Question answering system, question reception answering system, primary answer system, and question answering method using the same
The separation of the question answering system into components enables flexible and cost-effective operation and maintenance, addressing the challenges of conventional systems by allowing independent machine learning and language support.
Patent Information
- Application Number
- JP2024120851
- Authority / Receiving Office
- JP · JP
- Patent Type
- Applications
- Current Assignee / Owner
- Priority Date
- 2018-03-07
- Filing Date
- 2024-07-26
- Publication Date
- 2024-11-06
- Estimated Expiration
- Not applicable · inactive patent
AI Technical Summary
Conventional question answering systems using artificial intelligence require significant time, effort, and cost for knowledge data collection and machine learning, necessitate specialized human resources, and are difficult to evolve flexibly to meet changing customer needs, especially for small and medium-sized enterprises.
The system is separated into a question receiving and answering component and a primary answering system with artificial intelligence, allowing independent machine learning and maintenance, enabling flexible modification of answers without specialized knowledge, and supporting multiple languages.
Reduces system implementation costs, allows easy evolution and maintenance, and supports multiple languages, making it accessible to small enterprises and improving answer accuracy and flexibility.
Smart Images

Figure 2024156759000001_ABST
Abstract
Description
[Technical Field]
[0001] The present invention relates to a question-answering system that accepts questions input from the terminal devices of specific or unspecified end users and provides answers to those questions, and more particularly to a question-answering system that uses artificial intelligence, a question-receiving and answering system and a primary answering system equipped with artificial intelligence that cooperate to constitute the question-answering system, and a question-answering method that uses them. [Background technology]
[0002] In recent years, various response systems have been developed to assist specific end users, such as telephone operators, in answering questions from customers, such as answer assistance systems that use artificial intelligence to search past response history to provide more appropriate answers quickly in order to provide the most optimal answers to questions, and automatic response systems that use artificial intelligence to answer questions from unspecified end users (questioners) without human intervention.Furthermore, research is also underway on natural language processing technology that analyzes the syntax and meaning of questions entered by questioners to search for answers to questions, and on extracting optimal answers using artificial intelligence.
[0003] In the following explanations in this specification, an unspecified or specific end user who asks a question will be referred to as a "questioner," a person who introduces and uses the question answering system will be referred to as a "system user," and a person who provides the question answering system to the system user will be referred to as a "system provider." Furthermore, the terminal device used by end users to ask questions will be referred to as the "question terminal," and the system user's system, which serves as the platform for the question reception page through which the question-answering system is accessed from the question terminal, will be referred to as the "user system." Furthermore, the system administrator who, as a system user, performs the acquisition and analysis of log information and maintenance of the question-answering system used by their company will be referred to as the "user's system administrator," and the system provider who performs general management such as the operation, maintenance, and improvement of the entire system will be referred to simply as the "system administrator."
[0004] For example, Patent Document 1 discloses a question-answering system that searches for answer candidates for a question, calculates a score for each combination of question and answer candidate that indicates the degree to which each answer candidate is a correct answer to the question, and outputs the answer candidate that is most likely to be the correct answer to the question based on the calculated score. Furthermore, Patent Document 2 discloses a request content identification system that analyzes an input sentence in which a user's request is written in natural language and determines whether the input request content corresponds to a predefined request content.
[0005] Furthermore, Patent Document 3 discloses an information provision system that includes an information provision server and a suggestion server. When a user inputs a question, the suggestion server displays possible question candidates to assist the user in inputting the question. The system of Patent Document 3 is also configured so that the information provision system can be maintained using a system management terminal. [Prior art documents] [Patent documents]
[0006] [Patent Document 1] Japanese Patent Application Laid-Open No. 2013-171550 [Patent Document 2] Patent No. 5099367 specification [Patent Document 3] Japanese Patent Application Laid-Open No. 2016-99967 Summary of the Invention [Problem to be solved by the invention]
[0007] However, in the systems of Patent Documents 1 and 2, in order to improve the accuracy of answers to questions, it was necessary to collect a large amount of knowledge data and perform machine learning. Furthermore, even for the operation management of the question answering system after its introduction, a high level of knowledge about artificial intelligence etc. was required. Therefore, when introducing a question-answering system, collecting a huge amount of knowledge data and conducting machine learning before the system went into operation required a lot of time, effort, and cost. Furthermore, even after the system went into operation, machine learning to update the AI's knowledge and system maintenance required a lot of time and cost, and the operation and management of a question-answering system that uses AI required the securing of personnel with specialized knowledge of AI.
[0008] In view of problems in operational management after the system is put into operation, Patent Document 3 allows a system administrator to use a system management terminal in maintenance mode to receive response data logs from a response content determination means and to simulate input to verify the validity of knowledge data. Furthermore, the invention of Patent Document 3 supports a system administrator so that he or she can change expected question strings and response strings for the artificial intelligence in maintenance mode, even if he or she does not have highly specialized knowledge about artificial intelligence.
[0009] However, inputting anticipated questions and changing the corresponding responses alters the AI's knowledge data itself. Therefore, even in the system of Patent Document 3, while collecting and reviewing log data and simulating inputs do not require advanced knowledge, system administrators are required to have a certain level of specialized skills in AI (see the latter part of (0048) in Patent Document 3). Furthermore, changing the AI's knowledge data requires re-learning to reconstruct the neural network, which requires not only specialized knowledge in AI but also a great deal of time and effort. Therefore, in conventional AI-based question-answering systems, system users without specialized knowledge cannot directly participate in improving and gradually growing the system they have introduced. Furthermore, conventional AI-based question-answering systems lack a simple mechanism for flexibly changing the specific answers to questions as needed while the system is actually in operation, making it difficult to flexibly evolve the system to meet the needs of each system user.
[0010] Under these circumstances, even if small and medium-sized enterprises with limited resources wanted to implement a question-answering system, the hurdles to implementation were extremely high, due to the enormous implementation costs of conventional question-answering systems and the need to secure specialized personnel for system operation and management. Even if these implementation barriers were overcome, system users could not directly improve or grow the system to meet their company's changing needs. Furthermore, conventional question-answering systems could not handle questions in multiple languages simply by adding a simple configuration.
[0011] The present invention has been made in consideration of the problems and social needs of conventional question-answering systems as described above, and aims to provide a question-answering system that uses artificial intelligence, which can reduce system implementation costs, enable the final answer to be easily changed after the system has started operation, and enable the question-answering system to be improved and evolved, as well as a question receiving and answering system and a primary answering system that constitute the question-answering system, and a question-answering method using them. Another object of the present invention is to provide a question-answering system with a simple configuration that can accept questions in multiple languages and provide a final answer to a question terminal in a predetermined answer language depending on the language of the question, a question receiving and answering system and a primary answering system that constitute the question-answering system, and a question-answering method using them. Furthermore, another object of the present invention is to provide a question-answering system that utilizes artificial intelligence and can be shared by multiple users and reduce implementation costs, a question receiving and answering system and a first answering system that constitute the question-answering system, and a question-answering method using them. [Means for solving the problem]
[0012] In the present invention, a question answering system is configured by separating it into a question receiving and answering system that receives questions and returns final answers to the questions to the question terminal, and a primary answering system equipped with artificial intelligence, wherein the primary answering system provides a primary answer to the question as an identifier in response to a request from the question receiving and answering system, and the question receiving and answering system stores final answer candidates in association with the identifiers of the primary answers received from the primary answering system, reads out the final answer candidates based on the primary answers from the primary answering system, and returns the final answers to the question terminal device.
[0013] A question answering system according to a first aspect of the present invention is a question answering system comprising a question receiving and answering system using a computer that receives a question from a question terminal and returns a final answer to the question to the question terminal, and a first answering system that has artificial intelligence and returns a first answer to the question in response to a request, wherein: The first response system is an interface means for receiving a reply request for the first answer to the question from the question receiving and answering system, inputting the question received from the question receiving and answering system to the artificial intelligence, and transmitting the first answer to the question from the artificial intelligence to the question receiving and answering system; an artificial intelligence that performs machine learning on a plurality of expected questions and predetermined identifiers as first answers corresponding to the expected questions, and outputs a first answer to a question from the question receiving and answering system as the identifier in response to a request from the question receiving and answering system; The question receiving and answering system includes: a reception means for receiving a question from the question terminal; a first reply inquiry means for transmitting the question received by the question receiving means to the first reply system to request the first reply, and for receiving the first reply from the first reply system; a final answer storage unit that stores final answer candidates corresponding to the identifiers obtained by machine learning using the artificial intelligence in a 1:1 relationship with the identifiers; a final answer means for reading out the final answer candidate corresponding to the identifier of the primary answer from the final answer storage unit based on the primary answer received from the primary answer system, and returning a final answer to the question terminal based on the read final answer candidate; The present invention is characterized by comprising:
[0014] Here, a general-purpose artificial intelligence system provided for commercial service purposes can be used as the initial answer system that provides the initial answer, and the question receiving and answering system and the initial answering system can be connected via a communications network. By separating the question answering system into the question receiving and answering system and the initial answering system in this way, the question receiving and answering system is separated from the artificial intelligence, so that the machine learning of the artificial intelligence of the initial answering system can be designed separately from the final answer candidates of the question receiving and answering system. Furthermore, independent of the machine learning of the artificial intelligence, it is possible to independently modify, change, and add to the hit rate of answers to questions.
[0015] A question receiving and answering system according to a first aspect of the present invention is a question answering system that includes a question receiving and answering system using a computer that receives questions from a question terminal, and a first answering system that has artificial intelligence and learns by machine learning a plurality of expected questions and predetermined identifiers as first answers corresponding to the expected questions, and outputs first answers to questions from the question receiving and answering system as the identifiers, wherein the question receiving and answering system is a separate system, a question receiving means for receiving a question from the question terminal; a first reply inquiry means for transmitting the question received by the question receiving means to the first reply system to request the first reply, and for receiving the first reply from the first reply system; a final answer storage unit that stores final answer candidates corresponding to the identifiers, in a one-to-one association with the identifiers machine-learned in the first response system; final answer means for reading out the final answer candidate corresponding to the identifier of the primary answer from the final answer storage unit based on the primary answer received from the primary answer system, and returning a final answer to the question terminal based on the read final answer candidate; The present invention is characterized by comprising: The invention according to this aspect is a sub-combination invention that constitutes a question receiving and answering system that combines the question receiving and answering system and the primary answering system described above.
[0016] Furthermore, in another aspect of the present invention, the question answering system or question receiving and answering system is characterized in that the artificial intelligence of the primary answering system has machine-learned to use a large number of expected questions that make up multiple question groups, each group consisting of one or more expected questions with common attributes, and an identifier that corresponds one-to-one with the question group as the primary answer for the expected questions included in each question group.
[0017] In the above-mentioned question answering system or question reception and answering system, the question reception means can be further configured to acquire unique attribute information of the question accompanying the question from the question terminal, and the final answer storage unit can be configured to store different final answer candidates corresponding to the unique attribute information, as necessary. The unique attribute information includes identification information of the user system, language information indicating the type of language of the question, etc. The identification information of the user system can be obtained based on communication path information when accepting a question or information added by the user system along the communication path. The language information of the question can be configured to be specified when the question is entered, or can be obtained using well-known techniques, such as obtaining it from language specification information in the OS of the questioning terminal or an application for question answering.
[0018] A question answering system or a question receiving and answering system according to another aspect of the present invention acquires, based on the acquired unique attribute information, identification information of a user system through which the question was sent, the final answer storage unit stores, as each of the final answer candidates, a basic answer that is a base of the final answer and is stored in association with each of the identifiers, or a unique answer that is stored in association with each of the identifiers and identification information of the user system as necessary and has a different content from each of the basic answers, The final answer means is characterized in that it reads out the basic answer or the specific answer from the final answer memory unit based on the identifier of the primary answer and the identification information of the user system corresponding to the question, and transmits the read basic answer or the specific answer to the question terminal as a final answer. As described above, the identification information of the user system can be obtained as unique attribute information from the communication path information that the question took before reaching the question receiving and answering system or from information added during the communication process.
[0019] In another aspect of the present invention, the question answering system or question receiving and answering system is characterized in that the artificial intelligence of the primary answering system has machine-learned to output the same primary answer identifier for all of the anticipated questions consisting of multiple question groups, each of which includes multiple anticipated questions with a common attribute made up of different languages within one question group, and for the primary answers to the multiple anticipated questions made up of different languages included in the same question group, regardless of the type of language, and is configured to output the identifier corresponding to the question to the interface means as the primary answer to the question when a question is received from the question receiving and answering system. By having the artificial intelligence of the primary answer system trained to output the same primary answer for multiple assumed questions in different languages that have a common attribute, the same primary answer is output for questions in different languages that have a common attribute. Since final answer candidates are stored in association with the primary answers, if the primary answers are the same, the same final answer will be sent to the question terminal even if the question is in a different language. The answer language of the final answer may be configured to return all final answers in the same language regardless of the question language, or to return a final answer in English when the question is in a foreign language, or to return a final answer in the same answer language as the question language. Here, "plurality of assumed questions having a common attribute and written in different languages" includes assumed questions that have substantially the same meaning but written in different languages.
