Intent extraction device and program

Abstract

An intent extraction device according to an embodiment is provided with: a first determination unit that enters a first sentence in a first natural language that does not include a keyword of a first intent into a first model to determine whether or not the first sentence is associated with the first intent; and a first update unit that updates a parameter of the first model on the basis of the result of the determination by the first determination unit.

Description

Intent Extraction Device and Program

[0001] The embodiment relates to an intent extraction device and a program.

[0002] As a technology for providing communication services to users, an intent-based service provision system is known. In an intent-based service provision system, network resources and computing resources provided to a user are autonomously controlled based on the user's desires (intents).

[0003] Towards the realization of an intent-based service provision system, a method for extracting intents from natural language input by a user has been proposed. For example, a method for extracting keywords from natural language is known for extracting intents.

[0004] Y. Xiao, W. Quan, H. Zhou, M. Liu and K. Liu, "Lightweight Natural Language Driven Intent Translation Mechanism for Intent Based Networking," 2022 7th International Conference on Computer and Communication Systems (ICCCS). online <URL: https: / / ieeexplore.ieee.org / document / <9845995>

[0005] However, when the user does not have knowledge about the service they desire, the natural language input by the user may not contain keywords. If keywords are not extracted from the natural language, there is a possibility that the above method may not be able to correctly extract the intent.

[0006] The present invention has been made paying attention to the above circumstances, and an object thereof is to provide means for correctly extracting an intent.

[0007] ​An intent extraction device of one embodiment includes a first determination unit that inputs a first sentence in natural language that does not include a keyword of a first intent into a first model and determines whether the first sentence can be associated with the first intent, and a first update unit that updates parameters of the first model based on the determination result of the first determination unit.

[0008] According to the embodiment, it is possible to provide a means for correctly extracting intent.

[0009] FIG. 1 is a block diagram showing an example of a configuration of a service providing system according to an embodiment. FIG. 2 is a block diagram showing an example of a hardware configuration of an intent extraction device according to an embodiment. FIG. 3 is a block diagram showing an example of a configuration of a training function of the intent extraction device according to an embodiment. FIG. 4 is a diagram showing an example of a data structure of a service DB stored in the intent extraction device according to an embodiment. FIG. 5 is a block diagram showing an example of a configuration of an extraction function of the intent extraction device according to an embodiment. FIG. 6 is a flowchart showing an example of a training operation of a type determination model in the intent extraction device according to an embodiment. FIG. 7 is a flowchart showing an example of a training operation of a matching determination model in the intent extraction device according to an embodiment. FIG. 8 is a flowchart showing an example of an extraction operation in the intent device according to an embodiment.

[0010] Hereinafter, embodiments will be described with reference to the drawings. In the following description, components having the same functions and configurations will be denoted by the same reference numerals.

[0011] 1. Configuration 1.1 Service Providing System First, a configuration of a service providing system including an intent extraction device according to an embodiment will be described. Fig. 1 is a block diagram showing an example of the configuration of a service providing system according to an embodiment.

[0012] 1 , the service providing system 1 is an intent-based service providing system. That is, the service providing system 1 provides a communication service to the user U that is autonomously controlled based on the intent of the user U. The intent is information including, for example, the type of communication service (service type) that the user U desires to receive.

[0013] The service providing system 1 includes an intent extraction device 100 and a service providing device 200. The intent extraction device 100 and the service providing device 200 are managed by a service provider (not shown).

[0014] The intent extraction device 100 is, for example, an information processing device such as a server. The intent extraction device 100 is configured to generate question information indicating a question to be included in a dialogue C so that the chatbot B can efficiently extract an intent from the user U, and input the question information to the chatbot B. The intent extraction device 100 is also configured to extract the intent of the user U based on dialogue information indicating the content of the dialogue C between the user U and the chatbot B.

[0015] The service providing device 200 is, for example, an information processing device such as a server. The service providing device 200 determines the content of a communication service to be provided to the user U based on the intent of the user U extracted by the intent extraction device 100. The content of the communication service includes, for example, network resources, computational resources, and the like. Then, the service providing device 200 provides the user U with the communication service of the determined content.

[0016] 1.2 Hardware Configuration of Intent Extraction Apparatus Next, a hardware configuration of the intent extraction apparatus according to the embodiment will be described.

