Abstract generation apparatus, control method and system

Abstract

Provided is an information processing apparatus capable of generating a summary text that represents a representative topic. An information processing apparatus that generates a summary text from language data includes: a word processing section (112) that classifies words included in the language data and generates a plurality of word clusters in a manner such that words likely to be related to one topic belong to the same word cluster; a cluster estimation section (113) that selects, from the plurality of word clusters, a representative word cluster that includes words related to a topic that represents the content of the language data; and a summary section (114) that generates a summary text based on the representative word cluster from the language data.

Description

TECHNICAL FIELD

[0001] The present application relates to a technology of summarizing a document. BACKGROUND

[0002] Generally, in many companies and the like, a note made by a person is referred to, and a document editor is used to make a record of deliberations and meetings.

[0003] In view of this, a record making system has been proposed in which, in order to reduce the man-hours for making a record, a voice recognition technology is used to recognize a voice occurring in a record of deliberations, and automatically transcribe the voice one word at a time. Also, a technology has been proposed in which, by using a document obtained by transcription, a document structure analysis process is performed, and a document is summarized.

[0004] As these summarization generation methods, a known natural language processing summarization technology (see, for example, Non-Patent Literature 1) is used. In this summarization technology, frequent words included in a document to be summarized are detected, and a sentence including many frequent words is extracted from the document to be summarized as a summary.

[0005] PRIOR ART DOCUMENTS

[0006] NON-PATENT LITERATURE

[0007] Non-Patent Literature 1: Gunes Erkan et al., LexRank: Graph-based Lexical Centrality as Salience in Text Summarization, Internet <URL: https: / / www.cs.cmu.edu / afs / cs / project / jair / pub / volume22 / erkan04a-html / erkan04a.html> SUMMARY

[0008] PROBLEMS TO BE SOLVED BY THE INVENTION

[0009] According to the known summarization technology of Non-Patent Literature 1 and the like, since a sentence including many frequent words is extracted as a summary, in the case where a frequent word is a word that is far from a topic that represents the content of a document, there is a problem that the extracted summary cannot represent the representative topic.

[0010] An object of the present disclosure is to solve such a problem, and to provide a summarization generation apparatus, a control method, and a system that can generate a summary that represents a representative topic.

[0011] MEANS FOR SOLVING THE PROBLEMS

[0012] To achieve the above object, one embodiment of the present disclosure is an abstract generation device that generates an abstract text from language data, characterized by including: a word processing unit that classifies words included in the language data and generates a plurality of word clusters in a manner such that words likely to be related to one topic belong to the same word cluster; a selection unit that selects a representative word cluster that includes words related to a topic that represents a content described in the language data from among the plurality of word clusters; and a generation unit that generates an abstract text based on the representative word cluster from the language data.

[0013] Here, the word processing unit can estimate a probability indicating how likely it is that each word belonging to each of the generated word clusters belongs to the topic corresponding to the word cluster, and the selection unit can select the representative word cluster from among the plurality of word clusters using the probability of each word estimated for each of the plurality of word clusters.

[0014] Here, the selection unit can calculate an index value indicating a likelihood that the content described in the language data is represented by each word cluster by summing or multiplying together the probabilities estimated for a plurality of words included in the language data for each word cluster, and select the representative word cluster by comparing a plurality of index values calculated for a plurality of word clusters.

[0015] Here, the word processing unit can include: a morpheme analysis unit that analyzes the language data into a plurality of morphemes and estimates a part of speech of each morpheme; a part of speech filtering unit that extracts words that are nouns from among the plurality of morphemes generated by the morpheme analysis unit; a word cluster generation unit that classifies the extracted words and generates the plurality of word clusters; and a probability estimation unit that estimates the probability of each word belonging to each of the generated plurality of word clusters.

[0016] Here, the word cluster generation unit can sum up the frequency of occurrence of words extracted by the part of speech filtering unit for each word based on a positional relationship between words within the language data, and generate the plurality of word clusters using the summed frequency of occurrence and the positional relationship, and the probability estimation unit can estimate the probability of each word using the summed frequency of occurrence and the positional relationship.

[0017] Here, the device can further include a voice recognition unit that converts voice data into the language data, and the word processing unit can generate the plurality of word clusters from the generated language data.

[0018] In this case, a storage unit can also store in advance prior knowledge information indicating words related to a topic, and the word processing unit can classify the words included in the language data using the prior knowledge information.

[0019] In this case, a reception unit can also receive from a user a designation of the number of word clusters to be generated, and the word processing unit can generate the designated number of word clusters.

[0020] In this case, a storage unit can also store in advance outlier information indicating words unrelated to a topic desired by a user, and the word processing unit can exclude the words indicated by the outlier information when classifying the words included in the language data.

[0021] In this case, the language data can be composed of a plurality of documents, and the word processing unit can classify the words included in each of the following data units: the entire language data, a document included in the language data, a paragraph included in the document, a plurality of sentences included in the document, and one sentence included in the document, and generate a plurality of word clusters for each data unit, and the selection unit can select the representative word cluster from the plurality of word clusters for each data unit.

[0022] In this case, a reception unit can also receive from a user a designation of the data unit, and the word processing unit can classify each data unit received from the user.

[0023] In this case, the generation unit can generate the abstract text for each data unit based on the data unit.

[0024] In this case, an analysis unit can also determine the importance of each representative word cluster.

[0025] In this case, the generation unit can make the amount of data of the abstract text variable based on the determined importance.

[0026] In this case, a display unit and a reception unit can also be provided, the display unit can display the determined importance for each representative word cluster, the reception unit can receive from a user a change in the importance for each representative word cluster, and the analysis unit can change the importance of the representative word cluster to the importance received from the user.

[0027] In this case, the number of representative word clusters selected by the selection unit can be less than the number of word clusters generated by the word processing unit.

[0028] In this case, the language data can be constituted by a plurality of documents, and the selection means selects, for each of the plurality of documents, a representative word cluster containing words related to a topic representing the content described in the document, from among the plurality of word clusters, and the generation means generates an abstract based on the representative word cluster, from the plurality of topic documents that are the basis for generating a representative word cluster containing words related to the same topic.

[0029] In this case, the word processing means classifies the words contained in each data unit, and generates the plurality of word clusters for each data unit, the selection means selects the representative word cluster for each data unit, from among the plurality of word clusters, and the generation means generates the abstract based on the representative word cluster, from the plurality of topic documents.

[0030] In addition, one embodiment of the present disclosure is a system constituted by the above-described abstract generation apparatus and a server apparatus that generates language data from speech data, the server apparatus including: a communication means that receives speech data and transmits language data generated from the received speech data to the abstract generation apparatus; and a speech recognition means that generates the language data by converting the received speech data.

[0031] In addition, one embodiment of the present disclosure is a control method used in an abstract generation apparatus that generates an abstract from language data, including: a word processing step of classifying words contained in the language data and generating a plurality of word clusters in such a manner that words likely to be related to one topic belong to the same word cluster; a selection step of selecting a representative word cluster containing words related to a topic representing the content described in the language data, from among the plurality of word clusters; and a generation step of generating an abstract based on the representative word cluster, from the language data.

[0032] Effects of Invention

[0033] According to the above-described embodiment, it is possible to generate an abstract representing a representative topic based on a representative word cluster by selecting a representative word cluster containing words related to a topic representing the content described in language data, from among a plurality of word clusters. BRIEF DESCRIPTION OF DRAWINGS

[0034] Figure 1 FIG. 1 is a diagram showing the appearance of an information processing apparatus 10 according to an embodiment.

[0035] Figure 2is a block diagram showing the structure of the main unit 100 of the information processing apparatus 10.

[0036] Figure 3 shows the content of the document 152a.

[0037] Figure 4 shows the content of the noun phrase data 153a.

[0038] Figure 5 shows an example of the data structure of the word frequency matrix 154.

[0039] Figure 6 shows the word distribution 155a, 155b, 155c for each topic and the topic distribution 156a, 156b, 156c for each document.

[0040] Figure 7 (a) to (c) of shows the content of the word cluster 157a to 157c.

[0041] Figure 8 shows the content of the cluster probability data table 158a.

[0042] Figure 9 shows the content of the abstract 191.

[0043] Figure 10 shows the screen 201 displayed on the monitor 107 of the information processing apparatus 10.

[0044] Figure 11 (a) of shows the content of the prior knowledge data 221. Figure 11 (b) of shows the content of the outlier data 231.

[0045] Figure 12 is a flowchart showing the operation of the information processing apparatus 10.

[0046] Figure 13 is a flowchart showing the operation of the reference of the setting data.

[0047] Figure 14 is a flowchart showing the operation of the processing of the outlier.

[0048] Figure 15 is a flowchart showing the operation of the inheritance processing of the word cluster.

[0049] Figure 16 shows the case where the abstract 306 is generated from the document 1 (301) and the document 4 (304).

[0050] Figure 17 shows the case where the abstract 341 is generated from the 2 paragraphs 312, 315 in the document 1 (311) and the 1 paragraph 334 in the document 3 (331).

