Speech-to-text summarization system and method

Abstract

To provide a technology to create an easy-to-read summary by extracting a sentence from a voice text with a mix of general and specialized contents including inquiries at a customer center and so on.SOLUTION: A voice text summarization system according to the present invention which comprises a voice text summarization server has: a sentence extraction unit with a general extraction type configured in that a voice text summarization server extracts a sentence from a summarization object voice text consisting of multiple sentences in order while maintaining the order relation, using a general extraction type; a sentence extraction unit configured in that the voice text summarization server extracts the sentence from the summarization object voice text while maintaining the order relation, using a specialized extraction type; and a summarization voice text creating unit of combining the sentence extracted by the sentence extraction unit using the general extraction type and the sentence extracted by the sentence extraction unit using the specialized extraction type while maintaining the order relationship, to create the summarized voice text.SELECTED DRAWING: Figure 1

Description

[Technical Field]

[0001] The present invention relates to a system and method for summarizing speech to text. [Background technology]

[0002] Summaries, which briefly summarize the important parts of a text, are used in a variety of text genres, including news articles, papers, and books. Reading a summary allows you to grasp the main points in a short amount of time without having to read the entire text, so there is a high demand for summaries. It is generally known that there are two types of summaries: extractive and generative. Extractive algorithms are algorithms that create summaries by extracting representative sentences (or words) from the target text that fully explain the content of the text. On the other hand, generative algorithms are algorithms that create summaries using words and syntax that do not exist in the target text. Recently, methods have also been developed that extract important sentences and words using learning models created through machine learning.

[0003] For example, Patent Document 1 discloses a text summarization system that creates summary text from text to be summarized, and includes a learning preprocessing unit that performs predetermined preprocessing on a plurality of training texts to create training data, a learning model creation unit that creates a learning model for summarization using machine learning based on the training data, a preprocessing unit that performs predetermined preprocessing on the text to be summarized, a summary creation unit that creates summary text from the preprocessed text to be summarized based on the learning model, and a postprocessing unit that performs predetermined postprocessing on the summary text and outputs it, with the preprocessing involving processing such as replacing words contained in the training text and the text to be summarized with predetermined symbols, and the postprocessing involving processing the summary text by restoring the symbols replaced by the preprocessing unit to the original words, thereby reducing the processing load. [Prior art documents] [Patent documents]

[0004] [Patent Document 1] Japanese Patent Application Publication No. 2019-16181 Summary of the Invention [Problem to be solved by the invention]

[0005] The method disclosed in Patent Document 1, which involves preprocessing to replace words such as numbers and product names in text with symbols to create a learning model and a summary, and then restoring the symbols to the original words in postprocessing at the summary creation stage, is considered effective for texts such as economic news articles, which contain generally typical words. However, questions and answers at customer centers and other such places are based on immediate conversations between customers and operators, and tend to be a complex mix of sentences containing many general words intended to facilitate the conversation and many technical terms related to the products and functions of the inquiries. For such atypical text, the method of symbolizing words in preprocessing and then restoring them to the original words in postprocessing for summary creation is not effective. Therefore, an object of the present invention is to provide a technology for extracting sentences from speech text that contains a mixture of general and specialized content, such as inquiries at a customer center, and creating an easy-to-read summary. [Means for solving the problem]

[0006] In order to solve the above problems, one representative speech text summarization system of the present invention is a speech text summarization system equipped with a speech text summarization server, which has a general extraction type sentence extraction unit that extracts sentences from speech text to be summarized, which consists of a plurality of sentences arranged in order, using a general extraction type sentence extraction method while maintaining the order relationships, a specialized extraction type sentence extraction unit that extracts sentences from the speech text to be summarized using a specialized extraction type sentence extraction method while maintaining the order relationships, and a summarized speech text creation unit that combines the sentences extracted by the general extraction type sentence extraction unit and the sentences extracted by the specialized extraction type sentence extraction unit while maintaining the order relationships, to create a summarized speech text. [Effects of the Invention]

[0007] According to the present invention, it is possible to extract sentences from speech text that contains a mixture of general and specialized content, such as inquiries at a customer center, and create an easy-to-read summary. Problems, configurations, and effects other than those described above will become apparent from the following description of the preferred embodiments. [Brief explanation of the drawings]

