearranges the sentence types of the sentences 200a to 200d in the order of claim → reason → example, A statement statement 201 is generated.

  As a result, the generation system 100 can present the statement statements 201 to the user in an order in which the contents are consistent, and the subsequent interpretation of the data becomes easier, and the reliability of the interpretation is improved. Can do.

<System configuration example>
FIG. 3 is an explanatory diagram showing a system configuration example of the document processing system. The document processing system 300 is a system that constitutes a part of the intelligent processing system 1 shown in FIG. The document processing system 300 includes one or more computers 301 and a terminal 14. The computer 301 and the terminal 14 are connected by a network (not shown). The network is a LAN (Local Area Network) or a WAN (Wide Area Network), the Internet, or the like. Further, the computer 301 can access the document DB 302. The document DB 302 corresponds to the text data DB 112 shown in FIG. The computer 301 classifies a plurality of sentences in the document DB 302 and collects them into a plurality of sentence sets 303-1 to 303-n. The terminal 14 selects any sentence set 303-i and gives it to the computer 301.

<Example of Hardware Configuration of Computer 301>
FIG. 4 is a block diagram illustrating a hardware configuration example of the computer 301 illustrated in FIG. The computer 301 includes a processor 410, a memory 420, a local file system 430, an input device 440, an output device 450, a network device 460, and a bus 470. The processor 410 to the network device 460 are connected by a bus 470. The local file system 430 refers to a rewritable storage area that is built in or externally connected to the computer 301. Specifically, a storage device such as a hard disk drive (HDD), a solid state drive (SSD), or a RAM disk. It is.

  The input device 440 is an interface that receives input from the user, such as a keyboard and mouse, or a voice input device. The output device 450 is an interface that outputs the execution results of programs such as a display device and a printer or a speech synthesizer in a form that the user can read or hear. When operating remotely via a network, the input device 440 and the output device 450 can be omitted.

  FIG. 5 is an explanatory diagram showing an example of the contents stored in the document DB 302 shown in FIG. The document DB 302 is a database that stores a document divided into sentences. The document is, for example, the text data 13 shown in FIG. The document DB 302 includes a document ID field 501, a sentence ID field 502, and a body text field 503, and each field has a value in each field 501 to 503.

  A document ID field 501 is an area for storing a document ID as a value. The document ID is identification information that uniquely identifies a document. In this example, the document ID is a character string obtained by concatenating the newspaper name, date, and page number.

  The sentence ID field 502 is an area for storing a sentence ID. The sentence ID is identification information that uniquely identifies the position of the sentence in the document identified by the document ID. In this example, the order of appearance from the top position of the document is a sentence ID.

  The text field 503 is an area for storing a sentence existing at the position specified by the sentence ID. The document DB 302 reads a corresponding sentence by specifying a document ID and a sentence ID. When only the document ID is specified and no sentence ID is specified, all sentences having the same document ID are read. Further, if such a sentence or document acquisition method can be realized, the document may be stored in another data structure.

  FIG. 6 is an explanatory diagram of an example of the contents stored in the sentence set 303-i. Similarly to the document DB 302, the sentence set 303-i has a document ID field 501, a sentence ID field 502, and a body field 503, and each sentence has a value in each of the fields 501 to 503.

<Example of Functional Configuration of Document Processing System 300>
FIG. 7 is a block diagram illustrating a functional configuration example of the document processing system 300. The computer 301 includes an input unit 710, a search unit 711, and a classification unit 712. The computer 301 also includes an acquisition unit 720, an analysis unit 721, a sentence type determination unit 722, a template selection unit 723, a rearrangement processing unit 724, a calculation unit 725, a validity determination unit 726, and an output unit. 727 and a generation unit 728. Specifically, these functions are realized by causing the processor 410 to execute a program stored in the memory 420, for example. Further, the computer 301 stores issue ontology information 701, a template DB 702, an issue chain model 703, and a training document DB 704. Specifically, these are realized by, for example, KVS (Key Value Store), a relational database, or the local file system 430 shown in FIG.

(Pre-processing)
First, pre-processing will be described. The pre-processing is processing for classifying sentences existing in a document in the document DB 302 into a plurality of sentence sets 303-1 to 303-n. Specifically, the pre-processing corresponds to, for example, the input unit 102, the agenda analysis unit 103, and the search unit 104 illustrated in FIG. The input unit 102, the agenda analysis unit 103, and the search unit 104 are executed in cooperation with the data management system 101. In the pre-processing, an input unit 710, a search unit 711, and a classification unit 712 are used.

  The input unit 710 receives an input of a word that is a theme expression from the terminal 14. For example, the input unit 710 receives an input of a word indicating a theme expression such as “junk food”.

  The search unit 711 accepts a query input from the input unit 710. The query is a theme expression such as “junk food” described above. The search unit 711 searches the document DB 302 for a sentence related to the query. A sentence related to a query is a sentence including a character string of a theme expression that is a query.

  The classification unit 712 regroups sentences including various topics and contexts as search results into sentences of similar topics. Thereby, a sentence set 303-i corresponding to the topic is generated. When there are a plurality of topics as search results, sentence sets 303-1 to 303-n corresponding to the number n of topics are obtained.

  The classification unit 712 classifies sentences using machine learning, for example. As an example of machine learning, there is a k-means method using a Bag-of-words expression. The Bag-of-words expression is a method for numerically expressing a text only by information on how many words are included in the text. In the Bag-of-words expression, a vector having dimensions corresponding to the number of words in the dictionary is generated. In each dimension, the appearance frequency of the corresponding word is recorded.

  In the k-means method, texts are clustered into a predetermined number of clusters by comparing similarities between texts using a numerical vector of bag-of-words. As another example of machine learning, sentences can be classified by a method called a topic model such as latent semantic analysis, probabilistic latent semantic analysis, and latent Dirichlet distribution method.

(Sort process)
Next, the rearrangement process will be described. The rearrangement process is a process for outputting a consistent statement statement by rearranging a plurality of sentences. Specifically, the rearrangement process corresponds to, for example, the dispute determination unit 105 to the sentence rearrangement unit 107 shown in FIG. The issue determination units 105 to 107 are executed in cooperation with the data management system 101. In the rearrangement process, the acquisition unit 720, the analysis unit 721, the sentence type determination unit 722, the template selection unit 723, the rearrangement processing unit 724, the calculation unit 725, the validity determination unit 726, and the output unit 727. The generation unit 728, the issue point ontology information 701, the template DB 702, the issue chain model 703, and the training document DB 704 are used.

  The acquisition unit 720 receives an input of a character string as a theme from the terminal 14. The theme entered here is, for example, “Junk food is bad for your health. It is a sentence indicating a theme such as. Further, the acquisition unit 720 acquires any sentence set 303-i of the sentence set 303-i from the terminal 14.

  The analysis unit 721 uses the theme expression specifying unit 731 to specify the theme expression from the input theme. Specifically, the theme expression specifying unit 731 specifies the theme expression from the theme, for example, by a method based on the above-described rules or a method based on machine learning. For example, the theme is “Junk food is bad for your health. ], “Junk food” is specified as the theme expression. In the case of the rule-based method, a dictionary table for theme expression is used. The dictionary table stores words indicating theme expressions. A word indicating the theme expression may be stored for each topic that is a classification unit of the sentence set 303-i. In this case, the word corresponding to the topic of the sentence set 303-i acquired by the acquisition unit 720 is the target of the theme expression.

