Question Answering Device, Question Answering Method, and Program

Abstract

A question-answering device 100 for generating answer data with respect to question data comprises a question determination unit 10 that receives a predetermined question type into which a feature of the question data has been classified and passage data which is suitable for the question data, and determines, on the basis of the question type, whether or not machine reading comprehension is necessary. When machine reading comprehension is determined to be unnecessary, the question determination unit 10 sets the passage data as answer data. When machine reading comprehension is determined to be necessary, the question determination unit 10 sets, as answer data, answer text data which has been extracted from the passage data by machine reading comprehension on the basis of the question data and the passage data.

Description

Technical Field

[0001] The present disclosure relates to a question answering device, a question answering method, and a program.

Background Art

[0002] In recent years, question answering technology in which a computer automatically generates an answer to a question input in natural language by a user has attracted attention. As one such question answering technology, extraction-type machine reading that extracts an answer part from a document also described in natural language for a question input in natural language is known (see, for example, Non-Patent Document 1). In the question answering technology by machine reading, a neural network is used to collate a question with an answer part described in a document such as a manual, and it is known that in a question answering task of one question and one answer type, answer accuracy equal to or higher than that of a human can be achieved (see, for example, Non-Patent Document 2). On the other hand, in extraction-type neural machine reading, it is often necessary to equip a computer with a GPU (Graphic Processing Unit) not only during learning but also during inference, and generally, the processing time tends to be longer than that of question answering technology that does not use a neural network.

[0003] One application of question answering technology based on machine reading comprehension is operator support in a contact center. When an operator receives an inquiry from a customer, they search through various materials such as manuals and contracts to provide an answer. At this time, in order to shorten the search time, question answering technology based on machine reading comprehension is used. Answer candidates for the question are found using question answering technology based on machine reading comprehension and displayed to the operator. Since there are multiple target documents, it would take a long time to process all of these documents using machine reading comprehension with a neural network. Therefore, the documents are divided into fine-grained units such as paragraphs (passages), passages highly relevant to the question are narrowed down using information retrieval technology, and each is individually machine-read to extract the answer part. The extracted answer part is displayed as an answer candidate for the operator, or is highlighted and displayed together with the passage (see, for example, Non-Patent Document 3). Such a technology that combines information retrieval technology and question answering technology based on machine reading comprehension is hereinafter referred to as "large-scale machine reading comprehension."

Prior Art Documents

Non-Patent Documents

[0004]

Non-Patent Document 1

Non-Patent Document 2

Non-Patent Document 3

Summary of the Invention

Problems to be Solved by the Invention

[0005] In machine reading using a neural network, since a learning model of the neural network is generated with a document, a question, and a concise answer to the question as teacher data, the result of machine reading also tends to be output briefly.

[0006] However, in a contact center, an operator receives various types of questions from customers. For example, there are questions of the type that can be concisely answered in one to several words and are suitable for extractive neural machine reading, such as the name of a product that satisfies certain conditions, the expiration date of a product, whether a specific option is included in a service, etc. On the other hand, there are also questions of the type where an explanatory answer is expected that needs to be expressed in multiple sentences or text at the paragraph level, such as the overview of a product unknown to the customer, the reason why a desired option is not included in the contracted service, etc.

[0007] In extractive neural machine reading comprehension, for questions of the type where an explanatory answer like the latter is expected, it may output only a part of the expected answer. Specifically, for example, for a question such as "What are the features of XX?", only the text "The features of XX are as follows." may be extracted from the passage, and the subsequent specific feature explanation part in the document may be missed. Or, for example, for a true / false question such as "Can △△ do XX?", when the conditions for the basis are complex and consist of multiple sentences, although it may be possible to extract the text serving as the basis for true / false from the passage, it may not be possible to extract all of it. Therefore, for questions where an explanatory answer is expected, it is more appropriate to output the passage itself including the answer as it is without performing extractive neural machine reading comprehension.

[0008] That is, in the conventional question-and-answer technology, since it was not possible to appropriately determine the necessity of extractive neural machine reading comprehension according to the type of question, there was a problem that the accuracy of the answer to the question became low.

[0009] In view of such circumstances, an object of the present disclosure is to provide a question-and-answer device, a question-and-answer method, and a program that improve the accuracy of answers to questions.

Means for Solving the Problem

[0010] A question-and-answer device according to an embodiment is a question-and-answer device that generates answer data for question data, and includes a question determination unit that receives a predetermined question type in which the characteristics of the question data are classified and passage data that matches the question data, determines whether machine reading comprehension is necessary based on the question type, when it is determined that machine reading comprehension is not necessary, sets the passage data as the answer data, and when it is determined that machine reading comprehension is necessary, sets answer text data extracted from the passage data by the machine reading comprehension based on the question data and the passage data as the answer data.