[0008] [Figure 1] FIG. 1 is a system configuration diagram of a speech text summarization system according to the first embodiment. [Figure 2] FIG. 2 is a functional configuration diagram of the main unit. [Figure 3] FIG. 3 is a functional configuration diagram of the voice information management server. [Figure 4] FIG. 4 shows an example of information content stored in the voice information storage unit. [Figure 5] FIG. 5 is a functional block diagram of the speech text summarizing server. [Figure 6] FIG. 6 is an example of a list of learning models stored in the learning model storage unit 47 for specialized knowledge. [Figure 7] FIG. 7 shows an example of the sequence operation of the speech-to-text summarization system. [Figure 8] Figure 8 is a flowchart for creating a learning model of specialized knowledge. [Figure 9] FIG. 9 is a flowchart of the speech-to-text summarization process. [Figure 10] FIG. 10 shows the entire speech text to be summarized in the embodiment. [Figure 11] FIG. 11 shows the summarized speech text in the embodiment. DETAILED DESCRIPTION OF THE INVENTION

[0009] Hereinafter, an embodiment of the present invention will be described with reference to the drawings. Note that the present invention is not limited to this embodiment. In addition, in the description of the drawings, the same parts are designated by the same reference numerals.

[0010] [First embodiment] <Speech-to-text summarization system> FIG. 1 is a system configuration diagram of a speech text summarization system according to the first embodiment. The speech text summarization system 1 comprises a main unit 12, a speech recognition server 13, a speech information management server 14, and a speech text summarization server 15. An outside telephone 10 is connected to an extension telephone 11 via the main unit 12 so that they can communicate with each other. Taking a company's customer center as an example, the outside telephone 10 is located on the customer side, and the extension telephone 11 is located on the operator (company) side. The main unit 12 works in conjunction with the extension telephone 11 to establish or release a call path between the extension telephone 11 and the outside telephone 10, or between the extension telephones 11, through call control. It also saves recorded information during calls and transmits the recorded information to the speech information management server 14. The extension telephone 11 works in conjunction with the main unit 12 to connect to the outside telephone 10 or between the extension telephones 11 so that they can communicate with each other.

[0011] The speech recognition server 13 creates speech text from recorded information (speech data). The speech information management server 14 manages information on conversations between the external telephone 10 and the internal telephone 11 and their summaries (recorded information, speech text, summarized speech text). The speech text summarization server 15 creates summarized speech text from speech text using two sentence extraction methods (general extraction type sentence extraction and specialized extraction type sentence extraction), as described below.

[0012] The operation terminal 16 is a terminal operated by an operator and is connected to the voice information management server 14 via a web browser or the like. The LAN 17 is an internal network under the control of the main unit 12. The Internet 18 is an external network that connects the outside telephones 10 and various servers.

[0013] Some of the components that make up the speech text summarization system of this embodiment will be further described below. (Main device) FIG. 2 is a functional configuration diagram of the main device 12. The WAN interface unit 20 is an interface through which the main unit 12 connects to the Internet 18. The LAN interface unit 21 is an interface through which the main unit 12 connects to the LAN 17. The call control unit 22 establishes and releases a call path between the extension telephone 11 and the outside telephone 10, or between the extension telephones 11. The recorded information storage unit 23 stores recorded information (voice data) during a call. When the call ends, the recorded information transmission unit 24 transmits the stored recorded information during the call to the voice information management server 14. After the recorded information is transmitted, the recorded information in the recorded information storage unit 23 may be erased.

[0014] (Voice information management server) 3 is a functional configuration diagram of the voice information management server 14. The functions of the voice information management server 14 may be incorporated into the main unit 12. The network interface unit 30 is an interface that connects the voice information management server 14 to the Internet 18. The operation terminal interface unit 31 is an interface that an operator operates using a web browser or the like. For example, the operator can view and operate the recorded information stored in the recorded information storage unit 33, and can request the creation of a voice text or a summarized voice text.