[0017] 2 is a block diagram illustrating an example of a hardware configuration of the intent extraction device 100 according to the embodiment. As illustrated in FIG. 2, the intent extraction device 100 includes a control circuit 11, a storage 12, a communication module 13, an interface 14, a drive 15, and a storage medium 16.

[0018] The control circuit 11 is a circuit that controls all of the components of the intent extraction device 100. The control circuit 11 includes a central processing unit (CPU), a random access memory (RAM), a read only memory (ROM), and the like.

[0019] The storage 12 is an auxiliary storage device of the intent extraction device 100. The storage 12 is, for example, a hard disk drive (HDD), a solid state drive (SSD), or a memory card. The storage 12 stores data used in the training operation and the extraction operation. The storage 12 may also store programs for executing the training operation and the extraction operation.

[0020] The extraction operation is an operation for extracting the intent of the user U from the dialogue information. The training operation is an operation for updating the parameters of various models used for extracting the intent. The training operation and the extraction operation will be described in detail later.

[0021] The communication module 13 is a circuit used to obtain dialogue information from a dialogue C between a user U and a chatbot B.

[0022] The interface 14 is a circuit for communicating information between the service provider and the control circuit 11. The interface 14 includes input devices and output devices. The input devices include, for example, a touch panel and operation buttons. The output devices include, for example, an LCD (Liquid Crystal Display) or an EL (Electroluminescence) display. The interface 14 converts the input from the service provider into an electrical signal and then transmits it to the control circuit 11. The interface 14 outputs the execution result based on the input from the service provider to the service provider.

[0023] The drive 15 is a device for reading software stored in the storage medium 16. The drive 15 includes, for example, a CD (Compact Disk) drive and a DVD (Digital Versatile Disk) drive.

[0024] The storage medium 16 is a medium that stores software electrically, magnetically, optically, mechanically, or chemically. The storage medium 16 may store a program for executing the training operation and the extraction operation.

[0025] 1.3 Functional Configuration of Intent Extraction Device Next, a functional configuration of the intent extraction device according to the embodiment will be described. The intent extraction device 100 has a training function for executing a training action and an extraction function for executing an extraction action.

[0026] 1.3.1 Training Function FIG. 3 is a block diagram illustrating an example of the configuration of a training function of the intent extraction device according to the embodiment.

[0027] The CPU of the control circuit 11 loads into the RAM a program related to the training movement stored in the storage 12 or the storage medium 16. The CPU of the control circuit 11 then interprets and executes the program loaded into the RAM. As a result, the control circuit 11 functions as a computer including a type determination label generation unit 24, a request information generation unit 25, a type determination unit 26, a type determination model update unit 27, a consistency determination label generation unit 28, an explanation information acquisition unit 29, a consistency determination unit 30, and a consistency determination model update unit 31. The storage 12 also stores a service DB 21, a type determination model 22, and a consistency determination model 23.

[0028] The service DB 21 is a database that stores information about services that can be provided to the user U by the service providing device 200. Fig. 4 is a diagram illustrating an example of a data structure of the service DB stored in the intent extraction device according to the embodiment.

[0029] As shown in Fig. 4, the service DB 21 stores service types of services that can be provided to the user U by the service providing device 200, categorized in a tree structure. The service types are the names of the services provided to the user U. In the example of Fig. 4, three major classification nodes S1, S2, and S3 are connected to a root node. The major classification nodes S1, S2, and S3 correspond to higher-level conceptual service types that include multiple minor classification nodes.

[0030] Three subdivision nodes S11, S12, and S13 are connected to major division node S1. Two subdivision nodes S21 and S22 are connected to major division node S2. Three subdivision nodes S31, S32, and S33 are connected to major division node S3. Subdivision nodes S11 to S13, S21 to S22, and S31 to S33 correspond to specific service types provided to user U.

[0031] The service DB21 stores explanatory information D11-D13, D21-D22, and D31-D33 in association with the minor classification nodes S11-S13, S21-S22, and S31-S33, respectively. The explanatory information D11-D13, D21-D22, and D31-D33 includes one or more keywords for the service types corresponding to the minor classification nodes S11-S13, S21-S22, and S31-S33, respectively, and defines the content of the service types in natural language. The keywords for the service types are words that can identify the service type using only the keywords, such as the name of the service type or its synonyms.