  In the rearrangement process, the theme expression specifying unit 731 causes the sentence set 303 when there is a word that matches the theme expression specified from the theme in each sentence of the sentence set 303-i acquired by the acquiring unit 720. A tag is added as an annotation to the word in the sentence i. Specifically, for example, in the case where the theme expression is “junk food”, “junk food is a food that lacks nutritional balance significantly. "Junk food" matches the theme expression. Thus, “<theme-meid“ health ”> junk food </ theme> is significantly lacking in nutritional balance. The theme expression identification unit 731 adds a tag indicating the theme tag to “junk food” as an annotation. The theme-id “health” is an attribute value and is a topic corresponding to the theme expression.

  In the case of machine learning, the theme expression specifying unit 731 also specifies specific words and phrases related to the theme expression such as a higher concept and a lower concept. For example, when the theme expression is “junk food”, “food” as a higher concept and “fast food” and “soft drink” as lower concepts are specified from each sentence in the sentence set 303-i.

  The analysis unit 721 uses the issue expression specifying unit 732 to specify the issue expression from each sentence of the acquired sentence set 303-i. Specifically, the issue expression specifying unit 732 specifies an issue expression from each sentence of the sentence set 303-i by, for example, the above-described rule-based method or machine learning-based method. The issue expression specifying unit 732 uses a dictionary table prepared in advance. The dictionary table stores words that are points of dispute. As the dictionary table, for example, dispute point ontology information 701 is used.

  FIG. 8 is an explanatory diagram showing an example of the data structure of the dispute point ontology information 701. The issue ontology information 701 includes an issue field 801 and an issue expression field 802, and indicates a correspondence between an issue and an issue expression for each entry. The issue field 801 is an area for storing issues. The issue expression field 802 stores an issue expression that is a subordinate concept of the issue. The issue expression specifying unit 732 specifies a word that matches the issue expression of the issue ontology information 701 as an issue expression from each sentence of the sentence set 303-i.

  When a method based on machine learning is applied, the point-of-interest specifying unit 732 specifies, for example, a word / phrase corresponding to the point-of-interest expression from each sentence of the sentence set 303-i using a learning model.

  In addition, when there is a word that matches the specified issue expression in each sentence of the sentence set 303-i acquired by the acquisition unit 720, the issue expression specifying unit 732 adds a tag as an annotation to the word. Specifically, for example, when “dangerous drug” is a disputed expression, “Japan prohibited dangerous drugs. "Japan has banned <issue issue-id" health "> dangerous drug </ issue>. ], The issue expression specifying unit 732 assigns the issue tag to “dangerous drug” as an annotation. The issue-id “health” is an attribute value and is an issue corresponding to the issue expression.

  As will be described later, the issue tag may be assigned when the generation unit 728 generates the issue chain model 703. Therefore, in this case, the issue expression specifying unit 732 may simply specify the character string sandwiched between the assigned issue tags as the issue expression.

  The analysis unit 721 uses the affirmation / negative expression specifying unit 733 to specify, from each sentence of the theme or sentence set 303-i, an affirmative expression that affirms the theme expression or issue expression or a negative expression that denies the theme expression or issue expression. . Specifically, the affirmation / negative expression specifying unit 733 specifies affirmation / negative expression from each sentence of the theme or sentence set 303-i, for example, by a method based on the above-described rules or a method based on machine learning. The positive / negative expression specifying unit 733 uses a dictionary table prepared in advance. In the dictionary table, words indicating positive expressions and negative expressions are stored in association with attributes of positive or negative. Specifically, for example, when there is a disputed expression in a certain sentence, the affirmation / negative expression specifying unit 733 specifies a word / phrase that is a destination of the disputed expression and searches the dictionary table for the specified word / phrase. The positive / negative expression specifying unit 733 specifies the attribute (positive or negative) of the word / phrase when the matching word / phrase exists in the dictionary table.

  For example, “Junk food is bad for <issue issue-id“ health ”> health </ issue>. In the case of the sentence, “health” is the issue. If “bad”, which is a “health” dependency, is stored in the dictionary table with a negative attribute, the affirmative / negative expression specifying unit 733 specifies “bad” as a negative expression.

  In addition, when a method based on machine learning is applied, the affirmation / negative expression specifying unit 733, for example, by using a learning model, a phrase (for example, Identify “prohibited”).

  The affirmative / negative expression specifying unit 733 says, “Junk food is bad for <issue issue-id“ health ”> health </ issue>. To the sentence, "Junk food is <issue issue-id" health "> health </ issue> <negative> bad </ negative>. "

  The analysis unit 721 uses the promotion / suppression expression specifying unit 734 to specify whether the theme expression of each sentence in the sentence set 303-i corresponds to the promotion expression or the suppression expression. Specifically, the promotion / suppression expression specifying unit 734 specifies the promotion / suppression expression for the theme expression from each sentence of the theme or sentence set 303-i by, for example, the above-described rule-based method or machine learning-based method. To do. The promotion / suppression expression specifying unit 734 uses a dictionary table prepared in advance. In the dictionary table, words indicating the promotion expression and the suppression expression are stored in association with the attribute of promotion or suppression. Specifically, for example, when a theme expression exists in a certain sentence, the promotion / suppression expression specifying unit 734 specifies a predicate of the theme expression, and suppresses whether the phrase included in the predicate is a promotion expression. Whether it is an expression is determined by referring to the dictionary table.

  For example, “<the theme-id“ health ”> junk food </ theme> is significantly lacking in nutritional balance. In the case of the sentence, the predicate of the theme expression “junk food” is “significantly lacking nutritional balance”. In this case, if “missing” is stored as the suppression expression in the dictionary table, the relationship that the theme expression “junk food” suppresses “nutrition balance” can be specified.

  Then, the promotion / suppression expression specifying unit 734 indicates that “<theme-me-id“ health ”> junk food </ theme> significantly lacks nutritional balance. ”And annotated,“ <theme-meid ”health> junk food </ theme> has a marked nutritional balance <suppress> </ suppress>. "

  The analysis unit 721 uses the authority specifying unit 735 to specify a word indicating authority from each sentence in the sentence set 303-i. The word indicating authority indicates the subject of the sentence, such as an organization name or a person name. The authority specifying unit 735 uses a dictionary table in which an organization name and a person's name (or a suffix such as “Meeting” or “Mr.” attached to the end of the person's name) and an attribute indicating authority are associated with each other. The authority specifying unit 735 specifies a word indicating authority from each sentence in the sentence set 303-i with reference to the dictionary table.

  For example, “We are considering a new tax on fast food and soft drinks,” said Congressman A. ], The authority specifying unit 735 specifies “person A” as the authority. More specifically, instead of simply specifying an organization name or person name, if the predicate of the organization name or person name is a verb indicating that he / she has stated, such as “stated” or “shown”, it is specified as an authority. It is good as well. In this case, the authority specifying unit 735 adds an annotation to a word indicating authority. For example, “We are considering a new tax on fast food and soft drinks,” said Congressman A. "" I am considering a new tax on fast food and soft drinks, "said <authority> A legislator </ authority>. It becomes.