[0020] In another aspect of the present invention, the question answering system or question receiving and answering system includes the steps of: acquiring question language information included in the unique attribute information of a received question by the question receiving means; and storing the final answer candidates in a plurality of languages having substantially the same content in an answer language determined based on the question language information, in association with the identifier and the type of the answer language; The final answering means acquires information on the answer language corresponding to the question based on the acquired question language information, reads out the final answer candidate consisting of the answer language corresponding to the question based on the acquired identifier of the first answer and the answer language information, and returns the final answer in the answer language corresponding to the question language as the final answer corresponding to the question. With this configuration, if the question is in a foreign language, the answer language can be changed to English or another language, and the final answer can be returned in the same answer language as the question.
[0021] A question answering system or question receiving and answering system according to another aspect of the present invention is characterized in that the primary answering system extracts a primary answer to the question and outputs the answer certainty, and the question receiving and answering system reads out from the final answer storage unit the basic answer or the unique answer that is stored in association with the primary answer according to the answer certainty received from the primary answering system, and transmits the final answer to the question to the question terminal.
[0022] In addition, a question answering system or question receiving and answering system according to another aspect of the present invention is characterized in that it further comprises an operation history memory unit that stores various information from question reception to final answer, such as communication records with the question terminal, communication records with the first answer system, and unique attribute information of the question, and a maintenance means that, depending on authority, enables a user system to view the various information stored in the operation history memory unit and perform maintenance including correcting, adding, and changing the final answer candidates or unique answers stored in the final answer memory unit. This configuration makes it easy to maintain the initial answer system and the question answering system. In particular, maintenance of the question answering system does not require knowledge of artificial intelligence, and the system can be easily improved by changing the basic answers and unique answers that are final answer candidates. In addition, the maintenance means can be configured to allow the contents of the final answer candidates, basic answers, or unique answers stored in the final answer storage unit to be updated by adding, correcting, changing, deleting, etc., depending on authority, without being limited by the expected questions of the primary answer system.
[0023] The operation history storage unit can include information such as the number of questions and answers for the entire system and for each system user, the average answer confidence, the number and percentage of questions with answer confidence below a reference value, the number of questions and answers and the average answer confidence for each initial answer identifier, and a list of questions and answers with low answer confidence. This information can be viewed by a maintenance unit, making it easy to understand the system's operational status, such as system usage and answer accuracy. Furthermore, because the question receiving and answering system and the final answer storage unit are separated from the initial answering system, even those without specialized knowledge of artificial intelligence can use the maintenance unit to change and modify the final answers stored in the final answer storage unit (in this specification, unless otherwise expressly stated, the term "modify" includes not only partial correction or modification but also addition). Therefore, not only system administrators but also system users without advanced knowledge of artificial intelligence can modify, change, and add final answers, such as basic answers and unique answers, within the scope of their assigned authority.
[0024] The maintenance means can also be configured to add a function to correct, change, or add expected questions, common attributes, question groups, or first answers, and to have the first answering system perform machine learning again. Such an operation history storage unit and maintenance means make it possible to output a list of questions with low answer certainty and answer records extracted, or to consolidate or eliminate expected questions or question groups with low answer certainty according to analytical data from tabulations and statistical data, thereby improving and evolving the question answering system to provide more appropriate answers.
[0025] A first reply system according to a first aspect of the present invention is a question answering system that includes a question receiving and answering system using a computer that receives questions from a question terminal and returns final answers to the questions to the question terminal, and a first reply system that includes artificial intelligence that performs machine learning to generate a plurality of expected questions and predetermined identifiers as first answers corresponding to the expected questions, and outputs first answers to the questions from the question receiving and answering system as the identifiers, wherein the first reply system is in a separate system, an interface means for receiving a question and a request for the first answer from the question receiving and answering system, inputting the question to the artificial intelligence, and transmitting a first answer to the question from the artificial intelligence to the question receiving and answering system; The system is characterized by comprising: artificial intelligence that has machine-learned a large number of expected questions in different languages, a plurality of question groups each of which groups a plurality of expected questions that have a common attribute regardless of language, and an identifier that has a one-to-one correspondence with each of the question groups as a primary answer common to the plurality of expected questions belonging to each of the question groups; and when a question is received from the question receiving and answering system, outputs the primary answer to the question as the identifier to the interface means.
[0026] This invention relates to a first reply system that constitutes the above-mentioned question receiving and replying system, and is a sub-combination invention of the question answering system. According to this configuration, the primary answering system is trained to output a common primary answer (identifier) for anticipated questions that share common attributes even when the question language is different, thereby outputting the same identification number as the primary answer for questions that share common attributes even when the question language is different. Therefore, the same final answer can be returned to the question terminal even when the question language is different. Since it is relatively difficult to create an accurate question document in a language other than one's native language, being able to at least ask the question in one's native language makes it possible to ask an accurate question and obtain an accurate answer. Meanwhile, document reading ability is generally higher than document creation ability, so there is a high possibility that the answer language can be understood even if it is not one's native language. In addition, by determining the response language for the final response in advance according to the language of the question and storing final response candidates for each response language, it is possible to configure the system so that the final response is returned in a predetermined response language according to the language information of the question.
[0027] A question answering method for a question answering system according to a first aspect of the present invention is a question answering method for a question answering system including: a question receiving and answering system that receives questions from a question terminal by a computer and returns answers to the questions to the question terminal; and a primary answering system that has artificial intelligence and performs machine learning to learn a plurality of expected questions and predetermined identifiers that are answers to the plurality of expected questions as primary answers, and returns the identifiers as primary answers in accordance with an input question, (a) in the question receiving and answering system, receiving a question from a question terminal and transmitting the question to a first answering system; (b) in the first reply system, accepting the question from the question receiving and replying system, inputting the question to the artificial intelligence, outputting the first reply to the input question as the identifier, and transmitting the identifier of the first reply output from the artificial intelligence to the question receiving and replying system; (c) in the question receiving and answering system, when the identifier is received as the first answer from the first answer system, reading out the final answer candidate corresponding to the received identifier from a final answer storage unit that stores final answer candidates linked to the identifier that the artificial intelligence has machine-learned, and returning the read out final answer candidate to the question terminal as a final answer; The present invention is characterized by comprising:
[0028] A question answering method for a question receiving and answering system according to a first aspect of the present invention is a question answering method for a question answering system including: a question receiving and answering system that receives questions from a question terminal by a computer and returns answers to the questions to the question terminal; and a primary answering system that is equipped with artificial intelligence and performs machine learning to learn a plurality of expected questions and predetermined identifiers that are answers to the plurality of expected questions as primary answers, and returns the identifiers as primary answers in accordance with an input question, the method comprising: (a1) in the question receiving and answering system, receiving a question from a question terminal and transmitting the question to the first answering system; (b1) when receiving the identifier as the first answer from the first answer system, reading out the final answer candidate corresponding to the received identifier from a final answer storage unit that stores final answer candidates linked to the identifier that the artificial intelligence has machine-learned, and returning the read out final answer candidate to the question terminal as a final answer; The present invention is characterized by the following features.
[0029] A question answering system or a question answering method for a question receiving and answering system according to another aspect of the present invention further comprises the steps of: (a-1) receiving a question from a question terminal; (a-2) acquiring and storing unique attribute information of the question associated with the question; (a-3) sending the question to the first response system; Equipped with The step (c) or (b1) (c-1) receiving the identifier as the first response from the first response system; (c-2) acquiring identification information of the user system from the unique attribute information stored in step (a-1); (c-3) reading out the corresponding basic answer or the specific answer from an answer storage unit that stores final answer candidates corresponding to the identifier, divided into basic answers and specific answers specific to the user, based on the identifier of the primary answer and the identification information of the user system; (c-4) returning the read basic answer or the unique answer to the question terminal; The present invention is characterized by comprising:
[0030] A method for answering multilingual questions in a question answering system according to a first aspect of the present invention is a method for answering multilingual questions in a question answering system comprising: a question receiving and answering system that receives questions from a question terminal by a computer and returns answers to the questions to the question terminal; and a primary answering system that is equipped with artificial intelligence and that has machine learning to output primary answers with the same identifier for all expected questions configured from a plurality of question groups, each of which includes a plurality of expected questions in different languages and having a common attribute within one question group, and for first answers to all of the expected questions in different languages included in the same question group, regardless of the type of language, and that returns the identifier as a primary answer in response to an input question, (a) in the question receiving and answering system, receiving a question from a question terminal, acquiring and storing language information of the question from unique attribute information of the question associated with the question, and, when the question and the first answering system require the language information of the question, transmitting the question together with the language information of the question to the first answering system; (b) in the first reply system, accepting the question from the question receiving and replying system, inputting the question to the artificial intelligence, outputting the first reply to the input question as the identifier, and transmitting the identifier of the first reply output from the artificial intelligence to the question receiving and replying system; (c) in the question receiving and answering system, when the identifier is received as the first answer from the first answer system, reading out the final answer candidate corresponding to the received identifier from an answer storage unit that stores final answer candidates linked to the identifier that the artificial intelligence has machine-learned, and returning the read out final answer candidate to the question terminal as a final answer; The present invention is characterized by comprising: Here, "plurality of assumed questions having a common attribute and written in different languages" includes assumed questions that have substantially the same meaning but written in different languages.
[0031] A multilingual question receiving and answering method for a question receiving and answering system according to a first aspect of the present invention is a question receiving and answering method for a question answering system including: a question receiving and answering system that receives a question from a question terminal by a computer and returns an answer to the question to the question terminal; and a primary answering system that is equipped with artificial intelligence and that performs machine learning to output primary answers with the same identifier for all expected questions configured from a plurality of question groups, each of which includes a plurality of expected questions having a common attribute made up of different languages within one question group, and for first answers to the plurality of expected questions made up of different languages included in the same question group, regardless of the type of language, and that returns an identifier as a primary answer in accordance with an input question, comprising: (a1) in the question receiving and answering system, receiving a question from the question terminal, acquiring and storing language information of the question from unique attribute information of the question associated with the question, and, when the question and the first answering system require the language information of the question, transmitting the language information of the question together with the language information of the question to the first answering system; (b1) when receiving the identifier as the first answer from the first answer system, reading out the final answer candidate corresponding to the received identifier from a final answer storage unit that stores final answer candidates linked to the identifier that the artificial intelligence has machine-learned, and returning the read out final answer candidate to the question terminal as a final answer; The present invention is characterized by comprising: The present invention is a sub-combination invention of the invention according to claim 16. Here, the "plurality of hypothetical questions having a common attribute and written in different languages" includes hypothetical questions with substantially the same meaning but written in different languages.
[0032] In another aspect of the present invention, there is provided a question answering system or a method for receiving and answering questions in multiple languages for a question receiving and answering system, wherein the step (c) or (b1) comprises: (c-1) receiving the identifier as the first response from the first response system; (c-2) acquiring answer language information determined in accordance with a rule for determining an answer language based on the language information stored in step (a) or (a1); (c-3) A step of reading out final answer candidates stored in a final answer storage unit in the answer language specified by the answer language information, linked to the identifier and the answer language information learned by machine learning, the final answer candidates corresponding to the identifier received as the first answer and the acquired answer language information; (c-4) returning the read final answer candidate in the answer language to the question terminal as a final answer; The present invention is characterized by comprising:
[0033] A first aspect of the present invention provides an answering method for a primary answering system in a question answering system that uses a question receiving and answering system that receives a question from a question terminal and returns an answer to the question terminal, and a primary answering system that includes artificial intelligence that outputs a primary answer to the question as an identifier in response to a request from the question receiving and answering system, the method comprising: (a) a step of machine learning a large number of expected questions constituting a plurality of question groups, each of which contains a plurality of expected questions having a common attribute regardless of language, and an identifier corresponding one-to-one to each of the plurality of question groups as a first answer to each of the expected questions, by the artificial intelligence; (b) inputting the question received from the question receiving and answering system into the artificial intelligence, and outputting the identifier as a first answer from the artificial intelligence; (c) transmitting the identifier output as the first answer to the question to the question receiving and answering system as a first answer to the question; characterized by comprising In step (a), the plurality of possible questions having a common attribute included in each question group are preferably configured to include possible questions that have substantially the same meaning but are written in different languages. The present invention is a sub-combination invention of the invention according to claim 16. [Effects of the Invention]
[0034] The question-answering system of the present invention is separated into a question receiving / answering system and a primary answering system having artificial intelligence. Final answer candidates for a question are stored in the question receiving / answering system in association with the primary answer of the primary answering system. In response to a primary answer from the primary answering system, the final answer candidates from the question receiving / answering system are read and a final answer is returned. This allows the content of the final answer candidates to be freely modified and changed independently of the artificial intelligence. In other words, even if one does not have specialized knowledge of artificial intelligence, the final answer can be freely changed within a predetermined scope of authority. For example, even if the answer confidence of the primary answer from the primary answering system is low, by changing the final answer to the question to a final answer tailored to the question, improvements and maintenance can be made, such as substantially increasing the answer confidence without machine learning.