[0032] Specifically, for example, if the service type is "web conferencing," the description information corresponding to the service type may be a sentence in a natural language such as the following: The description information shown below includes, for example, the words "web conferencing" enclosed in < > as a keyword:<Web conferencing> is an online service that allows you to hold live meetings, conferencing, presentations, and training via the internet, particularly on TCP / IP connections. You can connect to the conference either by telephone or using your computer's speakers and microphone through a VoIP connection." Furthermore, for example, if the service type is "smart traffic system," the explanatory information corresponding to that service type may be a natural language sentence such as the following. In the explanatory information shown below, for example, the word "smart traffic system" enclosed in <> is included as a keyword. Furthermore, the word "intelligent transportation system," which is a synonym of "smart traffic system," may also be included as a keyword.“A<smart traffic system> is or<intelligent transportation system> "Intelligent Transportation Systems (ITS) uses advanced technologies like sensors, artificial intelligence, and communication systems to improve road safety, manage traffic flow, reduce congestion, and enhance the efficiency of transportation networks. Features may include adaptive traffic signals, real-time traffic information, vehicle-to-everything (V2X) communication, and smart parking solutions. These systems are integral to smart city initiatives, aiming to create safer, more efficient urban transportation environments." Each of the category determination model 22 and the consistency determination model 23 is a parameter of a mathematical model that includes, for example, a neural network based on BERT (bidirectional encoder representations from transformer) as a base model. Examples of parameters include weighting coefficients and bias terms in the neural network. The category determination model 22 and the consistency determination model 23 are used in the category determination process by the category determination unit 26 and the consistency determination process by the consistency determination unit 30, respectively.

[0033] The type determination process analyzes the meaning of a natural language sentence (request information) requesting a certain service and determines the service type of the service. In the type determination process, for example, the probability that a certain service type corresponds to the request information is calculated for each service type stored in the service DB 21. In other words, the output of the type determination process is a data string of probability values ​​having a number of elements corresponding to the number of service types.

[0034] The consistency determination process is a process for determining whether or not natural language information requesting a certain service (request information) and description information of a service type determined by the type determination process to be most suitable for the request information semantically match. In the consistency determination process, for example, a probability value between 0 and 1 indicating the consistency between the meaning of the request information and the description information is calculated.

[0035] The type determination label generation unit 24 generates correct labels (type determination labels) to be used in the training operation of the type determination model 22 based on the service DB 21. Specifically, for example, the type determination label generation unit 24 selects one service type from the service types stored in the service DB 21 as the correct service type. The type determination label generation unit 24 generates a data sequence {1, 0, 0 ..., 0} in which the probability value corresponding to the correct service type is "1" and the probability value corresponding to the incorrect service type is "0". The type determination label generation unit 24 inputs the generated type determination labels to the request information generation unit 25 and the type determination model update unit 27.

[0036] The request information generating unit 25 generates request information corresponding to the correct service type selected by the type determination label generating unit 24. The request information generating unit 25 inputs the generated request information to the type determining unit 26.

[0037] Specifically, for example, if the correct service type is "web conferencing," the request information generation unit 25 generates the following request information: "As the HR manager for a mid-sized company, I want to provide a reliable web conferencing solution for our new employee orientation program. We typically have around 30 participants joining from different locations. Location: Online. Estimated workloads: 30 attendees. Time: The first Monday of every month at 10 am. Our budget limit is $200 per month." Note that the request information does not necessarily include keywords related to the correct service type, as described above. Specifically, the request information may also include cases where keywords related to the correct service type are not included, such as when the correct service type is "smart traffic management" as shown below. "I want to implement a system that uses cameras and sensors to monitor traffic flow and optimize traffic signals in real time. The system should be able to detect accidents, congestion, and reroute traffic accordingly. I estimate the workload to be moderate and the location to be a busy intersection in the city. My budget limit is $100,000." The type determination unit 26 inputs the desired information into the type determination model 22 and calculates the probability that the desired information is met for each service type.The type determination unit 26 inputs the data string of the probability values ​​calculated as the output result from the type determination model 22 to the type determination model update unit 27 .

[0038] The type determination model update unit 27 compares the output result from the type determination model 22 with the type determination label to update the parameters of the type determination model 22. Specifically, the type determination model update unit 27 updates the parameters so as to minimize the error between the type determination label and the output result from the type determination model 22 (i.e., so that the probability value corresponding to the correct service type approaches "1" and the probability value corresponding to the incorrect service type approaches "0"). When updating the parameters, the type determination model update unit 27 uses, for example, an error backpropagation algorithm.