[0011] A question answering method according to an embodiment is a question answering method for generating answer data for question data, comprising receiving a predetermined question type into which the characteristics of the question data are classified and message data that conforms to the question data, determining whether machine reading is necessary based on the question type, when it is determined that machine reading is not necessary, using the message data as the answer data, and when it is determined that machine reading is necessary, using, as the answer data, answer text data extracted from the message data by the machine reading based on the question data and the message data.

[0012] A program according to an embodiment causes a computer to function as the above-described question answering device.

Advantages of the Invention

[0013] According to the present disclosure, it is possible to provide a question answering device, a question answering method, and a program that improve the accuracy of answers to questions.

Brief Description of the Drawings

[0014]

Figure 1

Figure 2

Figure 3

Figure 4

Figure 5

Figure 6

Figure 7

Figure 8

Figure 9

Figure 10

Figure 11

Figure 12

[0015] Hereinafter, embodiments for carrying out the present invention will be described in detail with reference to the drawings.

[0016] [First Embodiment] <Configuration of Question Answering Device> With reference to FIGS. 1 to 8, an example of the configuration of the question answering device 100 according to the first embodiment will be described.

[0017] The question answering device 100 includes a control unit (controller) 110, a storage unit 120, an input unit 130, and an output unit 140.

[0018] The control unit 110 may be configured by dedicated hardware such as an ASIC (Application Specific Integrated Circuit) or an FPGA (Field-Programmable Gate Array), or may be configured by a general-purpose processor or a processor specialized for specific processing, or may be configured to include both. The control unit 110 includes a question determination unit 10, a question classification unit 20, a message search unit 30, and an extraction reading comprehension unit 40.

[0019] The storage unit 120 includes one or more memories, and may include, for example, a semiconductor memory, a magnetic memory, an optical memory, etc. Each memory included in the storage unit 120 may function as, for example, a main memory device, an auxiliary memory device, or a cache memory. Each memory does not necessarily have to be provided inside the question answering device 100, and may be configured to be provided outside the question answering device 100. The storage unit 120 stores any information used for the operation of the question answering device 100. The storage unit 120 stores, for example, a document database 121. In addition to this, the storage unit 120 stores, for example, various programs or data, etc.

[0020] The input unit 130 receives the input of various information. The input unit 130 may be any device as long as a predetermined operation by the user is possible, and examples thereof include a microphone, a touch panel, a keyboard, a mouse, etc. For example, when the user performs a predetermined operation using the input unit 130, question data is input to the control unit 110. The input unit 130 may be provided outside the question answering device 100, or may be integrated with the question answering device 100.

[0021] Question data is data in which a question asking what the user wants to know is expressed, data such as the Q of FAQ (Frequently Asked Questions) that is prepared in advance by predicting what the user may want to know, etc. Examples of the types of question data include, for example, data directly input by the user using a computer terminal or the like, output data of a computer provided in an external device, data selected by the user from a plurality of question data stored in an external database, data in which the voice content spoken by the user to an AI (Artificial Intelligence) agent, an AI speaker, etc. is texturized, etc.

[0022] Specifically, as shown in FIG. 2, the question data may be, for example, "What is the transmission speed of ISDN in kbps?", "Please teach me about ISDN.", "What is the technology for performing high-speed digital data communication using a twisted pair cable communication line?", etc.

[0023] The output unit 140 outputs various types of information. The output unit 140 is, for example, a speaker, a liquid crystal display, an organic EL (Electro-Luminescence) display, etc. The output unit 140 outputs, for example, answer data for the question data. The output unit 140 may be provided outside the question answering device 100 or may be integrated with the question answering device 100.

[0024] The answer data is data that expresses the content that the user who asked the question wants to know. Specifically, as shown in FIG. 3, the answer data may be, for example, "64 kbps", "ISDN (Integrated Services Digital Network, a public switched telephone network that is digitized all the way from the switch, relay line, and subscriber line and can also be used for packet communication and circuit-switched data communication. Since voice is transmitted within the ISDN network by circuit switching at 64 kbps in the range of 0.3 - 3.4 kHz, the voice quality is more stable than VoIP.)", "digital subscriber line", etc.

[0025] The question determination unit 10 outputs the question data input from the input unit 130 to the question classification unit 20. The question determination unit 10 acquires from the question classification unit 20 a predetermined question type into which the question text data indicating the characteristics of the question data is classified.

[0026] The question text data is data obtained by converting the question data into some feature quantity. Specifically, as shown in FIG. 4, the question text data may be, for example, "how many kbps", "about", "what is", etc.

[0027] The predetermined question type is a predetermined classification name given to each question text data. In other words, the predetermined question type is a classification name given to the features of the question data. Specifically, as shown in FIG. 4, the predetermined question type may be, for example, "quantity" for "how many kbps", "explanation" for "about", and "name" for "what is". Other predetermined question types include, but are not limited to, "5W1H", "true or false", and "other".

