Summary generation device, control method and system
The summary generation device addresses the issue of thematic relevance in summaries by generating word clusters and selecting a representative cluster to create a summary that accurately reflects the topic, with adjustable length based on importance.
Patent Information
- Application Number
- JP2020172358
- Authority / Receiving Office
- JP · JP
- Patent Type
- Patents
- Current Assignee / Owner
- Filing Date
- 2020-10-13
- Publication Date
- 2024-12-03
- Estimated Expiration
- 2040-10-13
AI Technical Summary
Existing summarization techniques often fail to represent the representative topic in summarized texts, as they primarily extract sentences with frequently occurring words, leading to summaries that lack thematic relevance.
A summary generation device that estimates multiple topics from linguistic data, generates word clusters related to these topics, selects a representative word cluster, and generates a summary sentence based on this cluster, adjusting the summary's data amount according to the importance level.
This approach creates a summary sentence that effectively represents the representative topic of the original content, ensuring thematic relevance and allowing for customizable summary length based on importance levels.
Smart Images

Figure 0007593043000001 
Figure 0007593043000002 
Figure 0007593043000003
Abstract
Description
[Technical field]
[0001] The present disclosure relates to techniques for summarizing documents. [Background technology]
[0002] In general, in many companies, minutes of meetings and conferences are prepared using a document editor, referring to notes prepared manually.
[0003] In response to this, in order to reduce the man-hours required for manually creating minutes, a system for creating minutes has been proposed that uses speech recognition technology to recognize speech uttered during a meeting and automatically transcribe every word. Furthermore, a technology has been proposed for summarizing documents by performing document structure analysis processing on the documents obtained by transcription.
[0004] These summary generation methods use well-known natural language processing summarization techniques (see, for example, Non-Patent Document 1). These summarization techniques detect frequently occurring words contained in a document to be summarized, and extract sentences containing many of the frequently occurring words from the document to be summarized as summary sentences. [Prior art documents] [Non-patent literature]
[0005] [Non-Patent Document 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 of the Invention [Problem to be solved by the invention]
[0006] According to known summarization techniques such as Non-Patent Document 1, sentences containing many frequently occurring words are extracted as summaries. Therefore, if the frequently occurring words are words that have little relation to the topic that represents the contents of the document, the extracted summary does not represent the representative topic. This causes a problem.
[0007] An object of the present disclosure is to provide a summary generation device, a control method, and a system that can solve such problems and generate a summary sentence that expresses a representative topic. [Means for solving the problem]
[0008] In order to achieve the above object, one aspect of the present disclosure is a summary generation device that generates a summary from language data, comprising: a word processing means that infers a plurality of topics from a plurality of words included in the language data, generates, for each topic, a word cluster collecting a plurality of words that may be related to the topic, and estimates a probability indicating a likelihood that each word belonging to each generated word cluster belongs to the topic corresponding to the word cluster; a selection means that selects, from the plurality of word clusters, a representative word cluster including words related to the topic that is representative of the description content of the language data, using the probability of each word estimated for each of the plurality of word clusters; and a generation means that generates a summary from the language data based on the representative word cluster. an analysis means for determining the importance of each representative word cluster; Equipped with The generating means varies the amount of data of the summary sentence according to the determined importance. It is characterized by:
[0010] Here, the selection means may sum or multiply the probabilities estimated for each of the multiple words included in the language data for each word cluster to calculate an index value indicating the likelihood that the word cluster represents the content of the language data, and select the representative word cluster by comparing the multiple index values calculated for the multiple word clusters.
[0011] Here, the word processing means may include a morphological analysis means for performing morphological analysis of the language data to generate a plurality of morphemes and estimating the part of speech of each morpheme, a part of speech filtering means for extracting words that are nouns from the plurality of morphemes generated by the morphological analysis means, a word cluster generation means for classifying the extracted words and generating the plurality of word clusters, and a probability estimation means for estimating the probability of each word belonging to each of the plurality of generated word clusters.
[0012] Here, the word cluster generation means may determine a positional relationship between words within the language data, aggregate the occurrence frequency of each word extracted by the part-of-speech filtering means, and generate the multiple word clusters using the determined positional relationship and the aggregated occurrence frequency, and the probability estimation means may estimate the probability of each word using the determined positional relationship and the aggregated occurrence frequency.
[0013] Here, the system may further comprise a voice recognition means for converting voice data to generate the language data, and the word processing means may generate the plurality of word clusters from the generated language data.
[0014] Here, the present invention further includes a storage means for storing prior knowledge information indicating words related to a certain topic in advance, and the word processing means uses the prior knowledge information to process words included in the language data. Collected by topic You may do so.
[0015] Here, the system may further comprise a receiving means for receiving a designation of the number of word clusters to be generated from a user, and the word processing means may generate the designated number of word clusters.
[0016] Here, the present invention further includes a storage means for storing in advance outlier information indicating words that are not related to a topic desired by a user, and the word processing means processes the words included in the language data. Collect by topic In this case, words indicated by the outlier information may be excluded.
[0017] Here, the language data is made up of a plurality of documents, and the word processing means treats 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, or a single sentence included in the document as a data unit, and processes the words included in each data unit as follows: Collect by topic The plurality of word clusters may be generated for each data unit, and the selection means may select the representative word cluster from the plurality of word clusters for each data unit.
[0018] The present invention further comprises a receiving means for receiving designation of the data unit from a user, and the word processing means performs the following for each data unit received from the user: Collecting words You may do so.
[0019] Here, the generating means may generate the summary from each data unit.
[0020] Here, the system may further comprise an analysis means for determining the importance of each representative word cluster.
[0021] Here, the generating means may vary the data amount of the summary text depending on the determined importance.
[0022] Another aspect of the present disclosure is a summary generation device that generates a summary from language data, comprising: a word processing means that infers a plurality of topics from a plurality of words included in the language data, generates, for each topic, a word cluster that collects a plurality of words that may be related to the topic, and estimates a probability that each word belonging to each generated word cluster is likely to belong to the topic corresponding to the word cluster; a selection means 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, using the probability of each word estimated for each of the plurality of word clusters; a generation means that generates a summary from the language data based on the representative word cluster; and an analysis means that determines the importance of each representative word cluster. A display means and a receiving means for receiving input from a user. and, the display means displays the determined importance for each representative word cluster, the reception means receives a change in importance from a user for each representative word cluster, and the analysis means changes the importance of the representative word cluster to the importance received by the user. The generating means is characterized in that the data amount of the summary sentence or the number of documents is variable depending on the importance. .
[0023] Here, the number of the representative word clusters selected by the selection means may be less than the number of the word clusters generated by the word processing means.
[0024] Here, the language data consists of a plurality of documents, and the selection means selects, for each of the plurality of documents, from the plurality of word clusters, a representative word cluster containing words related to a topic that is representative of the contents of the document, and when there are a plurality of topic documents that serve as the basis for generating representative word clusters containing words related to the same topic, the generation means may generate a summary from the plurality of topic documents based on the representative word cluster.
[0025] Here, the word processing means treats 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, or a single sentence included in the document as a data unit, and processes the words included in each data unit as follows: Collect by topic The plurality of word clusters may be generated for each data unit, the selection means may select the representative word cluster from the plurality of word clusters for each data unit, and the generation means may generate the summary sentence from a plurality of data units in the plurality of topic documents.
[0026] Moreover, one aspect of the present disclosure is a system comprising the above-mentioned summary generation device and a server device that generates language data from voice data, wherein the server device is characterized in having a communication means for receiving voice data and transmitting language data generated from the received voice data to the summary generation device, and a voice recognition means for converting the received voice data to generate the language data.