[0035] Furthermore, to improve the accuracy of responses to questions, it is desirable to train as many expected questions as possible from the beginning of system implementation. However, increasing the number of expected questions and answers to be trained increases system implementation costs. In contrast, the present invention stores final answer candidates in a separate question receiving and answering system separate from the primary answering system, making it possible to store final answer candidates with a broader range of answer content than those directly responding to expected questions. This allows answers to expected questions that have not been trained by AI to be included in the final answer candidates, to the extent that the primary answer achieves a certain degree of answer confidence, thereby increasing satisfaction with the answers to questions. Therefore, the present invention makes it possible to enrich the answer content of the final answer candidates while reducing the number of expected questions and answers to be trained by AI in the primary answering system compared to conventional systems. In other words, it is possible to reduce the number of expected questions to be trained by AI compared to conventional systems, thereby reducing the implementation cost of the question-answering system at the start of operation and reducing the machine learning load on the AI at the start of operation.
[0036] In another aspect of the question-answering system of the present invention, the question receiving and answering system and the primary answering system are separated, and different final answer candidates for each system user are stored together, allowing multiple system users to share a single question-answering system. This allows the system implementation costs for each system user to be reduced.
[0037] Furthermore, a question answering system according to another aspect of the present invention separates the question receiving and answering system from the primary answering system, and performs machine learning to generate the same primary answer for multiple expected questions in different languages that share common attributes, and stores final answer candidates in different answering languages in the question receiving and answering system, thereby enabling questions in different languages to be accepted with a simple configuration.The answering language can be configured to answer in one language, or to answer in multiple predetermined answering languages according to the language of the question. [Brief explanation of the drawings]
[0038] [Figure 1] 1 is a conceptual diagram for explaining a usage mode of a question answering system according to a first embodiment of the present invention. [Figure 2] FIG. 10 is a schematic diagram showing a configuration example of another embodiment (second embodiment) of the question answering system of the present invention. [Figure 3] 3(a) is a functional block diagram illustrating an embodiment of the question receiving and answering system of the present invention, and FIG. 3(b) is a functional block diagram illustrating an example of the configuration of the final answering means shown in FIG. 3(a). [Figure 4] FIG. 4 is a functional block diagram showing an example of the configuration of the maintenance unit shown in FIG. 3. [Figure 5] 1 is a flowchart illustrating an example of a flow of responses to questions according to the present invention. [Figure 6] FIG. 2 is a diagram illustrating specific examples of common attributes and the correspondence between the first answers and the common attributes for a group of questions Q1 to QN that are trained by a first answering system such as an AI used in the present invention. [Figure 7] FIG. 2 is a diagram illustrating an example of a question group and a first answer according to the present invention. [Figure 8] FIG. 10 is a diagram illustrating another question group and a first answer according to the present invention. [Figure 9] FIG. 10 is a diagram illustrating further question groups and first responses. [Figure 10] FIG. 2 is a diagram illustrating an example of the relationship between a first answer identifier and a final answer stored in a final answer storage unit according to the present invention. [Figure 11] FIG. 11 is a diagram illustrating the content of a basic answer of a final answer corresponding to the identifier "NNN" shown in FIG. 10. [Figure 12] 10 is a flowchart illustrating the basic procedure for initial setup. [Figure 13] 10 is a flowchart illustrating a procedure for a company to maintain the final response in the final response storage unit. [Figure 14] 10 is a flowchart illustrating an example of a maintenance procedure for regularly or irregularly reviewing re-learning and basic answers of the question answering system. [Figure 15] FIG. 1 is a diagram showing specific examples of common attributes of expected questions in each question group Q1 to QN in an administrative agency (for example, a city hall or ward office), and the correspondence between the first answers corresponding to the question groups. [Figure 16] FIG. 10 is a diagram showing the relationship between specific examples of expected questions in a multilingual question group Q1 and their first answers. [Figure 17] FIG. 10 is a diagram showing the relationship between specific examples of expected questions in multilingual question group Q2 and their first answers. [Figure 18] FIG. 10 is a diagram showing the relationship between specific examples of expected questions in a multilingual question group QN and their first answers. [Figure 19] FIG. 10 is a diagram illustrating the relationship between a first response identifier, a language identifier, and a final response in multiple languages. [Figure 20] 10 is a flowchart illustrating a multilingual processing procedure. [Figure 21] FIG. 10 is a diagram illustrating specific examples of common attributes for a group of questions Q1 to QN for explaining entities, and the correspondence between the common attributes and first answers. [Figure 22] This figure shows the relationship between the expected questions and first answers in question group Q1 to explain entities. [Figure 23] This figure shows the relationship between the expected questions and first answers in question group Q2 to explain entities. [Figure 24] This figure shows the relationship between the expected questions and first answers in question group Q3 to explain entities. [Figure 25] A diagram showing the relationship between the expected questions and first answers of question group QN-1 to explain entities. [Figure 26] This figure shows the relationship between the expected questions and first answers of the question group QN to explain entities. [Figure 27] FIG. 10 is a diagram illustrating a relationship between a first response identifier and a final response. [Figure 28] FIG. 10 is a diagram illustrating correspondence between first answers and intent question groups. DETAILED DESCRIPTION OF THE INVENTION
[0039] The question answering system, question receiving and answering system, initial reply system, and question answering method according to the present invention will be described below with reference to the drawings. Fig. 1 is a diagram for schematically explaining the usage pattern and overall schematic configuration of a question answering system according to a first embodiment of the present invention. Fig. 1 illustrates a question receiving and answering system 10 and initial reply system 30 that constitute a question answering system 25 according to the present invention, as well as some of the various devices and communication networks that cooperate from the time a question sent from a question terminal 90 is received by the question answering system 25 of the present invention until a final reply is returned to the question terminal 90.
[0040] A question from a question terminal 90 is transmitted to the question receiving and answering system 10 via a question receiving page on a system (hereinafter referred to as the "user system") 40, which serves as a platform for users (e.g., businesses) connected via a communication network 91. If the question terminal 90 is a terminal device such as a desktop PC or smartphone connected to the user system 40, the question is transmitted directly from the user system 40 to the question receiving and answering system 10. Here, the communication network 91 is assumed to include a wireless communication network, relay station equipment, and multiple networks (not shown) (including the system provider's LAN). Note that while FIGS. 1 and 2 show a configuration in which the user system 40 is connected to an external network, it may also be connected to the local area network of the provider that provides the question answering system.
[0041] 1 to 14, businesses A to N are exemplified as users of the question answering system. However, users of the question answering system of the present invention are not limited to so-called commercial or industrial businesses. In other words, users are not limited to for-profit corporations or organizations, and any organization or individual, including, for example, universities, hospitals, local governments, and associations, can be users of the question answering system of the present invention (system users).
[0042] In FIG. 1, reference numeral 25 denotes a question answering system according to a first embodiment of the present invention, which includes a question receiving and answering system 10 and a primary answering system 30. The primary answering system 30 includes a knowledge database such as artificial intelligence, and a final answer storage unit 18, which constitutes part of the question receiving and answering system, stores final answer candidates, which are answers to questions. The final answer storage unit 18 can be configured as a storage device such as a server or a database. The final answer storage unit 18 may be an internal storage device of the question receiving and answering system 10, connected via a LAN, or connected via an external communication network. In FIG. 1, the final answer storage unit 18 is indicated by a dashed line, meaning that it may exist as a storage device internal to the question receiving and answering system 10 or as a storage device connected to a network. FIG. 3 shows a functional block diagram as an example of a configuration illustrating the main functions of the question receiving and answering system 10.
[0043] The question receiving and answering system 10 is composed of hardware such as a server equipped with a CPU, ROM, RAM, etc., and basic software and various other software, and is a computer system equipped with a question receiving means for receiving questions via a communication network 91, a first answer inquiry means for inquiring about a first answer from the first answering system 30, a final answering means for returning a final answer to the question terminal 90, and a final answering section, etc. The first reply system 30 is a computer system that includes an interface means for accepting questions and returning answers from artificial intelligence, and artificial intelligence for outputting the first reply as an identifier by machine learning anticipated questions and answers.
[0044] In conventional question-answering systems, it was necessary to machine-learn a huge amount of expected questions and answers to increase satisfaction with the answers. This required a lot of time and effort to collect a huge amount of expected question and answer data for machine learning in AI, and to machine-learn this data. Furthermore, machine learning in AI was essential to improve and update the accuracy of the question-answering system. In contrast, in the question answering system 25 of the present invention, the question receiving and answering system 10 and the initial answering system 30 equipped with artificial intelligence are clearly separated, and the initial answer by the initial answering system 30 is separated from the final answer that is actually returned. This allows the initial answering system 30 equipped with artificial intelligence and the question receiving and answering system 10 that returns the final answer to be separated and improved and evolved separately. That is, the question answering system 25 of the present invention can improve and evolve the system by subjecting the artificial intelligence to machine learning. In addition, it is possible to evolve and improve the question answering system 25 so that it can provide more appropriate answers to questions by appropriately adding and changing the content of the final answer candidates of the question answering system 25 based on an analysis of a question answering log, etc.
[0045] Furthermore, since the present invention allows for the system to be built on the premise that the final answer candidates can be easily changed and improved, it is possible to enrich the final answer candidates by reducing the number of expected questions and the number of initial answers that the AI learns from machine learning at the time of implementation compared to conventional systems. This reduces the implementation costs and the machine learning load at the time of implementation.
[0046] First, a hypothetical question that is subjected to machine learning by the primary response system 30 equipped with artificial intelligence, and a primary response that is output from the primary response system 30 as a response to the question will be described. Conventional artificial intelligence is trained through machine learning (supervised learning) to learn multiple expected questions and answers to those expected questions. Using the example of a city hall question-answering system, typically, in response to questions such as "Where should I go to pay my taxes?" or "Which counter do I use to pay my taxes?", the system is trained to output an answer such as "Counter 3 on the second floor is the tax counter." However, the primary response system 30 of the present invention is trained to output a specific identifier as the primary response, rather than a specific answer to the expected question. The identifier may take any form as long as it is unique enough to identify the primary response, but a short and simple configuration is desirable as long as it can identify each individual primary response. In this specification, a simple three-digit numeric identifier such as "100" is output.
[0047] In other words, the primary response system 30 performs machine learning so that for multiple anticipated questions that have some common attribute (hereinafter simply referred to as a "common attribute"), one and the same identifier is output as the primary response for each anticipated question. The size and range of the common attribute can be set arbitrarily.
[0048] In this specification, a bundle of anticipated questions that share one common attribute is referred to as a question group. That is, multiple anticipated questions that share a common meaning belong to one question group, and machine learning is performed to output one common first answer (identifier) for a concept (attribute) common to each anticipated question included in one question group. In this case, the number of first answers (identifiers) will be the same as the number of question groups, and the same first answer (identifier) will be output for each anticipated question that belongs to the same question group. For example, in the above example of anticipated questions about tax payment, if answers corresponding to questions related to "tax" are grouped into one question group, machine learning will be performed to output one common identifier as the first answer for all anticipated questions related to taxes.
[0049] What is considered a common attribute can be determined depending on the question group. For example, in the example of a possible question about "paying taxes" mentioned above, "place to pay taxes" is a common attribute. Here, it is possible to freely decide whether to put "tax refunds" and "tax consultations" in the same question group or in separate question groups. For example, it is possible to broaden the scope of common attributes and make everything related to taxes a single answer. In this case, one question group is formed with "tax" as a single common attribute, and the first answers to many possible questions about taxes are all output as a single identifier. When a question is received from the question receiving and answering system 10, the primary answering system 30 returns a predetermined identifier together with the answer certainty to the question as a primary answer to the question generated by the artificial intelligence.
[0050] In the question receiving and answering system 10, final answer candidates are stored in the final answer storage unit 18, linked to identifiers (first answers) learned by machine learning in the first answering system. Each final answer candidate preferably includes all answers to multiple expected questions with common attributes included in at least each question group. Furthermore, it is possible to configure the final answer candidate to include not only direct answers to the expected questions learned by machine learning but also answers to related questions. This makes it possible to reduce the system implementation period and implementation costs while maintaining satisfaction with the answers. Therefore, since the final answer candidates include many answer contents, such as answers to multiple expected questions and other predicted answers, the final answer candidates in the present invention tend to be long. When the question receiving and answering system 10 receives a first answer (identifier) from the first answering system 30, the final answer candidate corresponding to the identifier is retrieved, linked to each identifier, and stored in the final answer storage unit 18.