[0015] The recorded information receiving unit 32 receives recorded information transmitted by the main unit 12 from the recorded information transmitting unit 24. The recorded information storage unit 33 stores the recorded information received by the recorded information receiving unit 32. The speech text creation request transmitting unit 34 is operated by the operator via the operation terminal interface unit 31 to send a speech text creation request (a request to create speech text from recorded information (speech data)) to the speech recognition server 13. The speech text creation result receiving unit 35 receives the speech text creation request result sent by the speech recognition server 13. The speech text storage unit 36 ​​stores the speech text received by the speech text creation result receiving unit 35. The summarized speech text creation request transmitting unit 37 is operated by the operator via the operation terminal interface unit 31 to send a summarized speech text creation request (a request to create summarized speech text from recorded information (speech text)) to the speech text summarization server 15. The summarized speech text creation result receiving unit 38 receives the summarized speech text creation result sent by the speech text summarization server 15. The summarized speech text storage unit 39 stores the summarized speech text received by the summarized speech text creation result receiving unit 38. The speech information storage unit 40 includes three storage units: the recorded speech information storage unit 33, the speech text storage unit 36, and the summarized speech text storage unit 39. However, the speech information storage unit 40 is not limited to a configuration in which each storage unit is located in a specific area, and may also be configured to be distributed.

[0016] FIG. 4 shows an example of information content stored in the voice information storage unit 40. When the voice information management server 14 acquires the recorded information created by the main unit 12, it organizes it into the following items shown in FIG. 4 and stores it in the recorded information storage unit 33 of the voice information storage unit 40. Date and time: The date and time the call started Customer phone number: Phone number of external telephone 10 Customer name: A name registered in the telephone book of the main unit 12 and associated with the customer telephone number. Outgoing and incoming calls: Calls made from an internal telephone are considered outgoing calls, and calls received by an internal telephone are considered incoming calls. Extension number: The extension number from which the call was made Responder: The name of the operator registered in the telephone directory of the main unit 12, which associates the extension number of the extension telephone 11 with the name of the operator. Recording Time: The duration of the call recording Recording file name: The name of the file that records the call.

[0017] The speech text is data in which conversations are converted from a recording file into text by the speech recognition server 13, and information on each sentence and the speech start position (for example, speech start time) of each sentence are associated with each other. Note that information on the person speaking may be added for each sentence. The summarized speech text is data obtained by summarizing the speech text by the speech text summarizing server 15. Note that information about the speaker may be added for each sentence.

[0018] (Speech-to-text summarization server) FIG. 5 is a functional block diagram of the speech text summarizing server 15. As shown in FIG. The network interface unit 41 is an interface that connects the speech text summarization server 15 to the Internet 18. The operation terminal interface unit 42 is an interface that is operated by an administrator of the specialized knowledge learning model using a web browser or the like. For example, the operation terminal interface unit 42 is used to view and operate the learning data stored in the learning data storage unit 45 and the information on the learning model stored in the specialized knowledge learning model storage unit 47.

[0019] Next, we will explain how to create a learning model for specialized knowledge using Figure 5. The specialized knowledge learning model creation request receiving unit 43 receives the specialized knowledge learning model creation request sent by the operation terminal 16 through the network interface unit 41 and the operation terminal interface unit 42. The specialized knowledge learning model creation request can also include learning text. The learning data creation unit 44 creates learning data for the learning text read in upon receiving a request to create a learning model for specialized knowledge, based on the category (type) of the learning text. The learning data storage unit 45 stores the learning data created by the learning data creation unit 44. The specialized knowledge learning model creation unit 46 creates a learning model for specialized knowledge based on the stored learning data. The specialized knowledge learning model storage unit 47 stores the learning model for specialized knowledge created by the specialized knowledge learning model creation unit 46. The specialized knowledge learning model creation result transmission unit 48 transmits the results of the request to create a learning model for specialized knowledge to the operation terminal 16 via the operation terminal interface unit 42 and the network interface unit 41.

[0020] FIG. 6 is an example of a list of learning models stored in the specialized knowledge learning model storage unit 47. Learning texts can be divided into specialized document categories such as FAQs, inquiry examples, instruction manuals, construction and maintenance manuals, and design documents. They can also be sorted into product units such as Product A and Product B. By sorting specialized documents into categories and product units in this way, learning texts that contain a large amount of predetermined specialized knowledge for each unit can be obtained, and by using such learning texts, learning models can be created for each unit of predetermined specialized knowledge. The training text can be pre-recorded with information such as category and product name so that units of specialized knowledge can be mechanically distinguished. Therefore, when the training data creation unit 44 reads the training text, the machine can automatically distinguish the training text by specialized knowledge unit and sort the training text by each specialized knowledge unit. The specialized knowledge learning model creation unit 46 performs learning in units of training data sorted by specialized knowledge in this way, and as a result, a learning model is created for each specialized knowledge unit. Each learning model created for each specialized knowledge is then stored in the specialized knowledge learning model storage unit 47.