[0028] For details on question types, see, for example, the following non-patent literature: Hiroaki Sugiyama, Toyomi Meguro, Ryuichiro Higashinaka, "Large-scale collection and analysis of questions that examine the personality of dialogue systems," Transactions of the Japanese Society for Artificial Intelligence, DSF-518, 2016 Masaaki Nagata, Kuniko Saito, Yoshihiro Matsuo, "Japanese Natural Language Retrieval System Web Answers", 12th National Conference of the Association for Natural Language Processing, 2006

[0029] The question determination unit 10 outputs the question data input from the input unit 130 to the passage search unit 30. The question determination unit 10 acquires passage data (answer candidates) that match the question data from the passage search unit 30.

[0030] The passage data is text data in which a document is divided into small units (passages) such as paragraphs. The units of the passages are not particularly limited, but may be any units that can be divided automatically or manually, such as paragraphs, sections, subsections, etc. The passage data is stored in the document database 121.

[0031] Specifically, as shown in FIG. 5, the message data may be, for example, ID1: "ISDN (Integrated Services Digital Network) is a public switched telephone network that is digitized all the way to the switch, relay line, and subscriber line and can be used for packet communication and circuit-switched data communication. Since voice is transmitted within the ISDN network at 64 kbps circuit switching for 0.3 - 3.4 kHz, the voice quality is more stable than VoIP."; ID2: "Fiber To The Home, or simply FTTH, is a network configuration method of access optical communication that directly draws optical fiber into ordinary private homes as a transmission path. Also, not limited to ordinary private homes, it may also be referred to as FTTP (Fiber To The Premises) including small-scale offices that receive services in the same form."; ID3: "Digital Subscriber Line (DSL) refers to a technology for high-speed digital data communication using a twisted pair cable communication line or a telecommunications service." and so on.

[0032] The question determination unit 10 determines whether machine reading is necessary based on a predetermined question type acquired from the question classification unit 20. For example, when the question type is "explanation" or "other", the question determination unit 10 determines that the user expects an explanatory answer, and in this case, determines that machine reading is unnecessary. For example, when the question type is "quantity" or "name", the question determination unit 10 determines that the user does not expect an explanatory answer, and in this case, determines that machine reading is necessary.

[0033] Alternatively, the question determination unit 10 may determine that machine reading is necessary for those other than those expected to have a long character count in the classification of the question type analyzed in the dataset created for machine reading research (for example, Table 2 of the following non-patent document) (for example, Clause and Other). P. Rajpurkar, J. Zhang, K. Lopyrev, and P. Liang. Squad: 100,000+ questions for machine comprehension of text. In EMNLP, pp. 2383-2392, 2016

[0034] Alternatively, when using any classification method, the question determination unit 10 may calculate in advance the average number of characters of the machine reading comprehension results for each classification, and determine that machine reading comprehension is necessary for classifications with a number of characters less than a preset number of characters.

[0035] When the question determination unit 10 determines that machine reading comprehension is necessary, it outputs message data that matches the question data and the question data obtained from the message search unit 30 to the extraction reading unit 40. The question determination unit 10 obtains the answer text data extracted from the message data by machine reading comprehension from the extraction reading unit 40. The question determination unit 10 outputs the answer text data as answer data to the output unit 140. The answer text data may be, for example, "64 kbps", "digital subscriber line", etc. (see FIG. 3).

[0036] When the question determination unit 10 determines that machine reading comprehension is not necessary, it outputs the message data that matches the question data obtained from the message search unit 30 as answer data to the output unit 140. The message data may be, for example, "ISDN (Integrated Services Digital Network) is a public switched telephone network in which everything from switches, relay lines to subscriber lines is digitized and can be used for packet communication and circuit-switched data communication. Since voice is transmitted within the ISDN network by circuit switching at 64 kbps in the range of 0.3 - 3.4 kHz, the voice quality is more stable than VoIP." etc. (see FIG. 3). In this case, the question determination unit 10 does not output the message data that matches the question data and the question data obtained from the message search unit 30 to the extraction reading unit 40.

[0037] As described above, the question determination unit 10 receives (inputs) a predetermined question type into which the characteristics of the question data are classified and message data that matches the question data, determines whether machine reading comprehension is necessary based on the question type, and when it is determined that machine reading comprehension is not necessary, uses the message data as the answer data, and when it is determined that machine reading comprehension is necessary, uses, as the answer data, answer text data extracted from the message data by machine reading comprehension based on the question data and the message data. For this reason, it becomes possible to efficiently generate highly accurate answer data. In addition, highly accurate large-scale machine reading comprehension becomes possible.

[0038] The question classification unit 20 classifies question text data indicating the characteristics of the question data input from the question determination unit 10 into predetermined question types. The question classification unit 20 outputs the predetermined question types to the question determination unit 10. The question classification unit 20 may set the predetermined question types in advance.

[0039] The method by which the question classification unit 20 extracts the characteristics of the question data is not particularly limited. For example, the question classification unit 20 divides the question data into fine units and converts each unit into a set of characteristics. For example, the question classification unit 20 may use a set of individual characters included in the question data as the set of characteristics. For example, the question classification unit 20 may incorporate a morphological analyzer and convert the question data into a set of words, a set of morphemes, a set of parts of speech, etc., and use these as the set of characteristics.