  FIG. 9 is an explanatory diagram illustrating Example 1 of the sentence analysis result. The sentence analysis result 900 is expressed in a table format as an example. The sentence analysis result 900 includes an item number field 901, a sentence field 902, an issue field 903, an affirmation / denial field 904, a promotion / suppression field 905, an authority field 906, and a sentence type field 907. The combination of the values of the fields 901 to 907 constitutes one sentence entry. Since the sentence is given a tag as an annotation in the analysis unit 721, the issue point, the positive expression, the negative expression, the promotion expression, the suppression expression, and the authority in the sentence can be specified.

  The item number field 901 is an area for storing an item number. The item number is identification information that uniquely identifies the sentence analyzed by the analysis unit 721. A sentence field 902 is an area for storing a sentence analyzed by the analysis unit 721. The issue field 903 is a shield for storing an issue included in the sentence in the sentence field 902. In the case where the issue expression is described instead of the issue in the sentence, the issue that is the attribute (issue-id) of the issue tag is stored.

  The affirmation / denial field 904 is an area for storing information indicating the presence or absence of an affirmative expression or a negative expression for the issue in the sentence of the sentence field 902. “Yes” is stored when a positive expression (positive tag) is present, “Negative” is stored when a negative expression (negative tag) is present, and “None” is stored when neither is present.

  The promotion / suppression field 905 is an area for storing information indicating the presence / absence of the promotion expression and the suppression expression for the theme expression in the sentence in the sentence field 902. When the theme expression is “promoting” a certain object, “promotion” is stored, and when the theme expression is “promoted” by a target object, “promoted” is stored. The same applies to suppression and suppression.

  The sentence type field 907 is an area for storing the sentence type of the sentence in the sentence field 902. The sentence type that is the value of the sentence type field 907 is stored by the sentence type determination unit 722.

  Returning to FIG. 7, the sentence type determination unit 722 determines the sentence type of each sentence analyzed by the analysis unit 721. Sentence types are, for example, types of sentences such as claims, reasons, examples, current explanations, importance of claims, and quotations. The sentence type is determined based on the rhetorical characteristics of the sentence. For example, it is set from the viewpoint of in which order the rearranged sentences are arranged. The sentence type is used in a template described later. Because of this characteristic, the setting of the sentence type depends on the statement sentence to be generated.

  The sentence type determination unit 722 determines a sentence type using a rule-based method or a machine learning-based method. In the case of a rule-based method, the sentence type determination unit 722 uses a dictionary table. Appropriate rules are associated with each dictionary type in the dictionary table. For example, in the case of the sentence type of assertion, a sentence that states affirmation or denial of the theme expression can be used as the assertion sentence. Therefore, the rule “sentence is asserted as a sentence whose theme expression is affirmative or negative” is set. In addition, for example, a rule may be set that “a sentence that is a content that promotes a negative expression for a dispute point is a sentence type: a reason”. In addition, a rule may be set that “a sentence type: example is a sentence in which there is no positive expression or negative expression for an issue, no promotion expression or suppression expression for a theme expression, and there is a word indicating authority”.

  In the case of a method based on machine learning, a sentence that is manually distributed to each sentence type in advance is prepared as training data, and the computer 301 creates a sentence type determination model. Specifically, the sentence type determination unit 722 gives the training data to the analysis unit 721, and generates a numerical vector with the theme expression, the dispute point, the positive / negative expression, and the promotion / suppression relationship for the training data. The computer 301 creates a sentence type determination model for estimating the sentence type corresponding to the numerical vector by giving this numerical vector to SVM or logistic regression.

  The sentence type determination unit 722 generates, for each sentence in the sentence set 303-i, a numeric vector characterized by the theme expression, issue, affirmation / negative expression, and promotion / suppression relation identified by the analysis unit 721, as in the training data. To do. The sentence type determination unit 722 gives the generated numerical vector to the sentence type determination model, and determines the corresponding sentence type.

  FIG. 10 is an explanatory diagram illustrating Example 2 of the sentence analysis result 900. FIG. 10 shows a state in which the sentence type that is the value of the sentence type field 907 is stored in the sentence analysis result 900 of FIG.

  Returning to FIG. 7, the template selection unit 723 selects one of the templates from the template DB 702. The template is information that specifies the rhetorical arrangement of sentences. For example, the template defines a formal order that is common in many cases: first assertion, then reason, and finally introduce the case. The template has multiple slots. A slot means a blank in which a sentence should be placed. One sentence type is designated in one slot. When entering a sentence, only the sentence that matches the sentence type can be input.

  Each slot has properties of importance and necessity. The importance indicates the relative importance between slots. For major slots that change the meaning of the sentence significantly, such as an assertion or reason, specify a high importance level, and even if there is no sentence like the second or third case, the argument is large. For slots that do not change, the importance is set low. In addition, in the property indicating whether it is essential, when the slot is not filled, it can be set that the output document is meaningless.

  FIG. 11 is an explanatory diagram illustrating an example of a data structure of a template. In FIG. 11, two types of templates T1 and T2 (hereinafter collectively referred to as templates T) are shown as an example. The template T is expressed in a table format as an example. The template T is information that is manually set in advance.

  The template T includes an order field 1101, a sentence type field 1102, an importance field 1103, a required field 1104, and a slot 1105. The order field 1101 is an area for storing values indicating the order of applied sentences in ascending order. The sentences are rearranged in the order indicated by the order value.

  The sentence type field 1102 is an area for storing the sentence type of the sentence to be applied. The importance field 1103 is an area for storing the importance of sentence types. The importance is used for calculation of an evaluation value in a calculation unit 725 described later. The mandatory field 1104 is an area for storing information indicating whether or not a sentence type is essential. When the sentence type is essential, “True” is stored, and when it is not essential, “False” is stored. A slot 1105 is an area in which a sentence corresponding to each entry is embedded.

  The template T1 is a template T for arranging a maximum of three sentences. The sentence type of the first sentence is “claim”, the sentence type of the second sentence is “reason”, and the sentence type of the last sentence is “ It is rearranged to become “example”. Since the value of the required field 1104 of the sentence type “example” of the last sentence is “False”, the last sentence may be omitted.

  The template T2 is a template T for arranging a maximum of five sentences. The first sentence sentence type is “Claim”, the second sentence sentence type is “Current Status Description”, and the third sentence sentence type. Are rearranged so that the sentence type of the fourth and fifth sentences becomes “quote”. Since the value of the second to fifth sentence required fields 1104 is “False”, it may be omitted.

  In addition, when the feature of the template T to be selected in advance is set by the user, the template selection unit 723 selects only the template T corresponding to the feature. For example, it is possible to freely set only the template T in which all sentence type claims, reasons, and examples are set to “True”.

  The rearrangement processing unit 724 includes a combination processing unit 741 and a dispute point selection unit 742. The combination processing unit 741 embeds a sentence in each slot 1105 of the selected template T. The combination processing unit 741 calls the issue selection unit 742 when filling the slot 1105.