[0027] One aspect of the present disclosure is a control method used in a summary generation device that generates a summary from language data, comprising a word processing step of estimating a plurality of topics from a plurality of words included in the language data, generating, for each topic, a word cluster comprising a plurality of words that may be related to the topic, and estimating a probability indicating a likelihood that each word belonging to each generated word cluster belongs to the topic corresponding to the word cluster; a selection step of selecting, from the plurality of word clusters, a representative word cluster including words related to the topic that represents the content of the language data, using the probability of each word estimated for each of the plurality of word clusters; and a generation step of generating a summary from the language data based on the representative word cluster. and an analysis step of determining, for each representative word cluster, the importance of the representative word cluster. Including The generating step varies the amount of data of the summary sentence according to the determined importance. It is characterized by: Another aspect of the present disclosure is a control method used in a summary generation device that generates a summary from language data, the method including: a word processing step of inferring a plurality of topics from a plurality of words included in the language data, generating, for each topic, a word cluster including a plurality of words that may be related to the topic, and estimating a probability that each word belonging to each generated word cluster is likely to belong to the topic corresponding to the word cluster; a selection step of selecting, from the plurality of word clusters, a representative word cluster including words related to a topic that represents a description content of the language data, using the probability of each word estimated for each of the plurality of word clusters; and a generation step of generating a summary from the language data based on the representative word cluster, the language data being composed of a plurality of documents, the selection step selecting, for each of the plurality of documents, a representative word cluster including words related to a topic that represents a description content of the document, and the generation step generating a summary from the plurality of topic documents based on the representative word cluster when there are a plurality of topic documents that are used as sources for generating representative word clusters including words related to the same topic. and a step of receiving a change in importance for each representative word cluster from the representative word cluster. The step of generating a summary from the language data includes: a word processing step of inferring a plurality of topics from a plurality of words included in the language data, generating a word cluster for each topic, the word cluster including a plurality of words that may be related to the topic, and estimating a probability that each word belonging to each generated word cluster is likely to belong to the topic corresponding to the word cluster; a selection step of selecting a representative word cluster from the plurality of word clusters using the probability of each word estimated for each of the plurality of word clusters, the representative word cluster including words related to a topic that represents the content of the language data; a generation step of generating a summary from the language data based on the representative word cluster; an analysis step of determining an importance of each representative word cluster for each representative word cluster; Effect of the Invention
[0028] According to the above aspect, by selecting a representative word cluster from a plurality of word clusters, which contains words related to a topic that represents the content of the language data, a summary sentence expressing the representative topic can be generated based on the representative word cluster. [Brief description of the drawings]
[0029] [Figure 1] 1 is a diagram showing the external appearance of an information processing device 10 according to an embodiment. [Diagram 2] 1 is a block diagram showing a configuration of a main unit 100 of an information processing device 10. FIG. [Diagram 3] The contents of document 152a are shown. [Figure 4] The contents of the noun phrase data 153a are shown below. [Diagram 5] 2 shows an example of a data structure of the word frequency matrix 154. [Figure 6] The word distributions 155a, 155b, 155c per topic and the topic distributions 156a, 156b, 156c per document are shown. [Figure 7] (a) to (c) show the contents of word clusters 157a to 157c. [Figure 8] 1 shows the contents of a cluster probability data table 158a. [Figure 9] The contents of summary sentence 191 are shown. [Figure 10] A screen 201 displayed on the monitor 107 of the information processing device 10 is shown. [Figure 11] 2A shows the contents of the prior knowledge data 221. FIG. 2B shows the contents of the outlier data 231. [Figure 12] 4 is a flowchart showing an operation of the information processing device 10. [Figure 13] 13 is a flowchart showing an operation of referring to setting data. [Figure 14] 13 is a flowchart showing an operation of processing an outlier. [Figure 15] 13 is a flowchart showing the operation of inheritance processing of a word cluster. [Figure 16] 3 shows how an abstract 306 is generated from a document 1 (301) and a document 4 (304). [Figure 17] 1 shows how an abstract 341 is generated from two paragraphs 312 and 315 of document 1 (311) and one paragraph 334 of document 3 (331). DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS
[0030] 1. Embodiment An information processing device 10 according to an embodiment of the present disclosure will be described.
[0031] 1.1 Information processing device 10 The information processing device 10 (summary generation device) is a general computer system, for example, a personal computer. The information processing device 10 generates a summary from a document. As shown in Fig. 1, the information processing device 10 is configured by connecting a monitor 107, a keyboard 108, and a mouse 109 to a main unit 100.
[0032] Here, the document is a newspaper article, a literary work, a thesis, a book, etc., written in a language such as Japanese, English, French, etc., and is generally composed of a plurality of chapters, a plurality of sections, and a plurality of paragraphs, and each chapter, each section, and each paragraph includes a plurality of sentences. Each sentence is composed of a plurality of words arranged in a row. In order to enable handling by the information processing device 10, the document is expressed and stored in a computer-readable format. Here, one or more documents, one or more chapters, one or more sections, one or more paragraphs, and one or more sentences are language data. In particular, in this specification, a collection of multiple documents may be referred to as language data. Note that a document may include figures, photographs, graphs, etc., but the information processing device 10 does not consider these figures, photographs, graphs, etc. as targets for generating a summary.
[0033] 1.2 Main unit 100 As shown in FIG. 2, the main device 100 is composed of a CPU 101, a ROM 102, a RAM 103, a storage circuit 104, an input / output circuit 105, and the like.
[0034] The CPU 101, the ROM 102 and the RAM 103 constitute a control unit 106.
[0035] The RAM 103 stores various control variables and set parameters, and also provides a work area when the CPU 101 executes a program.
[0036] ROM 102 stores a control program (computer program) for executing operations in main device 100 and the like.
[0037] Main device 100 may further include a hard disk unit, and the hard disk unit may store a control program for executing operations in main device 100, and the like.
[0038] The CPU 101 operates according to a control program stored in the ROM 102 or the hard disk unit.
[0039] When the CPU 101 operates according to the control program, the control unit 106 functionally configures an overall 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 (word processing means) includes a morphological analysis unit 121, a part-of-speech filtering unit 122, a word cluster generation unit 123, and a cluster probability estimation unit 124. In addition, the cluster estimation unit 113 includes a representative word cluster estimation unit 125.
[0040] The general control unit 110 integrally controls the speech recognition unit 111, the word processing unit 112, the cluster estimation unit 113, the summarization unit 114, and the like.
[0041] Furthermore, the overall control unit 110 receives user operations from the keyboard 108 (reception means) via the input / output circuit 105. User operations include an instruction to start summarization processing, the setting of various parameters, etc. When the user operation is, for example, an instruction to start summarization processing, the overall control unit 110 collectively controls the speech recognition unit 111, word processing unit 112, cluster estimation unit 113, summarization unit 114, etc. to generate a summary.
[0042] 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 unit 110. The input / output circuit 105 also receives data for display, for example, data for a screen, from the general control unit 110, and outputs the received data to the monitor 107 to cause the monitor 107 to display it.
[0043] The memory circuit 104, the speech recognition unit 111, the word processing unit 112, the cluster estimation unit 113 and the summarization unit 114 will be described below.
[0044] 1.3 Memory circuit 104 The memory circuit 104 (storage means) is composed of, for example, a non-volatile semiconductor memory, etc. Of course, the memory circuit 104 may be composed of a hard disk.
[0045] The memory circuit 104 has areas for storing speech data 151a, 151b, 151c, documents 152a, 152b, 152c, noun phrase data 153a, 153b, 153c, word frequency matrix 154, word distributions by topic 155a, 155b, 155c, topic distributions by document 156a, 156b, 156c, word clusters 157a, 157b, 157c, cluster probability data tables 158a, 158b, 158c, representative word clusters 159 and abstracts 160.
[0046] (1) Audio data 151a, 151b, 151c The audio data 151a, 151b, and 151c are recording data of the proceedings of different conferences, and are configured in an audio file format, for example, MP3 (MPEG-1 Audio Layer-3).
[0047] (2) Documents 152a, 152b, and 152c The documents 152a, 152b, and 152c are composed of text data generated by applying voice recognition processing to the voice data 151a, 151b, and 151c by the voice recognition unit 111, respectively.
[0048] Documents 152a, 152b, and 152c are identified by document IDs "D001," "D002," and "D003," respectively.
[0049] An example of document 152a is shown in Fig. 3. As shown in this figure, document 152a is made up of text data including a plurality of sentences.
[0050] (3) Noun phrase data 153a, 153b, 153c Noun phrase data 153a, 153b, and 153c are data including noun phrases extracted by applying morphological analysis and part-of-speech filtering to text data included in documents 152a, 152b, and 152c by morphological analysis unit 121 and part-of-speech filtering unit 122. Each of noun phrase data 153a, 153b, and 153c includes a plurality of noun phrases.
[0051] An example of the noun phrase data 153a is shown in Fig. 4. As shown in this figure, the noun phrase data 153a is made up of a plurality of noun phrases extracted from the document 152a.
[0052] (4) Word Frequency Matrix 154 Word frequency matrix 154 is a data table obtained by word cluster generating section 123 tallying up the frequency of each noun phrase (hereinafter referred to as word) included in noun phrase data 153a, 153b, and 153c.
[0053] As an example, as shown in FIG. 5, the word frequency matrix 154 includes a plurality of pairs each consisting of a document ID and word frequency information, and each piece of word frequency information includes the frequency of each word included in the document.
[0054] As shown in this figure, for example, noun phrase data 153a generated from document 152a identified by document ID "D001" contains five occurrences of the word "Diet," four occurrences of the word "Deliberation," and no occurrences of the word "Olympics."
[0055] Also, for example, noun phrase data 153b generated from document 152b identified by document ID "D002" contains one occurrence of the word "Diet," two occurrences of the word "Discussion," and no occurrence of the word "Olympics."
[0056] Furthermore, for example, noun phrase data 153c generated from document 152c identified by document ID "D003" does not include the word "Diet" or the word "Discussion", but includes the word "Olympics" seven times.
[0057] (5) Word distribution by topic 155a, 155b, 155c Word distributions 155a, 155b, and 155c for each topic are data tables showing the distribution of word frequencies generated by the word cluster generating unit 123 using the word frequency matrix 154 aggregated for each topic from the noun phrase data 153a, 153b, and 153c and the positional relationships between words in each document. A classification method for extracting topics using the positional relationships between words in each document will be described later.
[0058] Here, the positional relationship between words in a document refers to, for example, the relationship between multiple words contained in one sentence in the document. Multiple words contained in one sentence are related to one topic (for example, a topic related to "politics") and can be said to be closely related to each other.