[0021] Next, the creation of the summarized speech text will be explained using FIG. The summarized speech text creation request receiving unit 49 receives a summarized speech text creation request from an operator, which is transmitted from the speech information management server 14. The summarized speech text creation request also receives speech text (hereinafter also referred to as "speech text to be summarized") from the speech management information server 14. The summarized speech text acquisition unit 50 acquires the speech text to be summarized received by the summarized speech text creation request receiving unit 49.

[0022] The general extraction type sentence extraction unit 51 can extract sentences using a summarization algorithm without using a learning model, or it can extract sentences using a general-purpose Japanese language model, and can extract sentences with general content that is neutral to specialized knowledge. When using a summarization algorithm without using a learning model, it is possible to use well-known summarization algorithm techniques such as algorithms that use TF-IDF similarity calculation methods or algebraic approaches, such as LexRank, LSA, Reduction, Luhn, SumBasic, and KL. When using a general-purpose Japanese language model, a vast amount of text covering a wide range of fields, such as Japanese Wikipedia and Japanese dictionaries, is used as training data, and well-known technologies such as deep learning, RNN (Recurrent Neural Network), and BERTSUMExt model are used as appropriate for the algorithm. In addition to Word2Vec, models such as BERT and GPT are used. When extracting sentences as summaries, information about the start time of a conversation in one sentence of the conversation may be deleted because it is not necessary during analysis to extract the sentence.

[0023] The expertise extraction-based sentence extraction unit 52 uses a learning model of expertise to calculate the similarity for each sentence in the conversational text to be summarized. Then, sentences with a similarity higher than a predetermined threshold are extracted. The similarity may be calculated for each unit of expertise. In this case, a single similarity calculated comprehensively from the similarities calculated for each unit of expertise for a given sentence can be compared with the threshold. The method of aggregating the similarities into a single similarity is not particularly limited. The summarized speech text creating unit 53 combines the sentences extracted by the general extraction type sentence extracting unit 51 and the sentences extracted by the specialized extraction type sentence extracting unit 52 in chronological order. The summarized speech text creation result sending unit 54 sends the summarized speech text, which is the result of the summarized speech text creation request, to the speech information management server 14 .

[0024] Figure 7 shows an example of the sequence of operations of the speech-to-text summarization system. It shows the operations and transmissions performed by the external telephone 10, main unit 12, internal telephone 11, speech information management server 14, operation terminal 16, speech recognition server 13, and speech-to-text summarization server 15. The details of these operations overlap with those described above, so a detailed explanation will be omitted.

[0025] <Procedure for creating summaries using expert extraction> Next, a procedure for creating a learning model of specialized knowledge in the speech text summarization server 15 and using it to create a summary of the speech text will be described. (Flowchart for creating a learning model of specialized knowledge) Figure 8 is a flowchart for creating a learning model of specialized knowledge. Each step is explained below. Receive a request to create a learning model of specialized knowledge (S101) The specialized knowledge learning model creation request receiving unit 43 receives a specialized knowledge learning model creation request from the operation terminal 16 via a terminal interface such as the Web. · Reading learning text (S102) When receiving a request to create a learning model for specialized knowledge, the specialized knowledge learning model creation request receiving unit 43 reads the learning text included in the creation request.

[0026] -Determine the type of learning text (S103) Since the learning points differ depending on the learning text, the learning text is classified by type (category). - Set the learning target for each "function" (S104) In the case of instruction manuals, construction and maintenance manuals, and design documents, the content (chapters, item numbers) is divided by "function." When creating summaries using a specialized knowledge learning model, the "function" content of each sentence in the speech text to be summarized can be compared to calculate the degree of similarity. - Set the learning target for each "inquiry content" (S105) In the case of FAQ / inquiry cases, the content is divided into "inquiry content" (for example, inquiries about products). When creating summaries using a learning model of specialized knowledge, the "inquiry content" is compared for each sentence of the speech text to be summarized, allowing the similarity to be calculated.