[0040] For details of the morphological analyzer, for example, the following non-patent documents can be referred to. Taku Kudo, Kaoru Yamamoto, Yuji Matsumoto: Applying Conditional Random Fields to Japanese Morphological Analysis, Proceedings of the 2004 Conference on Empirical Methods in Natural Language Processing (EMNLP-2004), pp.230-237 (2004)

[0041] The question classification unit 20 determines the question type corresponding to each feature. The method by which the question classification unit 20 determines the question type is not particularly limited. For example, the question classification unit 20 associates in advance words (e.g., "when", "how many") that are typical features appearing in questions with question types (e.g., "Quantity"). Then, when "when" or "how many" appears when the words obtained by splitting the question are regarded as a set of features, the question classification unit 20 associates "Quantity" with these as the question type. For example, the question classification unit 20 collects a large number of question data, and for each piece of question data, manually determines which question type the set of features belongs to and associates them in advance. Further, the question classification unit 20 may create a classification model of a classifier by machine learning using the set of features and the question type as teacher data. The classifier is not particularly limited, and for example, any classifier of supervised machine learning such as a linear classifier, SVM (Support Vector Machine), or neural network may be used. Then, the question classification unit 20 estimates the question type corresponding to the set of features by using the created classifier with the set of features, which are the words obtained by splitting the question, as the input.

[0042] The message search unit 30 searches for message data that matches the question data input from the question determination unit 10. The message search unit 30 generates a search request from the question data, and based on the generated search request, acquires message data that matches the search request from the document database 121. The number of message data acquired by the message search unit 30 is not particularly limited. For example, when the search request is text, the message search unit 30 searches the document database 121 for message data that includes the character string constituting the text. For example, when the search request is a set of words, the message search unit 30 searches the document database 121 for message data that includes the word. The search request may be data in which the question data is expressed. For example, it may be the question data itself, all of the divided data obtained by dividing the question data into words, or a part of the divided data obtained by dividing the question data into words. The message search unit 30 outputs the message data that matches the question data to the question determination unit 10. If there is no message data in the document database 121 that matches the question data, the message search unit 30 outputs a search result indicating that there is no message data that matches the question data to the question determination unit 10.

[0043] The message search unit 30 extracts question keyword data from the question data. The message search unit 30 incorporates, for example, a morphological analysis system that divides the question data into a word sequence and assigns part-of-speech tags to the divided word sequence, and extracts the noun words included in the question data as question keyword data.

[0044] Specifically, as shown in FIG. 6, for the question keyword data, for example, for "What is the transmission speed of ISDN in kbps?", 'ISDN', 'transmission','speed', 'kbps'; for "Please tell me about ISDN", 'ISDN'; for "What is the technology for high-speed digital data communication using twisted pair cable communication lines?", 'twist', 'pair', 'cable', 'communication line', 'high speed', 'digital', 'data', 'communication', 'technology', etc. may be used.

[0045] The message search unit 30 extracts message keyword data from the message data. The message search unit 30 incorporates, for example, a morphological analysis system that splits the message data into a word sequence and assigns part-of-speech tags to the split word sequence, and extracts the noun words included in the message data as message keyword data.

[0046] Specifically, as shown in FIG. 7, the message keyword data includes, for the message data of message ID: ID1, 'ISDN', 'Integrated Services Digital Network', 'Integrated Services Digital Network', 'Service', 'Comprehensive', 'Digital', 'Network', 'Switch', 'Relay', 'Line', 'Subscriber', 'Line', 'Digitization', 'Packet', 'Communication', 'Line', 'Exchange', 'Data', 'Communication', 'Public', 'Exchange', 'Telephone', 'Network', 'Voice', '0.3', '3.4', 'kHz', '64', 'kbps', 'Line', 'Exchange', 'ISDN', 'Network', 'Transmission', 'VoIP', 'Voice', 'Quality', 'Stability'; for the message data of message ID: ID2, 'Fiber To The Home', 'FTTH', 'Fiber To The Premises', 'Optical', 'Fiber', 'Transmission', 'Path', 'General', 'Individual', 'Home', 'Access', 'System', 'Optical', 'Communication', 'Network', 'Configuration', 'Method', 'General', 'Individual', 'Home', 'Form', 'Service', 'Provision', 'Office', 'FTTP', 'Fiber To The Premises', 'Site'; for the message data of message ID: ID3, 'Digital', 'Subscriber', 'Line', 'DSL', 'Digital Subscriber Line', 'Twisted', 'Pair', 'Cable', 'Communication', 'Line', 'Path', 'Digital', 'Data', 'Communication', 'Technology', 'Telecommunication', 'Service', etc.

[0047] The message search unit 30 uses the question keyword data and the message keyword data to obtain message data with a high degree of match between the question keyword data and the message keyword data as message data that conforms to the question data.