[0059] The positional relationship between words in a document refers to, for example, the relationship between a word contained in one sentence and a word contained in a sentence that is adjacent to the sentence and is located before or after the sentence. Words contained in two consecutive sentences are also related to one topic and can be said to be closely related.
[0060] For example, word distribution 155a is a data table that compiles the frequency of words related to a topic (e.g., a topic related to "politics") for each word from all of noun phrase data 153a, 153b, and 153c. Word distribution 155a is shown graphically in Fig. 6. Word distribution 155a shown in this figure is represented by a bar graph, with the words included in the data table arranged on the horizontal axis and the frequency of each word arranged on the vertical axis.
[0061] Also, for example, word distribution 155b is a data table generated from all of noun phrase data 153a, 153b, and 153c, similar to word distribution 155a, for a topic different from the above topic (for example, a topic related to "sports"). Figure 6 shows word distribution 155b graphically.
[0062] Furthermore, for example, word distribution 155c is a data table generated from all of noun phrase data 153a, 153b, and 153c, similar to word distribution 155a, for a topic (e.g., a topic related to "economy") different from the two topics. Figure 6 shows word distribution 155c graphically.
[0063] (6) Topic distribution per document 156a, 156b, 156c Topic distributions for each document 156a, 156b, and 156c are data tables showing the distribution of topic index values (described later) generated by the cluster estimation unit 113 from the noun phrase data 153a, 153b, and 153c for documents 152a, 152b, and 152c, respectively.
[0064] For example, topic distribution 156a is a data table obtained by aggregating the index values of each topic from noun phrase data 153a for one document 152a. Topic distribution 156a is shown graphically in Fig. 6. Topic distribution 156a shown in this figure is represented by a bar graph, with the topics included in the data table arranged on the horizontal axis and the probability (described later) of words belonging to each topic arranged on the vertical axis.
[0065] Also, for example, topic distribution 156b is a data table generated for one document 152b in the same manner as topic distribution 156a. Figure 6 shows topic distribution 156b graphically.
[0066] Further, for example, topic distribution 156c is a data table generated for one document 152c in the same manner as topic distribution 156a. FIG. 6 shows topic distribution 156c graphically.
[0067] (7) Word clusters 157a, 157b, and 157c Word cluster 157a is a group of words generated by word cluster generation unit 123 for document 152a by collecting, as a single topic, words that are similar in meaning, i.e., closely related (words that may be related to a single topic) from noun phrase data 153a according to the frequency of words appearing in noun phrase data 153a and the relative positions of the words within document 152a.
[0068] An example of the word cluster 157a is shown in Fig. 7(a) As shown in this figure, the word cluster 157a includes words such as "Diet," "Deliberation," "Cabinet," and "Election."
[0069] Similarly to word cluster 157a, word cluster 157b is a group of words generated by word cluster generation unit 123 for document 152a by collecting words that are similar in meaning, i.e., closely related, from noun phrase data 153a and putting them together as a single topic.
[0070] An example of word cluster 157b is shown in Fig. 7(b) As shown in this figure, word cluster 157b includes words such as "Olympics," "National Stadium," "Host country," and "Olympic flame."
[0071] Similarly to word cluster 157a, word cluster 157c is a group of words generated by word cluster generation unit 123 for document 152a by collecting words that are similar in meaning, i.e., closely related, from noun phrase data 153a and putting them together as one topic.
[0072] An example of word cluster 157c is shown in Fig. 7(c) As shown in this figure, word cluster 157c includes words such as "business climate," "economy," "balance of payments," and "trade."
[0073] Although three word clusters are generated from one document in the above description, the present invention is not limited to this and one, two, or four or more word clusters may be generated from one document.
[0074] Similarly, one or more word clusters are generated for each of documents 152b and 152c.
[0075] (8) Cluster Probability Data Tables 158a, 158b, 158c Cluster probability data table 158a contains the probability of each noun phrase (word) obtained by adjusting noun phrase data 153a for document 152a by word cluster generation unit 123 based on the frequency of words appearing in noun phrase data 153a (i.e., the frequency of being included in word frequency matrix 154) and the positional relationship between words within document 152a.
[0076] Here, the word probability indicates the likelihood that each word belonging to each word cluster belongs to the topic corresponding to that word cluster.
[0077] As shown in FIG. 8, cluster probability data table 158a includes a plurality of pieces of probability information, each of which includes a noun phrase and a probability.
[0078] Here, the noun phrase is a noun phrase included in noun phrase data 153a, and the probability is the probability of the noun phrase.
[0079] Similarly to cluster probability data table 158a, cluster probability data table 158b is configured by word cluster generating unit 123 by associating noun phrases generated from noun phrase data 153b for document 152b with probabilities.
[0080] Furthermore, like cluster probability data table 158a, cluster probability data table 158c is configured by word cluster generating unit 123 by associating noun phrases generated from noun phrase data 153c for document 152c with probabilities.
[0081] (9) Representative word cluster 159 Representative word cluster 159 is one word cluster selected by representative word cluster inference unit 125 from among word clusters 157a, 157b, and 157c for document 152a.
[0082] Representative word clusters are also selected for documents 152b and 152c, respectively.
[0083] (10) Abstract 160 The abstract 160 is an abstract including a summary sentence generated for the document 152a by the summarizing unit 114. The abstract 160 includes one or more sentences.
[0084] FIG. 9 shows an example of a summary sentence 191 extracted from the document 152a.
[0085] Abstracts are also generated for documents 152b and 152c, respectively.
[0086] 1.4 Voice Recognition Unit 111 The voice recognition unit 111 (voice recognition means) converts utterances (voice data) made during discussions at a meeting or the like into text data to generate a document (language data) made of the text data, and writes the generated document into the memory circuit 104.
[0087] Under the control of the general control unit 110, the speech recognition unit 111 identifies sections that match the phoneme patterns of vowels, consonants, and mora syllabaries for each piece of speech data stored in the storage circuitry 104, and generates a sequence of identifiers representing the phonemes (for example, a kana string). Next, the speech recognition unit 111 searches an internal dictionary for a word written in a mixture of kanji and kana that corresponds to the generated kana string, and replaces the obtained word with the generated kana string to generate text data.
[0088] In this way, the voice recognition unit 111 generates a document made up of text data from the voice data, and writes the generated document into the storage circuit 104 together with a document ID for identifying the document.
[0089] 1.5 Word Processing Unit 112 The word processing unit 112 (word processing means) performs morphological analysis on the document, performs part-of-speech filtering, generates word clusters, and estimates the probability of each word, as described below, under the control of the overall control unit 110. The word processing unit 112 classifies words contained in the document (language data) and generates multiple word clusters so that words that may be related to one topic belong to the same word cluster. The word processing unit 112 also estimates the probability that each word belonging to each generated word cluster is likely to belong to the topic corresponding to that word cluster.
[0090] As described above, the word processing unit 112 includes a morphological analysis unit 121, a part-of-speech filtering unit 122, a word cluster generation unit 123, and a cluster probability estimation unit .
[0091] As described later, the word cluster generating unit 123 and the cluster probability estimating unit 124 use, for example, LDA (Latent Dirichlet Allocation) as a known natural language processing technique.
[0092] (1) Morphological analysis section 121 The morphological analysis unit 121 reads out each document (language data) from the memory circuitry 104 under the control of the general control unit 110. Next, the morphological analysis unit 121 performs morphological analysis on the text data included in the read out document, and breaks down the text data to generate a plurality of morphemes. Furthermore, the morphological analysis unit 121 estimates the part of speech for each of the generated morphemes, and assigns part of speech information indicating the estimated part of speech to each morpheme (word).
[0093] The morpheme analysis unit 121 extracts morphemes from each document and outputs the morphemes to which part-of-speech information indicating the part of speech is added to the part-of-speech filtering unit 122 .
[0094] (2) Part of speech filtering unit 122 The part-of-speech filtering unit 122 receives morphemes (words) to which part-of-speech information indicating the part of speech has been added from the morpheme analysis unit 121.
[0095] When receiving the morphemes with parts of speech, the part-of-speech filtering unit 122 extracts morphemes (words) with part-of-speech information indicating a noun as a part of speech from the morphemes with part-of-speech information indicating a part of speech by the morpheme analysis unit 121. The extracted morphemes are called noun phrases. Note that the nouns include proper nouns.
[0096] The part-of-speech filtering unit 122 writes noun phrase data consisting of the extracted noun phrases into the memory circuit 104.
[0097] In this manner, noun phrase data 153a, 153b, and 153c are written into the memory circuit 104 for the documents 152a, 152b, and 152c, respectively.
[0098] (3) Word cluster generator 123 (Counting the frequency of occurrence of words) The word cluster generation unit 123, under the control of the overall control unit 110, tallies up the occurrence frequency of each word from the noun phrase data 153a for the document 152a. The word cluster generation unit 123 writes the word occurrence frequency tallied for each word into the word frequency information 171 in the word frequency matrix 154 that corresponds to the document 152a (i.e., the document ID "D001").