[0051] The final answer storage unit 18 stores, as final answer candidates, at least basic answers that form the basis of answers to all expected questions belonging to each question group, linked to identifiers corresponding to the primary answers of each question group. When only basic answers are stored as final answer candidates corresponding to each identifier, the basic answers become the final answers for that identifier. The content of the basic answers preferably includes answers to all expected questions related to the question group corresponding to the linked identifier. More preferably, the basic answers also include answers to expected questions with the same common attributes as the question group, even if they are not included in the expected questions of the question group. The final answer storage unit 18 can store unique answers in addition to basic answers as final answer candidates. Unique answers are final answer candidates that have the same basic framework as basic answers, but that replace basic answers or add new information to basic answers. Therefore, when the question receiving and answering system 10 receives a first answer (identifier) to a question from the first answer system 30, it reads out either a basic answer or a unique answer from the final answer candidates in the final answer storage unit 18 as the final answer candidate corresponding to the received identifier, and returns it to the question terminal 90 as the final answer to the question.
[0052] As will be explained later using specific examples, unique answers are used when the content of the answer to a question varies depending on the system user. For example, information identifying the system user (such as an identification number) is obtained based on the unique attribute information of the question sent from the question terminal 90, and a "basic answer or unique answer" linked to the primary answer (identifier) received from the primary answer system 30 and the system user's identification number is read from the final answer storage unit 18. This makes it possible to read out the unique answer corresponding to the system user's identification number and return a different final answer to each system user, even if the primary answers have the same identifier.
[0053] Here, "question-specific attribute information" refers to information specific to each question received from the question terminal 90, and may include various information such as information about the question terminal 90, information other than the question entered by the questioner, communication information and communication path information from the question terminal 90 to the question answering system, information about the person asking the question (such as the system user's identification number), and information about the language of the question and the language of the answer. This "question-specific attribute information" may be automatically acquired by the question receiving and answering system 10 or the like from communication information and communication path information, or may be configured to prompt the questioner to enter information as needed. Furthermore, the system user's identification number may be added to the system user's identification information in the user system 40 when the question is transmitted via the user system 40.
[0054] A typical use of a unique answer is when it is necessary to return a final answer that includes business hours and location specific to a system user, which differ from the contents of the basic answer. For example, when a system user who is a real estate agent asks a question about a real estate agent from a question terminal, it is necessary to inform the system user (real estate agent) of the business hours and location specific to that system user. In such a case, a unique answer that includes the business hours and location of the system user (real estate agent) in addition to the basic answer is stored in the final answer storage unit as an answer corresponding to the primary answer. As a result, the system user's unique answer is read based on the primary answer and the system user's identification number, and the system user's business hours and location are returned to the question terminal 90 as a final answer.
[0055] Furthermore, as will be explained in detail later, by including anticipated questions in different languages with common attributes in a question group in the artificial intelligence of the initial answering system and using machine learning to generate anticipated questions and initial answers, it is possible to accept questions in multiple question languages. The language of the answers to questions does not necessarily have to be the same as the language of the question; the final answer may be returned in a language set as the default for the question answering system. Because document reading ability is relatively high compared to document writing ability, many people can read and understand answers even if it is not their native language, so there is no need to answer in the same language as the question. To reduce the cost of creating answers and shorten the implementation period, it is useful to multilingualize only the question language. If there is more leeway in implementation costs and time, it is also useful to configure the system so that answer languages are limited to languages that are likely to be used, such as Japanese, English, or Chinese. It is also possible to configure the system to provide the final answer in the same language as the question. By linking the final answer in each answer language to each identifier of the primary answer and answer language information corresponding to the question language, and storing the final answers in various answer languages in the final answer storage unit 18, it becomes possible to return final answers in multiple answer languages.
[0056] As is clear from the above explanation, the present invention separates the initial reply system 30 from the question receiving and answering system 10, and configures the initial reply system 30 to perform machine learning of the artificial intelligence and the creation and updating of final answers in the question receiving and answering system 10 separately and independently. Furthermore, the initial reply system 30 and the question receiving and answering system 10 are connected by the concept of common attributes of expected questions and the identifier of the initial answer. Therefore, the expected questions to be machine-learned, the initial answers (identifiers), and the final answers are aligned, allowing for a structure in which the expected questions and final answer candidates are connected to each other. Therefore, it is possible to construct flexible final answer candidates for the expected questions to be machine-learned, and it is also possible to perform maintenance, improvements, and updates of the question answering system 25 separately and independently of each other.
[0057] As described above, by separating the first reply system 30 and the question receiving and reply system 10, final answer candidates can be flexibly constructed, which enables various ideas and efforts to reduce the system implementation costs and load depending on the field of the question answering system to be applied. For example, by narrowing down to a specialized field or a specific field in a specific industry, the predicted questions and the required answers become relatively clear, so the number of expected questions can be reduced and machine-learned by the first answering system 30, and the final answer candidates of the question receiving and answering system can include answers corresponding to questions that have not been machine-learned by the artificial intelligence. This makes it possible to improve the quality of the final answers while suppressing the number of machine-learned expected questions and the number of first answers. In other words, the burden at the time of implementation can be significantly reduced, making it possible to shorten the implementation cost and the time until operation begins. In addition, the number of expected questions can be narrowed down by narrowing the range of target questions to questions about specialized fields, specific industries, and specific fields. Narrowing down to specialized fields or specific fields in specific industries makes it relatively easy to predict questions, so the number of types of common attributes can be reduced.
[0058] Furthermore, the final answer candidates basically include all answers to anticipated questions with common attributes, and also include answers related to the anticipated questions. Therefore, for one question, a final answer including many answers with high question frequency in fields related to the question is simultaneously provided for one question. Therefore, there is a high possibility that the final answer will also include the answer to the question the asker had planned to ask next. In that case, there will be no need to ask a second or subsequent question, which also has the effect of suppressing questions.
[0059] By configuring the initial answering system 30 and the question receiving and answering system 10 separately in this way, the number of anticipated questions and initial answers can be reduced, thereby reducing the load on machine learning and lowering the costs of building and implementing the question answering system 25. Furthermore, by narrowing the range of acceptable questions and organizing the anticipated questions and narrowing the number of common attributes, the number of anticipated questions and initial answers (identifiers) to be machine-learned can be significantly reduced. For example, in order to increase the accuracy rate, the number of anticipated questions and answers, which conventionally required at least tens of thousands, can be narrowed down to just 500 to 1,000 anticipated questions and initial answers by narrowing the range of questions to specific tasks in specific industries. This allows the question answering system to begin operation by performing machine learning on these anticipated questions and initial answers.
[0060] Furthermore, by narrowing down the industry or business, it is possible to create highly accurate anticipated questions that are standardized to correspond to the industry or business and build a question-answering system, which makes it possible for the same question-answering system to be used jointly by multiple companies in the same industry.If multiple companies in the same industry can share the use of one question-answering system, the introduction costs can be shared, further lowering the barrier to introducing the system. With regard to the shared use of one question answering system 25 by fellow businesses, as mentioned above, by storing unique answers in the final answer storage unit 18 and configuring it so that a different final answer can be returned for each user business based on the unique attribute information of the question, it becomes possible to provide an individual final answer that suits the individuality of each user, thereby making it possible to realize an even more preferable form of shared use. The above-described "common attributes," "anticipated questions," "primary answers," "basic answers," and "unique answers" will be explained in more detail later using specific examples.
[0061] In addition, in the first response system 30 that provides the first response of the present invention, WATSON (WATSON is a registered trademark) by International Business Machines Corporation or the like can be used as a general-purpose artificial intelligence service.
[0062] In the above description, an example of a configuration in which the primary reply system 30 is connected to the primary reply system 30 via the communication network 91 has been shown as the first embodiment of the question answering system. However, as shown as an example in Fig. 2, a question answering system 25' according to another embodiment of the present invention may connect the question receiving and answering system 26 and the primary reply system 27 via an internal line or a dedicated communication line, without going through an external communication network. The configuration in Fig. 2 differs only in that the question receiving and answering system 26 and the primary reply system 27 are connected via an internal line (such as a dedicated line), and the basic configuration and functions of the question receiving and answering system 26 and the primary reply system 27 are the same as those of the question receiving and answering system 10 and the primary reply system 30 in Fig. 1, respectively.
[0063] 1 shows an example in which a management terminal (system management terminal 50) of a system provider that operates and manages question answering system 25 is connected to communication network 91. As is clear from the explanation that communication network 91 also includes a LAN, system management terminal 50 can also be configured to be connected to question answering system 25 (25') via an internal communication system (LAN).
[0064] Next, an example of the main configuration of the question receiving and answering system 10 (26) will be described with reference to Fig. 3(a). Fig. 3(a) is a functional block diagram illustrating the configuration of the question receiving and answering system 10 (26) of the present invention, which is composed of, for example, hardware such as a CPU, bus control and registers and other control circuits, input / output devices, memory, and hard disk, as well as software and various programs that control this hardware. The question receiving and answering system 10 (26) basically comprises a communication interface 11 with a transmission and reception function, a receiving means 12 for receiving questions from outside, a question terminal 90 and an operation history storage unit 13 for storing information related to questions, a first response inquiry means 14, and a final response means 15. It is also desirable to have a maintenance means 16.
[0065] The receiving means 12 of the question receiving and answering system 10 (26) receives a question from a question terminal 90 and stores the unique attribute information of the question in, for example, the operation history storage unit 13. The first answer inquiry means 14 inquires of the first answer system 30 (27) about the question received from the receiving means 12 and receives a first answer and, if possible, its answer confidence level from the first answer system 30 (27). When the first answer inquiry means 14 receives the first answer and the answer confidence level from the first answer system 30 (27), it outputs this information to the final answering means 15. When the first answer received from the first answering means satisfies a predetermined condition (for example, when the answer confidence level is equal to or greater than a predetermined threshold), the final answering means 15 reads out a final answer candidate (basic answer or unique answer) corresponding to the first answer and the unique attribute of the question from the final answer storage unit 18 (28) and transmits it to the question terminal 90. The operation history storage unit 13 stores the operation history of each process from receiving a question to providing a final answer, such as the question's unique attributes, question content, first answer, answer confidence level, final answer, etc. as needed. Note that in the above and following explanations, an example is shown in which the answer confidence level is always received together with the first answer from the first answer system 30 (27), but it is not essential to receive the answer confidence level from the first answer system 30 (27). Also, a configuration may be adopted in which other similar information is received instead of the answer confidence level.
[0066] The above example shows receiving one primary answer (identifier) from the primary answer system 30 (27), but it is also possible to receive a list of multiple primary answer candidate identifiers from the primary answer system, select the only appropriate primary answer (identifier) in the question receiving and answering system 10 (26), and obtain the final answer.
[0067] The maintenance means 16 receives requests for collection, analysis, and maintenance of log information and various types of operation information from the system administrator or system users, and executes processes such as information collection and analysis, maintenance, etc. The maintenance means 16 allows even a system user who does not have specialized knowledge of artificial intelligence, for example, to independently manage, modify, or change part of the contents of the final answers in the final answer storage unit 18 (28) within a predetermined scope of authority, separately and independently from the system administrator of the question answering system 25 (25').
[0068] FIG. 3(b) is a functional block diagram showing an example of the configuration of the final response means 15 of FIG. 3(a). When a first answer is received from the first answer system 30 (27), the first answer is transmitted from the first answer inquiry means 14 to the final answer means 15. When the final answer control unit 35 of the final answer means 15 receives the first answer, it acquires the unique attribute information of the question stored in the operation history storage unit 13 via the unique attribute information acquisition unit 36 and transmits it to the final answer control unit 35. Based on the identifier of the first answer and the unique attribute information (information on the system user, etc.), the final answer control unit 35 commands the final answer acquisition unit 37 to acquire a basic answer or a unique answer corresponding to the first answer from the final answer storage unit 18 (28). The basic answer or unique answer acquired by the final answer acquisition unit 37 is sent to the final answer transmission unit 38 and, under the control of the final answer control unit 35, is transmitted as a final answer to the question terminal 90 via the communication interface 11.
[0069] Fig. 4 is a functional block diagram showing an example of the configuration of the maintenance means 16 shown in Fig. 3(a). The maintenance means 16 includes an initial setting means 21, an analysis means 22, and a data update means 23. The maintenance means 16 receives operations from the system management terminal 50 via the initial setting means 21, and performs tasks such as inputting and setting multiple expected questions and primary answers (identifiers) when building the system and setting machine learning data for the primary answer system 30 (27), and stores final answer candidates (basic answers and unique answers as needed) in the final answer storage unit 18 (28). The analysis means 22 creates various statistical data such as statistical information on questions, first answers, final answers, answer certainty, etc. stored in the operation history storage unit 13, for example, the number of questions and answers for each system user, the number of questions and answers for each answer group, the average answer certainty, the number and ratio of cases where the answer certainty is below a reference value, etc. These statistical data can only be monitored by persons with predetermined authority.