[0048] For example, for the question data "What is the transmission speed of ISDN in kbps?", the message search unit 30 searches for message data that contains as many of the question keyword data 'ISDN', 'transmission','speed', and 'kbps' as possible. For example, in the message data with message ID: ID1, the number of matches between the question keyword data and the message keyword data is "3". For example, in the message data with message ID: ID2, the number of matches between the question keyword data and the message keyword data is "1". For example, in the message data with message ID: ID3, the number of matches between the question keyword data and the message keyword data is "0". Therefore, the message search unit 30 determines that the message data that conforms to the question data "What is the transmission speed of ISDN in kbps?" is the message data with message ID: ID1.

[0049] The document database 121 is a database that stores message data extracted from a plurality of documents such as manual documents and documents extracted from free encyclopedias on the Internet (e.g., Wikipedia), and is divided into fine units for comparison with the documents. An index may be created in the document database 121 in advance. The message search unit 30 can search for message data that conforms to the question data at high speed by referring to the index. Note that the document database 121 is stored in the storage unit 120 in advance.

[0050] The extraction and reading comprehension unit 40 extracts answer text data from the passage data by using machine reading comprehension based on a neural network, based on the pair of question data and passage data input from the question determination unit 10. For example, the extraction and reading comprehension unit 40 uses, as teacher data, a large number of data prepared in advance, which are based on the question data, the passage data, and the positions (start point and end point) of character strings in the text data included in the passage data, to generate a learning model of the neural network. Note that the learning model is not limited to a model using a neural network. Then, the extraction and reading comprehension unit 40 uses the learning model to extract, as the answer text data, the most appropriate character string as the answer to the question, based on unknown question data and passage data. The extraction and reading comprehension unit 40 outputs the answer text data to the question determination unit 10.

[0051] FIG. 8 is a diagram showing an example of the output of machine reading comprehension using the BERT model (hereinafter referred to as "BERT reading comprehension"). For example, the extraction and reading comprehension unit 40 extracts, as the output of BERT reading comprehension, answer text data of "64 kbps" based on the question data "What is the transmission speed of ISDN in kbps?" and the passage data of passage ID: ID1 that matches the question data. For example, the extraction and reading comprehension unit 40 extracts, as the output of BERT reading comprehension, answer text data of "It is a public switched telephone network in which all of the switch, relay line, and subscriber line are digitized and can also be used for packet communication and circuit-switched data communication." based on the question data "Please tell me about ISDN." and the passage data of passage ID: ID1 that matches the question data. For example, the extraction and reading comprehension unit 40 extracts, as the output of BERT reading comprehension, answer text data of "Digital Subscriber Line" based on the question data "What is the technology for performing high-speed digital data communication on a twisted pair cable communication line?" and the passage data of passage ID: ID3 that matches the question data.

[0052] The question answering device 100 according to the first embodiment can generate highly accurate answer data by appropriately determining the necessity of machine reading according to the question type. As a result, a question answering device 100 with improved accuracy of answers to questions can be realized. In addition, since unnecessary machine reading does not need to be performed, the computational cost by the computer's GPU can be significantly reduced. Further, a question answering device 100 capable of highly accurate large-scale machine reading can be realized.

[0053] <Question Answering Method> With reference to FIG. 9, an example of the question answering method according to the first embodiment will be described.

[0054] In step 101, the question determination unit 10 acquires, from the question classification unit 20, the question type into which the question text data included in the question data is classified.

[0055] In step 102, the question determination unit 10 acquires, from the message search unit 30, the message data IDs i (i = 1, 2, ··· n) that match the question data.

[0056] In step 103, the question determination unit 10 determines whether i is smaller than n. When the question determination unit 10 determines that i is less than or equal to n (step 103 → YES), it performs the process of step 104. When the question determination unit 10 determines that i is larger than n (step 103 → NO), it performs the process of step 108.

[0057] In step 104, the question determination unit 10 determines whether the question type is included in a predetermined list, that is, whether machine reading comprehension is required. When the question determination unit 10 determines that the question type is included in the predetermined list, that is, machine reading comprehension is not required (step 104 → YES), the process of step 105 is performed. When the question determination unit 10 determines that the question type is not included in the predetermined list, that is, machine reading comprehension is required (step 104 → NO), the process of step 106 is performed. The predetermined list is, for example, a list including the above-mentioned "Explanation" and "Others" (see FIG. 4). That is, the fact that the question type is included in the predetermined list means that the question type corresponds to, for example, "Explanation" or "Others". Also, the fact that the question type is not included in the predetermined list means that the question type corresponds to, for example, "Quantity" or "Name".

[0058] In step 105, the question determination unit 10 uses the message data ID i that matches the question data obtained from the message search unit 30 as the answer data.

[0059] In step 106, the question determination unit 10 outputs the question data and the message data ID i that matches the question data obtained from the message search unit 30 to the extraction and reading comprehension unit 40, and uses the answer text data r i extracted from the message data ID i obtained from the extraction and reading comprehension unit 40 as the answer data. i as the answer data.