[0099] Similarly to the above, the word cluster generation unit 123 calculates the frequency of occurrence of each word from the noun phrase data 153b and 153c for the documents 152b and 152c, and writes the calculated frequency into the word frequency information 171 in the word frequency matrix 154 corresponding to the documents 152b and 152c.
[0100] (Extracting words with a specific positional relationship) Next, the word cluster generating unit 123 extracts a plurality of words having the above-mentioned positional relationship from each sentence included in the documents 152a, 152b, and 152c.
[0101] As an example, the word cluster generating unit 123 extracts one sentence from the document 152a. Next, the word cluster generating unit 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 are considered to have one positional relationship. For example, the word cluster generating unit 123 extracts words such as "Diet", "Deliberation", "Olympics", "Related", and "Bill" from one extracted sentence. The extracted words such as "Diet", "Deliberation", "Olympics", "Related", and "Bill" are considered to have one positional relationship. The word cluster generating unit 123 executes the above process for all sentences included in the document 152a. The word cluster generating unit 123 also executes the above process for each sentence included in the documents 152b and 152c.
[0102] As another example, the word cluster generation unit 123 may extract two sentences arranged consecutively from the document 152a. Next, the word cluster generation unit 123 extracts a plurality of words included in the two extracted sentences from the noun phrase data 153a. In this way, the extracted plurality of words are considered to have one positional relationship. The word cluster generation unit 123 executes the above process for all sentences included in the document 152a. The word cluster generation unit 123 also executes the above process for each of the two sentences arranged consecutively included in the documents 152b and 152c.
[0103] A plurality of words contained in one sentence, or a plurality of words contained in two sentences arranged one after the other, have a common meaning, a common way of using, a common field of use, etc. In other words, a plurality of words contained in one sentence or the above two sentences have a single positional relationship, and these multiple words are considered to be related to a single topic.
[0104] If the multiple words having the positional relationships generated as described above are considered to belong to one group, the same number of groups as the number of generated positional relationships are generated as described above. When one positional relationship is generated from one sentence, the number of generated groups is equal to the number of sentences contained in the document. Also, when one positional relationship is generated from two sentences arranged consecutively one after the other, the number of generated groups is equal to the number of combinations of the above two sentences selected from the document.
[0105] (Classification for topic extraction) The word cluster generating unit 123 associates the first group with the second group when one word included in the first group and one word included in the second group are the same among the multiple groups generated as described above. For example, when the first group includes the word "Diet" and the second group also includes the word "Diet", the first group and the second group are associated with each other. This association is performed for all groups generated as described above. The two groups associated in this way are treated as a new group. For example, the first group including the word "Diet" and the second group including the word "Diet" are treated as a new group. This association is performed for all generated groups, and the process of associating groups with each other is repeated until the groups can be consolidated into several groups as a whole.
[0106] In addition, the first group and the second group may be associated when a first word included in the first group is the same as a second word included in the second group, and a third word included in the first group is the same as a fourth word included in the second group. For example, when the first group includes the word "Diet", the second group also includes the word "Diet", the first group includes the word "Deliberation", and the second group also includes the word "Deliberation", the first group and the second group are associated.
[0107] Furthermore, the first group and the second group may be associated with each other by increasing the number of words for association between the first group and the second group.
[0108] In this way, the words included in one finally generated group may relate to one topic (e.g., "politics"), and the words included in another group may relate to another topic (e.g., "sports"). Here, the finally generated groups are called topic groups.
[0109] (Generating word distributions by topic) Next, the word cluster generation unit 123 uses the word frequency matrix 154 and the above-mentioned positional relationships to count the frequency of words appearing in one of the multiple topic groups for all of the documents 152a, 152b, 152c (and the noun phrase data 153a, 153b, 153c), and generates, for example, a word distribution 155a for each topic as shown in FIG. 6.
[0110] Similarly, the word cluster generating unit 123 generates, for example, word distributions 155b and 155c for each topic shown in FIG. 6 for other topic groups.
[0111] (Generating word clusters) The word cluster generating unit 123 selects, for example, word distribution 155a from among word distributions 155a, 155b, and 155c. Next, the word cluster generating unit 123 extracts words included in document 152a (i.e., noun phrase data 153a) from among the words listed in the horizontal direction of the selected word distribution 155a. Next, the word cluster generating unit 123 generates a word cluster consisting of the extracted words. FIG. 7(a) shows word cluster 157a as an example. This word cluster includes many words related to, for example, "politics."
[0112] Furthermore, the word cluster generating unit 123 selects, for example, word distribution 155b. Next, the word cluster generating unit 123 extracts words included in document 152a (i.e., noun phrase data 153a) from among the words listed in the horizontal direction of the selected word distribution 155b. Next, the word cluster generating unit 123 generates a word cluster consisting of the extracted words. FIG. 7(b) shows word cluster 157b as an example. This word cluster includes many words related to "sports."
[0113] Furthermore, the word cluster generating unit 123 selects, for example, word distribution 155c. Next, the word cluster generating unit 123 extracts words included in document 152a (i.e., noun phrase data 153a) from among the words listed in the horizontal direction of the selected word distribution 155c. Next, the word cluster generating unit 123 generates a word cluster consisting of the extracted words. An example of word cluster 157c is shown in FIG. 7(c). This word cluster includes many words related to "economy."
[0114] Here, the word cluster generating unit 123 may extract a predetermined number of words with high frequencies from among the words listed in the selected word distribution.
[0115] Moreover, the word cluster generator 123 generates one or more word clusters for the documents 152b and 152c in the same manner as described above.
[0116] As described above, the word cluster generator 123 generates one or more word clusters for each document.
[0117] In this way, the word cluster generation unit 123 classifies the extracted words and generates a plurality of word clusters. In addition, the word cluster generation unit 123 counts the occurrence frequency of the extracted words for each word, obtains the positional relationship between the words in the document (language data), and generates a plurality of word clusters using the counted occurrence frequency and the obtained positional relationship.
[0118] (4) Cluster Probability Estimation Unit 124 The cluster probability estimation unit 124 first randomly assigns probabilities to each word in each word cluster, and then corrects the randomly assigned probabilities for the words using the word distributions corresponding to the word clusters and the positional relationships described above.
[0119] In this manner, for example, for word cluster 157a in document 152a, cluster probability estimation unit 124 generates cluster probability data table 158a as shown in Fig. 8. Similarly, cluster probability estimation unit 124 generates cluster probability data tables for other word clusters in document 152a. Cluster probability estimation unit 124 writes the probability set for each word into the cluster probability data table in memory circuit 104.
[0120] In this way, the cluster probability estimation unit 124 estimates the probability of each word belonging to each of the generated word clusters. The cluster probability estimation unit 124 also estimates the probability of each word using the determined positional relationship and the counted occurrence frequency.
[0121] 1.6 Cluster Estimation Unit 113 The cluster estimation unit 113 (selection means) selects, for each document (language data), from a plurality of word clusters, a representative word cluster including words related to a topic that represents the contents of the document. The cluster estimation unit 113 also selects a representative word cluster from the plurality of word clusters using the probability of each word estimated for each of the plurality of word clusters.
[0122] As described above, the cluster inference unit 113 includes the representative word cluster inference unit 125 .
[0123] In the above example, for each document, multiple word clusters are generated.
[0124] When multiple word clusters are generated for each document, the representative word cluster inference unit 125 selects one word cluster as a representative word cluster from the multiple generated word clusters as follows.
[0125] For example, for one word cluster in document 152a, representative word cluster inference unit 125 multiplies the probabilities included in cluster probability data table 158a for a plurality of words included in noun phrase data 153a.
[0126] 4, noun phrase data 153a includes the words "Diet," "Deliberation," "Olympics," "Related," "Bill," .... Also, according to cluster probability data table 158a, the probabilities of the words "Diet," "Deliberation," "Olympics," "Related," "Bill," ... are P1, P2, P3, P4, P5, ....
[0127] Each probability is expressed as a percentage (%).
[0128] In this case, the total product (index value) is calculated as P1×P2×P3×P4×P5×, . . .
[0129] Here, noun phrase data 153a includes, for example, three occurrences of the word "bill," so that the probability corresponding to the word "bill" is multiplied three times in the calculation of the total product value.
[0130] The representative word cluster estimation unit 125 calculates the total product value for other word clusters for the document 152a in the same manner as described above.
[0131] Thus, for example, a multiplicative value is calculated for multiple word clusters in document 152a.
[0132] In this way, the representative word cluster inference unit 125 generates, for example, for document 152a, a topic distribution 156a for each document as shown in Fig. 6. In topic distribution 156a, the horizontal axis indicates topics, and the vertical axis indicates the total value (index value) for each word cluster, that is, for each topic.
[0133] The representative word cluster estimation unit 125 selects the largest total product value from the multiple 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.
[0134] The representative word cluster estimation unit 125 also calculates the total product value (index value) for each word cluster for the documents 152b and 152c in the same manner as described above.
[0135] For example, for documents 152b and 152c, representative word cluster inference unit 125 generates topic distributions 156b and 156c for each document as shown in Fig. 6. In topic distributions 156b and 156c, the horizontal axis indicates topics, and the vertical axis indicates the total value (index value) for each word cluster, that is, for each topic.