  The dispute point selection unit 742 selects a sentence including the dispute point from the sentence set 303-i using the sentence sequence embedded in the slot 1105, the template T, and the sentence set 303-i. The issue selection unit 742 refers to the issue chain model 703 and selects a sentence including the issue from the sentence set 303-i.

  FIG. 12 is an explanatory diagram showing an example of the data structure of the issue chain model 703. The issue chain model 703 is structured data in which continuity about the contents of a sentence is arranged. The issue chain model 703 is structured data in which the content continuous state is organized. The issue chain model 703 is used to infer what is likely to be taken up in the next sentence when the issue A is taken up in one sentence. The issue chain model 703 is generated by the generation unit 728. Details of the generation processing by the generation unit 728 will be described later.

  The issue chain model 703 has an issue field 1201 of the preceding sentence, an issue field 1202 of the succeeding sentence, a location information field 1203, and a frequency field 1204. The issue chain is formed by combining the values of the fields 1201 to 1204. Specify the entry. Here, the preceding sentence and the succeeding sentence are two consecutive sentences, and the succeeding sentence appears after the preceding sentence.

  The preceding sentence issue field 1201 is an area for storing the issue of the preceding sentence. The issue of the preceding sentence is an issue included in the preceding sentence.

  The subsequent sentence issue field 1202 is an area for storing the subsequent sentence issue. The issue of the succeeding sentence is the issue included in the succeeding sentence.

  The location information field 1203 is an area for storing location information of issues in the preceding sentence and issues in the subsequent sentence. The location information is composed of a combination of a document ID and a sentence ID. The document ID is identification information that uniquely identifies a document in which a preceding sentence and a succeeding sentence exist. The sentence ID is identification information that uniquely identifies the preceding sentence and the succeeding sentence in the document. The sentence ID is, for example, the order of appearance of the preceding sentence (or succeeding sentence) in the document with the document ID. When there are a plurality of identical combinations of issues in the preceding sentence and issues in the subsequent sentence, location information is stored for each combination.

  The frequency field 1204 is an area for storing the frequency. The frequency is the number of times a combination of the previous sentence issue and the subsequent sentence issue appears. Specifically, for example, it is the number of combinations stored in the location information. The frequency may be expressed as a probability. When expressed as a probability, the value of the sum of the values in the frequency field 1204 of the issue chain model 703 as the denominator and the number of location information of each entry as the numerator is the probability of each entry.

  The entry on the first line of the issue chain model 703 shows that the issue “Health” of the preceding sentence and the issue “Tax” of the succeeding sentence are the 10th page of the 1st page of the Y newspaper, year Y, month Z, in the population. It appears in the 42nd sentence (preceding sentence) of the 8th page of the N newspaper P year Q month R day. Further, since the number of appearances is 2, the frequency is “2”.

  FIG. 13 is an explanatory diagram showing a chain of dispute points. In FIG. 13, two sentences are given as examples. Sentences 1301 and 1302 are successive sentences of the same document. Sentence 1301 is a sentence “Junk food may be harmful to health.” Sentence 1302 is “The government is considering taxing junk food for health promotion measures”. It is a sentence.

  The issue expression of sentence 1301 is “health”, and the issue expression of sentence 2 is “health” and “taxation”. If this is mapped to the issue, the issue of the sentence 1301 is “health”, and the issue of the sentence 1302 is “health” and “tax”. In this example, since “tax” is used as a means for promoting “health”, it can be seen that the topic has changed from “health” to “tax” related to health. Since “health” and “tax” are far from each other in terms of meaning, it is difficult to model this change of topic in terms of similarity in meaning. The issue chain model 703 is a model for directly using topic change to rearrange sentences.

  FIG. 14 is an explanatory diagram illustrating an example of use of the issue chain model 703. In FIG. 14, a usage example of the issue chain model 703 will be described using a state transition diagram. The circle in the state transition diagram is a node and indicates a point of issue. The arrows in the state transition diagram are edges, which indicate transitions between nodes and their weights. The weight is the frequency shown in FIG. Sentence 1 to sentence N are sentences selected from the sentence set 303-i, and the combination processing unit 741 selects and selects sentence 1, sentence 2,..., Sentence N in this order from the start end to the end end. It is added to the slot 1105 of the template T. N is the number of slots 1105 of the selected template T.

  For example, in sentence 1, it is assumed that a sentence having “health” as an issue is selected. In addition, when selecting the sentence 1, the dispute point selection unit 742 may sequentially select the sentence as the sentence 1 from the sentence set 303-i, or may select only the sentence including a specific issue point. The specific issue may be an issue designated in advance by the user, or may be an issue whose number of occurrences is a predetermined number or more in the sentence set 303-i. When the “health” of the sentence 1 is selected, the issue selection unit 742 identifies an entry whose value of the issue field 801 of the preceding sentence is “healthy” in the issue chain model 703.

  When selecting sentence 2, issue point selection unit 742 selects the next issue point by following the edge from the “healthy” node of statement 1. Specifically, the issue selection unit 742 identifies the value of the issue field 1202 of the subsequent sentence with reference to the entry whose value of the issue field 1201 of the preceding sentence is “healthy”. When searching with brute force, the issue point selection unit 742 specifies the value of the issue field 1202 of the subsequent sentence that is the transition destination, and searches to the end point.

  Further, the issue selection unit 742 may select the issue of the subsequent sentence having a specific frequency with reference to the value of the frequency field 1204. The specific frequency may be a maximum value or a frequency equal to or higher than a predetermined value among the frequencies from the issue point of the preceding sentence as the transition source to the issue point of the subsequent sentence as the transition destination. When selecting the maximum value, for example, dynamic programming is used. The predetermined value may be an average value or a median value of the values in the frequency field 1204, or may be a value arbitrarily set by an administrator. As a result, priority is given to combinations of sentences that are likely to cause transition of dispute points.

  In this way, the issue selection unit 742 selects the issue of the i (i = 1 to N−1) th sentence and the i + 1th issue. In addition, since the number of sentences is limited, there may be a case where there are no transition destination nodes that match on the way. In that case, the dispute point selection unit 742 can set the transition destination node as a passage prohibition node, backtrack, and continue the search again.

  The calculation unit 725 calculates an evaluation value for evaluating the sequence of sentence combinations embedded in the slot 1105 of the selected template T by the combination processing unit 741. As a viewpoint of evaluation, for example, the satisfaction status of the selection template T can be cited. The evaluation value regarding the satisfaction status of the selected template T is a value indicating how many slots 1105 of the selected template T are filled. For example, when 2 out of 3 are filled, it becomes 2/3. In addition, since importance is associated with each slot 1105, the calculation unit may obtain a numerator value by a weighted linear sum with the importance as a weight. For example, when all the slots 1105 of the template T1 are filled, the evaluation value regarding the satisfaction status of the selected template T1 is not 3/3 but (1 × 1 + 1 × 1 + 1 × 0.8) /3=2.8. / 3.

  If the slot 1105 of the required field 1104 of the selection template T is empty, the calculation unit 725 sets an insufficient flag for the sentence combination. The calculation unit 725 may set the insufficient flag without calculating the evaluation value related to the satisfaction status of the selected template T.