[0051] Explanation of reference numerals

[0052] 10 information processing apparatus

[0053] 100 main body apparatus

[0054] 101 CPU

[0055] 102 ROM

[0056] 103 RAM

[0057] 104 storage circuit

[0058] 105 input / output circuit

[0059] 106 control section

[0060] 107 monitor

[0061] 108 keyboard

[0062] 109 mouse

[0063] 110 overall control section

[0064] 111 speech recognition section

[0065] 112 word processing section

[0066] 113 cluster estimation section

[0067] 114 abstract section

[0068] 121 morpheme analysis section

[0069] 122 part-of-speech filtering section

[0070] 123 word cluster generation section

[0071] 124 cluster probability estimation section

[0072] 125 representative word cluster estimation section DETAILED DESCRIPTION

[0073] 1 Embodiment

[0074] An information processing apparatus 10 as an embodiment of the present disclosure will be described.

[0075] 1.1 Information processing apparatus 10

[0076] The information processing apparatus 10 (abstract generation apparatus) is a general computer system, for example, a personal computer. The information processing apparatus 10 generates an abstract text from a document. As shown in FIG. 1, the information processing apparatus 10 includes a CPU 101, a ROM 102, a RAM 103, a storage circuit 104, an input / output circuit 105, a control section 106, a monitor 107, a keyboard 108, a mouse 109, an overall control section 110, a speech recognition section 111, a word processing section 112, a cluster estimation section 113, an abstract section 114, a morpheme analysis section 121, a part-of-speech filtering section 122, a word cluster generation section 123, a cluster probability estimation section 124, and a representative word cluster estimation section 125. Figure 1As shown, the information processing device 10 is configured such that a monitor 107, a keyboard 108, and a mouse 109 are connected to the main device 100.

[0077] Here, the document refers to news reports, literary works, papers, books, etc., expressed in languages ​​such as Japanese, English, and French. It generally consists of multiple chapters, sections, and paragraphs, with each chapter, section, and paragraph containing multiple sentences. Sentences are formed by arranging multiple words. The document is presented and stored in a computer-readable form so that it can be operated by the information processing device 10. Here, one or more documents, one or more chapters, one or more sections, one or more paragraphs, and one or more sentences are considered language data. In particular, in this specification, there are cases where a collection of multiple documents is referred to as language data. Furthermore, the document may contain graphics, photographs, charts, etc., but the information processing device 10 does not use these graphics, photographs, charts, etc., as objects for generating summary text.

[0078] 1.2 Main Unit 100

[0079] like Figure 2 As shown, the main device 100 consists of a CPU 101, a ROM 102, a RAM 103, a storage circuit 104, an input / output circuit 105, etc.

[0080] CPU101, ROM102 and RAM103 constitute the control unit 106.

[0081] RAM103 stores various control variables and set parameters, and provides a working area for CPU101 to execute programs.

[0082] ROM102 stores control programs (computer programs) used to execute actions in the main device 100.

[0083] The main unit 100 may also include a hard disk unit, which stores control programs and other data used to execute actions in the main unit 100.

[0084] CPU101 operates according to the control program stored in ROM102 or hard disk unit.

[0085] The CPU 101 operates according to the control program, thereby the control unit 106 functionally comprises a general control unit 110, a speech recognition unit 111, a word processing unit 112, a cluster estimation unit 113, and a summarization unit 114. The word processing unit 112 includes a morpheme analysis unit 121, a part-of-speech tagging unit 122, a word cluster generation unit 123, and a cluster probability estimation unit 124. Furthermore, the cluster estimation unit 113 includes a representative word cluster estimation unit 125.

[0086] The general control section 110 performs unified control of the voice recognition section 111, the word processing section 112, the cluster estimation section 113, the abstract section 114, and the like.

[0087] In addition, the general control section 110 receives the user's operation from the keyboard 108 (receiving mechanism) via the input / output circuit 105. The user's operation includes an instruction to start the abstract processing, a setting of various parameters, and the like. In the case where the user's operation is an instruction to start the abstract processing, for example, the general control section 110 performs unified control of the voice recognition section 111, the word processing section 112, the cluster estimation section 113, the abstract section 114, and the like, and causes the abstract book to be generated.

[0088] The input / output circuit 105 receives an input signal from the keyboard 108 or the mouse 109, and outputs the received input signal to the general control section 110. In addition, the input / output circuit 105 receives data for display, such as screen data, from the general control section 110, and outputs the received data to the monitor 107, and displays it on the monitor 107.

[0089] The following describes the storage circuit 104, the voice recognition section 111, the word processing section 112, the cluster estimation section 113, and the abstract section 114.

[0090] 1.3 Storage Circuit 104

[0091] The storage circuit 104 (storage mechanism) is constituted by, for example, a nonvolatile semiconductor memory or the like. Of course, the storage circuit 104 can also be constituted by a hard disk.

[0092] The storage circuit 104 has areas for storing the voice data 151a, 151b, 151c, the documents 152a, 152b, 152c, the noun phrase data 153a, 153b, 153c, the word frequency matrix 154, the word distribution 155a, 155b, 155c for each theme, the theme distribution 156a, 156b, 156c for each document, the word cluster 157a, 157b, 157c, the cluster probability data table 158a, 158b, 158c, the representative word cluster 159, and the abstract book 160.

[0093] (1) Voice Data 151a, 151b, 151c

[0094] The voice data 151a, 151b, 151c are each recording data that has recorded the proceedings of a different meeting. The voice data 151a, 151b, 151c are constituted by, for example, a voice file format of MP3 (MPEG-1 Audio Layer-3).

[0095] (2) Documents 152a, 152b, 152c

[0096] Documents 152a, 152b, and 152c are each composed of text data, which is generated by the speech recognition unit 111 performing speech recognition processing on the speech data 151a, 151b, and 151c.

[0097] Documents 152a, 152b, and 152c are identified by their document IDs “D001”, “D002”, and “D003”, respectively.

[0098] For example, document 152a Figure 3 As shown in the figure, document 152a consists of text data containing multiple sentences.

[0099] (3) Noun phrase data 153a, 153b, 153c

[0100] Noun phrase data 153a, 153b, and 153c are data containing noun phrases. These noun phrases are extracted by the morpheme analysis unit 121 and the part-of-speech filtering unit 122 through morpheme analysis and part-of-speech filtering of the text data contained in documents 152a, 152b, and 152c. Noun phrase data 153a, 153b, and 153c each contain multiple noun phrases.

[0101] For example, a noun phrase data 153a Figure 4 As shown in the figure, the noun phrase data 153a consists of multiple noun phrases extracted from document 152a.

[0102] (4) Word frequency matrix 154

[0103] The word frequency matrix 154 is a data table obtained by the word cluster generation unit 123 by summing the frequencies of the noun phrases (hereinafter referred to as words) contained in each noun phrase data 153a, 153b, 153c.

[0104] like Figure 5 As shown, as an example of word frequency matrix 154, it contains multiple groups consisting of document IDs and word frequency information, where each word frequency information contains the frequency of each word contained in the document.

[0105] As shown in the figure, for example, the noun phrase data 153a generated from document 152a identified by document ID "D001" contains 5 words "Congress", contains 4 words "deliberation", and does not contain the word "Olympics".

[0106] Additionally, for example, the noun phrase data 153b generated from document 152b identified by document ID "D002" contains one word "Congress", two words "deliberation", and no word "Olympics".

[0107] Further, for example, the noun phrase data 153c generated from the document 152c identified by the document ID "D003" does not include the word "national" and the word "review", and includes 7 words "Olympics".

[0108] (5) Word distribution 155a, 155b, 155c for each topic

[0109] The word distribution 155a, 155b, 155c for each topic (topic) is a data table indicating a distribution of frequencies of words generated by the word cluster generation unit 123 using the word frequency matrix 154 aggregated from the noun phrase data 153a, 153b, 153c and the positional relationship between words within each document for each topic. Further, the classification method for extracting a topic using the positional relationship between words within each document will be described later.

[0110] Here, the positional relationship between words within a document refers to, for example, the relationship of a plurality of words included in one sentence within a document. If the plurality of words included in one sentence are related to one topic (for example, a topic related to "politics"), it can be said that the relationship is close.

[0111] In addition, the positional relationship between words within a document refers to, for example, the relationship between the words included in one sentence and the words included in a sentence that is continuously arranged before or after the sentence within a document. If the words included in the two consecutive sentences are both related to one topic, it can be said that the relationship is close.

[0112] For example, the word distribution 155a is a data table in which the frequency of words related to one topic (for example, a topic related to "politics") is aggregated for each word from the entire noun phrase data 153a, 153b, 153c. In Figure 6 In the drawing, the word distribution 155a is illustrated. The word distribution 155a illustrated in the drawing is expressed by a bar graph in which the words included in the data table are arranged on the horizontal axis and the frequency of each word is arranged on the vertical axis.

[0113] In addition, for example, the word distribution 155b is a data table generated from the entire noun phrase data 153a, 153b, 153c, like the word distribution 155a, with respect to another topic different from the above-mentioned topic (for example, a topic related to "sports"). Figure 6 In the drawing, the word distribution 155b is illustrated.

[0114] Furthermore, for example, word distribution 155c is about a different topic than the two topics mentioned above (e.g., a topic involving "economics"), and like word distribution 155a, it is a data table generated from the noun phrase data 153a, 153b, and 153c as a whole. Figure 6 The diagram shows the word distribution at 155c.

[0115] (6) The topic distribution of each document is 156a, 156b, and 156c.

[0116] The topic distributions 156a, 156b, and 156c for each document are data tables representing the distribution of index values ​​(described later) of topics generated by the cluster estimation unit 113 for documents 152a, 152b, and 152c respectively using noun phrase data 153a, 153b, and 153c.