[0070] The data update means 23 accepts access from the management terminal of the user system 40 and the system management terminal 50, and updates the desired data. The system management terminal 50 can update the knowledge data of the primary answer system 30 (27) through machine learning, and can update the contents of the final answer candidates in the final answer storage unit 18 (28). For example, when the answer certainty of the primary answer to a question is low, new expected questions and primary answers (identifiers) can be added and machine learning can be performed on the artificial intelligence of the primary answer system 30 (27). In addition, it is possible to create final answer candidates corresponding to new identifiers in the final answer storage unit of the question receiving and answering system, and to add, modify, or change existing basic answers and unique answers. The initial setting means 21 can also have its functions integrated into the data update means 23.
[0071] A system user can read the access log information of the system user from the management terminal of the user system (40), view the operation history and analysis data, and correct or change the final response data stored in the final response memory unit 18 (28) within a predetermined permitted range. This allows system users (vendors), regardless of whether they have specialized knowledge of artificial intelligence, to change, modify, or add to the contents of the final answers stored in the final answer storage unit 18 (28) themselves. In other words, system users (vendors, etc.) can modify and update the final answers (basic answers and unique answers) stored in the final answer storage unit 18 within a predetermined permitted range using the management terminal of the user system 40 connected via a communication network, without having to request the system provider.
[0072] A basic flow of an answer to a question from a question terminal 90 in a question-answering system 25 (25') according to one embodiment of the present invention will be described with reference to Figure 5. Note that, in the following description, an example of a processing procedure for providing a unique answer using unique attribute information of a question will be described, but in the question-answering system 25 (25'), it is possible to freely select whether or not to use unique attribute information and whether or not to provide a unique answer. When a question q1 is received from the question terminal 90 (step S1), the reception means 12 stores the question's unique attribute information and question content, etc. in the operation history memory unit 13 (or another memory unit) (step S2), and then the first answer inquiry means 14 queries the first answer system 30 (27) for the first answer to the question q1 (step S3).
[0073] The primary reply system 30 (27) extracts a primary reply (identifier of the final answer: a1 to an in FIG. 5) to the question q1 and returns it to the primary reply inquiry means 14. The primary reply system 30 (27) transmits one or more primary answers (identifiers) in descending order of answer certainty, for example, and the answer certainty corresponding to the primary answer, to the primary reply inquiry means 14. When the primary reply inquiry means 14 receives a primary answer from the primary reply system 30 (27) (step S4), it sends it to the final reply means 15. The final reply means 15 verifies whether the primary answer (identifier) satisfies a predetermined condition for returning a reply to the questioner, for example, whether the answer certainty of the primary answer with the highest answer certainty among the received primary answers exceeds a predetermined threshold (step S5).
[0074] If the primary answer satisfies the predetermined condition (Step S5; Yes), the final answering means 15 acquires the unique attribute information of the question and obtains the system user's identification information (Step S6). The unique attribute information of the question is stored in the operation history storage unit 13, etc., and by checking the communication path (origin) of the question from the unique attribute information, for example, it is possible to determine which system user (vendor) the question passed through and identify the system user (vendor). Next, based on the acquired primary answer and the information (identification information) of the system user (vendor), a basic answer candidate corresponding to the primary answer or a unique answer candidate corresponding to that system user (vendor) is acquired (Step S7). If there are multiple received primary answers, the final answer candidates (basic answer candidates or unique answer candidates) to be returned as the final answer are narrowed down. For example, the basic answer candidate or unique answer candidate corresponding to the highest answer certainty of the received primary answer (identifier) is acquired from the final answer storage unit 18 (28) (Step S7). Thereafter, the acquired final answer candidates are transmitted from final answer means 15 to question terminal 90 which has asked the question (step S9).
[0075] If the primary answer does not satisfy the predetermined condition in step S5 (step S5; No), for example, if the answer certainty is below a predetermined threshold, a predetermined answer prepared for when there is no appropriate final answer candidate for question q1, such as "Please ask more specifically," is obtained (step S8) and transmitted from final answering means 15 to question terminal 90 (step S9). Here, if the primary answer does not satisfy the predetermined condition in step S5, the final answer can be associated with a special identification number and stored in final answer storage unit 18 (28), or can be embedded as data in the program. Note that the order of the processing steps of steps S5 and S6 in FIG. 5 may be reversed.
[0076] When the final answer is transmitted to the question terminal 90, various information including history and other log data relating to the question and the final answer is stored in the operation history storage unit 13 (step S10). When a plurality of primary answers are presented from the primary answer system 30 (27) in descending order of answer certainty, it is desirable that the operation history storage unit 13 store all answer certainty or answer certainty up to a predetermined rank.
[0077] As is clear from the above explanation, the primary answer system 30 (27) outputs a primary answer (identifier) corresponding to the question group to which the expected question that is similar to the input question belongs, and, if the system has the function, the answer certainty. Incidentally, the questioner believes that he or she has asked a question to the system user (one of businesses A to N) that he or she accessed. Therefore, depending on the content of the question, it may be necessary to return a different answer depending on the user system 40 of business A to N that relayed the question from the question terminal 90. For example, in the case of a question regarding business hours, it is necessary to reply with the correct business hours for each business that the question has passed through. In such a case, unique answer candidates for each system user are prepared, and the system user's identification number is obtained based on the unique attribute information of the question stored in the operation history storage unit 13, etc., and a unique answer corresponding to the system user (business A to N) that the question has passed through is obtained, and a final answer informing the business hours for each system user is returned.
[0078] 6 to 11, we will explain using concrete examples the expected questions that the primary answer system 30 (27) will machine-learn, the primary answers associated with question groups, and the basic answers and unique answers that are stored as final answer candidates in the final answer storage unit 18 (28). In the examples of Figures 6 to 11, the term "agent" refers to a system user who wishes to use the question-answering system and is an agent engaged in real estate transactions. In the examples of Figures 6 to 11, multiple agents (agents A to N) have signed contracts to use the question-answering system 25 (25'), and each agent A to N can access the question-answering system 25 (25') via the respective question pages of their user systems 40. Figures 6 to 9 also show specific examples of question groups, and Figures 10 and 11 show examples of basic answer candidates and unique answer candidates stored in the final answer storage unit 18 (28).
[0079] The question answering system 25 (25') identifies the system user (vendor) related to the question based on information about which user system 40 each question passed through (communication route information: part of the question's unique attribute information), and acquires unique answer candidates specific to the system user. If the system user (vendor) does not request unique additional information, basic answer candidates are acquired and sent to the question terminal 90 as the final answer. Although not shown in the examples of FIGS. 6 to 11, it is also possible to have a simple configuration in which unique answers for each system user (vendor) are not recognized, and only basic answers are used as final answers for all questions.
[0080] First, with reference to Figures 6 to 9, the expected questions and the first answers to be learned by the first answer system 30 (27) will be described using specific examples. Note that the following description shows an example in which one question group includes multiple expected questions. FIG. 6 is a diagram illustrating the common attributes of each question group in question groups Q1 to QN, and the relationship between the first answers and each question group Q1 to QN in a one-to-one correspondence. In the following diagrams, question groups are referred to as "question groups" where appropriate. In FIG. 6, the first answers are given identifiers 101 to NNN. These identifiers can be freely expressed as long as they can identify the first answers. Each question group Q1 to QN has multiple expected questions organized under the concept of common attributes for each group, and FIGS. 7 to 9 show specific examples of some of the multiple expected questions included in each question group Q1 to QN.
[0081] In FIG. 6, question group Q1 is a group with a common attribute of "real estate agents," and the first answer has the identifier "101." Question group Q2 has a common attribute of "guidelines for mortgage loans," and the first answer has the identifier "102." Similarly, question group QN has a common attribute of "points to consider when gathering local information," and the first answer has the identifier "NNN." This identifier is the same as the identifier used when creating a final answer in the question receiving and answering system 10 or 26 described later, and is an identifier associated with a final answer candidate described later with reference to FIGS.
[0082] 7 to 9 show specific examples of multiple expected questions for question groups Q1, Q2, and QN. FIG. 7 shows specific examples of multiple expected questions Q1(1) to Q1(n) corresponding to the common attribute of question group Q1, "real estate agent." FIG. 8 shows specific examples of multiple expected questions Q2(1) to Q2(9) corresponding to the common attribute of question group Q2, "guidelines for borrowing a home loan." Similarly, FIG. 9 shows specific examples of multiple expected questions QN(1) to QN(12) corresponding to the common attribute of question group QN, "key points to consider when gathering local information."
[0083] In FIG. 7, a single identifier "101" is assigned to all of the first answers to the n anticipated questions in question group Q1. That is, all of the first answers to the n anticipated questions in question group Q1 that share the common attribute "real estate agent" have a single identifier 101. Correspondingly, the final answer storage unit 18 (28) stores basic answer candidates corresponding to the identifiers 101 of the first answers, and the basic answer candidate corresponding to the identifier 101 is a bundle of answers that includes all answers corresponding to the common attribute of the multiple anticipated questions Q1(1) to Q1(n) in the question group. Specific examples of the final answer candidates stored in the final answer storage unit 18 (28) will be described later with reference to FIGS. 10 and 11.
[0084] Similarly, Figure 8 specifically illustrates multiple expected questions Q2(1) to Q2(9) corresponding to question group Q2, and the same single identifier "102" is assigned as the first answer to all of these multiple expected questions Q2(1) to Q2(9). Similarly, Figure 9 specifically illustrates 12 expected questions QN(1) to QN(12) related to question group QN about the common attribute "points to consider when gathering local information," and the same identifier "NNN" is assigned as the first answer to all of these 12 expected questions.
[0085] 10 and 11 show specific examples of basic answer candidates and unique answer candidates stored in the final answer storage unit 18 (28). The display in the primary answer (identifier) column shown in Fig. 10 and Fig. 11 corresponds to the identifiers 101 to NNN shown in Fig. 6 to 9, which are primary answers from the primary answer system 30 (27). The final answer storage unit 18 (28) stores "basic answer candidates" in correspondence with the identifiers 101 to NNN of the primary answers received from the primary answer system 30 (27), and, if desired by the system user (vendors A to N), "unique answer candidates" corresponding to the desired system user. In the user (vendor) identifier column in Figures 10 and 11, the system user identifiers are displayed as "vendor A," "vendor B," etc., but numbers, letters, symbols, etc. can be used freely as long as they can identify the system user (vendor). Note that in Figures 10 and 11, the information identifying the user is the system user identifier, but this can be any identifier that identifies each user, and may also be "identification information of the user system."
[0086] The system user (vendor) identifier is a unique identifier assigned to each vendor A to N, and the unique answer candidates for each vendor are stored in the final answer storage unit 18 (28) in association with the vendor's identifier. FIGS. 10 and 11 illustrate examples of basic answer candidates assigned as final answers. Referring to the primary answer identifiers "102" and "NNN" in FIGS. 10 and 11, the primary answer identifier "102" field does not contain a unique answer candidate corresponding to vendor B's identifier, and the primary answer identifier "NNN" field does not contain a unique answer candidate corresponding to vendor A's or vendor C's identifiers. Thus, the system user (vendor) does not need to set unique answer candidates; they can be set arbitrarily at the user's (vendor's) discretion. Furthermore, a configuration may be adopted in which all unique answer candidates are eliminated for all primary answers or for specific primary answers, leaving only basic answer candidates. The identifier of the system user (vendor) for each question is determined from the unique attributes of the question, and the basic answer candidate or the corresponding unique answer candidate corresponding to each question is read from the final answer memory unit 18 (28).
[0087] Specific examples of basic answer candidate content are shown in Figures 10 and 11. As can be seen from the content of these basic answer candidates, the basic answer candidates and unique answer candidates that become the final answers are relatively long answers that include all answers related to multiple expected questions in question groups with common attributes. This is based on the fact that the final answers are bundles of answers that include all answers related to multiple expected questions included in each question group. This configuration makes it possible to maintain a certain level of satisfaction for the asker with the final answer to the question, even if the number of expected questions and first answers that the first answer system learns through machine learning is reduced. Because the expected questions are carefully selected from questions that are likely to be asked, basic answers that include all answers to the expected questions contain a lot of useful information for the asker. Therefore, a basic answer to one question may also include an answer to the question that the asker had planned to ask next, further increasing the asker's satisfaction. By assigning a wide range of meanings to multiple anticipated questions in this way, it is possible to reduce the time required for creating learning data and for machine learning in AI to 1 / 10 or 1 / 100 or less, making it possible to provide AI-based question-answering services efficiently and at very low cost.