  Further, as an evaluation viewpoint, an evaluation value indicating the relevance between adjacent sentences embedded in the selected template T can be cited. The evaluation value indicating relevance is the sum of the frequencies from the issue of the preceding sentence that is the transition source to the issue of the subsequent sentence that is the transition destination.

  Further, when the insufficient flag is not set, the calculation unit 725 may calculate a weighted linear sum of the evaluation value regarding the satisfaction status and the evaluation value indicating the relevance as the evaluation value. The weight can be arbitrarily set by the administrator.

  The validity determination unit 726 determines whether to end the rearrangement process based on the evaluation value calculated by the calculation unit 725. That is, the validity determination unit 726 determines the validity of the combination of sentences embedded in the selection template T to be determined. Specifically, for example, the validity determination unit 726 determines whether or not an insufficient flag is set for a combination of sentences embedded in the selection template T to be determined. If the insufficient flag is set, it is necessary to reselect the template T, and it is determined that the rearrangement process is not terminated.

  In addition, the validity determination unit 726 ends the rearrangement process based on the evaluation value when the insufficient flag is not set for the combination of sentences embedded in the selection template T to be determined. It is determined whether or not. Specifically, for example, if the evaluation value related to the satisfaction status of the selected template T is equal to or greater than the threshold value, the validity determination unit 726 determines that the rearrangement process has ended, assuming that the chain of issues is consistent. To do.

  In addition, when the evaluation value indicating the relevance is equal to or greater than the threshold value, the validity determination unit 726 determines that the rearrangement process is finished, assuming that the chain of issues is consistent. The validity determination unit 726 performs the end determination using at least one of the evaluation value regarding the satisfaction status and the evaluation value indicating the relevance.

  When the evaluation value to be determined is equal to or greater than the threshold value, the validity determination unit 726 may determine that the rearrangement process has ended, assuming that the chain of issues is consistent. When both evaluation values are used, if it is determined that at least one of the two evaluation values should not be terminated, the validity determination unit 726 determines that the chain of issues is inconsistent and performs the reordering process. You may determine not to complete | finish. In addition, the validity determination unit 726 may determine that the rearrangement process is finished, assuming that the chain of issues is consistent when the weighted linear sum of both evaluation values is equal to or greater than the threshold value. In any case, if it is determined not to end, the template selection unit 723 selects an unselected template T.

  The output unit 727 outputs the determination result by the validity determination unit 726. Specifically, for example, a combination of sentences embedded in the selection template T, that is, a statement sentence is output. As an output method, for example, output to an output device (display on a display, print output on a printer, voice output by voice synthesis), transmission to another device (for example, the terminal 14) via a network device, local Includes storage in the file system.

  In the template selection unit 723, when there is no template T to be selected, for example, when there is no unselected template T in the template DB 702, or when the number of selections exceeds a predetermined number, there is no corresponding statement. , Output to that effect.

  FIG. 15 is an explanatory diagram illustrating an example of output information from the output unit 727. The output information 1500 is information indicating the result of rearranging the analysis result of FIG. 10 by the combination processing unit 741. Output information 1500 in FIG. 15 is an example in which the template T1 in FIG. 11 is applied to the analysis result in FIG.

  Returning to FIG. 7, the generation unit 728 generates the issue chain model 703 using the training document DB 704. A specific configuration example of the generation unit 728 will be described with reference to the drawings.

  FIG. 16 is a block diagram illustrating a detailed functional configuration example of the generation unit 728. The generation unit 728 includes an issue extraction model generation unit 1601, an issue identification model generation unit 1602, an issue annotation assignment unit 1603, and an issue chain analysis unit 1604.

  The training document DB 703 is a database that stores training documents. The training document is a document in which an annotation is added to the issue expression in the document manually. Here, “annotation” means an operation of recording a start character position and an end character position of a portion corresponding to the point expression. For example, annotating is an operation of enclosing the issue expression with issue tags such as <issue> to </ issue> using, for example, an XML (Extensible Markup Language) notation.

  As the annotation attribute, the issue point to which the issue expression belongs is recorded as the attribute of the issue tag, such as <issue issue-id = “health”> to </ issue>. Note that the attribute value “health” needs to be uniquely distinguishable between issues, and in practice, a unique name is given according to an identifier format such as a URI (Uniform Resource Identifier).

  The issue extraction model generation unit 1601 generates an issue extraction model 1610. The issue extraction model 1610 is a rule or pattern for extracting a part as an issue expression if the partial character string in the sentence in the training document is an issue issue expression. Specifically, for example, the dispute point extraction model generation unit 1601 refers to the dispute point ontology information 701, and if there is a character string that matches the dispute point expression of the dispute point ontology information 701, adds the issue tag as an annotation. In this case, the issue tag attribute is set to the issue point of the issue point ontology information 701.

  For example, when a character string “Junk food” appears in a sentence in the training document, the issue extraction model generation unit 1601 adds an annotation to “Junk food” and sets <issue issue-id = “health”> junk. “Food </ issue>”. The issue-id “health” is the issue of the issue expression “junk food” in the issue ontology information 701. When the issue chain model 703 is generated by a rule-based method, the issue extraction model 1610 is a set of training documents including annotated issue expressions. Although not shown, the issue extraction model generation unit 1601 associates the location information of the issue expression (the document ID of the training document and the sentence ID of the sentence including the issue expression) with the issue expression with the annotation.

  In the case of a method based on machine learning, the dispute point extraction model generation unit 1601 converts a character string in the training document including the above-described annotation expression with an annotation to a series of numerical vectors by feature extraction. First, the issue extraction model generation unit 1601 performs morphological analysis on the sentence. Each morpheme is set as a series node. Next, the point extraction model generation unit 1601 lists the surface layer character strings and parts of speech of each morpheme. For example, suppose that there are a total of Vs types of surface character strings and a total of Vp types of parts of speech in the training document.

  The dispute point extraction model generation unit 1601 adds each surface layer character string and each part of speech as a dictionary table key in order, and assigns natural numbers one by one so as not to overlap each other in the order of registration in the dictionary table. The dispute point extraction model generation unit 1601 converts the surface character string and part of speech of each node into numerical values using the dictionary table. Then, each node is expressed as a numeric vector in which the dimension corresponding to the surface character string and the part of speech is “1” and the other dimensions are “0”.

  This numeric vector is paired with a label indicating whether or not it is an issue. The issue extraction model generation unit 1601 gives an issue label if the node is an issue (if an annotation is given), and an O (other) label if the node is not an issue. As another way of giving a label, assuming a compound word issue, a B-issue label is assigned to the node at the start of the issue, an I-issue label is assigned to the node at the intermediate location, and so on. There is also a method for distinguishing issue labels. As a variant, there is also a method of distinguishing the end position as an E-issue label and the issue of a single node as a U-issue label. The dispute point extraction model generation unit 1601 inputs this numerical vector string and label set to SVM or CRF to learn. The model output as a result of this machine learning is the issue extraction model 1610. By applying the issue extraction model 1610, it is possible to specify the issue expression to which the given character string belongs.