[0117] For example, topic distribution 156a is a data table obtained by summing the index values ​​of each topic from the noun phrase data 153a for one document 152a. Figure 6 The figure shows topic distribution 156a. Topic distribution 156a is presented as a bar chart with the topics contained in the data table arranged on the horizontal axis and the probability of words belonging to each topic (described later) arranged on the vertical axis.

[0118] Additionally, for example, topic distribution 156b is a data table generated in the same way as topic distribution 156a, concerning one document 152b. Figure 6 The figure shows the distribution of themes 156b.

[0119] Furthermore, for example, topic distribution 156c is a data table generated in the same way as topic distribution 156a, concerning a single document 152c. Figure 6 The figure shows the distribution of the theme 156c.

[0120] (7) Word clusters 157a, 157b, 157c

[0121] Word cluster 157a is a group of words generated by word cluster generation unit 123 from noun phrase data 153a for document 152a. Based on the frequency of words appearing in noun phrase data 153a and the positional relationship of words in document 152a, words that are similar in meaning, that is, words that are closely related (words that may be related to a topic) are collected as a topic and generated.

[0122] For example, a word cluster 157a Figure 7 As shown in (a), word cluster 157a contains words such as “parliament,” “deliberation,” “cabinet,” and “election.”

[0123] Also, as with the word cluster 157a, the word cluster 157b is a group of words collected and generated by the word cluster generation section 123 as one theme from the noun phrase data 153a with respect to the document 152a, with words that are similar in meaning, i.e., words that are close in relationship.

[0124] An example of the word cluster 157b is shown in (b) of FIG. 17. As shown in this figure, the word cluster 157b contains the words "Olympics", "National Stadium", "host country", "torch", and so on. Figure 7

[0125] Also, as with the word cluster 157a, the word cluster 157c is a group of words collected and generated by the word cluster generation section 123 as one theme from the noun phrase data 153a with respect to the document 152a, with words that are similar in meaning, i.e., words that are close in relationship.

[0126] An example of the word cluster 157c is shown in (c) of FIG. 17. As shown in this figure, the word cluster 157c contains the words "economy", "economy", "balance of payments", "trade", and so on. Figure 7

[0127] Further, in the above, three word clusters were generated from one document, but this is not limiting. One, two, or four or more word clusters can be generated from one document.

[0128] Also, as with the above, one or more word clusters are generated with respect to the documents 152b and 152c as well.

[0129] (8) Cluster probability data tables 158a, 158b, 158c

[0130] The cluster probability data table 158a contains the probabilities of each noun phrase (word) adjusted by the word cluster generation section 123 with respect to the document 152a based on the noun phrase data 153a, from the frequency of the words appearing in the noun phrase data 153a (i.e., the frequencies contained in the word frequency matrix 154) and the positional relationship of the words within the document 152a.

[0131] Here, the probability of a word indicates the likelihood of each word belonging to the topic corresponding to the word cluster to which the word belongs.

[0132] As shown in (a) of FIG. 18, the cluster probability data table 158a contains a plurality of pieces of probability information, each piece of probability information containing a noun phrase and a probability. Figure 8 Here, the noun phrase is a noun phrase contained in the noun phrase data 153a, and the probability is the probability of the noun phrase.

[0133]

[0134] ​​​Also, as with the cluster probability data table 158a, the cluster probability data table 158b is configured so as to correspond the noun phrase generated by the word cluster generation section 123 for the document 152b to the probability according to the noun phrase data 153b.

[0135] Further, as with the cluster probability data table 158a, the cluster probability data table 158c is configured so as to correspond the noun phrase generated by the word cluster generation section 123 for the document 152c to the probability according to the noun phrase data 153c.

[0136] (9) Representative word cluster 159

[0137] The representative word cluster 159 is one word cluster selected by the representative word cluster estimation section 125 from among the word clusters 157a, 157b, 157c for the document 152a.

[0138] Likewise, the representative word cluster is also selected for the documents 152b, 152c, respectively.

[0139] (10) Abstract 160

[0140] The abstract 160 is an abstract containing the abstract text generated by the abstract section 114 for the document 152a. The abstract 160 contains one or more sentences.

[0141] Figure 9 An example of the abstract text 191 extracted from the document 152a is shown.

[0142] The abstract is also generated for the documents 152b, 152c, respectively.

[0143] 1.4 Speech recognition section 111

[0144] The speech recognition section 111 (speech recognition mechanism) converts the utterance content (speech data) in a conference discussion or the like into text data to generate a document (language data) composed of the text data, and writes the generated document into the storage circuit 104.

[0145] The speech recognition section 111 determines, by the control of the general control section 110, the section in accordance with the pattern of the phonemes of vowels, consonants, and diphthongs for each speech data stored in the storage circuit 104, and generates an arrangement of identifiers (e.g., a kana string) representing the phonemes. Next, the speech recognition section 111 searches for a sentence in a mixed Japanese-kanji expression corresponding to the generated kana string from an inbuilt dictionary, substitutes the obtained sentence to the generated kana string, and generates text data.

[0146] In this way, the speech recognition section 111 generates a document composed of text data from speech data, attaches a document ID for identifying the document to the generated document, and writes the document into the storage circuit 104.

[0147] 1.5 word processing section 112

[0148] The word processing section 112 (word processing mechanism) performs morphological analysis, performs part-of-speech filtering, generates word clusters, and estimates the probability of each word with respect to a document in accordance with the control of the general control section 110 as shown below. The word processing section 112 classifies the words included in a document (language data) and generates a plurality of word clusters in such a manner that words likely to be related to one theme belong to the same word cluster. In addition, the word processing section 112 estimates the probability indicating the degree of likelihood that each word belonging to each generated word cluster belongs to the theme corresponding to the word cluster.

[0149] As described above, the word processing section 112 includes a morphological analysis section 121, a part-of-speech filtering section 122, a word cluster generation section 123, and a cluster probability estimation section 124.

[0150] As described later, the word cluster generation section 123 and the cluster probability estimation section 124 use, for example, LDA (Latent Dirichlet Allocation) as a known natural language processing technique.

[0151] (1) Morphological analysis section 121

[0152] The morphological analysis section 121 reads each document (language data) from the storage circuit 104 under the control of the general control section 110. Next, the morphological analysis section 121 performs morphological analysis on the text data included in the read document, decomposes the text data, and generates a plurality of morphemes. Further, the morphological analysis section 121 estimates the part of speech for each of the generated plurality of morphemes, and attaches part-of-speech information indicating the estimated part of speech to each morpheme (word).

[0153] The morphological analysis section 121 outputs the morphemes extracted from each document and attached with part-of-speech information indicating the part of speech to the part-of-speech filtering section 122.

[0154] (2) Part-of-speech filtering section 122

[0155] The part-of-speech filtering section 122 receives morphemes (words) attached with part-of-speech information indicating the part of speech from the morphological analysis section 121.

[0156] If a morpheme attached with a part of speech is received, the part-of-speech filtering section 122 extracts a morpheme (word) attached with part-of-speech information indicating a noun as a part of speech from the morphemes attached with part-of-speech information indicating the part of speech by the morphological analysis section 121. The extracted morpheme is referred to as a noun phrase. In addition, proper nouns are included in nouns.

[0157] The part-of-speech filter 122 writes the noun phrase data composed of the extracted noun phrases to the storage circuit 104.

[0158] Thus, the noun phrase data 153a, 153b, 153c is written to the storage circuit 104 for the documents 152a, 152b, 152c, respectively.

[0159] (3) Word cluster generation section 123

[0160] (Summing of the frequency of occurrence of words)

[0161] The word cluster generation section 123 sums the frequency of occurrence of words for each word from the noun phrase data 153a for the document 152a by the control of the summarization control section 110. The word cluster generation section 123 writes the frequency of occurrence of words summed for each word to the word frequency information 171 corresponding to the document 152a (i.e., the document ID "D001") in the word frequency matrix 154.

[0162] In addition, the word cluster generation section 123 sums the frequency of occurrence of words for each word from the noun phrase data 153b, 153c for the documents 152b, 152c, and writes the summed frequency to the word frequency information 171 corresponding to the documents 152b, 152c in the word frequency matrix 154, similarly to the above.

[0163] (Extraction of words having a prescribed positional relationship)

[0164] Next, the word cluster generation section 123 extracts a plurality of words having the above-described positional relationship for each sentence included in the documents 152a, 152b, 152c.

[0165] As an example, the word cluster generation section 123 extracts one sentence from the document 152a. Next, the word cluster generation section 123 extracts a plurality of words included in the extracted sentence from the noun phrase data 153a. In this way, the extracted plurality of words is considered to have one positional relationship. For example, the word cluster generation section 123 extracts the words "National Diet", "deliberate", "Olympic Games", "related", "bill", and the like from the extracted one sentence. It can be considered that the extracted words "National Diet", "deliberate", "Olympic Games", "related", "bill", and the like have one positional relationship. The word cluster generation section 123 performs the above-described processing for all the sentences included in the document 152a. In addition, the word cluster generation section 123 also performs the above-described processing with respect to each sentence included in the documents 152b, 152c.