[0088] Furthermore, the primary response system 30 (27) always outputs the same primary response (identifier) for expected questions belonging to the same question group, regardless of the unique attributes of the question, such as which user system the question passed through. In contrast, the question receiving and response system 10 (26) reads out a unique answer candidate specific to the system user as needed, replacing the basic answer candidate stored in association with the primary response, and transmits the read unique answer to the question terminal 90 as the final answer. For this reason, the unique answer candidate is stored in association with the identifier of the primary response and the identification information of the system user (system user identifier). The unique answer candidate is a basic answer candidate corresponding to the identifier of the primary response that has been replaced to suit the system user. In this way, the final answer candidates (basic answer candidates and unique answer candidates) stored in the final answer storage unit 18 (28) are stored data separate from the artificial intelligence, so their contents can be changed or added as appropriate regardless of machine learning for the artificial intelligence. Therefore, maintenance work such as changing or adding final answer candidates stored in the final answer storage unit 18 (28) does not require advanced specialized knowledge of artificial intelligence. However, since the basic answer candidates correspond to the first answer of the first answer system, form the basis for the final answer, and are answer contents common to all system users, it is preferable to restrict their changes by system users.
[0089] As is clear from the above description, the question-answering system 25 (25') according to the present invention allows multiple system users (e.g., vendors) in the same industry or business to jointly use data content, such as anticipated questions and final answer candidates, accumulated through experience and track record in order to efficiently respond to questions from end users or potential end users, as well as equipment for automatically providing appropriate answers in response to questions, such as advanced computer systems and databases, such as artificial intelligence systems. Building a question-answering system that allows such joint use not only significantly reduces the construction costs when introducing a question-answering system for the first time, but also shortens the system implementation period. Furthermore, when a vendor sharing the same data or system later adopts the system, the data and system environment are already configured, allowing for the reliable introduction of an excellent question-answering system in an extremely short period of time. Furthermore, because the question-answering system is shared by many users, it can respond to a wide variety of questions in a short period of time that would not be possible for a single company to handle. This allows for the system to be improved in a short period of time.
[0090] As described above, the question receiving and answering system 10 (26) according to the present invention is equipped with the maintenance means 16, and the initial setting of the primary answering system 30 (27) and the final answer storage unit 18 (28) is performed by the initial setting means 21. The initial setting procedure of the primary answering system 30 (27) and the final answer storage unit 18 (28) will be described with reference to FIG. 12.
[0091] 12 is a flowchart illustrating the procedure for the basic work involved in the initial setup of the primary response system 30 (27) and the final response storage unit 18 (28). To build the system, first, anticipated questions, common attributes of anticipated questions, question groups, primary responses, final responses (basic answers, unique answers), and unique attributes of questions are determined (step S21). To achieve this, the work of system users (industries) is analyzed to create multiple expected questions, and answers are created for all of the expected questions. Expected questions with common attributes are grouped into a single question group, and the primary answers are assigned identifiers with a one-to-one correspondence with the question group. Of the final answer candidates, the basic answer candidates are designed to cover all answers related to the multiple expected questions within a single question group. In addition, unique attribute information is determined, and the range and content of the unique answer candidates to be linked to the unique attribute information is determined. At this time, language information can also be included in the unique attribute information. In this case, expected questions in various languages and different final answer candidates are prepared as necessary. Such final answer candidates are created for each and every question group.
[0092] Next, the primary answer system 30 (27) performs machine learning on all expected questions included in each question group Q1 to QN and the primary answers (identifiers) that correspond one-to-one to each question group (step S22). Furthermore, the final answer storage unit 18 (28) stores the identifiers of the primary answers in association with basic answer candidates (step S23). As described above, in step S23, not only basic answer candidates for one identifier corresponding to the primary answer are stored, but also unique answer candidates that differ for each system user based on a common basic answer candidate, associated with each primary answer and the identifier of the system user. Note that steps S22 and S23 may be performed in reverse order or simultaneously.
[0093] As described above, the question-answering system 25 (25') according to the present invention is equipped with the maintenance means 16. This maintenance means 16 allows the user to periodically (or irregularly) check the question and answer history, create and analyze statistical information, and otherwise continuously monitor the effectiveness of system usage, thereby making it possible to correct or add basic answer candidates and unique answer candidates.
[0094] The question-answering system 25 (25') of the present invention can continuously grow by using machine learning in the primary answer system 30 (27) and maintaining the final answer candidates stored in the final answer storage unit 18 (28) based on monitoring results. Furthermore, regularly checking questions with low answer confidence can provide users with an opportunity to learn about the needs of existing and potential customers that they themselves may not be aware of. This not only fulfills the original purpose of the question-answering system 25 (25'), which is to simply answer questions, but can also serve as a means of gathering customer and market information necessary for business growth. Furthermore, as explained below, even system users without artificial intelligence knowledge can maintain the basic answer candidates and unique answer candidates under certain conditions, allowing system users themselves to continuously improve and grow the question-answering system they use. Furthermore, because this question-answering system can be shared by multiple system users, it can receive a much larger and more diverse range of questions than would be possible if it were used by a single company. Therefore, when the system provider performs maintenance on the entire question-answering system 25 (25'), it is expected that the expected questions and basic answer candidates will be revised and improved to address a wide variety of issues, accelerating the growth of the question-answering system. All shared users can benefit from the growth of this question-answering system.
[0095] 13 and 14, the procedure for maintenance of the primary response system 30 (27) and the final response storage unit 18 (28) by the system user (vendor) or system administrator will be described. First, the procedure for correcting and changing the final response storage unit 18 (28) by the system user (vendor) will be described.
[0096] FIG. 13 is a flowchart showing the processing steps for a system user (vendor) to modify or change the final answer specific to their own company. A system user (vendor) can access the maintenance means 16 (see Figures 3 and 4) of the question-answering system 25 (25') using a system user's management terminal connected to the question-answering system 25 (25') via any communication network, and check the status of the final answers, such as the distribution of answer confidence levels for questions and first answers, the average answer confidence level, etc., from records such as access logs of questions posed to the system user, and consider correcting, changing, or adding to the contents of the final answer candidates (basic answers and unique answers).If an appropriate final answer could not be provided to a question from a question terminal, or if a final answer was provided but the answer confidence level of the first answer was low, the system user (vendor) can correct or add to the final answer candidates stored in the final answer storage unit within the scope of their authority, or by requesting a system administrator.
[0097] As shown in FIG. 13, the system user (vendor) periodically or irregularly checks and analyzes the access log (step S31). In this case, if the system user (vendor) desires to modify or change the final answers (basic answers and unique answers) stored in the final answer storage unit 18 (28) (step S32; Yes), it is then confirmed whether the system user has the authority to modify or change the final answers (step S33). Although not shown in FIG. 13, even when the system user (vendor) views operation history information and analyzes logs, etc. (step S31), it is desirable to confirm in advance with the person in charge whether the person has access authority. If the system user (vendor) does not have the authority to modify or change, etc., the modification or change is not permitted and the maintenance process ends (step S33: No). If the authority to modify or change, etc. is permitted (step S33: Yes), the system user (vendor) can modify (including change or addition) the final answer candidates stored in the final answer storage unit 18 (28) within the system user's authority (step S34).
[0098] For example, if a question q1 from the question terminal 90 asks "I would like to know about business hours," and if a unique answer has already been set for that system user, information specific to the business, such as temporary closures, can be added to the existing unique answer candidates. If the system user (business) wishes to make a modification beyond the system user's authority based on the results of analyzing log data, etc., such as changing the basic answer candidates, adding a possible question to an existing question group in the primary answer system 30 (27), adding a question group with a new common attribute, or modifying or adding a basic answer candidate, he or she should request the system administrator of the question answering system to consider the request.
[0099] 14 is a flowchart illustrating an example of a maintenance procedure for the entire question-answering system 25 (25′) performed by a system administrator. This maintenance procedure is performed via the system management terminal 50 by a system administrator who manages the operation of the entire question-answering system. When starting maintenance (step S41), first, it is confirmed whether there is access authority to this operation history (step S42), and if there is no access authority, this processing flow ends (step S42; No).If there is access authority (step S42; Yes), the access status of the entire system, log data of questions and answers, and analytical data such as statistical data stored in the operation history storage unit 13 are confirmed and analyzed (step S43).
[0100] When correcting or changing (including adding, the same applies below) the final answer candidates in final answer storage unit 18 (28) (Step S44; Yes), it is confirmed whether the user has the authority to correct or change (Step S45). If the user has the authority to correct or change (Step S45: Yes), the basic answer or unique answer is changed, or a new unique answer of the vendor is added (Step S46). If the user does not have the authority to correct or change (Step S45: No), the process is terminated. Of the final answer candidates, the system administrator may be permitted to correct or change only the basic answer candidates, but not the unique answer candidates, and only the system user may have the authority to correct or change the unique answer candidates. This diagram shows an example in which the system administrator has separate authority over final answer candidates and machine learning, but both authorities can also be handled by a single authority.
[0101] If the final answer candidates are not to be corrected or changed (step S44; No), or when the corrections and changes of the final answer candidates in the final answer storage unit 18 (28) are completed (step S46), it is next confirmed whether to add expected questions, etc. to the first answer system 30 (27), which is an artificial intelligence system, and perform machine learning (step S47). If additional machine learning is not necessary (step S47; No), this maintenance process is terminated. If additional machine learning is to be performed (step S47; Yes), it is confirmed whether the first answer system 30 (27) has the authority to correct or change the expected questions or first answers (identifiers) (step S48). If the authority to correct or change is present (step S48; Yes), desired additions, corrections, or deletions are made to the expected questions and first answers, and machine learning is performed by the first answer system (step S49), and the system maintenance process is terminated. If the authority to correct is not present (step S47; No), machine learning is not performed and the maintenance process is terminated.
[0102] The maintenance process by the system user (vendor) (FIG. 13) and the maintenance process by the system administrator (FIG. 14) described above can be carried out by updating and changing data based on the following rules, for example, to gradually evolve the question-answering system 25 (25'), enabling it to steadily grow step by step into a system with high answer certainty.
[0103] <Example of response rules when the answer confidence is low> (1) If the question is completely unknown and there is no corresponding final answer, a pair of question and answer identifiers is added to the primary answer system 30 (27), and the final answer is also added to the final answer storage unit 18 (28). This maintenance is performed by the system administrator. (2) If the question is unknown but there is an appropriate existing final answer, the system administrator adds a pair of the question and the identifier of the existing final answer to the primary answer system 30 (27).
[0104] <Example of response rules when the answer certainty exceeds the standard value but the answer content is inappropriate> (1) If the content of the question is known but the content of the final answer is not appropriate, the system administrator corrects only the data of the final answer in the final answer storage unit 18 (28). (2) When it is necessary to modify a basic response, which is a common response for all related businesses, the system administrator can modify the basic response in the final response storage unit 18 (28). (3) If the basic answer, which is a common answer, does not match the content of some system users (vendors), the system administrator of the system user (vendor) can copy the basic answer to the final answer area corresponding to their own company's vendor identification number and modify the content of the basic answer. At this time, if a final answer only for their own company is to be modified without affecting the common answer, it can be added or modified for each vendor's identifier.
[0105] <Rules for when dividing question groups is expected to improve answer accuracy> (1) If there are answer identifiers with a large number of questions and answers, and it is expected that narrowing the range of answers and dividing the question group and final answer will improve user satisfaction, the currently registered question group and final answer will be divided. This maintenance process is performed by the system administrator. (2) If the length of the answer to a question becomes too long, it is desirable to divide the question group and the final answer, just as in the case where the number of questions is large. This maintenance process is performed by the system administrator.
[0106] <Support for multiple languages> In the above-described embodiment and specific example, the question answering system of the present invention has been described by exemplifying expected questions and final answers using Japanese as the question language and answer language. The question-answering system 25 (25') according to the present invention can handle questions in a variety of languages by grouping anticipated questions and utilizing unique attribute information of the questions. An example of a configuration for handling questions and answers in a variety of languages in the question-answering system 25 (25') according to the present invention will be described below with specific examples of anticipated questions and final answer candidates.
[0107] When the native languages of the asker and the respondent are different, smooth communication is difficult unless either the asker or the respondent is familiar with the other's native language. In such situations, a question-answering system that can accept questions in the asker's native language allows the asker to create clear questions in correct writing. Being able to ask questions in one's native language improves the quality of the questions, significantly increasing the probability that the question-answering system will provide an appropriate answer.
[0108] Therefore, a question-answering system is desirable to have a mechanism for accepting questions in various languages. Generally, compared with document creation ability, document reading ability is relatively high even in languages other than the native language, so answers to questions do not necessarily have to be in the questioner's native language. For example, from the perspective of cost-effectiveness, it is possible to accept a wide variety of question languages, while limiting the answer languages to a few languages widely used around the world, such as the major languages of the country where the question-answering system is installed, English, and Chinese.