[0027] ·Text Formatting (S106) Format your text to ensure consistency in writing, to reduce the chance of search results being affected by variations in writing style. For example, unify "full-width uppercase English letters, full-width lowercase English letters, half-width lowercase English letters" to "half-width uppercase English letters", or change "full-width numbers" to "half-width numbers". ·Morphological analysis (S107) For example, using the open source MeCab, morphological analysis is performed based on the MeCab system dictionary and technical dictionary. The technical dictionary is given priority in the morphological analysis. For example, the word "main device" is split into "main" and "device" in the system dictionary, so the word "main device" is registered in the technical dictionary. As a result, the text "main device" becomes a single word, "main device." Stop word removal (S108) Words that are unlikely to be used as index terms or search terms, or words that would reduce search efficiency if used, are removed from searches. Words other than "nouns," "verbs," and "adjectives" are determined to be function words and removed as stop words, and "symbols" are also removed as stop words. In addition, words that are too general or appear too frequently (e.g., "over there," "over there," "how many"), words that are insufficient to identify the document data to be searched (e.g., a single alphabetic character), or words that are not normally used as search terms (e.g., words that start with "lowercase hiragana" or "lowercase katakana") are removed as stop words. ·Convert technical words (S109) Morphological analysis using a system dictionary and a technical dictionary is used to separate the text into words, and then for each word, the system unifies the expressions of words with the same meaning that have multiple ways of writing them, or converts them into expressions of general word combinations (e.g., "send and receive" → "send and receive"). This reduces the chances of your search not finding results due to differences in spelling even if the meaning is the same. You can also leave the original word as it is. Morphological analysis word division (S110) Split into combinations of words used in the training text (e.g. split "app version" into "application version", then add the words before and after before splitting to get "app version application version app version"). This reduces the possibility that a search will not return a hit due to different combinations of words entered in the search sentence, even if the meaning is the same. For example, the word "app version" will also include the combination of the two words "application" and "version." Therefore, if the training text contains expressions such as "application" and "version," searching for "app version" will result in a high degree of similarity to the search target. The procedures from S103 to S110 are performed by the learning data creation unit 44.

[0028] Learning data storage (S111) The learning data storage unit 45 stores the results of the morphological analysis up to the process of dividing words as learning data for creating a learning model of specialized knowledge. ·Creating learning models for specialized knowledge (S112) The specialized knowledge learning model creation unit 46 vectorizes (quantifies) the words used in the learning text so that the similarity between the search statement and the query example can be calculated based on the proximity of the vectors. For word vectorization, for example, Word2Vec is used to create a learning model that associates words with vectors. Using the word vectors, the TF-IDF method is used to vectorize (quantify) each "function" or "query content" set as the search target. · Learning model memory of specialized knowledge (S113) The specialized knowledge learning model storage unit 47 stores the model created in the specialized knowledge learning model creation. Send the results of the expert knowledge learning model creation (S114) The specialized knowledge learning model creation result transmission unit 48 transmits the result of the specialized knowledge learning model creation request to the operation terminal 16.

[0029] (Flowchart for creating a speech-to-text summary) Next, we will explain the procedure for creating a summary of a speech text in the speech text summarization server 15 using the above-mentioned specialized knowledge learning model. Figure 9 is a flowchart of the speech text summary creation process. Figure 9(a) shows the overall procedure, and Figure 9(b) shows the detailed procedure for expertise-based sentence extraction performed by the expertise-based sentence extraction unit 52. First, the overall procedure will be explained based on FIG. 9(a). Receive a request to create a summary voice text (S201) The summarized speech text creation request receiving unit 49 receives a summarized speech text creation request from the speech information management server 14 . - Reading speech text to be summarized (S202) When the summarized speech text creation request receiving unit 49 receives a summarized speech text creation request, it reads the speech text to be summarized in the creation request, and the speech text to be summarized is acquired by the summarized speech text acquiring unit 50 .

[0030] · Sentence extraction using general extraction (S203) In the general extraction type sentence extraction unit 51, sentences are extracted from the speech text to be summarized based on the well-known Japanese summarization technique as described above (hereinafter also referred to as "Step 1"). When extracting sentences as a summary, information on the start time of a conversation in one sentence of the conversation may be deleted because it is not necessary for the analysis to extract the sentence. · Sentence extraction using specialized extraction methods (S204) The specialized extraction type sentence extraction unit 52 extracts sentences using a learning model of specialized knowledge (hereinafter also referred to as "Step 2"). Details will be explained separately with reference to Figure 9(b). Note that the general extraction type sentence extraction (S203) and the specialized extraction type sentence extraction (S204) are each performed based on the speech text to be summarized. - Summary speech text creation (S205) The summarized speech text creating unit 53 combines the results of the general extraction type sentence extraction and the specialized extraction type sentence extraction in chronological order (hereinafter also referred to as "step 3"). - Send summary voice text creation result (S206) The summarized speech text creation result sending unit 54 sends the result of the summarized speech text creation request to the speech information management server 14 .