[0166] Alternatively, as another example, the word cluster generation unit 123 can extract two consecutive sentences from document 152a. Then, the word cluster generation unit 123 extracts multiple words from the extracted noun phrase data 153a, considering that the extracted words have a positional relationship. The word cluster generation unit 123 performs the above processing on all sentences contained in document 152a. The word cluster generation unit 123 also performs the above processing on the two consecutive sentences contained in documents 152b and 152c, respectively.

[0167] Multiple words contained in a single sentence, or multiple words contained in two consecutive sentences, share common meanings, usages, or areas of application. That is, it can be considered that multiple words in a single sentence or the aforementioned two sentences have a positional relationship, and these multiple words are related to a single topic.

[0168] If we consider multiple words with positional relationships generated as described above belonging to one group, then, as mentioned above, the number of groups generated is the same as the number of positional relationships generated. When generating one positional relationship based on one sentence, the number of generated groups equals the number of sentences contained in the document. Furthermore, when generating one positional relationship based on two consecutively configured sentences, the number of generated groups equals the number of combinations of the two sentences selected from the document.

[0169] (Classification used for topic extraction)

[0170] If, in the multiple groups generated as described above, one word in the first group is the same as one word in the second group, the word cluster generation unit 123 establishes an association between the first and second groups. For example, if both the first and second groups contain the word "Congress," the first and second groups are associated. This association is established for all groups generated as described above. Thus, the two associated groups are treated as a new group. For example, the first group containing the word "Congress" and the second group containing the word "Congress" are treated as a new group. This association is established for all generated groups, and the process of establishing associations between groups is repeated as a whole until several groups can be aggregated.

[0171] Furthermore, a connection can be established between the first and second groups if the first word in the first group is the same as the second word in the second group, and the third word in the first group is the same as the fourth word in the second group. For example, if both the first and second groups contain the word "Congress," and both the first and second groups contain the word "deliberation," then the first and second groups can be associated.

[0172] Further, the number of words used to establish the association can be increased between the first group and the second group, thereby associating the first group with the second group.

[0173] In this way, the plurality of words included in the finally generated one group can be related to one theme (e.g., "politics"). In addition, the plurality of words included in another group can be related to another theme (e.g., "sports"). Here, the finally generated group is referred to as a theme group.

[0174] (Generation of word distribution for each theme)

[0175] Next, the word cluster generation section 123 aggregates the frequency of the words appearing in one of the plurality of theme groups, using the word frequency matrix 154 and the above-described positional relationship, with respect to the documents 152a, 152b, 152c (and the noun phrase data 153a, 153b, 153c) as a whole, and generates, for example, the word distribution 155a for each theme shown in FIG. 15A. Figure 6

[0176] Similarly, the word cluster generation section 123 generates, for example, the word distributions 155b, 155c for each theme shown in FIG. 15B and FIG. 15C, with respect to the other theme groups. Figure 6

[0177] (Generation of word cluster)

[0178] The word cluster generation section 123 selects, for example, the word distribution 155a from the word distributions 155a, 155b, 155c. Next, the word cluster generation section 123 extracts the words included in the document 152a (i.e., the noun phrase data 153a) from the words exemplified in the horizontal axis direction of the selected word distribution 155a. Next, the word cluster generation section 123 generates a word cluster composed of the extracted words. Figure 7 (a) of FIG. 17A indicates a word cluster 157a as an example. This word cluster includes many words related to "politics", for example.

[0179] In addition, the word cluster generation section 123 selects, for example, the word distribution 155b. Next, the word cluster generation section 123 extracts the words included in the document 152a (i.e., the noun phrase data 153a) from the words exemplified in the horizontal axis direction of the selected word distribution 155b. Next, the word cluster generation section 123 generates a word cluster composed of the extracted words. Figure 7 (b) of FIG. 17B indicates a word cluster 157b as an example. This word cluster includes many words related to "sports".

[0180] ​​Next, the word cluster generation unit 123 selects, for example, word distribution 155c. Then, the word cluster generation unit 123 extracts words contained in document 152a (i.e., noun phrase data 153a) from the words listed along the horizontal axis of the selected word distribution 155c. Then, the word cluster generation unit 123 generates word clusters composed of the extracted words. Figure 7 (c) represents word cluster 157c as an example. This word cluster contains many words related to "economy".

[0181] Here, the word cluster generation unit 123 can also extract a specified number of high-frequency words from the words listed in the selected word distribution.

[0182] In addition, similarly to the above, the word cluster generation unit 123 generates one or more word clusters for documents 152b and 152c.

[0183] As described above, the word cluster generation unit 123 generates one or more word clusters for a single document.

[0184] In this way, the word cluster generation unit 123 classifies the extracted words and generates multiple word clusters. In addition, the word cluster generation unit 123 sums the occurrence frequencies of the extracted words for each word, calculates the positional relationships between words in the document (language data), and uses the summed occurrence frequencies and the calculated positional relationships to generate multiple word clusters.

[0185] (4) Cluster probability estimation part 124

[0186] The cluster probability estimation unit 124 initially sets a probability randomly for each word within each word cluster. Then, the cluster probability estimation unit 124 uses the word distribution corresponding to that word cluster and the aforementioned positional relationship to correct the probability randomly set for the words.

[0187] Thus, as Figure 8 As shown, the cluster probability estimation unit 124 generates a cluster probability data table 158a for, for example, the word cluster 157a of document 152a. Similarly, the cluster probability estimation unit 124 generates cluster probability data tables for other word clusters of document 152a. The cluster probability estimation unit 124 writes the probability set for each word into the cluster probability data table of the storage circuit 104.

[0188] In this way, the cluster probability estimation unit 124 estimates the probability of each word belonging to each of the multiple generated word clusters. In addition, the cluster probability estimation unit 124 estimates the probability of each word using the calculated positional relationships and the total occurrence frequency.

[0189] 1.6 Cluster Prediction Section 113

[0190] The cluster estimation unit 113 (selection mechanism) selects, for each document (language data), a representative word cluster that contains words related to a topic that represents the content of the document from among a plurality of word clusters. In addition, the cluster estimation unit 113 selects a representative word cluster from among a plurality of word clusters using the probability of each word estimated for each of the plurality of word clusters.

[0191] As described above, the cluster estimation unit 113 includes a representative word cluster estimation unit 125.

[0192] In the above example, a plurality of word clusters are generated for each document.

[0193] In a case where a plurality of word clusters are generated for each document, the representative word cluster estimation unit 125 selects one word cluster as a representative word cluster from among the plurality of generated word clusters as shown below.

[0194] The representative word cluster estimation unit 125, for example, with respect to one word cluster of the document 152a, calculates a total product of the probabilities included in the cluster probability data table 158a with respect to the plurality of words included in the noun phrase data 153a.

[0195] For example, as shown in Figure 4 , the noun phrase data 153a contains the words "National Diet", "deliberate", "Olympic Games", "related", and "bill". In addition, according to the cluster probability data table 158a, the probabilities of the words "National Diet", "deliberate", "Olympic Games", "related", and "bill" are PI, P2, P3, P4, and P5, respectively.

[0196] Further, each probability is expressed in percentage (%).

[0197] In this case, the total product value (index value) = PI x P2 x P3 x P4 x P5 x... is calculated.

[0198] Here, as an example, the noun phrase data 153a contains three words "bill", and therefore, in the calculation of the total product value, the probability corresponding to the word "bill" is multiplied by three.

[0199] The representative word cluster estimation unit 125, for example, with respect to other word clusters of the document 152a, also calculates the total product value in the same manner as described above.

[0200] In this way, for example, the total product value is calculated with respect to a plurality of word clusters of the document 152a.

[0201] In this way, the representative word cluster estimation unit 125, for example, with respect to the document 152a, generates a topic distribution 156a for each document as shown in Figure 6 . In the topic distribution 156a, the horizontal axis represents topics, and the vertical axis represents the total product value (index value) of each word cluster, that is, each topic.

[0202] The representative word cluster estimation unit 125 selects the largest total product value from among the plurality of total product values calculated, and selects the word cluster that was the basis for calculating the selected total product value as the representative word cluster.

[0203] The representative word cluster estimation unit 125 also calculates the total product value (index value) for each word cluster with respect to the documents 152b, 152c, in the same manner as described above.

[0204] The representative word cluster estimation unit 125 generates, for example, the theme distribution 156b, 156c for each document with respect to the documents 152b, 152c, as illustrated. Figure 6 In each of the theme distributions 156b, 156c, the horizontal axis represents the theme, and the vertical axis represents the total product value (index value) for each word cluster, i.e., each theme.

[0205] The representative word cluster estimation unit 125 also selects the representative word cluster with respect to the documents 152b, 152c, in the same manner as described above.

[0206] Further, the representative word cluster estimation unit 125 calculates the total product value = P1xP2xP3xP4xP5x... as described above. However, the method is not limited to this.

[0207] Alternatively, the representative word cluster estimation unit 125 calculates the sum value (index value) = P1+P2+P3+P4+P5+..., selects the largest sum value from among the plurality of sum values calculated, and selects the word cluster that was the basis for calculating the selected sum value as the representative word cluster.

[0208] In this manner, the representative word cluster estimation unit 125 of the cluster estimation unit 113 (selection mechanism) calculates the index value that represents the likelihood that the word cluster represents the content described in the document by summing or multiplying the plurality of probabilities estimated for each of the plurality of words included in the document (language data) for each word cluster, and selects the representative word cluster by comparing the plurality of index values calculated for the plurality of word clusters.