  The issue identification model generation unit 1602 generates an issue identification model 1620. The issue identification model 1620 is a model for associating which entry in the issue ontology information 701 the target issue expression corresponds to. As a specific method, first, as a simple case, the issue identification model generation unit 1602 determines that a target issue expression appears only in one entry in the issue ontology information 701, Generate a rule to associate. Next, when the target issue expression appears in a plurality of entries in the issue ontology information 701, the issue identification model generation unit 1602 resolves the ambiguity from the surrounding context and associates it with one of the entries. Generate rules. Specifically, for example, it is a problem similar to the ambiguity resolution of natural language processing, entity identification, and bibliographic identification, and ambiguity can be resolved by using these known methods. A set of these rules is the issue identification model 1620.

  The issue annotation assigning unit 1603 uses the issue extraction model 1610 to extract the issue expression from the sentences in the document stored in the document DB 302. Specifically, for example, the issue annotation assigning unit 1603 applies the issue extraction model 1610 to identify the issue expression to which the character string included in the sentence in the document belongs. Next, the issue annotation assigning unit 1603 uses the issue identification model 1620 to associate the specified issue expression with the issue in the issue ontology information 701. Then, the issue annotation assignment unit 1603 assigns an issue tag to the character string corresponding to the issue expression. At this time, the issue annotation giving unit 1603 refers to the issue ontology information 701 and sets the issue associated with the issue expression using the issue identification model 1620 as the issue-ID. As a result, an annotation is added to the document in the document DB 302, and a document group 1630 with a dispute annotation is generated.

  The dispute chain analysis unit 1604 includes a problem that appears in the preceding sentence and a problem that appears in the subsequent sentence as a pair of two sentences (preceding sentence and succeeding sentence) that appear consecutively in each document of the document group 1630 with the annotated point. And the location information is identified. When the dispute point expression is specified from the sentence, the issue chain analysis unit 1604 refers to the issue point ontology information 701 and converts it into an associated issue point using the issue point identification model 1620. The issue chain analysis unit 1604 counts the number of appearances of the same combination. Then, the issue chain analysis unit 1604 adds, to the issue chain model 703, an entry having the location information and the frequency of appearance for each combination of the issue of the preceding sentence and the issue of the subsequent sentence. The issue chain model 703 as shown in FIG.

  Note that the issue chain analysis unit 1604 may identify the combination of the issue expression appearing in the preceding sentence and the issue expression appearing in the subsequent sentence and the location information thereof without collecting the issue expressions as issue points.

  Also, the issue chain analysis unit 1604 does not convert the issue expression that appears in the subsequent sentence into an issue, but uses the combination of the issue sentence and the issue expression that appears in the subsequent sentence as the issue (or issue expression) of the preceding sentence. The associated entry may be retained. In this case, the method of searching for the issue point appearing in the subsequent sentence differs from the issue point appearing in the preceding sentence. In this case, when the issues appearing in the preceding sentence are given in the issue chain model 703, a list of pairs of the succeeding sentence and the issue expression appearing in the following sentence is returned. When determining the issue point appearing in the subsequent sentence using the list, the issue chain analysis unit 1604 samples the combination of the issue expression and the sentence according to a certain probability distribution from the list.

  For example, when sampling the cases according to the uniform distribution, the dispute chain analysis unit 1604 samples the list entries with a probability equal to each other, and returns the issue expression appearing in the selected subsequent sentence. In this case, the issue expression that appears frequently in the succeeding sentence can be easily selected, and therefore, the same effect as when the frequency is summarized. On the other hand, the issue chain analysis unit 1604 may calculate the similarity between the issue expression and the theme expression that appear in each subsequent sentence, and may sample the cases with a probability distribution proportional to the magnitude of the similarity. In this case, the issue chain analysis unit 1604 selects an issue expression (or issue issue) having a high degree of similarity to the theme expression. Thereby, the issue expression (or issue issue) can be selected in consideration of the connection with the theme expression.

<Example of sorting process>
FIG. 17 is a flowchart illustrating an example of a reordering process procedure performed by the computer 301 of the document processing system 300. Prior to the sorting process, the document group in the document DB 302 is classified into a plurality of sentence sets 303-i manually or by the processing of the input unit 710, the search unit 711, and the classification unit 712. It is assumed that an arbitrary sentence set 303-i is selected as a target.

  First, the acquisition unit 720 acquires a sentence set 303-i and a theme to be rearranged (step S1701). Next, the analysis unit 721 specifies the theme expression, issue expression, positive expression, negative expression, promotion expression, suppression expression, and authority for each sentence in the acquired sentence set 303-i (step S1702). . Thereby, the sentence analysis result 900 as shown in FIG. 9 is obtained.

  Then, the sentence type determination unit 722 determines a sentence type for each sentence in the acquired sentence set 303-i (step S1703). Thereby, the sentence analysis result 900 as shown in FIG. 10 is obtained.

  Thereafter, the template selection unit 723 selects an unselected template T from the template DB 702 (step S1704). If the template selection unit 723 cannot select the template T (step S1705: No), an error is output and the process ends (step S1706). For example, when there is no unselected template T in the template DB 702, or when the number of selections exceeds a predetermined number, the template selection unit 723 cannot select the template T and outputs an error.

  On the other hand, when the template selection unit 723 can select the template T (step S1705: Yes), the combination processing unit 741 selects a point of dispute by using the point-of-interest selection unit 742, and selects a sentence including the selected point as a slot of the selected template T. It is embedded in 1105 (step S1707).

  Then, the calculation unit 725 calculates an evaluation value for the selected template T in which the sentence is embedded (step S1708). Thereafter, the validity determination unit 726 determines whether or not to end the rearrangement process based on the evaluation value for the selected template T calculated by the calculation unit 725 (step S1709). When it determines with not complete | finishing (step S1709: No), it returns to step S1704 and the template selection part 723 tries selection of the template T again (step S1704). On the other hand, if it is determined to end (step S1709: Yes), the output unit 727 outputs a combination of sentences embedded in the selected template T that is a determination result by the validity determination unit 726, that is, output information 1500 that is a statement sentence. Is output (step S1710). Thereby, a series of rearrangement processes are completed.

<Application example of rearrangement processing>
Here, an application example of the rearrangement process described above will be described. The rearrangement process can be implemented, for example, in a debate argument system.

  The theme at the debate is a so-called topic. For example, the topic “Junk food should be banned” is cited. The theme expression in the debate indicates a word appearing in the sentence of the topic and its related word. In debate, since there are two positions, the positive side and the negative side, it is important information whether the theme expression is positive or negative in the topic. Therefore, both the theme expression and the positive / negative attribute are treated as theme expressions.

  In addition, since there are two positions in the debate, the positive side and the negative side, it is necessary to extract the speaker when the speaker is specified in the sentence. The speaker is specified by, for example, the authority specifying unit 735 described above.

  In the following, the case where it is applied to the positive argument of debate is given as an example. Originally, in the case of debate argument, a plan, definition of terms, explanation of the current situation, explanation of importance, explanation of resolution, etc. are required as sentence types. In the case of an affirmative position, explain each sentence type.

  A plan is a sentence that represents a specific measure for realizing the topic. For example, a plan that shows specific measures that can be implemented, such as “set the upper limit of annual income to 10 million yen” for the topic “should set an upper limit on income”. It is.

  A term definition is a definition sentence of a word that appears in a topic or a word that is used in an argument. In the above example, the scope of discussion is designated such that “income” represents “income of a person living in Japan”.