[0031] Next, the procedure for extracting sentences by the specialized extraction method (S204) will be described in detail with reference to FIG. 9(b). Extract one sentence from the speech text to be summarized (S301) Each sentence of the conversational text that appears in the speech text to be summarized is vectorized (quantified). If there is a sentence to extract and one sentence is extracted, the process moves on to the text formatting process; if there is no sentence to extract, the process moves on to the "sentence extraction by similarity" process (S308). ·Text formatting (S302) This is the same as the process of "formatting text" (S106) explained in the flowchart for creating a learning model of specialized knowledge (Fig. 8). Furthermore, information about the start time of a conversation in one sentence of a conversation may be deleted because it is unnecessary. ·Morphological analysis (S303) This is the same as the "morphological analysis" (S107) process described in the flowchart for creating a learning model of specialized knowledge (Fig. 8). Stop word removal (S304) This is the same as the process of "removing stop words" (S108) explained in the flowchart for creating a learning model of specialized knowledge (FIG. 8). Convert technical words (S305) This is the same as the process of "Convert technical terms" (S109) explained in the flowchart for creating a learning model of specialized knowledge (Fig. 8). Morphological analysis word division (S306) This is the same as the process of "dividing words in morphological analysis" (S110) explained in the flowchart for creating a learning model of specialized knowledge (Fig. 8). Vectorization of a single conversation sentence (S307) Using word vectors, the TF-IDF method is used to vectorize (quantify) each sentence of the conversation in the speech text to be summarized. Sentence extraction by similarity (S308) The similarity is calculated for each sentence in the speech text to be summarized, and sentences with similarity higher than a predetermined threshold are extracted.

[0032] <Example> An example of creating a summarized speech text according to the first embodiment will be described. FIG. 10 shows the entire speech text to be summarized in the example. Information on the start time of the conversation is included for each sentence in the conversation. The similarity of the sentences when the specialized knowledge learning model of step 2 is applied is also displayed. In the example, a conversation between a customer center operator and a customer inquiring about the installation of a camera-equipped door phone is converted into text sentence by sentence.

[0033] FIG. 11 shows the summarized speech text in the embodiment. In FIG. 11(a), sentence extraction using the general extraction method (Step 1) is applied, and sentences with general content are extracted. Although a publicly known Japanese summarization technique without a learning model is used, extraction may also be performed using a Japanese learning model trained on cross-disciplinary content such as a Japanese dictionary. Note that information about the start time of a conversation in one sentence is deleted because it is not necessary during analysis to extract the sentence. Figure 11(b) shows an example of sentence extraction using the expertise extraction method (Step 2) and extracting sentences with a similarity of 91% or more. It can be seen that sentences with specialized content, such as product and model numbers, have been extracted. Increasing the threshold generally narrows the search to more specialized sentences, while lowering it includes sentences with a higher proportion of general content. Therefore, it is possible to adjust the proportion of specialized content included in the summary by adjusting the threshold value. Note that here too, information about the start time of a conversation sentence is deleted as it is not necessary for the analysis used to extract sentences. Figure 11(c) shows the final product, the summarized speech text, which is the result of combining the extraction results of Step 1 and Step 2 in chronological order. If the extraction results of Step 1 and Step 2 overlap, they are overwritten when combined. The extraction in Step 1 contains general content that often appears in conversation and helps facilitate the conversation, while the extraction in Step 2 contains important, specialized content that is expected to appear in the summary. It was confirmed that by combining these in Step 3, sentences that are easy to read as conversation and contain important content are extracted as the summary. In step 2, important sentences that are desired to appear in the summary are extracted according to their magnitude relative to the similarity threshold, but it is also possible to allow the operator to specify the summary volume in terms of the number of lines or a percentage (%). In this case, if the conditions are not met, measures can be taken such as issuing a warning, deleting the sentences extracted in step 1, or changing the similarity threshold to make them meet the conditions. Furthermore, since the extraction in step 2 often contains important and specialized content that is expected to appear in the summary, in step 3, the extracted results of step 2 can be emphasized by changing the color of the sentences extracted in step 2 or by making them bold.