[0209] Here, the number of representative word clusters selected by the cluster estimation unit 113 (selection mechanism) is less than the number of the plurality of word clusters generated by the word processing unit 112.

[0210] 1.7 Abstract Unit 114

[0211] The abstract unit 114 (extraction mechanism) generates an abstract based on the representative word cluster selected from each document (language data) by the representative word cluster estimation unit 125.

[0212] That is, the abstract section 114 extracts sentences containing words included in the representative word cluster from each document using the words included in the representative word cluster (i.e., sentences representing the contents of the documents), generates an abstract book containing the extracted sentences.

[0213] An example of the abstract text 191 generated by the abstract section 114 is shown in Fig. 17. The abstract text 191 shown in this figure contains all or part of the words included in the word cluster 157a selected as the representative word cluster as shown in (a) of Fig. 16. Figure 9 Figure 7

[0214] 1.8 Screen Example

[0215] Figure 10 A screen example displayed by the information processing apparatus 10 is shown.

[0216] The screen 201 shown in this figure is a screen for setting the hyperparameters set in the information processing apparatus 10, and requires user input.

[0217] As shown in this figure, the screen 201 contains input fields 202, 203, 204, and radio buttons 205, 206, 207 to 210.

[0218] The input fields 202, 203, 204 are fields for accepting input of the prior knowledge data, the number of word clusters, and the words of outliers, and are input by the user. The radio buttons 205, 206 are radio buttons for setting whether to inherit the word clusters. Either of the radio buttons 205, 206 is selected by the user. The radio buttons 207, 208, 209, 210 are radio buttons for selecting one of the document unit, the paragraph unit, the multiple sentence unit, and the one sentence unit as the data unit. Either of the radio buttons 207, 208, 209, 210 is selected by the user.

[0219] The general control section 110 generates the screen 201, and outputs the generated screen 201 to the monitor 107 via the input / output circuit 105. The monitor 107 displays the screen 201.

[0220] The keyboard 108 and the mouse 109 accept operation instructions from the user with respect to the input fields 202, 203, 204, and the radio buttons 205 to 210, and output instruction signals corresponding to the accepted operation instructions to the general control section 110 via the input / output circuit 105.

[0221] The general control section 110 performs processing corresponding to the received instruction signals.

[0222] Further, the prior knowledge data, the number of word clusters, the words of outliers, the data unit, and whether to inherit the word clusters will be described next. ​​

[0223] 1.9 Prior knowledge data

[0224] Prior knowledge data (prior knowledge information) is data indicating words related to a theme. In other words, the prior knowledge data is used to classify words so that words related to one theme belong to one word cluster. Like this, information indicating which word cluster a word belongs to is given in advance in accordance with the user's preference.

[0225] As an example shown in (a) of FIG. 21, prior knowledge data 221 is composed of word clusters 222, 223, 224. The word cluster 222 includes the words "diet", "deliberation", "law". The word cluster 223 includes the words "Olympic Games", "stadium". The word cluster 224 includes the words "economy", "interest". Figure 11 In the prior knowledge data 221, words to be registered in advance for each word cluster are described using JSON form as a data form.

[0226] The prior knowledge data is generated by the user in advance. The generated prior knowledge data is stored in the storage circuit 104 according to the user's operation instruction.

[0227] The word cluster generation section 123 classifies words included in a document using the prior knowledge data stored in the storage circuit 104, and generates a plurality of word clusters.

[0228] An explanation will be made using the prior knowledge data 221 shown in (a) of FIG. 21.

[0229] Figure 11 In a case where the words "diet", "deliberation", "law" included in the word cluster 222 are included in a document, the word cluster generation section 123 determines that the document corresponds to the word cluster 222. In addition, in a case where the words "Olympic Games", "stadium" included in the word cluster 223 are included in a document, the word cluster generation section 123 determines that the document corresponds to the word cluster 223. In addition, in a case where the words "economy", "interest" included in the word cluster 224 are included in a document, the word cluster generation section 123 determines that the document corresponds to the word cluster 224.

[0230] Like this, it is possible to prepare prior knowledge data indicating words related to a theme in advance, and generate word clusters based on words included in a document using the prior knowledge data as teaching data.

[0231] Specifically, it is possible to make the generated abstract close to the abstract desired by the user by giving words considered to be important in accordance with the user's preference to a word cluster as prior knowledge in advance in a case where the user who requested the abstract of the document of the meeting minutes participated in the target meeting.

[0232] Specifically, it is possible to make the generated abstract close to the abstract desired by the user by giving words considered to be important in accordance with the user's preference to a word cluster as prior knowledge in advance in a case where the user who requested the abstract of the document of the meeting minutes participated in the target meeting. ​

[0233] 1.10 Number of word clusters

[0234] The number of word clusters to be generated can also be accepted from the user by the keyboard 108. The keyboard 108 outputs the accepted number of word clusters to the summarization control section 110 via the input / output circuit 105.

[0235] The word processing section 112 receives the number of word clusters from the summarization control section 110. The word cluster generation section 123 generates the specified number of word clusters.

[0236] The case where the user can specify the number of word clusters is, for example, the case shown below.

[0237] There is a case where the user who attended the meeting as a summary target can specify the number of word clusters in advance when requesting a summary to the information processing apparatus 10. Specifically, it is the case where the agenda is decided in advance at the beginning of the meeting, and the meeting is progressed according to the agenda. As an example, it is, for example, the case where brainstorming within an implementation team is performed, (a) a certain period of time is spent for discussing ideas related to medicine, then, (b) ideas related to construction are discussed, and finally, (c) discussion is performed on the topic related to office solution. In this case, since the words appearing in the respective topics of (a), (b), (c) are largely different, the user can specify three word clusters in advance. The information processing apparatus 10 generates the specified number of word clusters.

[0238] 1.11 Words of outliers

[0239] The storage circuit 104 can also store, in advance, outlier data (outlier information) indicating words that are not related to the topic desired by the user (i.e., words that are not similar to the words indicating the topic desired by the user). In this case, the word cluster generation section 123 excludes the words indicated by the outlier data when classifying the words included in the document.

[0240] As an example, Figure 11 The outlier data 231 shown in (b) of FIG. 10 includes "thank you" and "agenda". The word cluster generation section 123 excludes the words "thank you" and "agenda" indicated by the outlier data when classifying the words included in the document.

[0241] Examples of such words include the sentence "Today's agenda is as follows." and "The meeting is over. Thank you." that appear fixedly within the meeting. It is appropriate to exclude the words "agenda" in "Today's agenda is as follows." and "thank you" in "The meeting is over. Thank you." from the classification since they are not related to the topic of the meeting.

[0242] As such, it is also possible to make a determination that, in a case where words within a document are classified so that related words belong to one word cluster, a word that is not similar to any word within the document is treated as an outlier and is not taken into consideration.

[0243] 1.12 Whether to inherit a word cluster

[0244] There is a case where one topic is continuously discussed in a plurality of meetings. As an example, there are weekly meetings, regular meetings, and the like in development meetings. In this case, it is possible to use a word cluster generated from a document representing the content of a first meeting again when generating a word cluster from a document representing the content of a second meeting.

[0245] Accordingly, the word cluster generation section 123 writes a word cluster generated from a document representing the content of a first meeting to the storage circuit 104 and saves it in advance according to an operation instruction from the user.

[0246] In this case, the representative word cluster estimated by the cluster estimation section 113 can also be written to the storage circuit 104.

[0247] According to an operation instruction from the user, the word cluster generation section 123 can also use a word cluster of a first meeting saved in the storage circuit 104 when generating a word cluster from a document representing the content of a second meeting.

[0248] By using a word cluster generated and saved in a previous meeting, it is possible to refer to a word cluster having a high importance degree in the previous meeting and to extract information (an abstract) having a higher importance degree in the current meeting.

[0249] In addition, the keyboard 108 (the reception mechanism) can also receive a designation of a word cluster to be deleted from the user. The keyboard 108 outputs the received designation of the word cluster to be deleted to the overall control section 110 via the input / output circuit 105.

[0250] The word processing section 112 receives the designation of the word cluster to be deleted from the overall control section 110. The word cluster generation section 123 deletes the designated word cluster from the storage circuit 104.

[0251] As an example of deleting information on a word cluster, there is a case of a single meeting. In a scenario of a brainstorming session, a casual conversation, and the like, there is a case where discussion ends in the meeting. In this case, it is also possible to delete a word cluster saved in the storage circuit 104 according to an operation instruction from the user.

[0252] In addition, in the case where a plurality of word clusters are generated and stored in the storage circuit 104 in the first meeting of the meetings that are successively held, the word cluster that the user definitely knows cannot be used in the next meeting among the plurality of word clusters can also be deleted from the storage circuit 104 according to the user's operation instruction.

[0253] 1.13 Data Unit

[0254] The word cluster generation section 123 can also take, as a data unit, one of the entire language data composed of a plurality of documents, the entire document included in the language data, a paragraph included in the document, a plurality of sentences included in the document, and one sentence included in the document, classify the words included in the data unit for each data unit, and generate a plurality of word clusters for each data unit. The data unit is a unit that determines a theme.

[0255] The cluster estimation section 113 selects a representative word cluster from the plurality of word clusters for each data unit.