[0060] In step 107, the question determination unit 10 adds 1 to i and performs the process of step 103 again.

[0061] In step 108, the question determination unit 10 outputs the message data ID i or the answer text data r i as the answer data to the output unit 140.

[0062] According to the above question answering method, by appropriately determining the necessity of machine reading according to the question type, highly accurate answer data can be generated. Thereby, a question answering method with improved accuracy of answers to questions can be realized.

[0063] 〔Second Embodiment〕 With reference to FIG. 10, an example of the configuration of a question answering apparatus 100A according to the second embodiment will be described.

[0064] The difference between the question answering apparatus 100A according to the second embodiment and the question answering apparatus 100 according to the first embodiment is that, before determining whether machine reading is necessary based on a predetermined question type acquired from the question classification unit 20, the question determination unit 10A first determines whether the message data is to be used as answer data based on the number of characters of the message data. Since the other configurations are the same, duplicate descriptions will be omitted.

[0065] The question answering apparatus 100A includes a control unit 110A, a storage unit 120, an input unit 130, and an output unit 140. The control unit 110 includes a question determination unit 10A, a question classification unit 20, a message search unit 30, and an extraction reading unit 40.

[0066] The question determination unit 10A measures the number of characters of the message data that matches the question data acquired from the message search unit 30. For example, the number of characters of the message data with message ID: ID1 is 190 characters. For example, the number of characters of the message data with message ID: ID2 is 191 characters. For example, the number of characters of the message data with message ID: ID3 is 98 characters.

[0067] The question determination unit 10A determines whether to use the measured message data as answer data based on whether the number of characters in the measured message data is less than a threshold value. The threshold value may be set in advance and may be arbitrarily set, for example, 30 characters, 100 characters, etc. The index used by the question determination unit 10A during determination may be the number of words in the message data. The index only needs to indicate at least the number of characters included in the message data and is not particularly limited.

[0068] When the number of characters in the measured message data is less than the threshold value, the question determination unit 10A determines to use the message data as answer data, and outputs the message data to the output unit 140 as answer data. For example, when the number of characters (98 characters) in the message data of message ID: ID3 is less than the threshold value (for example, 100 characters), the question determination unit 10A determines to use the message data of message ID: ID3 as answer data, and outputs the message data of message ID: ID3 to the output unit 140 as answer data (refer to the answer data corresponding to question data No. 3 in Fig. 10). In this case, the question determination unit 10A does not output the message data that conforms to the question data and the question data obtained from the message search unit 30 to the extraction and reading comprehension unit 40.

[0069] When the number of characters in the measured message data is greater than or equal to the threshold value, the question determination unit 10A determines not to use the message data as answer data, and further determines whether machine reading comprehension is required based on a predetermined question type obtained from the question classification unit 20. For example, when the number of characters (190 characters) in the message data of message ID: ID1 is greater than or equal to the threshold value (for example, 100 characters), the question determination unit 10A determines not to use the message data of message ID: ID1 as answer data. Also, for example, when the number of characters (191 characters) in the message data of message ID: ID2 is greater than or equal to the threshold value (for example, 100 characters), the question determination unit 10A determines not to use the message data of message ID: ID2 as answer data.

[0070] When the question determination unit 10A determines that machine reading is necessary, it outputs the message data that matches the question data obtained from the question data and the message search unit 30 to the extraction reading unit 40. The question determination unit 10A acquires the answer text data extracted from the message data by machine reading from the extraction reading unit 40. The question determination unit 10A outputs the answer text data as answer data to the output unit 140. For example, the question determination unit 10A outputs the answer text data "64 kbps" as answer data to the output unit 140 (see the answer data corresponding to question data No. 1 in FIG. 10).

[0071] On the other hand, when the question determination unit 10A determines that machine reading is unnecessary, it outputs the message data that matches the question data obtained from the message search unit 30 as answer data to the output unit 140. For example, the question determination unit 10A outputs the message data with message ID: ID2 as answer data to the output unit 140 (see the answer data corresponding to question data No. 2 in FIG. 10). In this case, the question determination unit 10A does not output the message data that matches the question data obtained from the question data and the message search unit 30 to the extraction reading unit 40.

[0072] The question answering device 100A according to the second embodiment can first determine whether to use the message data as answer data based on the number of characters of the message data, and then appropriately determine the necessity of machine reading according to the question type, thereby generating highly accurate answer data. As a result, a question answering device 100A with improved accuracy of answers to questions can be realized. For example, when a message is expressed in a few words to about one sentence on the screen of a computer terminal, the operator can quickly view and understand them, so there is no need to perform machine reading. However, in the prior art, machine reading was uniformly performed for all messages. By avoiding such problems and eliminating unnecessary machine reading, the computational cost of the computer's GPU can be significantly reduced. In addition, a question answering device 100A capable of high-precision large-scale machine reading can be realized.