[0034] [Second embodiment] The second embodiment differs from the first embodiment in that in step 2, a dictionary of technical terms registered in advance is used instead of a learning model of specialized knowledge. Using a technical dictionary, the words used within each sentence of the speech text to be summarized are searched for. After a word search (exact match), sentences in which a word is found most frequently are extracted. (effect) For example, a dictionary of technical terms related to products in need can be constructed in the form of a vocabulary list, and sentences that are exact matches can be extracted.With less load than the first embodiment, which uses a learning model, important sentences contained in the speech text to be summarized that you would like to appear in the summary will appear in the summary.

[0035] [Third embodiment] The third embodiment differs from the second embodiment in that, when a word search is performed in a dictionary of pre-registered technical terms in step 2, importance is calculated by weighting based on the number of matches of the word found as a result of the word search (full match, partial match), and sentences with high importance are extracted. The importance may be defined, for example, as follows: Importance = [(number of exact matches) x (weight of exact matches)] + [(number of partial matches) x (weight of partial matches)] (effect) For example, a dictionary of technical terms related to products in demand can be constructed in the form of a vocabulary list, and hit sentences can be extracted using weighting based on the conditions of exact match, partial match, and match.With less load than the first embodiment which uses a learning model, important sentences contained in the speech text to be summarized that you would like to appear in the summary will appear in the summary.

[0036] Although the embodiments of the present invention have been described above, the present invention is not limited to the above-described embodiments, and various modifications are possible within the scope of the gist of the present invention. For example, the order of sentences in the speech text to be summarized does not need to be in chronological order as in the embodiment, but may be arranged in a predetermined order, and the order (context) of the sentences may be maintained in steps 1 to 3. Furthermore, it goes without saying that the language of the speech and text is not limited to Japanese, and can be applied to other languages ​​as well.

[0037] The following are examples of items that may be included in the present invention, but the present invention is not limited to these. (Item 1) 1. A speech-to-text summarization system comprising a speech-to-text summarization server, the speech-to-text summarizing server: a general extraction type sentence extraction unit that extracts sentences from speech text to be summarized, which is made up of a plurality of sentences arranged in order, while maintaining the order of the sentences; a specialized-extraction-type sentence extraction unit that extracts sentences from the speech text to be summarized while maintaining their order; a condensed speech text creation unit that creates a condensed speech text by combining the sentences extracted by the general extraction type sentence extraction unit and the sentences extracted by the specialized extraction type sentence extraction unit while maintaining their order; A speech-to-text summarization system comprising: (Item 2) Item 1. A speech text summarization system according to item 1, wherein the expertise extraction type sentence extraction unit extracts sentences using a learning model of expertise. (Item 3) Item 1. A speech text summarization system according to item 1, wherein the specialized extraction type sentence extraction unit extracts sentences using a dictionary of specialized terms. (Item 4) Item 2: A speech text summarization system in which, in the specialized knowledge extraction type sentence extraction unit, extraction based on a learning model of the specialized knowledge performs a similarity search for each sentence to extract sentences that show a similarity above a predetermined threshold. (Item 5) 5. The speech text summarization system according to any one of items 1 to 4, wherein the speech text to be summarized is made up of a plurality of conversation sentences arranged in chronological order. (Item 6) a speech recognition server that creates speech text from speech data; a speech information management server that manages the speech data, the speech text, and the summarized speech text; 6. The speech text summarization system according to any one of items 1 to 5, further comprising: (Item 7) Step 1: Extracting sentences from speech text to be summarized, which consists of multiple sentences arranged in order, using a general extraction method while maintaining the order; Step 2 extracts sentences from the speech text to be summarized using a specialized extraction method while maintaining their order; Step 3: combining the sentences extracted in step 1 and the sentences extracted in step 2 while maintaining their order to create a summarized speech text; A speech-to-text summarization method comprising: (Item 8) 8. The speech text summarization method according to item 7, wherein the sentence extraction using the expertise extraction type extracts sentences using a learning model of expertise. (Item 9) 8. The speech text summarization method according to item 7, wherein the specialized extraction type sentence extraction uses a dictionary of specialized terms to extract sentences. (Item 10) Item 9. The speech text summarization method according to item 8, wherein in step 2, the extraction based on the learning model of the specialized knowledge performs a similarity search for each sentence to extract sentences that show a similarity above a predetermined threshold. (Item 11) 11. The speech text summarization method according to any one of items 7 to 10, wherein the speech text to be summarized is made up of a plurality of conversation sentences arranged in chronological order. [Explanation of symbols]