[0256] In addition, the specification of the data unit can also be accepted from the user by the keyboard 108 (the reception mechanism). The keyboard 108 outputs the accepted specification of the data unit to the summary control section 110 via the input and output circuit 105.

[0257] The word cluster generation section 123 classifies the words for each data unit accepted from the user.

[0258] In the present embodiment, in order to classify the topics with the discussion of one meeting as the object, as described above, the data unit can also be specified as the unit of a plurality of sentences or the unit of one sentence. In addition, in order to be able to cope with a discussion that suddenly occurs again, the sentences can also be cluster-classified for each sentence.

[0259] In addition, the abstract section 114 can also extract, as an abstract text, a sentence that represents the recorded content of the data unit from the data unit for each data unit.

[0260] 1.14 Action of Information Processing Apparatus 10

[0261] The action of the information processing apparatus 10 will be described using a flowchart.

[0262] (1) Outline Action of Information Processing Apparatus 10

[0263] The outline action of the information processing apparatus 10 will be described using the flowchart shown in FIG. 10. Figure 12

[0264] ​The overall control unit 110 waits for the user to input the setting of hyperparameters (step S101). When the setting of hyperparameters is not input (No in step S102), the overall control unit 110 moves the control to step S101 and repeats the process.

[0265] When the setting of hyperparameters is input (Yes in step S102), the overall control unit 110 refers to the setting data (step S103).

[0266] The speech recognition unit 111 obtains the converted text data from the speech data through speech recognition (step S104).

[0267] The morpheme analysis unit 121 performs morpheme analysis and generates morphemes (step S105).

[0268] The part-of-speech screening unit 122 extracts noun phrases from the morphemes (step S106).

[0269] The word cluster generation unit 123 generates word clusters, and the cluster probability estimation unit 124 estimates the probability for each word (step S107).​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​Next, the total control section 110 judges whether or not the number of word clusters is set by the user (step S134). In the case where the number of word clusters is set (YES in step S134), the total control section 110 writes the number of word clusters into the storage circuit 104 (step S135).

[0279] Next, the total control section 110 writes the data unit into the storage circuit 104 (step S136).

[0280] Up to this point, the explanation of the operation with reference to the set data is ended.

[0281] (3) Processing of an abnormal value

[0282] The processing of an abnormal value is explained using the flowchart shown in Fig. 17. Figure 14 The word cluster generation section 123 refers to the set data (hyperparameters) set by the user (step S151).

[0283] Next, the word cluster generation section 123 judges whether or not an abnormal value word is set (step S152). In the case where an abnormal value word is set (YES in step S152), the word cluster generation section 123 deletes the abnormal value word from the word cluster stored in the storage circuit 104 (step S153).

[0284] Up to this point, the explanation of the processing of an abnormal value is ended.

[0285] (4) Inheritance processing of a word cluster

[0286] The inheritance processing of a word cluster is explained using the flowchart shown in Fig. 18.

[0287] Figure 15 The total control section 110 refers to the set data (hyperparameters) set by the user (step S171).

[0288] Next, the total control section 110 judges whether or not the generated word cluster is inherited (step S172). In the case where inheritance is set (YES in step S172), the total control section 110 maintains the saved state of the generated word cluster in the storage circuit 104 (step S173). In the case where inheritance is not set (NO in step S172), the total control section 110 deletes the word cluster present in the storage circuit 104. In this case, it is also possible to delete all the word clusters present in the storage circuit 104, and it is also possible to delete only the word cluster designated by the user (step S174).

[0289] Up to this point, the explanation of the inheritance processing of a word cluster is ended.

[0290] (4) Inheritance processing of a word cluster

[0291] ​1.15 Embodiment (1)

[0292] In the above-described embodiments, the speech recognition unit 111 performs speech recognition processing on the speech data and generates text data.

[0293] However, the present application is not limited to this mode.

[0294] One mode of the present application can also be a system configured of the information processing apparatus 10 and a server apparatus. The information processing apparatus 10 and the server apparatus are connected via a network. The server apparatus provides speech recognition processing as one of cloud services. That is, the server apparatus receives speech data, and converts the received speech data into text data and generates a document.

[0295] The information processing apparatus 10 can also utilize the speech recognition processing provided by the server apparatus.

[0296] The information processing apparatus 10 is provided with a network communication circuit that is connected to the server apparatus via a network. The network communication circuit sends speech data to the server apparatus under the control of the general control unit 110, and entrusts the server apparatus with speech recognition processing on the speech data.

[0297] The server apparatus is provided with a network communication circuit (communication mechanism) and a speech recognition circuit (speech recognition mechanism). The speech recognition circuit of the server apparatus has the same structure as the speech recognition unit 111 of the information processing apparatus 10.

[0298] The network communication circuit of the server apparatus receives speech data and an entrustment of speech recognition processing on the speech data from the information processing apparatus 10 via the network. Upon acceptance of the entrustment and the speech data, the speech recognition circuit converts the received speech data into text data, and generates a document composed of the text data. The network communication circuit of the server apparatus sends the generated document to the information processing apparatus 10 via the network.

[0299] The network communication circuit of the information processing apparatus 10 receives the document from the server apparatus, and writes the received document into the storage circuit 104.

[0300] 1.16 Embodiment (2)

[0301] The morphological analysis unit 121 of the information processing apparatus 10 can also use a publicly known morphological analysis method. The morphological analysis unit 121 can also use, for example, any one of publicly known MeCab, JUMAN, KyTea, and ChaSen.

[0302] 1.17 Embodiment (3)

[0303] The word cluster generation section 123 and the cluster probability estimation section 124 can utilize various known natural language processing techniques. For example, LDA or the like can be cited as a method of estimating a latent topic of text data from words appearing in the text data. LDA is a document classification model that assumes that text data has multiple topics. This method estimates a latent topic by sequentially learning the frequency of appearance of a word, the positional relationship, and the like from text data as an object.

[0304] In addition, a latent topic refers to a set of important words clustered per topic, and has a topic probability (simply, probability) indicating how likely each of the words belongs to the topic. With the latent topic, sentences are classified per latent topic.

[0305] Accordingly, the word cluster generation section 123 generates a plurality of word clusters. A word cluster, which is an element capable of constituting a topic, is a set of words clustered from words.

[0306] In addition, the cluster probability estimation section 124 obtains a probability that each of the words clustered belongs to a topic.

[0307] 1.18 Embodiment (4)

[0308] The abstract section 114 can also utilize various known natural language processing techniques.

[0309] For example, LexRank or the like can be cited as a method of extracting important sentences by scoring sentences that can be important in text data according to words appearing in the text data.

[0310] When LexRank is applied, a word appearing in text data and a similarity to a word included in a word cluster are expressed in a graph to calculate an eigenvector centrality, and a relative importance of a text unit is calculated. That is, a word appearing frequently in other sentences and similar to an important word is considered to be important, and it can be considered that a sentence in which such an important word appears is important.

[0311] An abstract is generated by this method, and it is possible to obtain an abstract text in which a feature is given by the word cluster.

[0312] 1.19 Embodiment (5)

[0313] There is a case in which a plurality of word clusters are generated per data unit in a document as an object of generating a word cluster.

[0314] In a case in which one data unit originally includes a plurality of topics, a plurality of word clusters are generated from the data unit.

[0315] In addition, in a case where a plurality of word clusters are similar to each other, a plurality of word clusters are generated from one data unit. For example, a case where a word cluster associated with brain science and a word cluster associated with AI. Words belonging to two word clusters have a high possibility of containing the same content, and it can be said that the word clusters are similar to each other in the word cluster associated with brain science and the word cluster associated with AI.

[0316] 1.20 Embodiment (6)

[0317] The user can also specify the data amount (i.e., the amount of characters) of the abstract before the abstract is generated from the document. That is, the summary control section 110 can receive the data amount of the abstract from the keyboard 108 via the input / output circuit 105 according to the operation of the user.

[0318] The abstract section 114 generates the abstract within the range of the specified data amount of the abstract.

[0319] For example, in a case where the abstract section 114 initially generates a 500-character abstract composed of five sentences, if the user specifies 300 characters as the amount of characters, one sentence is deleted from the five sentences. The abstract section 114 counts the number of characters of the abstract after the deletion of one sentence. If the number of characters of the abstract after the deletion is 300 characters or less, the abstract section 114 determines the abstract after the deletion of one sentence. On the other hand, in a case where the number of characters of the abstract after the deletion exceeds 300 characters, one sentence is further deleted. As such, the above processing is repeated until the number of characters of the abstract after the deletion reaches 300 characters or less as specified by the user.

[0320] As such, by the user appropriately specifying the data amount of the abstract, it is possible to suppress a situation where the abstract cannot function as a concise summary due to an excessive data amount of the abstract, and a situation where the gist of the discussion cannot be understood from the abstract due to an insufficient data amount of the abstract.

[0321] Further, the summary control section 110 can also receive the number of sentences of the abstract from the keyboard 108 via the input / output circuit 105 according to the operation of the user. In this case, the abstract section 114 generates the abstract within the range of the specified number of sentences.

[0322] In this case, as with the above, in a case where the number of sentences exceeds the number specified by the user, the abstract section 114 repeatedly performs the deletion of sentences from the abstract until the number reaches the number specified by the user or less.

[0323] 1.21 Embodiment (7)

[0324] As described above, the word processing section 112 can also classify the words included in the data unit by each data unit, the data unit being one of the entire language data composed of a plurality of documents, the entire document included in the language data, a paragraph included in the document, a plurality of sentences included in the document, and one sentence included in the document.