[0073] <Question Answering Method> Referring to FIG. 11, an example of the question answering method according to the second embodiment will be described.

[0074] In step 201, the question determination unit 10A acquires, from the question classification unit 20, the question type to which the question text data indicating the characteristics of the question data is classified.

[0075] In step 202, the question determination unit 10A acquires the message data IDs i (i = 1, 2, ··· n) from the message search unit 30 that match the question data.

[0076] In step 203, the question determination unit 10A determines whether i is less than n. When the question determination unit 10A determines that i is less than or equal to n (step 203 → YES), it performs the process of step 204. When the question determination unit 10A determines that i is greater than n (step 203 → NO), it performs the process of step 208.

[0077] In step 204, the question determination unit 10A determines whether the number of characters of the message data that matches the question data acquired from the message search unit 30 is less than the threshold. When the number of characters of the message data that matches the question data acquired from the message search unit 30 is less than the threshold (step 204 → YES), it performs the process of step 206. When the number of characters of the message data that matches the question data acquired from the message search unit 30 is greater than or equal to the threshold (step 204 → NO), it performs the process of step 205.

[0078] In step 205, the question determination unit 10A determines whether the question type is included in a predetermined list, that is, whether machine reading comprehension is required. If the question determination unit 10A determines that the question type is included in the predetermined list, that is, machine reading comprehension is not required (step 205 → YES), the process of step 206 is performed. If the question determination unit 10A determines that the question type is not included in the predetermined list, that is, machine reading comprehension is required (step 205 → NO), the process of step 207 is performed.

[0079] In step 206, the question determination unit 10A uses the message data ID i that matches the question data obtained from the message search unit 30 as the answer data.

[0080] In step 207, the question determination unit 10A outputs the question data and the message data ID i that matches the question data obtained from the message search unit 30 to the extraction and reading comprehension unit 40, and uses the answer text data r i extracted from the message data ID i obtained from the extraction and reading comprehension unit 40 as the answer data. i as the answer data.

[0081] In step 208, the question determination unit 10A adds 1 to i and performs the process of step 203 again.

[0082] In step 209, the question determination unit 10A outputs the message data ID i or the answer text data r i to the output unit 140 as the answer data.

[0083] According to the above question answering method, first, based on the number of characters of the message data, it is determined whether the message data is used as the answer data, and then, according to the question type, the necessity of machine reading comprehension is appropriately determined, so that high-precision answer data can be generated. Thereby, a question answering method with improved accuracy of answers to questions can be realized.

[0084] <Modification Example> In this embodiment, the case where the question-and-answer devices 100 and 100A use extractive machine reading has been described as an example. However, the type of machine reading is not limited to this. For example, the question-and-answer devices 100 and 100A may use generative machine reading that does not limit the answer data to the range of message data.

[0085] Also, in this embodiment, the case where one piece of question data is input to the question-and-answer devices 100 and 100A has been described as an example. However, a plurality of pieces of question data may be input. When a plurality of pieces of question data are input, the question-and-answer devices 100 and 100A may repeatedly accept the input of the question data one by one and output the answer data.

[0086] <Program> The above-described question-and-answer devices 100 and 100A can be realized by a computer 201. Also, a program for causing the question-and-answer devices 100 and 100A to function may be provided. Further, the program may be stored in a storage medium or may be provided through a network. FIG. 12 is a block diagram showing a schematic configuration of a computer 201 that functions as the question-and-answer devices 100 and 100A, respectively. Here, the computer 201 may be a general-purpose computer, a dedicated computer, a workstation, a PC (Personal Computer), an electronic notebook, or the like. The program instructions may be program codes, code segments, etc. for executing necessary tasks.

[0087] As shown in FIG. 12, the computer 201 includes a processor 210, a ROM (Read Only Memory) 220, a RAM (Random Access Memory) 230, a storage 240, an input unit 250, an output unit 260, and a communication interface (I / F) 270. Each component is communicably connected to each other via a bus 280. Specifically, the processor 210 is a CPU (Central Processing Unit), MPU (Micro Processing Unit), GPU (Graphics Processing Unit), DSP (Digital Signal Processor), SoC (System on a Chip), etc., and may be composed of a plurality of processors of the same or different types.

[0088] The processor 210 controls each component and executes various arithmetic processes. That is, the processor 210 reads a program from the ROM 220 or the storage 240 and executes the program using the RAM 230 as a working area. The processor 210 performs control of each of the above components and various arithmetic processes according to a program stored in the ROM 220 or the storage 240. In the above-described embodiment, the program according to the present disclosure is stored in the ROM 220 or the storage 240.

[0089] The program may be stored in a computer-readable storage medium. By using such a storage medium, it is possible to install the program in the computer 201. Here, the storage medium in which the program is stored may be a non-transitory storage medium. The non-transitory storage medium is not particularly limited, and may be, for example, a CD-ROM, a DVD-ROM, a USB (Universal Serial Bus) memory, etc. Further, the program may be in a form downloaded from an external device via a network.