[0136] The representative word cluster estimation unit 125 also selects representative word clusters for the documents 152b and 152c in the same manner as above.
[0137] As described above, the representative word cluster estimation unit 125 calculates the total product value=P1×P2×P3×P4×P5×, . . . However, the present invention is not limited to this method.
[0138] The representative word cluster estimation unit 125 may calculate the sum value (index value) = P1 + P2 + P3 + P4 + P5 +, ..., select the largest sum value from the multiple calculated sum values, and select the word cluster used to calculate the selected sum value as the representative word cluster.
[0139] In this way, the representative word cluster estimation unit 125 of the cluster estimation unit 113 (selection means) sums or multiplies, for each word cluster, the multiple probabilities estimated for each of the multiple words contained in the document (language data), to calculate an index value indicating the likelihood that the word cluster represents the contents of the document, and selects a representative word cluster by comparing the multiple index values calculated for the multiple word clusters.
[0140] Here, the number of representative word clusters selected by the cluster estimation unit 113 (selection means) is smaller than the number of word clusters generated by the word processing unit 112.
[0141] 1.7 Summary 114 The summarizing unit 114 (extraction means) generates a summary based on the representative word clusters selected by the representative word cluster estimation unit 125 from each document (language data).
[0142] In other words, the summarization unit 114 uses the words contained in the representative word cluster to extract sentences from each document that contain the words contained in the representative word cluster (i.e., sentences that represent the contents of the document), and generates a summary that contains the extracted sentences.
[0143] An example of a summary generated by the summarizing unit 114 is shown in Fig. 9. The summary 191 shown in this figure contains all or some of the words contained in the word cluster 157a shown in Fig. 7(a) that is selected as the representative word cluster.
[0144] 1.8 Screenshots An example of a screen displayed on the information processing device 10 is shown in FIG.
[0145] A screen 201 shown in this figure is a screen for requesting input from a user in order to set hyperparameters to be set in the information processing device 10.
[0146] As shown in this figure, a screen 201 includes input fields 202, 203, and 204, and radio buttons 205, 206, and 207-210.
[0147] Input fields 202, 203, and 204 are fields where the user performs input operations to accept input of prior knowledge data, the number of word clusters, and outlier words, respectively. Radio buttons 205 and 206 are radio buttons for setting whether or not to inherit word clusters. Either one of radio buttons 205 and 206 is selected by the user. Radio buttons 207, 208, 209, and 210 are radio buttons for selecting one of document units, paragraph units, multiple sentence units, and single sentence units, respectively, as data units. Either one of radio buttons 207, 208, 209, and 210 is selected by the user.
[0148] The general control unit 110 generates a 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.
[0149] The keyboard 108 and mouse 109 receive operational instructions from the user in the input fields 202 , 203 , 204 and radio buttons 205 to 210 , and output instruction signals corresponding to the received operational instructions to the general control unit 110 via the input / output circuit 105 .
[0150] The general control unit 110 executes a process according to the received instruction signal.
[0151] The prior knowledge data, the number of word clusters, outlier words, whether or not to inherit word clusters, and data units will be described next.
[0152] 1.9 Prior knowledge data Prior knowledge data (prior knowledge information) is data that indicates words related to a topic. In other words, prior knowledge data is used to classify words so that words related to one topic belong to one word cluster. In this way, information indicating which word cluster a word belongs to is given in advance according to the user's preferences.
[0153] As shown in Fig. 11(a) as an example, prior knowledge data 221 is composed of word clusters 222, 223, and 224. Word cluster 222 includes the words "Diet", "Deliberation", and "Law". Word cluster 223 includes the words "Olympics" and "Stadium". Word cluster 224 includes the words "Economy" and "Interest rate".
[0154] The prior knowledge data 221 uses the JSON format as a data format, and describes words to be registered in advance for each word cluster.
[0155] The prior knowledge data is generated in advance by a user, and the generated prior knowledge data is stored in the storage circuitry 104 in response to an operational instruction from the user.
[0156] The word cluster generating unit 123 uses the prior knowledge data stored in the memory circuit 104 to classify words contained in a document and generate a plurality of word clusters.
[0157] The following description will be given using the prior knowledge data 221 shown in FIG.
[0158] If a document contains words "Diet," "Deliberation," and "Law" included in word cluster 222, the word cluster generation unit 123 determines that the document belongs to word cluster 222. Furthermore, if a document contains words "Olympics" and "Stadium" included in word cluster 223, the word cluster generation unit 123 determines that the document belongs to word cluster 223. Furthermore, if a document contains words "Economy" and "Interest rate" included in word cluster 224, the word cluster generation unit 123 determines that the document belongs to word cluster 224.
[0159] In this way, it is possible to prepare prior knowledge data indicating words related to a topic in advance, and use the prior knowledge data as training data to sequentially generate word clusters from the words contained in a document.
[0160] Specifically, if a user requesting a summary of a meeting minutes document participated in the target meeting, the generated summary can be made closer to the one intended by the user by providing word clusters in advance as prior knowledge of words that are considered to be important in accordance with the user's preferences.
[0161] 1.10 Number of word clusters The number of word clusters to be generated may be specified by the user via keyboard 108. Keyboard 108 outputs the received number of word clusters to general control unit 110 via input / output circuit 105.
[0162] The word processing unit 112 receives the number of word clusters from the general control unit 110. The word cluster generating unit 123 generates the designated number of word clusters.
[0163] Examples of situations in which the user may specify the number of word clusters include:
[0164] When a user who has participated in a meeting to be summarized requests the information processing device 10 to provide a summary, the user may be able to clearly specify the number of word clusters in advance. Specifically, this is the case when an agenda is determined in advance at the start of the meeting and the meeting proceeds according to the agenda. For example, a brainstorming session is conducted within a team, and (a) medical-related ideas are discussed for a certain period of time, then (b) architectural ideas are discussed, and finally, the agenda item (b) regarding office solutions is discussed. In this case, the words appearing in each of the topics (a), (b), and (c) are significantly different, so the user may specify three word clusters in advance. The information processing device 10 generates the specified number of word clusters.
[0165] 1.11 Outlier Words The storage circuitry 104 may 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 any words related to the topic desired by the user). In this case, the word cluster generating unit 123 excludes the words indicated by the outlier data when classifying the words included in the document.
[0166] 11B includes, as examples, “Otsukaresama” and “Agenda.” When classifying words included in a document, the word cluster generating unit 123 excludes the words “Otsukaresama” and “Agenda” indicated by the outlier data.
[0167] Examples of cases where such words appear include sentences that appear routinely in meetings, such as "Today's agenda is as follows," and "This concludes the meeting. Thank you for your hard work." The word "agenda" in "Today's agenda is as follows," and the word "Thank you for your hard work" in "This concludes the meeting. Thank you for your hard work" are considered to be unrelated to the topic of the meeting, and should be excluded from classification.
[0168] In this way, when classifying words in a document so that related words belong to the same word cluster, words that are not similar to any other words in the document may be treated as outliers and not taken into consideration.
[0169] 1.12 Inherit or not inherit word clusters There are cases where a single topic is discussed in multiple consecutive meetings. For example, in a weekly development meeting or a regular meeting, the word clusters generated from the document that represents the contents of the first meeting can be used again to generate word clusters from the document that represents the contents of the second meeting.
[0170] Therefore, in response to an operational instruction from the user, the word cluster generating unit 123 writes and stores the word clusters generated from the document expressing the contents of the first conference in the memory circuit 104.
[0171] In this case, the cluster estimation unit 113 may write the estimated representative word cluster into the memory circuitry 104 .
[0172] In response to a user's operational instruction, the word cluster generation unit 123 may use the word cluster of the first meeting stored in the memory circuit 104 when generating word clusters from a document expressing the contents of the second meeting.
[0173] By using word clusters generated and stored from previous meetings, it is possible to refer to word clusters that were deemed important in previous meetings and extract more important information (summary sentences) for the current meeting.
[0174] Alternatively, a designation of a word cluster to be deleted may be received from a user via keyboard 108 (reception means). Keyboard 108 outputs the received designation of a word cluster to be deleted to general control unit 110 via input / output circuit 105.
[0175] The word processing unit 112 receives the designation of the accepted word cluster to be deleted from the general control unit 110. The word cluster generating unit 123 deletes the designated word cluster from the memory circuit 104.
[0176] An example of a case where information on a word cluster is deleted is a one-off meeting. In a situation such as a brainstorming session or a chat, the discussion may be completed within the meeting. In such a case, the word cluster stored in the memory circuit 104 may be deleted by an operational instruction from the user.
[0177] In addition, in the case where a plurality of word clusters are generated and stored in the memory circuit 104 in the first meeting of successive meetings, if a user clearly knows that a word cluster among the plurality of word clusters cannot be used in the next meeting, that word cluster may be deleted from the memory circuit 104 in response to an operational instruction from the user.