  The current situation description represents the current situation where the plan is not adopted. For example, in the case of the above example, there is no upper limit of income, so the sentence describes a demerit such as “a large income bias”.

  When the reordering process is applied to debate affirmative arguments, the computer 301 does not generate all of the affirmative arguments. Select and arrange them in order.

  When the topic “should ban junk food” is given, the theme expression specifying unit 731 specifies “junk food” as the theme expression from the topic. The affirmation / negative expression specifying unit 733 specifies a positive expression or a negative expression from the topic. In this case, since “prohibited” is included in the topic, the affirmative / negative expression specifying unit 733 specifies “prohibited” as the negative expression. Therefore, the positive / negative attribute of the theme expression “junk food” is “negative”.

  In addition, a sentence set 303-i related to the topic is selected by the user as an analysis target. Here, it is assumed that the sentence set 303-i in which the topic is classified as “health” is selected.

  The computer 301 uses the obtaining unit 720 to obtain the sentence set 303-i and junk food denial as the theme expression. Therefore, in the rearrangement process, sentences indicating claims, reasons, and examples that “deny” “junk food” are collected, and the collected sentences are arranged in the order of claims, reasons, and examples.

  The analysis unit 721 performs sentence analysis on each sentence in the sentence set 303-i, and obtains a sentence analysis result 900 illustrated in FIG. In FIG. 9, four sentences are listed. The bold text in the sentence is the theme expression, and the underlined text is the issue expression. The theme expression includes not only the words that appear in the thesis, but also the upper and lower terms, and specific vocabulary. Therefore, “Fast Food” and “Soft Drink” are the theme expressions for the theme expression “Junk Food”. May be handled. Such a synonym is identified by the analysis unit 721 with reference to a synonym dictionary table (not shown). Also, the dispute point ontology information 701 may be used instead of the synonym dictionary table.

  In FIG. 9, in the issue field 903, the abstraction levels of issues appearing in the sentence are aligned. For example, the issue expression “tax” is aligned with the issue “tax” as a general concept, and the issue expression “obesity” is aligned with the issue “health” as a general concept. The analysis unit 721 executes this conversion process using the issue point ontology information 701.

  In the affirmation / denial field 904, the entry of item number 3 and item number 4 is negated. Item No. 3 describes the “risk” of “health” that is the issue. Therefore, the affirmation / negative expression specifying unit 733 determines that “risk”, which is the destination of the issue “health”, is “negative” for the issue “health” in order to suppress the issue “health”. On the other hand, when the subject of the issue “Health” promotes the issue “Health” such as “Merit”, the “Merit” is determined as an affirmative expression of the issue “Health”.

  In the promotion / suppression field 905, the entries of item number 2 and item number 4 respectively indicate suppression and promotion. Here, in the sentence of item number 2, the theme expression “junk food” is described as the subject in the passive state of the suppression expression “regulation”, that is, the object of suppression. Therefore, the value of the promotion / suppression field 905 in the entry of item number 2 is “suppressed”. In the sentence No. 4, the theme expression “junk food” is described as the subject of the active expression of the promotion expression “encouragement”, that is, the action subject of promotion. Therefore, the value of the promotion / suppression field 905 in the entry of item number 4 is “promotion”.

  Also, by executing the sentence type determination process of the sentence type determination unit 722, the sentence type is registered in the sentence type field 907 of FIG. 9, and the sentence analysis result 900 shown in FIG. 10 is obtained. Here, it is assumed that the entries of item number 1 and item number 2 are determined to be “example”, the item number 3 is determined to be “claim”, and the entry of item number 4 is determined to be “reason”.

  The sentence in No. 3 is an “assertion” because it expresses the opinion that the theme expression is denied. The sentence of No. 4 is the reason why the theme expression “junk food” promotes the negative expression of the issue expression “obesity”, that is, explains the adverse effects of the theme expression “junk food”. Lastly, the sentences of No. 1 and No. 2 have no positive / negative expressions concerning the theme expression “junk food”, and the influence of the theme expression “junk food” is neither promoted nor suppressed. However, from the sentences of No. 1 and No. 2, the authority specifying unit 735 specifies the words of authority such as “Member” and “Government”. Therefore, the sentence type determination unit 722 identifies the sentence types of the sentences of item number 1 and item number 2 as “examples”.

  In addition, the template selection unit 723 selects the template T in the template DB 702. In this example, when the template T1 including the assertion, reason, and example in the sentence type is selected, another template T (for example, “Current Status Description”) is selected. ”Or“ importance of claim ”is higher than“ True ”), the calculation unit 725 calculates an evaluation value.

  In addition, the combination processing unit 741 embeds a sentence in the selected template T. The combination processing unit 741 refers to the disputed chain model 703, gives priority to the arrangement of contents seen in the past, selects sentences from each sentence type, and rearranges the sentences. The rearranged sentence list becomes an output candidate.

  The calculation unit 725 confirms the satisfaction status and relevance of the selected template T. For example, when the selected template T is T1, the slots 1105 of the order 1 to the order 3 of the selection template T1 are filled, in particular, the slots 1105 of the order 1 and the order 2 where the mandatory field 1104 is mandatory are filled. Make sure. The calculation unit 725 has a relationship in which the chain of issues of each slot 1105 exists in the issue chain model 703 and the chain of issues is frequently “health” to “health” or “health” to “tax”. Confirm that the evaluation value is returned.

  The validity determination unit 726 compares, for example, the calculated evaluation value with a threshold value, and determines that the process is finished if it is equal to or greater than the threshold value. The output unit 727 outputs the output information 1500, which is the determination result as shown in FIG. This generates arguments in the debate.

  Thus, according to the present embodiment, it is possible to generate an output statement that is logically consistent. Further, when the target sentence has a relationship between the theme expression and the affirmative / negative expression that affirms or denies the issue expression, the sentence can be determined as a sentence type “claim”. When the theme expression promotes or suppresses positive / negative expressions, or when the positive / negative expression is promoted or suppressed by theme expressions, the sentence can be determined as a sentence type of “reason”. Further, when there is no affirmative / negative expression for the theme expression and the authority is specified, the sentence can be determined as a sentence type “example”.

  When the importance level is set for each sentence type of the template T, an evaluation value based on the importance degree can be calculated for a sentence combination embedded in the slot 1105 of the sentence type. Therefore, it is possible to objectively determine how much the sentence combination follows the template T.

  In addition, by applying the issue chain model 703, sentences are arranged based on the change of the past topic, so that the sentences can be rearranged consistently even when the topic changes. In addition, since the frequency that is the number of occurrences of the chain is set in each entry of the issue chain model 703, by selecting the frequency of the issue chain model 703, a combination of sentences that are likely to cause chain of issues is selected. Can do. Therefore, it is possible to generate an output sentence that is logically consistent with high accuracy. Also, by applying the issue point ontology information 701, specific issue points in the sentence can be summarized into issues with a high degree of abstraction, and the degree of freedom of combination of sentences can be improved.

  Although the present invention has been described in detail with reference to the accompanying drawings, the present invention is not limited to such specific configurations, and various modifications and equivalents within the spirit of the appended claims Includes configuration.