[0325] The cluster estimation section 113 (selection mechanism) can also include a cluster analysis section (analysis mechanism) that determines the importance of each representative word cluster.

[0326] In addition, the cluster analysis section can also total the number or amount of data units corresponding to each representative word cluster, and determine the importance of the representative word cluster based on the total value of each representative word cluster.

[0327] In addition, in a case where the total value of one representative word cluster exceeds a predetermined value, the cluster analysis section can also set the importance of the representative word cluster to a predetermined maximum value.

[0328] (1) For example, in a case where there are 8 data units corresponding to the representative word cluster a, 4 data units corresponding to the representative word cluster b, and 2 data units corresponding to the representative word cluster c, the cluster analysis section can set the importance of each representative word cluster to be representative word cluster a > representative word cluster b > representative word cluster c for the representative word clusters a, b, and c.

[0329] That is, the cluster analysis section can set the importance of each representative word cluster based on the number of data units corresponding to the representative word cluster. In this case, the more the number of data units, the higher the importance of the representative word cluster is set, and the less the number of data units, the lower the importance of the representative word cluster is set.

[0330] The importance set for each representative word cluster can also be presented to the user. That is, the cluster analysis section outputs the set importance to the monitor 107 via the summarization control section 110 and the input / output circuit 105 for each representative word cluster. The monitor 107 outputs the set importance for each representative word cluster.

[0331] (2) Even in a case where the cluster analysis section sets a low importance for a certain representative word cluster, the user can set a high importance for the representative word cluster when judging that the importance of the representative word cluster is high.

[0332] On the contrary, even in a case where the cluster analysis section sets a high importance for a certain representative word cluster, the user can set a low importance for the representative word cluster when judging that the importance of the representative word cluster is low.

[0333] In this way, it is possible to change the importance of the representative word cluster as desired by the user.

[0334] The total control section 110 can also instruct the input / output circuit 105 to accept a change in the importance of the representative word cluster from the keyboard 108 in accordance with a user operation.

[0335] The cluster analysis section changes the importance of the representative word cluster to the importance accepted in accordance with the user operation instruction.

[0336] In the case of the example of (1), in the case where the user judges that the importance of the representative word cluster c is the highest, the cluster analysis section changes the setting of the importance of each representative word cluster to the importance of the representative word cluster c > the importance of the representative word cluster a > the importance of the representative word cluster b for the representative word clusters a, b, and c.

[0337] In this way, the importance of each representative word cluster is changed to the importance desired by the user, and as a result, it is possible to select a more appropriate representative word cluster that reflects the desires of the user.

[0338] (3) The amount of data of the abstract text or the number of sentences described in Embodiment (6) can also be made variable in accordance with the importance of the representative word cluster. That is, the abstract section 114 can also make the amount of data of the abstract text variable in accordance with the decided importance. Here, the abstract section 114 can also make the amount of text included in the abstract text or the number of sentences included in the abstract text variable.

[0339] For example, the amount of data of the abstract text or the number of sentences can be increased in the case where the importance of the representative word cluster is high, and the amount of data of the abstract text or the number of sentences can be decreased in the case where the importance of the representative word cluster is low.

[0340] As in the case of the example of (1), in the case where the importance of each representative word cluster is set to the importance of the representative word cluster a > the importance of the representative word cluster b > the importance of the representative word cluster c, the abstract text of the representative word cluster a is set to 4 sentences, the abstract text of the representative word cluster b is set to 2 sentences, and the abstract text of the representative word cluster c is set to 1 sentence.

[0341] In this way, the amount of data of the abstract text or the number of sentences is made variable in accordance with the importance of the representative word cluster, and the reason for this is that, because the amount of information possessed by a main representative word cluster is large, the likelihood of being able to provide the information (abstract text) desired by the user increases with respect to such a representative word cluster.

[0342] 1.22 Embodiment (8)

[0343] As described above, the word cluster generation unit 123 takes one of the entire language data composed of a plurality of documents, the entire document included in the language data, a paragraph included in the document, a plurality of sentences included in the document, and one sentence included in the document as a data unit. Here, a specific description will be made regarding the case where the entire document and the paragraph included in the document are taken as data units, respectively.

[0344] (1) Case where the entire document is taken as a data unit

[0345] A description will be made regarding the case where the entire document is taken as a data unit using the following. Figure 16

[0346] As shown in the figure, it is assumed that language data composed of a plurality of documents 301, 302, 303, 304, and 305 is set as an object of summarization. In addition, it is assumed that representative word clusters with "politics", "sports", "economy", "politics", and "economy" as topics are respectively selected by the cluster estimation unit 113 with respect to the plurality of documents 301, 302, 303, 304, and 305.

[0347] In this case, with respect to the two documents 301 and 304, it is estimated that each of them has "politics" as a topic, and therefore the summarization unit 114 can generate one summary text from the documents 301 and 304. In this case, the generated summary is composed of one or more sentences.

[0348] As described above, in the case where it is estimated that a plurality of documents have the same topic, one summary text is generated from the plurality of documents, and therefore the generated summary text can become an appropriate summary text that succinctly represents the contents of the plurality of documents having the same topic.

[0349] The information processing apparatus 10 (summarization generation apparatus) can generate a summary text from language data composed of a plurality of documents. The cluster estimation unit 113 can select, with respect to each of the plurality of documents, a representative word cluster that contains words related to a topic of a description of the document from a plurality of word clusters. In the case where there are a plurality of topic documents that are the basis for generating a representative word cluster containing words related to the same topic, the summarization unit 114 can generate a summary text based on the representative word cluster from the plurality of topic documents.

[0350] (2) Case where a paragraph included in a document is taken as a data unit

[0351] A description will be made regarding the case where a paragraph included in a document is taken as a data unit using the following. Figure 17

[0352] ​​As shown in the figure, it is assumed that language data composed of a plurality of documents 311, 321, 331 is set as a subject of abstract generation. Further, it is assumed that the document 311 is composed of a plurality of paragraphs 312, 313, 314, 315, 316, the document 321 is composed of a plurality of paragraphs 322, 323, 324, 325, 326, and the document 331 is composed of a plurality of paragraphs 332, 333, 334, 335, 336.

[0353] Here, it is assumed that, with respect to the paragraphs 312, 313, 314, 315, 316 of the document 311, the representative unit clusters having "politics", "economy", "economy", "politics", and "economy" as a theme are respectively selected by the cluster estimation section 113.

[0354] Further, it is assumed that, with respect to the paragraphs 322, 323, 324, 325, 326 of the document 321, the representative unit clusters having "sports", "economy", "sports", "economy", and "sports" as a theme are respectively selected by the cluster estimation section 113.

[0355] Further, it is assumed that, with respect to the paragraphs 332, 333, 334, 335, 336 of the document 331, the representative unit clusters having "economy", "economy", "politics", "economy", and "sports" as a theme are respectively selected by the cluster estimation section 113.

[0356] In this case, with respect to the paragraphs 312, 315 of the document 311 and the paragraph 334 of the document 331, it is estimated that they have "politics" as a theme, respectively, and therefore the abstract section 114 can generate one abstract text from the paragraphs 312, 315 of the document 311 and the paragraph 334 of the document 331. In this case, the generated abstract is composed of one or a plurality of sentences.

[0357] As described above, in a case where, with respect to a plurality of documents including a plurality of paragraphs, a plurality of paragraphs having the same theme are estimated to exist, one abstract text is generated from the plurality of paragraphs, and therefore the generated abstract text can become an appropriate abstract text that represents the contents of the plurality of paragraphs having the same theme concisely.

[0358] Further, in a case where, with respect to one document including a plurality of paragraphs, a plurality of paragraphs having the same theme are estimated to exist, one abstract text can be generated from the plurality of paragraphs.

[0359] The word processing section 112 can classify words included in each data unit and generate a plurality of word clusters for each data unit. The cluster estimation section 113 can select a representative word cluster from a plurality of word clusters for each data unit.

[0360] As described above, the abstract section 114 can extract an abstract text from a plurality of data units.

[0361] In addition, the abstract section 114 can also extract an abstract from a plurality of data units of a plurality of documents.

[0362] 1.23SUMMARY

[0363] As explained above, it is possible to generate an abstract text expressing a representative topic by selecting a representative word cluster including a word related to the topic of the contents of the representative document from among the generated plurality of word clusters by the cluster estimation section 113, and thereby generating an abstract text expressing a representative topic using the representative word cluster.

[0364] INDUSTRIAL APPLICABILITY

[0365] The abstract generation device according to the present application has the effect that an abstract text expressing a representative topic can be generated, and is therefore useful as a technique for abstracting a document.

Claims

1. A summary generation device for generating abstracts based on linguistic data, characterized in that, have: The word processing mechanism classifies the words contained in the language data and generates multiple word clusters in a manner that makes words that may be related to a topic belong to the same word cluster. Select an organization to select a representative word cluster from the plurality of word clusters that contains words related to the topic of the recorded content representing the language data; as well as The generator, based on the language data and the representative word clusters, generates a summary text. The word processing mechanism estimates the probability that each word belonging to each generated word cluster is likely to belong to the topic corresponding to that word cluster. The selection mechanism calculates the sum or product of the probabilities of each word in the language data for each word cluster, calculates an index value representing the likelihood of the recorded content of the language data represented by that word cluster, compares the multiple index values ​​calculated for multiple word clusters, and selects the representative word cluster from the multiple word clusters.