[0178] 1.13 Data Units The word cluster generation unit 123 may treat the entire language data consisting of multiple documents, the entire document included in the language data, a paragraph included in a document, multiple sentences included in a document, or a single sentence included in a document as a data unit, classify the words included in each data unit, and generate multiple word clusters for each data unit. The data unit is a unit for determining a topic.
[0179] The cluster estimation unit 113 selects a representative word cluster from a plurality of word clusters for each data unit.
[0180] Also, a data unit designation may be accepted from the user through keyboard 108 (accepting means). Keyboard 108 outputs the accepted data unit designation to general control unit 110 via input / output circuit 105.
[0181] The word cluster generating unit 123 classifies words for each data unit received from the user.
[0182] In this embodiment, in order to classify topics based on discussions in one meeting, the data unit may be a specified number of sentences or a single sentence, as described above. Also, in order to be able to handle discussions that suddenly reoccur, text may be classified into clusters for each sentence.
[0183] Furthermore, summarization section 114 may extract, for each data unit, a sentence representative of the description content of that data unit from that data unit as a summary sentence.
[0184] 1.14 Operation of the information processing device 10 The operation of the information processing device 10 will be described with reference to a flowchart.
[0185] (1) Overview of the overall operation of the information processing device 10 The overall operation of the information processing device 10 will be described with reference to the flowchart shown in FIG.
[0186] The overall control unit 110 waits for the user to input hyperparameter settings (step S101). If the hyperparameter settings have not been input ("NO" in step S102), the overall control unit 110 transfers control to step S101 and repeats the process.
[0187] If the hyperparameter settings have been input ("YES" in step S102), the general control unit 110 refers to the setting data (step S103).
[0188] The voice recognition unit 111 acquires text data converted from the voice data by voice recognition (step S104).
[0189] The morpheme analysis unit 121 performs morpheme analysis to generate morphemes (step S105).
[0190] The part-of-speech filtering unit 122 extracts noun phrases from the morphemes (step S106).
[0191] The word cluster generation unit 123 generates word clusters, and the cluster probability estimation unit 124 estimates the probability for each word (step S107).
[0192] The cluster estimation unit 113 selects a representative word cluster from among a plurality of word clusters for each data unit (step S108).
[0193] The summarizing unit 114 generates a summary for each document using the representative word cluster selected by the representative word cluster estimation unit 125 (step S109).
[0194] This completes the description of the operation of the information processing device 10.
[0195] (2) Operation of referencing setting data The operation of referring to the setting data will be described with reference to the flowchart shown in FIG.
[0196] The operation described here is a detailed procedure of step S103 in FIG.
[0197] The general control unit 110 refers to the setting data (hyperparameters) set by the user (step S131).
[0198] The central control unit 110 judges whether or not the prior knowledge data is set (step S132). If the prior knowledge data is set ("YES" in step S132), the central control unit 110 writes the prior knowledge data into the storage circuit 104 (step S133).
[0199] Next, the central control unit 110 determines whether the number of word clusters has been set by the user (step S134). If the number of word clusters has been set (YES in step S134), the central control unit 110 writes the number of word clusters into the memory circuit 104 (step S135).
[0200] Next, the general control unit 110 writes the data unit into the memory circuit 104 (step S136).
[0201] This concludes the explanation of the operation for referring to setting data.
[0202] (3) Treatment of outliers The process of the outlier will be described with reference to the flowchart shown in FIG.
[0203] The word cluster generating unit 123 refers to setting data (hyperparameters) set by the user (step S151).
[0204] Next, the word cluster generating unit 123 determines whether or not an outlier word has been set (step S152). If an outlier word has been set ("YES" in step S152), the word cluster generating unit 123 deletes the outlier word from the word cluster stored in the memory circuitry 104 (step S153).
[0205] This concludes the explanation of outlier processing.
[0206] (4) Inheritance of word clusters The word cluster inheritance process will be described with reference to the flowchart shown in FIG.
[0207] The general control unit 110 refers to the setting data (hyperparameters) set by the user (step S171).
[0208] Next, the central control unit 110 judges whether or not a setting is made to inherit the already generated word clusters (step S172). If the setting is made to inherit ("YES" in step S172), the central control unit 110 maintains the state in which the generated word clusters are stored in the storage circuit 104 (step S173). If the setting is not made to inherit ("NO" in step S172), the central control unit 110 deletes the word clusters present in the storage circuit 104. In this case, all word clusters present in the storage circuit 104 may be deleted, or only the word clusters designated by the user may be deleted (step S174).
[0209] This completes the explanation of the word cluster inheritance process.
[0210] 1.15 Example (1) In the above embodiment, the voice recognition unit 111 performs voice recognition processing on the voice data to generate text data.
[0211] However, the present invention is not limited to this form.
[0212] One aspect of the present disclosure may be a system including an information processing device 10 and a server device. The information processing device 10 and the server device are connected via a network. The server device provides voice recognition processing as one of the cloud services. In other words, the server device receives voice data and converts the received voice data into text data to generate a document.
[0213] The information processing device 10 may use a voice recognition process provided by a server device.
[0214] The information processing device 10 includes a network communication circuit connected to a server device via a network. The network communication circuit transmits voice data to the server device under the control of the general control unit 110, and requests the server device to perform voice recognition processing on the voice data.
[0215] The server device includes a network communication circuit (communication means) and a voice recognition circuit (voice recognition means). The voice recognition circuit of the server device has a similar configuration to the voice recognition unit 111 of the information processing device 10.
[0216] The network communication circuit of the server device receives, via the network, from the information processing device 10, the voice data together with a request for voice recognition processing of the voice data. Upon receiving the voice data together with the request, the voice recognition circuit converts the received voice data into text data and generates a document consisting of the text data. The network communication circuit of the server device transmits the generated document to the information processing device 10 via the network.
[0217] The network communication circuit of the information processing device 10 receives a document from the server device and writes the received document into the memory circuit 104 .
[0218] 1.16 Example (2) The morphological analysis unit 121 of the information processing device 10 may use a known morphological analysis method. For example, the morphological analysis unit 121 may use any one of the known MeCab, JUMAN, KyTea, and ChaSen.
[0219] 1.17 Example (3) The word cluster generation unit 123 and the cluster probability estimation unit 124 may use various known natural language processing techniques. For example, LDA is a method for estimating latent topics of text data from words that appear in the text data. LDA is a document classification model that assumes that the text data has multiple topics. This method estimates latent topics by sequentially learning the frequency of appearance and positional relationships of words only from the target text data.
[0220] A latent topic is a set of important words clustered by topic, and each word has a topic probability (simply, probability) that indicates how likely it is to belong to that topic. Using these latent topics, sentences are classified into each latent topic.
[0221] In this way, the word cluster generating unit 123 generates a plurality of word clusters. A word cluster is a set of words obtained by clustering words as elements that can constitute a topic.
[0222] In addition, the cluster probability estimation unit 124 obtains the probability that each of the clustered words belongs to a topic.
[0223] 1.18 Example (4) The summarization unit 114 may utilize various known natural language processing techniques.
[0224] For example, LexRank is a method for extracting important sentences by scoring potentially important sentences from text data based on words that appear in the text data.
[0225] When applying LexRank, for the words included in the representative word cluster, the eigenvector centrality of the words that appear in the text data is calculated using a graph representation to calculate the relative importance of each text unit. In other words, words that appear frequently in other sentences and are similar to important words are considered important, and sentences in which those important words appear can be considered important.
[0226] By generating an abstract using this method, a summary sentence that is characterized by the word clusters can be obtained.
[0227] 1.19 Example (5) There are cases where multiple word clusters are generated for each data unit from a document that is the subject of word cluster generation.
[0228] When a single data unit originally contains multiple topics, multiple word clusters are generated from the data unit.
[0229] Furthermore, when multiple word clusters are similar to each other, multiple word clusters are generated from one data unit. For example, consider a word cluster related to neuroscience and a word cluster related to AI. There is a high probability that the words belonging to both word clusters contain the same ones, so it can be said that the word clusters related to neuroscience and the word cluster related to AI are similar to each other.
[0230] 1.20 Example (6) Before generating a summary from a document, the user may specify the amount of data in the summary (i.e., the amount of characters in the summary). In other words, the general control unit 110 can receive the amount of data in the summary from the keyboard 108 via the input / output circuit 105 in response to an instruction from the user.
[0231] The summarizing unit 114 generates a summary within the range of the specified data amount of the summary.
[0232] For example, if summarization unit 114 initially generates a summary of 500 characters consisting of five sentences, and the user has specified 300 characters as the character length, summarization unit 114 deletes one of the five sentences. Summarization unit 114 counts the number of characters in the summary after the one sentence has been deleted. If the number of characters in the summary after the deletion is within 300 characters, summarization unit 114 confirms the summary after the one sentence has been deleted. On the other hand, if the number of characters in the summary after the deletion exceeds 300 characters, summarization unit 114 further deletes one sentence. In this manner, the above process is repeated until the number of characters in the summary after the deletion is within the 300 characters specified by the user.
[0233] In this way, by allowing the user to appropriately specify the amount of data in the summary text, it is possible to prevent situations in which the amount of data in the summary text is too large and it does not fulfill its role as a summary, which should be concise, or the amount of data in the summary text is too small and the main points of the discussion cannot be understood from the summary text.