[0090] The ROM 220 stores various programs and various data. The RAM 230 temporarily stores programs or data as a working area. The storage 240 is composed of an HDD (Hard Disk Drive) or an SSD (Solid State Drive), and stores various programs and various data including an operating system.

[0091] The input unit 250 includes one or more input interfaces that receive a user's input operation and acquire information based on the user's operation. For example, the input unit 250 is a pointing device, a keyboard, a mouse, etc., but is not limited thereto.

[0092] The output unit 160 includes one or more output interfaces that output information. For example, the output unit 160 is a display that outputs information as video, or a speaker that outputs information as audio, but is not limited thereto. Note that when the output unit 160 is a touch panel type display, it also functions as the input unit 250.

[0093] The communication interface (I / F) 270 is an interface for communicating with an external device.

[0094] Regarding the above embodiments, the following supplementary notes are further disclosed.

[0095] (Supplementary Note 1) A question answering apparatus that generates answer data for question data, a memory, at least one controller connected to the memory, comprising, the controller, receives a predetermined question type in which the characteristics of the question data are classified and message data that matches the question data, determines whether machine reading comprehension is necessary based on the question type, When it is determined that machine reading is not required, the message data is used as the answer data. When it is determined that machine reading is required, based on the question data and the message data, the answer text data extracted from the message data by the machine reading is used as the answer data. Question answering device. (Appended claim 2) The controller classifies the characteristics of the question data into the question types, searches the message data, and extracts answer text data from the message data by machine reading based on the question data and the message data. The question answering device according to appended claim 1. (Appended claim 3) Before determining whether machine reading is necessary, the controller determines whether the number of characters in the message data is less than a threshold value. If the number of characters is less than the threshold value, the message data is used as the answer data. The question answering device according to appended claim 1 or 2. (Appended claim 4) A question answering method by a question answering device that generates answer data for question data, comprising the steps of receiving a predetermined question type in which the characteristics of the question data are classified and message data that matches the question data, determining whether machine reading is necessary based on the question type, when it is determined that machine reading is not required, using the message data as the answer data, when it is determined that machine reading is required, using, based on the question data and the message data, the answer text data extracted from the message data by the machine reading as the answer data, A question answering method including the above. (Appended claim 5) A non-transitory storage medium storing a program executable by a computer, A non-transitory storage medium storing a program that causes the computer to function as the question-and-answer device according to any one of claims 1 to 3.

[0096] Although the above-described embodiments have been described as representative examples, it is obvious to those skilled in the art that many changes and substitutions can be made within the spirit and scope of the present disclosure. Therefore, the present invention should not be construed as being limited by the above-described embodiments, and various modifications and changes are possible without departing from the scope of the claims. For example, it is possible to combine a plurality of configuration blocks described in the configuration diagrams of the embodiments into one, or to divide one configuration block. Also, it is possible to combine a plurality of steps described in the flowcharts of the embodiments into one, or to divide one step.

Description of Reference Numerals

[0097] 10, 10A Question determination unit 20 Question classification unit 30 Message search unit 40 Extraction and reading comprehension unit 100, 100A Question-and-answer device 110, 110A Control unit 120 Storage unit 121 Document database 130 Input unit 140 Output unit

Claims

1. A question answering apparatus for generating answer data for question data, comprising: a question determination unit that receives a predetermined question type into which the characteristics of the question data are classified and message data that matches the question data, determines whether machine reading for collating the question data and the message data using a neural network is necessary based on the question type, sets the message data as the answer data when it is determined that the machine reading is unnecessary, and sets, as the answer data, answer text data extracted from the message data by the machine reading based on the question data and the message data when it is determined that the machine reading is necessary; Question answering apparatus.

2. A question classification unit that classifies the characteristics of the question data into the question type; a message search unit that searches for the message data; and an extraction reading unit that extracts answer text data from the message data by machine reading based on the question data and the message data. The question answering apparatus according to claim 1 further comprises: The question answering apparatus according to claim 1.

3. Before determining whether the machine reading is necessary, the question determination unit determines whether the number of characters in the message data is less than a threshold value, and sets the message data as the answer data when the number of characters is less than the threshold value. The question answering apparatus according to claim 1 or 2. The question answering apparatus according to claim 1 or 2.

4. A question answering method by a question answering apparatus for generating answer data for question data, comprising: receiving a predetermined question type into which the characteristics of the question data are classified and message data that matches the question data; determining whether machine reading for collating the question data and the message data using a neural network is necessary based on the question type; when it is determined that the machine reading is unnecessary, setting the message data as the answer data; and when it is determined that the machine reading is necessary, setting, as the answer data, answer text data extracted from the message data by the machine reading based on the question data and the message data. The question answering method includes: Including, a question answering method.

5. A program for causing a computer to function as the question answering apparatus according to any one of claims 1 to 3.

Images