[0109] Such a multilingual question-answering system can significantly reduce the human burden of various consultation and guidance services at government agencies, universities, hospitals, etc. For example, in local governments, even if they can automatically answer simple questions in various languages via a question-answering system, such as where the counters and facilities for various procedures, such as the counter for moving-in procedures for new residents, the counter for providing resident registration certificates, and the counter for paying taxes, the labor required for individual responses by humans can be significantly reduced. In particular, during first-time consultations or in the early stages of consultations, many of the questions asked are formal or simple and basic, so a question-answering system that can accept questions in other languages is extremely useful for both foreign questioners and the institutions that use the system. Furthermore, the fact that questioners can use their smartphones or other devices to ask questions in advance from a remote location makes this a very useful system for both questioners and the institutions using the system, compared to the previous system in which questions had to be asked on-site.
[0110] In the question answering system 25 (25') according to the present invention, a question answering system 25 (25') that can handle a wide variety of languages can be constructed simply by adding a simple configuration to the basic configuration of the present invention. The question answering system 25 (25') according to the present invention is configured such that if the common attributes of the questions are the same, they belong to the same question group even if the question language is different, and by recognizing the difference in the type of question language as part of the unique attributes of the question, it is possible to handle questions and answers in a variety of languages. To this end, multiple anticipated questions that have the same meaning but are in different languages are machine-learned in the primary answer system 30 (27) as belonging to the same question group with the same common attribute. At the same time, the final answer storage unit 18 (28) stores final answer candidates (basic answer candidates or unique answer candidates) for each answer language type. Answers in the default language set in the question answering system 25 (25') are registered as basic answer candidates, and answers in other languages are registered as unique answer candidates.
[0111] When responding to questions in multiple languages, it is necessary to extract a correct primary answer according to the language of the question and to respond in an appropriate language corresponding to the language of the question. To determine the question language and the answer language, the questioner can input their desired question language and answer language. The simplest method is to provide a language selection button for specifying the type of question language and answer language on the question input screen of the question terminal when the question answering system 25 (25') is accessed from the question terminal 90. In this case, it is desirable to select the desired language using a pull-down menu or radio buttons, etc., so that only languages supported by the question answering system can be specified. Furthermore, if the question receiving and answering system 10 (26) or the primary answering system 30 (27) has the function of inferring the type of question language from the form of the question sentence (such as the bit configuration of the character code), the question receiving and answering system 10 (26) or the primary answering system 30 (27) can also determine the type of language from the question sentence. Furthermore, the type of question language can also be obtained from common information required for communication between the question terminal and the question answering system. Therefore, it is possible to infer the question language even without input from the questioner. The answer language can be handled by initially setting it to be determined according to a predetermined rule depending on the question language.
[0112] The desired question language and answer language types are sent to the question receiving and answering system 10 (26) along with the question and stored as part of the question's unique attributes, for example, in the operation history storage unit 13. Here, configuring the answer language to automatically be the same as the input language will improve user convenience. However, in light of the fact that document reading ability is significantly higher than document creation ability, and from the perspective of reducing the burden at the time of system implementation and the response load on the question answering system 25 (25'), the answer languages may be configured to be fewer than the number of languages in the question, such as limited to English or Japanese.
[0113] As described above, it is first necessary to receive questions in multiple languages from the primary reply system 30 (27) and obtain an appropriate primary reply regardless of the language of the question. To this end, multiple expected questions with the same common attribute are created according to the type of language, and the expected questions in multiple languages with one common attribute are grouped together into one question group, and machine learning is performed on the primary reply system so that a primary reply that corresponds one-to-one with each question group can be obtained. For example, question groups Q1 to QN are created in which expected questions in multiple languages, such as Japanese, English, and Chinese, that share a common attribute are grouped together. Here, the expected questions for each language that share the same common attribute may be expected questions with the same meaning, obtained by directly or freely translating expected questions created in one language into other languages. However, they may not necessarily be identical, but may have approximately the same meaning taking into account the various linguistic cultures. Next, the question groups Q1 to QN, which include expected questions in various languages, are trained by machine learning so that a first answer that corresponds one-to-one with each question group Q1 to QN is obtained. As a result, the first answer system 30 (27) obtains the same first answer (identifier) for expected questions that belong to the same question group, regardless of the type of question language.
[0114] In the question receiving and answering system 10 (26), final answer candidates in various languages are stored in a one-to-one correspondence with the primary answers in accordance with each answer language in the final answer storage unit 18. Therefore, when a primary answer is received from the primary answering system 30 (27), a final answer candidate in an appropriate answer language can be read based on the identifier of the primary answer and the language information (identifier) of the question, and the final answer can be returned to the question terminal 90.
[0115] Combination with unique answer candidates other than language, such as the vendor identifier mentioned above, is also possible. By preparing basic answer candidates corresponding to the first answer and unique answer candidates that differ for each system user (vendor) according to the type of language, it is possible to read out the unique answer candidate for each system user in the appropriate language according to the question language and use it as the final answer. Whether or not to prepare unique answer candidates corresponding to combinations of multiple types of unique attributes as the final answer can be set appropriately as needed. When a single question-answering system is not used by multiple institutions, such as when used independently by a government agency or university, unique answer candidates are not necessarily required.
[0116] When the question receiving and answering system 10 (26) receives a question from the question terminal 90, the receiving means 12 acquires the language information of the question from the unique attribute information of the question and stores it in, for example, the operation history storage unit 13. The storage location is not limited to the operation history storage unit 13, and a separate location for storing the individual attribute information of the question may be provided. When the question is received, the first answer inquiry means 14 requests a first answer corresponding to the question from the first answer system 30 (27). In this case, if the first answer system 30 (27) requests the language information of the question, the question receiving and answering system 10 (26) sends the question text and its language information. If the language information of the question is not requested, only the question text is sent. When the question receiving and answering system 10 (26) receives a first answer from the first answer system 30 (27), it acquires an answer language identifier based on the language information stored in the operation history storage unit 13. Then, based on the identifier of the first answer and the answer language information (identifier), it acquires a final answer candidate in an appropriate answer language.
[0117] 15 to 19, question groups, expected questions, and final answers when building a question-answering system that can answer questions in various languages will be explained using a specific example of questions and answers in an administrative agency. Figures 15 to 19 show simplified examples of question groups Q1 to QN, expected questions Q1(1) to QN(8) in Japanese, English, and Chinese for each question group, and final answer candidates (basic answer candidates and unique answer candidates) in Japanese, English, and Chinese.
[0118] Fig. 15 shows common attributes of expected questions in question groups Q1 to QN and specific examples of first answers corresponding to the question groups, assuming that the question-answering system according to the present invention is introduced in a city hall and questions in multiple languages are accepted. In the example of Fig. 15, the common attribute of question group Q1 is "moving in procedures," the common attribute of question group Q2 is "obtaining a resident registration card," and the common attribute of question group QN is "paying taxes," and the first answers are "101" to "NNN," respectively. Figs. 16 to 18 show expected questions and first answers (identifiers) included in question groups Q1, Q2, and QN, respectively, and Fig. 19 shows specific examples of possible final answers. In FIG. 19, for the sake of simplicity, only the final answers for each language are shown, but it is also possible to provide unique answers other than languages for each type of language.
[0119] Figure 16 shows an example in which the number of expected questions is four for Japanese, three for English, and two for Chinese. The number of expected questions for each language can be the same, but as shown in Figure 16, the number of expected questions for each language does not necessarily have to be the same. In the early stages of introducing a question-answering system, prioritizing the ability to answer basic questions and being able to respond to questions in English or Chinese rather than overly preparing a large number of expected questions often leads to improved convenience for both questioners and user institutions. This is because it is expected that there will be a need to accommodate foreign questioners, as foreigners from different cultures often do not understand basic matters. Furthermore, questions may have the same meaning in each language, or they may have different meanings and nuances in each language to suit customs, culture, etc.
[0120] In question group Q2 shown in FIG. 17, "obtaining a resident registration card" is selected as the common attribute, and the first answer for question group Q2 is "102." FIG. 17 shows specific examples of expected questions in multiple languages as expected questions belonging to question group Q2. In question group QN, "paying taxes" is selected as the common attribute, and the first answer for question group QN is "NNN." Furthermore, FIG. 18 shows specific examples of expected questions in multiple languages as expected questions for question group QN.
[0121] 16, 17 and 18 show specific examples of question sentences in Japanese and English, and for Chinese, they are shown in a schematic manner using Japanese. As can be seen from the diagram, for example, the first answer to questions in question group Q1, as shown in Figure 16, is always the same, with identifier "101," regardless of the language of the question. That is, although the question text includes expected questions in Japanese, English, and Chinese, the first answer to question group Q1 is always the same, with identifier "101," regardless of the language of the expected question. This is also true for question groups Q2 and QN, as shown in Figures 17 and 18; regardless of the language of the expected question, the first answer to question group Q2 is "102" and the answer to question group QN is "NNN." In the primary answer system, all expected questions and primary answers belonging to each of these question groups are machine-learned, and the primary answers from the primary answer system are configured to all be the same answers, "101" to "NNN," regardless of the language.
[0122] 19 shows a specific example of a primary answer, a language identifier that is part of the unique attributes of a question, and final answer candidates stored in the final answer storage unit 18 (28). Note that in the example of FIG. 19, in order to clearly explain how to handle various languages, unique answer candidates other than language information are not shown, and only final answer candidates in various languages are shown. However, as in the examples described with reference to FIGS. 5 to 11, by preparing different unique answer candidates for each type of answer language for each system user (vendor) in addition to final answer candidates for each language, it is possible to configure the system so that a final answer in an appropriate answer language is returned to the question terminal 90.
[0123] When the question answering system acquires a first answer, the final answer means 15 acquires the language identifier of the answer language from the unique attribute information of the question stored in the operation history storage unit 13, and reads out a final answer candidate corresponding to the first answer and the language identifier from the final answer storage unit 18. For example, if the identifier "101" is acquired as the first answer and the language identifier of the answer language is English, a final answer candidate in English as shown in FIG. 19 is read out from the final answer storage unit 18 and transmitted to the question terminal 90 as the final answer. In the example shown in Figure 19, there are no unique answer candidates for each system user (vendor), so the final answer candidate in English is returned as the final answer. However, by providing unique answer candidates for each system user (vendor) in each language, it is also possible to return the English unique answer candidate as the final answer.
[0124] In the above explanation, an example was shown in which the questioner individually specified the question language and the answer language. However, as mentioned above, information regarding the question language can also be automatically determined by the question receiving and answering system 10 (26) or the primary answering system 30 (27) from the form of the question sentence. Furthermore, even without input from the questioner, the information can be obtained as common information required for communication between the question terminal and the question answering system based on language information set in the OS, web browser, etc. of the question terminal. Therefore, the question language and answer language can be configured to be set by the user for each question, or the question receiving and answering system 10 (26) can automatically recognize the question language based on information available and respond in the default language, or a configuration combining the questioner's specification and the default setting is possible.
[0125] It is preferable to determine the question language by prioritizing the language specified by the questioner, language recognition by the question receiving and answering system 10 (26) or the primary answering system 30 (27) using an automatic language recognition function from the question text, or language recognition based on language settings in the web browser, OS, etc. The answer language can also be set to the same language as the question language, or can be initially set to a language different from the question language. Rules for determining the answer language can be adopted as appropriate based on the convenience of the questioner and system efficiency, etc.
[0126] FIG. 20 shows an example of a processing procedure when the types of query languages are different. Although the information handled in the unique attribute information of the question differs slightly, steps S1 to S4 and S9 to S10 are the same as the steps shown in FIG. 5, and therefore the same numbers as in FIG. 5 are used in FIG. 20 to simplify the explanation. As explained in FIG. 5, when a question q1 is received from the question terminal 90 (step S1), the unique attribute information of the question (such as the language information of the question and the system user information) is acquired and stored, for example, in the operation history storage unit 13 (step S2). This step S2 differs from step S2 in FIG. 5 in that the language information of the question is acquired and stored as the unique attribute information. Thereafter, the question received from the question terminal 90 is transmitted to the primary reply system 30 (27) to request a primary reply (step S3). When the primary reply system 30 (27) receives the question q1, it extracts and outputs one or more primary replies corresponding to the question, and the question receiving and replying system 10 (26) receives the primary replies from the primary reply system 30 (27) (step S4). The processing steps up to this point are the same as those described in FIG. 5, except that in step S2, the language information of the question is acquired and stored as unique attribute information of the question.