[0234] The overall control unit 110 may receive the number of sentences in the summary from the keyboard 108 via the input / output circuit 105 in response to a user's operation instruction. In this case, the summarizing unit 114 generates a summary within the range of the specified number of sentences.
[0235] In this case, similarly to the above, if the number of sentences exceeds the number specified by the user, summarization section 114 repeats deleting sentences from the summary until the number becomes equal to or less than the number specified by the user.
[0236] 1.21 Example (7) As described above, the word processing unit 112 may treat the entire language data consisting of multiple documents, the entire document contained in the language data, a paragraph contained in a document, multiple sentences contained in a document, or a single sentence contained in a document as a data unit, and classify the words contained in each data unit.
[0237] The cluster estimation unit 113 (selection means) may include a cluster analysis unit (analysis means) that determines the importance of each representative word cluster.
[0238] Furthermore, the cluster analysis unit may count, for each representative word cluster, the number or amount of data units that correspond to that representative word cluster, and determine the importance of the representative word cluster according to the count value for each representative word cluster.
[0239] Furthermore, when the count value of one representative word cluster exceeds a predetermined value, the cluster analysis unit may set the importance of the representative word cluster to a predetermined maximum value.
[0240] (1) For example, if there are eight data units corresponding to representative word cluster a, four data units corresponding to representative word cluster b, and two data units corresponding to representative word cluster c, the cluster analysis unit may set the importance of each representative word cluster for representative word clusters a, b, and c such that the importance of representative word cluster a > the importance of representative word cluster b > the importance of representative word cluster c.
[0241] That is, the cluster analysis unit may set the importance of a representative word cluster according to the number of data units corresponding to the representative word cluster, in which case the more the number of data units, the higher the importance of the representative word cluster is set, and the fewer the number of data units, the lower the importance of the representative word cluster is set.
[0242] In this way, the importance level set for each representative word cluster may be presented to the user. That is, the cluster analysis unit outputs the importance level set for each representative word cluster to the monitor 107 via the general control unit 110 and the input / output circuit 105. The monitor 107 outputs the importance level set for each representative word cluster.
[0243] (2) Even if the cluster analysis unit sets a low importance for a certain representative word cluster, if the user determines that the representative word cluster is important to the user, the user may set the importance of the representative word cluster to be high.
[0244] Conversely, even if the cluster analysis unit sets a high importance for a certain representative word cluster, if the user determines that the representative word cluster is of low importance to the user, the importance of the representative word cluster may be set to low.
[0245] In this way, the importance of the representative word cluster may be changed according to the user's intention.
[0246] The general control unit 110 may accept a change in the importance of the representative word cluster from the keyboard 108 via the input / output circuit 105 in response to an operational instruction from the user.
[0247] The cluster analysis unit changes the importance of the representative word cluster to the importance received in response to an operational instruction from the user.
[0248] In the case of example (1), if the user determines that the importance of representative word cluster c is the highest, the cluster analysis unit changes the importance settings of each representative word cluster for representative word clusters a, b, and c, such that the importance of representative word cluster c > the importance of representative word cluster a > the importance of representative word cluster b.
[0249] In this way, the importance of each representative word cluster is changed to the importance intended by the user, and as a result, a more appropriate representative word cluster that reflects the user's intention can be selected.
[0250] (3) The amount of data of the summary sentence described in the embodiment (6) or the number of sentences may be varied according to the importance of the representative word cluster. That is, the summarization unit 114 may vary the amount of data of the summary sentence according to the determined importance. Here, the summarization unit 114 may vary the amount of characters included in the summary sentence or the number of sentences included in the summary sentence.
[0251] For example, when the importance of the representative word cluster is high, the amount of data in the summary sentence or the number of sentences may be increased, whereas when the importance of the representative word cluster is low, the amount of data in the summary sentence or the number of sentences may be decreased.
[0252] As in the case of example (1), if the importance of each representative word cluster is set as follows: importance of representative word cluster a > importance of representative word cluster b > importance of representative word cluster c, the summary sentence of representative word cluster a will be four sentences, the summary sentence of representative word cluster b will be two sentences, and the summary sentence of representative word cluster c will be one sentence.
[0253] In this way, the amount of data or number of sentences in the summary text is made variable depending on the importance of the representative word cluster. This is because major representative word clusters contain a large amount of information, and this increases the likelihood of being able to provide the information (summary text) that the user intends to obtain regarding such representative word clusters.
[0254] 1.22 Example (8) As described above, the word cluster generating unit 123 uses, as a data unit, any of the following: the entire language data consisting of multiple documents, the entire document included in the language data, a paragraph included in a document, multiple sentences included in a document, and one sentence included in a document. Here, a specific description will be given of the cases where the entire document and a paragraph included in a document are used as data units.
[0255] (1) When the entire document is treated as a data unit The case where the entire document is treated as a data unit will be described with reference to FIG.
[0256] As shown in this figure, assume that the target of summary generation is language data consisting of multiple documents 301, 302, 303, 304, and 305. Also assume that the cluster estimation unit 113 has selected representative word clusters with the topics "politics," "sports," "economy," "politics," and "economy" for each of the multiple documents 301, 302, 303, 304, and 305.
[0257] In this case, since it is estimated that the topic of each of the two documents 301 and 304 is "politics," the summarization unit 114 may generate one summary sentence from the documents 301 and 304. In this case, the generated summary sentence is made up of one or more sentences.
[0258] In this way, when multiple documents are estimated to have the same topic, a single summary is generated from these multiple documents, so that the generated summary can be an appropriate and concise representation of the contents of the multiple documents that have the same topic.
[0259] The information processing device 10 (summary generation device) may generate a summary sentence from language data consisting of a plurality of documents. The cluster estimation unit 113 may select, for each of the plurality of documents, from a plurality of word clusters, a representative word cluster including words related to a topic that represents the contents of the document. When there are a plurality of topic documents that serve as the basis for generating representative word clusters including words related to the same topic, the summarization unit 114 may generate a summary sentence from the plurality of topic documents based on the representative word cluster.
[0260] (2) When the paragraphs in a document are treated as data units The case where a paragraph included in a document is treated as a data unit will be described with reference to FIG.
[0261] As shown in this figure, the target of summary generation is language data consisting of multiple documents 311, 321, and 331. Also, document 311 is composed of multiple paragraphs 312, 313, 314, 315, and 316; document 321 is composed of multiple paragraphs 322, 323, 324, 325, and 326; and document 331 is composed of multiple paragraphs 332, 333, 334, 335, and 336.
[0262] Here, it is assumed that the cluster estimation unit 113 has selected representative unit clusters with the topics "politics", "economy", "economy", "politics" and "economy" for paragraphs 312, 313, 314, 315 and 316 of document 311, respectively.
[0263] It is also assumed that the cluster estimation unit 113 has selected representative unit clusters with the topics "sports", "economy", "sports", "economy" and "sports" for paragraphs 322, 323, 324, 325 and 326 of document 321, respectively.
[0264] Furthermore, it is assumed that the cluster estimation unit 113 has selected representative unit clusters with the topics "economy", "economy", "politics", "economy" and "sports" for paragraphs 332, 333, 334, 335 and 336 of document 331, respectively.
[0265] In this case, since it is estimated that the topic of paragraphs 312 and 315 of document 311 and paragraph 334 of document 331 is "politics," summarization unit 114 may generate one summary sentence from paragraphs 312 and 315 of document 311 and paragraph 334 of document 331. In this case, the generated summary sentence consists of one or more sentences.
[0266] In this way, for multiple documents each containing multiple paragraphs, if there are multiple paragraphs that are estimated to have the same topic, a single summary sentence is generated from these multiple paragraphs, so that the generated summary sentence can be an appropriate one that concisely represents the content of the multiple paragraphs that have the same topic.
[0267] In addition, in a case where a document includes a plurality of paragraphs, and a plurality of paragraphs are estimated to have the same topic, a single summary sentence may be generated from these paragraphs.
[0268] The word processing unit 112 may classify the words included in each data unit and generate a plurality of word clusters for each data unit. The cluster estimation unit 113 may select a representative word cluster from the plurality of word clusters for each data unit.
[0269] As described above, the summarizing section 114 may extract a summary from a plurality of data units.
[0270] Furthermore, the summarizing section 114 may extract summaries from a plurality of data units of a plurality of documents.