  The present invention is not limited to the above-described embodiments, and includes various modifications and equivalent configurations within the scope of the appended claims. For example, the above-described embodiments have been described in detail for easy understanding of the present invention, and the present invention is not necessarily limited to those having all the configurations described. A part of the configuration of one embodiment may be replaced with the configuration of another embodiment. Moreover, you may add the structure of another Example to the structure of a certain Example. In addition, for a part of the configuration of each embodiment, another configuration may be added, deleted, or replaced.

  In addition, each of the above-described configurations, functions, processing units, processing means, etc. may be realized in hardware by designing a part or all of them, for example, with an integrated circuit, and the processor realizes each function. It may be realized by software by interpreting and executing the program to be executed.

  Information such as programs, tables, and files that realize each function can be stored in a storage device such as a memory, a hard disk, and an SSD (Solid State Drive), or a recording medium such as an IC card, an SD card, and a DVD.

  Further, the control lines and the information lines are those that are considered necessary for the explanation, and not all the control lines and the information lines that are necessary for the mounting are shown. In practice, it can be considered that almost all the components are connected to each other.

Claims (9)

A statement rearrangement method executed by a computer having a processor for executing a program and a memory for storing the program,
The processor accesses a storage device that stores a template that defines a set of sentences, a plurality of sentence types that specify a sentence type, and an arrangement order of the plurality of sentence types, and a rule or learning model related to a phrase. Is possible,
The processor is
Based on a rule or a learning model for a word that defines a theme expression that is a word constituting an arbitrary theme sentence or a phrase related to the phrase, the theme expression is obtained from each sentence of a plurality of sentences in the sentence set. A theme expression identification process to be identified;
A sentence type determination process for determining a sentence type for each sentence including the theme expression specified by the theme expression specifying process based on a rule or learning model relating to the relationship between the theme expression and the sentence type ;
From the plurality of sentences whose sentence types have been determined by the sentence type determination process, a set of sentences corresponding to the plurality of sentence types specified in the template is selected, and the selected set of sentences is A sorting process for sorting according to the sorting order;
An output process for outputting a set of sentences after sorting by the sorting process;
A method for rearranging sentences, characterized in that The processor is
From each sentence , based on a rule or a learning model relating to a word that defines a dispute expression that is a word that exemplifies a dispute, a dispute expression specifying process for specifying the dispute expression;
Based on rules or learning models for positive-negative expression is a word that positive or negative, from said each statement, identifying the positive or to that affirmative constant-negation denied issue expressions identified by the issue representation specifying process A positive / negative expression specifying process,
In the sentence type determination process, the processor determines that the theme expression included in each sentence is the affirmation based on a rule or a learning model regarding the relationship between the theme expression, the positive / negative expression and the sentence type. 2. The sentence type of a sentence that is a target of an affirmative / negative expression for the point expression included in each sentence specified by the negative expression specifying process is determined for each sentence. How to rearrange sentences. The processor is
Rules for promoting a positive / negative expression, a phrase that promotes the theme expression, a phrase that suppresses the theme expression, or a word that defines a promotion / suppression expression that is a phrase that suppresses the theme expression, or Based on the learning model, for the theme expression included in each sentence, for each sentence, execute a promotion / suppression expression specifying process for specifying the promotion / suppression expression,
In the sentence type determination process, the processor determines whether the theme expression included in each sentence is the promotion based on a rule or a learning model regarding the relationship between the theme expression, the promotion / suppression expression, and the sentence type. The sentence sequence according to claim 2, wherein a sentence type of a sentence that is a target of a promotion / suppression expression included in each sentence specified by the suppression expression specifying process is determined for each sentence. Replacement method. The processor is
From each of the sentences, execute an authority specifying process that specifies an authority that is a word indicating a speaking subject,
In the sentence type determination process, the processor determines a sentence type for each sentence including an authority specified by the authority specifying process based on a rule or a learning model relating to a word defining the authority. The sentence rearrangement method according to claim 1. The template defines the importance of sentence types,
The processor is
Satisfaction based on the satisfaction status of the template with respect to the set of sentences after rearrangement by the rearrangement processing and the importance regarding the sentence type of the sentence specified by the arrangement order given by the template. A calculation process for calculating an evaluation value for the situation ;
A validity determination process for determining the validity of the set of sentences after the rearrangement based on the evaluation value calculated by the calculation process;
2. The output process according to claim 1, wherein in the output process, the processor outputs a statement sentence that is a set of the rearranged sentences when the validity determination process determines that the process is valid. How to sort sentences. The storage device stores a dispute chain model in which a plurality of issues are chained and a chain frequency is set between the chained issues.
In the rearrangement process, the processor selects a set of sentences corresponding to the plurality of sentence types defined in the template from the plurality of sentences for which the sentence type is determined according to the issue chain model . Select and rearrange subsequent sentences that contain issues linked to the issues of the previous sentence ,
The processor is
For a set of statements of the rearranged by the rearrangement processing, the issue given by chaining model, based on the sum of the frequencies of the chain between the preceding sentence and the subsequent statements specified by the arrangement order , A calculation process for calculating an evaluation value indicating a relationship between the preceding sentence and the succeeding sentence ;
A validity determination process for determining the validity of the set of sentences after the rearrangement based on the evaluation value calculated by the calculation process;
2. The sentence rearrangement according to claim 1, wherein, in the output process, the processor outputs the set of rearranged sentences when it is determined to be valid by the validity determination process. 3. Method. The storage device stores an issue chain model that defines a state in which multiple issues are chained,
In the rearrangement process, the processor selects a set of sentences corresponding to the plurality of sentence types defined in the template from the plurality of sentences for which the sentence type is determined according to the issue chain model . 2. The method of rearranging sentences according to claim 1, wherein a subsequent sentence including an issue chained from an issue of the preceding sentence is selected and rearranged. The storage device stores issue ontology information that defines the correspondence between issue and issue expression,
In the rearrangement process, the processor converts a point expression included in each sentence of the plurality of sentences for which the sentence type has been determined into a corresponding point by referring to the point ontology information, and is defined in the template. The set of sentences corresponding to the plurality of sentence types is selected and rearranged according to the issue chain model using issues converted from issue expressions included in each sentence, or rearranged. The sentence rearrangement method according to 7. A computer having a processor for executing a program and a memory for storing the program,
The processor accesses a storage device that stores a template that defines a set of sentences, a plurality of sentence types that specify a sentence type, and an arrangement order of the plurality of sentence types, and a rule or learning model related to a phrase. Is possible,
The processor is
Based on a rule or a learning model for a word that defines a theme expression that is a word constituting an arbitrary theme sentence or a phrase related to the phrase, the theme expression is obtained from each sentence of a plurality of sentences in the sentence set. A theme expression identification process to be identified;
A sentence type determination process for determining a sentence type for each sentence including the theme expression specified by the theme expression specifying process , based on a rule or a learning model regarding the relationship between the theme expression and the sentence type ;
From the plurality of sentences whose sentence types have been determined by the sentence type determination process, a set of sentences corresponding to the plurality of sentence types specified in the template is selected, and the selected set of sentences is A sorting process for sorting according to the sorting order;
An output process for outputting a set of sentences after sorting by the sorting process;
A computer characterized by executing.

Images