2. The abstract generation apparatus as described in claim 1, characterized in that, The word processing mechanism has the following features: The morpheme analysis mechanism performs morpheme analysis on the language data, generates multiple morphemes, and infers the part of speech of each morpheme; A part-of-speech filtering mechanism extracts words that can be used as nouns from a plurality of morphemes generated by the morpheme parsing mechanism; A word cluster generation mechanism classifies the extracted words and generates the multiple word clusters; as well as A probability estimation mechanism estimates the probability of each word belonging to each of the generated multiple word clusters.

3. The abstract generation apparatus as described in claim 2, characterized in that, The word cluster generation mechanism calculates the positional relationships between words in the language data, sums the frequency of occurrence of each word against the frequency of occurrence of words extracted by the part-of-speech tagging mechanism, and uses the calculated positional relationships and the summed frequency of occurrence to generate the multiple word clusters. The probability estimation mechanism uses the calculated positional relationships and the total frequency of occurrence to estimate the probability of each word.

4. The abstract generation apparatus as described in claim 1, characterized in that, It also includes a speech recognition mechanism to convert speech data into the language data. The word processing mechanism generates the multiple word clusters based on the generated language data.

5. The abstract generation apparatus as described in claim 1, characterized in that, It also has a storage mechanism that stores prior knowledge information about words related to a topic in advance. The word processing mechanism uses the prior knowledge information to classify the words contained in the language data.

6. The abstract generation apparatus as described in claim 1, characterized in that, It also has the function of accepting the request, specifying the number of word clusters to be generated based on the user's request. The word processing mechanism generates a specified number of word clusters.

7. The abstract generation apparatus as described in claim 1, characterized in that, It also has a storage mechanism that pre-stores outlier information representing words unrelated to the user's desired topic. When classifying the words contained in the language data, the word processing mechanism excludes words represented by outlier information.

8. The abstract generation apparatus as described in claim 1, characterized in that, The language data consists of multiple documents. The word processing mechanism takes one of the following as data units: the entire language data, the documents contained in the language data, the paragraphs contained in the documents, the multiple sentences contained in the documents, and the single sentence contained in the documents. It then classifies the words contained in each data unit and generates the multiple word clusters for each data unit. The selection mechanism selects the representative word cluster from the plurality of word clusters for each data unit.

9. The abstract generation apparatus as described in claim 8, characterized in that, It also has an accepting agency, which accepts the data units designated by the user. The word processing mechanism categorizes each data unit received from the user.

10. The abstract generation apparatus as described in claim 8, characterized in that, The generating mechanism generates the summary text based on each data unit.

11. The abstract generation apparatus as claimed in claim 1, characterized in that, It also has an analysis mechanism that determines the importance of each representative word cluster.

12. The abstract generation apparatus as described in claim 11, characterized in that, The generation mechanism can vary the amount of data in the abstract text based on the determined importance.

13. The abstract generation apparatus as described in claim 11, characterized in that, It also has: Display mechanism; and The accepting agency, from the user's acceptance input, The display mechanism shows the determined importance for each representative word cluster. The receiving agency, based on changes in the importance of each representative word cluster, The analysis agency changed the importance of the word cluster to the importance received from the user.

14. The abstract generation apparatus as claimed in claim 1, characterized in that, The number of representative word clusters selected by the selection mechanism is less than the number of multiple word clusters generated by the word processing mechanism.

15. The abstract generation apparatus as claimed in claim 1, characterized in that, The language data consists of multiple documents. For each of the plurality of documents, the selection mechanism selects a representative word cluster from the plurality of word clusters that contains words related to the topic representing the content of that document. When multiple topic documents exist, which serve as the basis for generating representative word clusters containing words related to the same topic, the generating mechanism generates a summary based on the representative word clusters from the multiple topic documents.

16. The abstract generation apparatus as described in claim 15, characterized in that, The word processing mechanism takes one of the following as data units: the entire language data, the documents contained within the language data, the paragraphs contained in the documents, the multiple sentences contained in the documents, and a single sentence contained in the documents. For each data unit, it classifies the words contained within that data unit and generates the multiple word clusters for each data unit. The selection mechanism selects the representative word cluster from the plurality of word clusters for each data unit. The generating organization generates the summary text based on multiple data units within the multiple topic documents.

17. A system comprising the summary generation apparatus of claim 1 and a server apparatus for generating language data based on speech data, characterized in that, The server device includes: A communication mechanism receives voice data and sends language data generated based on the received voice data to the summary generation device; and A speech recognition mechanism converts the received speech data to generate the language data.

18. A control method for use in a summarization apparatus that generates a summary text based on language data, characterized in that, include: The word processing step involves classifying the words contained in the language data and generating multiple word clusters in a manner that ensures that words that may be related to a topic belong to the same word cluster. In the selection step, from the plurality of word clusters, a representative word cluster containing words related to the topic representing the recorded content of the language data is selected; as well as The generation step involves generating a summary text based on the language data and the representative word clusters. In the word processing step, the probability of each word belonging to each generated word cluster belonging to the topic corresponding to that word cluster is estimated. In the selection step, by calculating the sum or product of the probabilities of each of the multiple words contained in the language data for each word cluster, an index value representing the likelihood of the recorded content of the language data represented by the word cluster is calculated, and the multiple index values ​​calculated for multiple word clusters are compared to select the representative word cluster from the multiple word clusters.

19. The control method as described in claim 18, characterized in that, The word processing steps include: The morpheme parsing step involves parsing the language data to generate multiple morphemes and inferring the part of speech of each morpheme. The part-of-speech filtering step extracts words that are nouns from the multiple morphemes generated by the morpheme parsing step. The word cluster generation step involves classifying the extracted words and generating the multiple word clusters. as well as The probability estimation step estimates the probability of each word belonging to each of the generated multiple word clusters.

20. The control method as described in claim 19, characterized in that, In the word cluster generation step, the positional relationships between words within the language data are determined. For each word, the frequency of occurrence of the words extracted in the part-of-speech tagging step is summed. The determined positional relationships and the summed frequency of occurrence are then used to generate the multiple word clusters. In the probability estimation step, the probability of each word is estimated using the calculated positional relationships and the total frequency of occurrence.

21. The control method as described in claim 18, characterized in that, It also includes a speech recognition step, which converts speech data to generate the language data. In the word processing step, the plurality of word clusters are generated based on the generated language data.

22. The control method as described in claim 18, characterized in that, The summary generation device also includes a storage mechanism that stores prior knowledge information about words related to a topic. In the word processing step, the prior knowledge information is used to classify the words contained in the language data.

23. The control method as described in claim 18, The summary generation device also includes a receiving mechanism that receives a user's specification of the number of word clusters to be generated. In the word processing step, a specified number of word clusters are generated.

24. The control method as described in claim 18, characterized in that, The summary generation device also includes a storage mechanism that stores in advance outlier information representing words unrelated to the user's desired topic. In the word processing step, when classifying the words contained in the language data, words represented by outlier information are excluded.

25. The control method as described in claim 18, characterized in that, The language data consists of multiple documents. In the word processing step, one of the following is taken as a data unit: the entire language data, the document contained in the language data, the paragraph contained in the document, multiple sentences contained in the document, and a single sentence contained in the document. For each data unit, the words contained within that data unit are classified, and for each data unit, the multiple word clusters are generated. In the selection step, the representative word cluster is selected from the plurality of word clusters for each data unit.

26. The control method as described in claim 25, characterized in that, The summary generation device also includes a receiving agency, which receives the data unit specified by the user. In the word processing step, data is categorized according to each unit received from the user.

27. The control method as described in claim 25, characterized in that, In the generation step, the summary text is generated for each data unit.

28. The control method as described in claim 18, It also includes an analysis step that determines the importance of each representative word cluster.

29. The control method as described in claim 28, In the generation step, the amount of data in the abstract can be varied depending on the determined importance.

30. The control method as described in claim 28, characterized in that, The summary generation device also includes: a display mechanism; and The accepting agency, from the user's acceptance input, The display mechanism shows the determined importance for each representative word cluster. The receiving agency, based on changes in the importance of each representative word cluster, In the analysis step, the importance of the word cluster is changed to the importance received from the user.

31. The control method as described in claim 18, The number of representative word clusters selected in the selection step is less than the number of multiple word clusters generated in the word processing step.

32. The control method as described in claim 18, characterized in that, The language data consists of multiple documents. In the selection step, for each of the plurality of documents, a representative word cluster is selected from the plurality of word clusters that contains words related to the topic representing the content of that document. In the generation step, if there are multiple topic documents that serve as the basis for generating representative word clusters containing words related to the same topic, a summary text is generated based on the representative word clusters according to the multiple topic documents.

33. The control method as described in claim 32, characterized in that, In the word processing step, one of the following is taken as a data unit: the entire language data, the document contained in the language data, the paragraph contained in the document, multiple sentences contained in the document, and a single sentence contained in the document. For each data unit, the words contained within that data unit are classified, and for each data unit, the multiple word clusters are generated. In the selection step, the representative word cluster is selected from the plurality of word clusters for each data unit. In the generation step, the summary text is generated based on multiple data units within the multiple topic documents.

Images