[0038] 1...Speech text summarization system, 10...External telephone, 11...External telephone, 12...Main unit, 13...Speech recognition server, 14...Speech information management server, 15...Speech text summarization server, 16...Operation terminal, 17...LAN, 18...Internet, 20...WAN interface unit, 21...LAN interface unit, 22...Call control unit, 23...Recorded information storage unit, 24...Recorded information transmission unit, 30...Network interface unit, 31...Operation terminal interface unit, 32...Recorded information reception unit, 33...Recorded information storage unit, 34...Speech text creation request transmission unit, 35...Speech text creation result reception unit, 36...Speech text storage unit, 37...Summarized speech text creation request Transmitter, 38...summarized speech text creation result receiver, 39...summarized speech text storage unit, 40...speech information storage unit, 41...network interface unit, 42...operation terminal interface unit, 43...specialized knowledge learning model creation request receiver, 44...learning data creator, 45...learning data storage unit, 46...specialized knowledge learning model creation unit, 47...specialized knowledge learning model storage unit, 48...specialized knowledge learning model creation result transmitter, 49...summarized speech text creation request receiver, 50...summarized target speech text acquirer, 51...general extraction type sentence extraction unit, 52...specialized knowledge extraction type sentence extraction unit, 53...summarized speech text creation unit, 54...summarized speech text creation result transmitter

Claims

1. 1. A speech-to-text summarization system comprising a speech-to-text summarization server, the speech-to-text summarizing server: a general extraction type sentence extraction unit that extracts sentences of general content from speech text to be summarized, which is made up of multiple sentences arranged in order, by applying a summary creation algorithm without a learning model or a general-purpose language model while maintaining the order of the sentences; a specialized knowledge extraction type sentence extraction unit that extracts sentences of specialized content from the speech text to be summarized while maintaining order, based on a similarity equal to or greater than a predetermined threshold obtained by performing a similarity search for each sentence of the speech text to be summarized using a learning model of specialized knowledge, or based on a match obtained by performing a word search for each sentence of the speech text to be summarized using a dictionary of specialized terms; a condensed speech text creation unit that creates a condensed speech text by combining, while maintaining the order of the sentences of general content extracted by the general extraction type sentence extraction unit and the sentences of specialized content extracted by the specialized extraction type sentence extraction unit; A speech-to-text summarization system comprising:

2. A speech text summarization system as described in claim 1, wherein the degree of match obtained by the word search is the number of exact matches, or an importance obtained by adding weighting to the number of exact matches or the number of partial matches.

3. 2. The speech text summarization system according to claim 1, wherein the speech text to be summarized comprises a plurality of conversation sentences arranged in chronological order.

4. a speech recognition server that creates speech text from speech data; a speech information management server that manages the speech data, the speech text, and the summarized speech text; The speech-to-text summarization system according to any one of claims 1 to 3, further comprising:

5. Step 1: Extracting sentences of general content from speech text to be summarized, which is composed of a plurality of sentences arranged in order, by a general extraction type sentence extraction unit, applying a summary creation algorithm without a learning model or a general-purpose language model while maintaining the order of the sentences; Step 2: extracting sentences with technical content from the speech text to be summarized while maintaining order using a sentence extraction unit based on a similarity equal to or greater than a predetermined threshold obtained by performing a similarity search for each sentence of the speech text to be summarized using a learning model of specialized knowledge, or on a match obtained by performing a word search for each sentence of the speech text to be summarized using a dictionary of technical terms; Step 3: creating a summarized speech text by combining the sentences with general content extracted in step 1 and the sentences with specialized content extracted in step 2 while maintaining their order using a summarized speech text creation unit; A speech-to-text summarization method comprising:

6. 6. The speech text summarization method according to claim 5, wherein the speech text to be summarized is made up of a plurality of conversation sentences arranged in chronological order.

Images