[0271] 1.23 Summary As described above, by using the cluster estimation unit 113 to select a representative word cluster that contains words related to a topic that represents the contents of the document from the multiple word clusters that have been generated, a summary that expresses the representative topic can be generated using the representative word cluster. [Industrial Applicability]
[0272] INDUSTRIAL APPLICABILITY The summary generation device according to the present disclosure has an effect of being able to generate a summary sentence expressing a main topic, and is useful as a technique for summarizing a document. [Explanation of symbols]
[0273] 10. Information processing device 100 Main unit 101 CPU 102 ROM 103 RAM 104 Memory circuit 105 Input / Output Circuit 106 Control section 107 Monitor 108 Keyboard 109 Mouse 110 General control section 111 Voice Recognition Unit 112 Word Processing Unit 113 Cluster Estimation Unit 114 Summary 121 Morphological analysis section 122 Part of speech filtering section 123 Word Cluster Generation Unit 124 Cluster Probability Estimation Unit 125 Representative word cluster estimation unit
Claims
1. A summary generation device for generating a summary sentence from language data, comprising: a word processing means for inferring a plurality of topics from a plurality of words included in the language data, generating a word cluster for each topic that includes a plurality of words that may be related to the topic, and estimating a probability indicating a likelihood that each word belonging to each generated word cluster belongs to the topic corresponding to the word cluster; a selection means for selecting, from the plurality of word clusters, a representative word cluster including words related to a topic representative of the content described in the language data, using the probability of each word estimated for each of the plurality of word clusters; a generation means for generating a summary sentence from the language data based on the representative word cluster; an analysis means for determining the importance of each representative word cluster; Equipped with The generating means varies the amount of data of the summary sentence according to the determined importance. A summary generating device comprising:
2. A summary generation device for generating a summary sentence from language data, comprising: a word processing means for inferring a plurality of topics from a plurality of words included in the language data, generating a word cluster for each topic that includes a plurality of words that may be related to the topic, and estimating a probability indicating a likelihood that each word belonging to each generated word cluster belongs to the topic corresponding to the word cluster; a selection means for selecting, from the plurality of word clusters, a representative word cluster including words related to a topic representative of the content described in the language data, using the probability of each word estimated for each of the plurality of word clusters; a generating means for generating a summary sentence from the language data based on the representative word cluster; Equipped with The language data includes a plurality of documents, the selection means selects, for each of the plurality of documents, from the plurality of word clusters, a representative word cluster including words related to a topic representative of the contents of the document; When there are a plurality of topic documents from which representative word clusters including words related to the same topic are generated, the generating means generates a summary sentence from the plurality of topic documents based on the representative word clusters. A summary generating device comprising:
3. The selection means calculates, for each word cluster, a sum or product of the probabilities estimated for each of the multiple words included in the language data, to calculate an index value indicating the likelihood that the word cluster represents the content of the language data, and selects the representative word cluster by comparing the multiple index values calculated for the multiple word clusters.
3. An apparatus for generating a summary according to claim 1 or 2.
4. The word processing means includes: a morpheme analysis means for performing a morpheme analysis on the language data to generate a plurality of morphemes and estimating a part of speech of each morpheme; a part-of-speech filtering means for extracting words which are nouns from the plurality of morphemes generated by the morphological analysis means; a word cluster generating means for classifying the extracted words to generate the plurality of word clusters; a probability estimation means for estimating the probability of each word belonging to each of the generated word clusters; 3. The apparatus according to claim 1, further comprising:
5. The word cluster generating means obtains a positional relationship between words in the language data, counts, for each word, an occurrence frequency of the words extracted by the part-of-speech filtering means, and generates the plurality of word clusters using the obtained positional relationship and the counted occurrence frequency.
5. Apparatus according to claim 4, characterized in that:
6. Further, a voice recognition means for converting voice data to generate the language data is provided, The word processing means generates the plurality of word clusters from the generated linguistic data.
3. An apparatus for generating a summary according to claim 1 or 2.
7. The system further includes a storage means for storing prior knowledge information indicating words related to one topic in advance, The word processing means uses the prior knowledge information to collect words included in the language data by topic.
3. An apparatus for generating a summary according to claim 1 or 2.
8. The method further includes a reception unit for receiving a designation of the number of word clusters to be generated from a user, The word processing means generates a specified number of word clusters.
3. An apparatus for generating a summary according to claim 1 or 2.
9. The system further includes a storage means for storing outlier information indicating words that are not related to a topic desired by a user in advance, The word processing means excludes words indicated by outlier information when collecting the words included in the language data by topic.
3. An apparatus for generating a summary according to claim 1 or 2.
10. The language data includes a plurality of documents, the word processing means treats any one of 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 a single sentence included in the document as a data unit, and for each data unit, collects words included in the data unit by topic and generates the plurality of word clusters for each data unit; The selection means selects the representative word cluster from the plurality of word clusters for each data unit.
3. An apparatus for generating a summary according to claim 1 or 2.
11. The apparatus further includes a receiving means for receiving a designation of the data unit from a user, The word processing means collects words for each data unit received from a user.
11. Apparatus according to claim 10, characterized in that:
12. The generating means generates the summary from each data unit.
11. Apparatus according to claim 10, characterized in that:
13. A summary generation device for generating a summary sentence from language data, comprising: a word processing means for inferring a plurality of topics from a plurality of words included in the language data, generating a word cluster for each topic that includes a plurality of words that may be related to the topic, and estimating a probability indicating a likelihood that each word belonging to each generated word cluster belongs to the topic corresponding to the word cluster; a selection means for selecting, from the plurality of word clusters, a representative word cluster including words related to a topic representative of the content described in the language data, using the probability of each word estimated for each of the plurality of word clusters; a generation means for generating a summary sentence from the language data based on the representative word cluster; an analysis means for determining the importance of each representative word cluster; A display means; A reception means for receiving an input from a user, the display means displays the determined importance for each representative word cluster; the accepting means accepts a change in importance from a user for each representative word cluster; The analysis means changes the importance of the representative word cluster to the importance accepted by the user, The generating means varies the amount of data of the summary sentence or the number of documents according to the importance. A summary generating device comprising:
14. The number of the representative word clusters selected by the selection means is less than the number of the word clusters generated by the word processing means.
3. An apparatus for generating a summary according to claim 1 or 2.
15. the word processing means treats any one of 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 a single sentence included in the document as a data unit, and for each data unit, collects words included in the data unit by topic and generates the plurality of word clusters for each data unit; the selection means selects the representative word cluster from the plurality of word clusters for each data unit; The generating means generates the summary sentence from a plurality of data units in the plurality of topic documents.
3. Apparatus according to claim 2, characterized in that:
16. A system comprising the summary generation device according to claim 1, 2 or 13 and a server device for generating language data from speech data, The server device includes: a communication means for receiving voice data and transmitting language data generated from the received voice data to the summary generation device; a voice recognition means for converting the received voice data to generate the language data; A system comprising:
17. 1. A control method for use in a summary generation device that generates a summary sentence from language data, comprising: a word processing step of inferring a plurality of topics from a plurality of words included in the language data, generating a word cluster for each topic that includes a plurality of words that may be related to the topic, and estimating a probability indicating a likelihood that each word belonging to each generated word cluster belongs to the topic corresponding to the word cluster; a selection step of selecting, from the plurality of word clusters, a representative word cluster including words related to a topic representative of the content described in the language data, using the probability of each word estimated for each of the plurality of word clusters; a generation step of generating a summary sentence from the language data based on the representative word cluster; and an analysis step of determining, for each representative word cluster, a level of importance of the representative word cluster; The generating step varies the amount of data of the summary sentence according to the determined importance. A control method comprising:
18. A control method for use in a summary generation device that generates a summary sentence from language data, comprising: a word processing step of inferring a plurality of topics from a plurality of words included in the language data, generating a word cluster for each topic that includes a plurality of words that may be related to the topic, and estimating a probability indicating a likelihood that each word belonging to each generated word cluster belongs to the topic corresponding to the word cluster; a selection step of selecting, from the plurality of word clusters, a representative word cluster including words related to a topic representative of the content described in the language data, using the probability of each word estimated for each of the plurality of word clusters; generating a summary sentence from the language data based on the representative word cluster; Including, The language data includes a plurality of documents, The selection step includes selecting, for each of the plurality of documents, a representative word cluster including words related to a topic that represents the contents of the document from the plurality of word clusters; In the generating step, when there are a plurality of topic documents from which representative word clusters including words related to the same topic are generated, a summary is generated from the plurality of topic documents based on the representative word clusters. A control method comprising:
19. A control method for use in a summary generation device that generates a summary sentence from language data, comprising: a word processing step of inferring a plurality of topics from a plurality of words included in the language data, generating a word cluster for each topic that includes a plurality of words that may be related to the topic, and estimating a probability indicating a likelihood that each word belonging to each generated word cluster belongs to the topic corresponding to the word cluster; a selection step of selecting, from the plurality of word clusters, a representative word cluster including words related to a topic representative of the content described in the language data, using the probability of each word estimated for each of the plurality of word clusters; a generation step of generating a summary sentence from the language data based on the representative word cluster; an analysis step of determining the importance of each representative word cluster; displaying the determined importance for each representative word cluster; receiving a change in importance from a user for each representative word cluster; Including, The analyzing step includes changing the importance of the representative word cluster to an importance accepted by a user; The generating step varies the amount of data of the summary sentence or the number of documents according to the importance. A control method comprising:
Citation Information
Patent Citations
Automatic abstract generation method based on concept semantic unit
CN105320642A
Keyword extraction
JP2002245061A
Summary sentence generation device and program
JP2012203460A
Techniques for user-centered document summarization.
JP2019514120A
System and Method for Analysis of Discussion
KR1020200114214A