[0127] As described above, the language information of the question may be obtained by input from the questioner, or may be automatically obtained by the question receiving and answering system 10 (26) based on communication information between the question terminal 90 and the question receiving and answering system 10 (26). If the questioner specifies a response language, it is preferable that the response language information be stored as one piece of language information in step S2. This language information is stored as a predetermined language identifier. In the following explanation, the language of the question is Chinese, the questioner does not specify any answer language, and if the question language is Chinese, the final answer will also be in Chinese, the same language as the question. It is also possible to configure the system to answer in English in order to reduce the number of answer languages for questions in multiple languages, or when the Chinese answer cannot be prepared in time and the final answer in Chinese is not available in the final answer storage unit. In this case, various configurations are possible, such as a configuration in which English final answer candidates are stored in the final answer storage unit 18 (28) in association with Chinese identification information, or a configuration in which the system is initially set to read out English final answer candidates from the final answer storage unit 18 (28) when the answer language is Chinese.
[0128] When a first response is received from the first response system 30 (27) (step S4), the system acquires the identifier of the response language and the identifier of the system user stored in the operation history storage unit 13, etc. (step S55). Currently, the question language is Chinese, and the questioner has not specified the response language, so the system acquires the identifier of Chinese as the response language according to the initial setting. At the same time, it is confirmed whether the first response satisfies predetermined conditions (step S56).
[0129] If the primary answer satisfies a predetermined condition, such as having a certain answer certainty (S56: Yes), a final answer candidate corresponding to the answer language is read from the final answer storage unit 18 based on the identifier of the primary answer and the identifier of the answer language (step S57). Here, if unique answer candidates other than language are stored in the final answer storage unit 18 (28), a unique answer candidate corresponding to the system user, etc. is read based on the system user's identifier, etc. For example, referring to the examples of FIGS. 15 to 19, if a question in Chinese is "I've come to pick up my resident registration card," this corresponds to the expected question Q2(7) illustrated in FIG. 17, so the question receiving and answering system 10(26) receives a first answer identifier "102" from the first answering system 30 (step S4). Also, since the answer language information is not specified at the time of the question, in step S55, a Chinese language identifier is acquired as the answer language. Then, as shown in FIG. 19, based on the first answer identifier "102" and the Chinese language identifier, a Chinese final answer candidate with the first answer identifier "102" is read from the final answer storage unit 18(28) (step S57). While FIG. 19 does not show specific documents of the Chinese final answer candidates, in reality, Chinese final answers are stored.
[0130] In the example shown in FIG. 19, since no unique answer candidates other than those in the corresponding language are prepared, the final answer shown in FIG. 19 becomes the final answer as is (step S57) and is transmitted to the question terminal 90 (step S9). Here, if there are no unique answer candidates in languages other than Chinese, but there are unique answer candidates in a corresponding predetermined language (for example, Japanese), the unique answer candidate in that language (Japanese) may be returned as the final answer. Finally, the final answer and other information returned to the questioner are stored in the operation history storage unit 13 (step S10).
[0131] If the primary answer does not satisfy the predetermined conditions (S56: No), a final answer candidate prepared for each answer language, such as "Please ask a more specific question," is obtained from the final answer storage unit 18 (28) based on the answer language information (step S58) and transmitted to the question terminal 90 (step S9). Alternatively, a similar final answer may be generated by a program (step S58) and transmitted to the question terminal 90 (step S9). Here, rather than preparing individual final answer candidates for each answer language, it is also possible to prepare multiple documents in all answer languages stating, for example, "Please ask more specific questions" as one final answer candidate, and configure the system so that if the initial answers do not satisfy the specified conditions, that final answer candidate is read out and returned for all questions. Thereafter, various information relating to the question, from the question to the final answer, is stored in the operation history storage unit (step S10).
[0132] <Question-Answering System According to Another Aspect of the Present Invention: Entity Function> Further, another embodiment of the present invention will be described. If the primary response system 30 (27) equipped with artificial intelligence has the function of informing the target object (entity) contained in the question, it is possible to significantly reduce the number of expected questions to be trained by machine learning by pre-registering the specified entity in the primary response system. If the primary reply system 30 (27) has the function of outputting the entities included in the question, by registering the entities in advance, it is possible to receive the entities included in the question along with the primary reply (identifier) from the primary reply system. To utilize this entity output function, a question group with the "question intent" as a common attribute in the expected questions and multiple expected questions with the objects (entities) included in the question intent as variables are created, and the necessary entities are registered together in the primary answer system 30 (27). In this way, by registering multiple entities corresponding to each question group and preparing multiple expected questions with the entities as variables, the number of expected questions prepared is the product of the number of registered entities and the number of created expected questions, making it possible to significantly reduce the number of expected questions to actually be machine-learned. Here, entities can also be registered as entities common to the system, rather than for each question group.
[0133] 21 to 27, a method for generating expected questions and final answers in a question-answering system using entities will be described with reference to a specific example. FIGS. 21 to 27 show an example of a question-answering system 25 (25') in a city hall. FIG. 21 illustrates common attributes of expected questions belonging to each of question groups Q1 to QN and identifiers of first answers. In the example shown in FIG. 21, question groups Q1 to QN-2 are composed of expected questions with the same common attributes as before.
[0134] For each question group Q1 to QN-2, as shown in Fig. 22 to Fig. 26, the same expected questions as those described above and the corresponding first answers "101" to "NNN-2" are prepared. Therefore, when the first answer system 30 (27) receives a question and the question corresponds to one of the expected questions in question groups Q1 to QN-2, the corresponding first answer "101" to "NNN-2" is provided from the first answer system 30 (27) to the question receiving and answering system 10 (26). When the question receiving and answering system 10 (26) receives one of "101" to "NNN-2" as the first answer, it acquires a final answer candidate (basic answer candidate or unique answer candidate) corresponding to the identifier of the received first answer, as shown in Fig. 27, and returns it to the question answering system.
[0135] In contrast, the common attributes of question groups QN-1 and QN in Figure 21 are "how to get to the counter" and "obtain necessary documents," respectively, and do not specify specific entities such as destinations or necessary documents. Furthermore, as shown in Figures 25 and 26, the expected questions included in each question group QN-1 and QN do not include specific entities such as which counter the person wants to go to or what documents they want to obtain, but rather the variables "xxx." That is, in question groups QN-1 and QN, the common attribute is "intention to do something," and the entities (the "counter" the person wants to go to or the "documents" the person wants) that are the "objects of what the person wants to do" are variables. Hereinafter, such question groups QN-1 and QN are referred to as "intent question groups." On the other hand, if the question from the asker is a request for a destination or document, the question must include a specific object. Therefore, when the primary answering system outputs "NNN-1" or "NNN" as the primary answer, it must also output the entity contained in the question.
[0136] When the question receiving and answering system receives the intent question group QN-1, or the identifier "NNN-1" or "NNN" corresponding to QN, it must create and send an appropriate final answer candidate corresponding to the question based on the intent contained in the question. For this purpose, the question receiving and answering system has a correspondence table showing final answer identifiers for selecting a final answer corresponding to a registered entity, as shown in the example of FIG. 28, and references the entity included in the question to acquire the identifier of the final answer candidate that stores the final answer candidate corresponding to the entity. If the corresponding final answer candidate exists among the identifiers "101" to "NNN-2" corresponding to the normal question groups Q1 to QN-2, these identifiers are specified as the identifier of this final answer candidate. If there is no appropriate final answer among the identifiers "101" to "NNN-2", a configuration may be adopted in which separate final answer identifiers and final answer candidates are provided.
[0137] If the primary answer is "NNN-1" or "NNN" corresponding to the intent question group, the question receiving and answering system obtains the identifier of the corresponding final answer candidate from the primary answer (identifier) received from the primary answer system and the entity identifier, and obtains the final answer. For example, suppose the question from the question terminal is, "I'm here to complete the procedures for moving in." Since there is no directly applicable expected question for this question in question groups QN-1 and QN, it is assumed that the intent question group QN-1(1) is the first answer "NNN-1." In this case, the identifier of the first answer with the highest answer certainty is "NNN-1" and the entity "moving in procedures" are provided. When the question receiving and answering system receives the first answer identifier "NNN-1" and the entity identifier "moving in procedures," it obtains the identifier "101" of the final answer candidate from the correspondence table illustrated in FIG. 28, reads out the final answer candidate corresponding to the identifier "101" of the first answer illustrated in FIG. 27, and transmits it to the question terminal as the final candidate.
[0138] If a first answer "NNN-1" or "NNN" corresponding to the intent question group is received but no entity exists, a final answer candidate cannot be obtained, and an answer such as "I don't know which counter you want to go to" or "I don't know what documents are required" is returned to the questioning terminal as the final answer, as shown in Figure 28. Such cases are expected when a counter or required documents that are not registered as entities are specified in the question from the questioning terminal. By collecting and analyzing such log information and changing the registration of expected questions and entities, it is possible to increase the accuracy rate for questions and evolve the question-answering system 25 (25').
[0139] In the above explanation, we have explained examples in which the final answer candidates are "basic answer candidates" and "unique answer candidates" as they are, and examples in which the final answer candidates are "basic answer candidates and unique answer candidates in various languages," and examples in which the final answer is "basic answer candidates obtained from a combination of a primary answer (identifier) and an entity identifier," but other combinations are also possible, and all configurations obtained by combining the described configurations with each other also fall within the scope of the present invention.
[0140] Although the embodiment described above only illustrates an example of a question group based on natural language processing, the questions handled by the question-answering system 25 (25') according to the present invention are not limited to questions based on natural language formats. The question-answering system 25 (25') may also be configured to incorporate an image analysis and image recognition system and transmit a final answer related to the image by receiving image data from a question terminal. A response to a question based on an image means, for example, when image data of a tangible object such as a specific building, structure, animal, or person is received from a question terminal as a question, information related to the image data is returned as a final answer. Specifically, when a question is received containing a photo of the exterior of a facility such as a kitchen handled by a renovation company, information about the facility can be provided as a first response. The question answering system 25 (25') can register the information about the facility as a unique response for the company, thereby enabling the company to return an appropriate response to the questioner.
[0141] Furthermore, by receiving image data from the questioning terminal, a reply containing information about the image can be sent back in a language suitable for the questioner. A specific application example is its use in tourist guides. In this case, an image of a destination can be received from the query terminal, and directions to the destination (such as how to transfer between trains) can be provided using image recognition technology and a search for transportation options. By acquiring language information from the photo or communication information, it can also be configured to return an explanation of how to transfer in an appropriate language. Similarly, explanations of artworks at art museums and other venues will be available in a variety of languages. [Explanation of symbols]
[0142] 10, 26 Question Reception and Answering System 11 Communication Interface 12. Question acceptance method 13 Operation history memory unit 14. First response inquiry method 15 Final answer method 16 Maintenance methods 18, 28 Final answer storage section (final answer database) 25 25' Question Answering System 27, 30 First response system 35 Final response control section 36 Unique attribute information acquisition unit 37 Final answer acquisition part 38 Final Response Transmission Department 40 User System 90 Question Terminal 91 Communication networks (complex networks including LANs: both wired and wireless)
Claims
1. A question answering system comprising a question receiving and answering system using a computer that receives a question in a language or a question in the form of image data sent from a question terminal as question data and returns a final answer to the question based on the question data to the question terminal, and a first answering system that has artificial intelligence and returns a first answer to the question data from the question receiving and answering system in response to a request, The first response system is transmit, to the question receiving and answering system, in response to a request from the question receiving and answering system, a list of the first answers or first answer candidate identifiers corresponding to the question data from the question receiving and answering system, by machine learning the first answers consisting of a plurality of expected questions and predetermined identifiers corresponding to the expected questions; The question receiving and answering system comprises: storing a final answer candidate corresponding to the identifier of the first answer; generating a final answer based on the final answer candidate corresponding to the first answer or the list of first answer candidate identifiers received from the first answer system from among the stored final answer candidates, and sending the final answer to the question terminal that sent the question; A question answering system comprising:
2. the plurality of expected questions are composed of a plurality of question groups, each of which is composed of a bundle of a predetermined plurality of the expected questions each having a common attribute; The question answering system according to claim 1, characterized in that the first answer system outputs the same first answer for the expected questions belonging to the same question group through the machine learning.
3. The question receiving and answering system comprises: a basic answer and / or a unique answer having a different format or content from the basic answer is stored in correspondence with each identifier as the final answer candidate corresponding to the identifier of the first answer, The question answering system according to claim 1 or 2, characterized in that the basic answer or the unique answer among the final answer candidates is sent to the question terminal based on unique attribute information corresponding to the question received from the question terminal or the question terminal that sent the question.
Citation Information
Patent Citations
Human-machine intelligence chatting method with artificial intelligence and device therefor
JP2017010517A
JP1975099367A
Non-factoid question answering system and computer program
JP2013171550A
Information providing system, information providing method, and program
JP2016099967A
Cited By
Reception system and program
JP2026133859A