[0039] Every time the parameters are updated, the type determination model update unit 27 applies the updated parameters to the type determination model 22 used by the type determination unit 26. When the update termination condition for the type determination model 22 is satisfied, the type determination model update unit 27 stores the most recently updated parameters in the storage 12 as the trained type determination model 22.

[0040] The condition for ending the update of the type determination model 22 may be, for example, that the number of type determination labels input to the type determination unit 26 in the training motion exceeds a threshold. The condition for ending the update of the type determination model 22 may be, for example, that the number of parameter updates exceeds a threshold.

[0041] The consistency determination label generation unit 28 generates correct labels (consistency determination labels) to be used in the training operation of the consistency determination model 23 based on the service DB 21. Specifically, for example, the consistency determination label generation unit 28 selects one service type from the service types stored in the service DB 21 as a service type to be subjected to consistency determination. The consistency determination label generation unit 28 randomly generates correct labels when the service types corresponding to the request information and description information input to the consistency determination model 23 are different from each other, and correct labels when the service types correspond to each other. The correct label when the service types corresponding to the request information and description information are different from each other has a probability value of "0," and the correct label when the service types correspond to each other has a probability value of "1." The consistency determination label generation unit 28 inputs a set of the service type to be subjected to consistency determination and the consistency determination label to the request information generation unit 25, the description information acquisition unit 29, and the consistency determination model update unit 31.

[0042] The request information generation unit 25 generates request information corresponding to the service type to be subjected to the match determination selected by the match determination label generation unit 28. The request information generation unit 25 inputs the generated request information to the match determination unit 30. Note that if the request information generated for the purpose of training operations for the type determination process can be reused as request information generated for the purpose of training operations for the match determination process, the request information generation unit 25 does not need to generate new request information.

[0043] The explanation information acquisition unit 29 acquires explanation information from the service DB 21 based on the pair of the service type to be determined for consistency and the consistency determination label. Specifically, if the consistency determination label has a probability value of "1", the explanation information acquisition unit 29 extracts explanation information corresponding to the service type to be determined for consistency from the service DB 21. If the consistency determination label has a probability value of "0", the explanation information acquisition unit 29 randomly extracts explanation information corresponding to a service type different from the service type to be determined for consistency from the service DB 21. The explanation information acquisition unit 29 inputs the acquired explanation information to the consistency determination unit 30.

[0044] The consistency determination unit 30 calculates the probability that the service types corresponding to the request information and the explanatory information match, using the request information input from the request information generation unit 25 and the explanatory information input from the explanatory information acquisition unit 29 as input to the consistency determination model 23. The consistency determination unit 30 inputs the probability value calculated as the output result from the consistency determination model 23 to the consistency determination model update unit 31.

[0045] The consistency determination model update unit 31 compares the output result from the consistency determination model 23 with the consistency determination label to update the parameters of the consistency determination model 23. Specifically, the consistency determination model update unit 31 updates the parameters so as to minimize the error between the consistency determination label and the output result from the consistency determination model 23. When updating the parameters, the consistency determination model update unit 31 uses, for example, an error backpropagation method.

[0046] Every time the parameters are updated, the consistency determination model update unit 31 applies the updated parameters to the consistency determination model 23 used by the consistency determination unit 30. When the update termination condition for the consistency determination model 23 is satisfied, the consistency determination model update unit 31 stores the most recently updated parameters in the storage 12 as the trained consistency determination model 23.

[0047] The update termination condition for the consistency determination model 23 may be, for example, that the number of consistency determination labels input to the consistency determination unit 30 in the training operation exceeds a threshold. The update termination condition for the consistency determination model 23 may be, for example, that the number of parameter updates exceeds a threshold.

[0048] 1.3.2 Extraction Function FIG. 5 is a block diagram showing an example of the configuration of the extraction function of the intent extraction device according to the embodiment.

[0049] The CPU of the control circuit 11 loads into the RAM a program related to the extraction operation stored in the storage 12 or the storage medium 16. The CPU of the control circuit 11 then interprets and executes the program loaded into the RAM. As a result, the control circuit 11 functions as a computer including a type determination unit 26, an explanation information acquisition unit 29, a consistency determination unit 30, a dialogue information acquisition unit 32, a request information extraction unit 33, a question generation unit 34, and an output unit 35.

[0050] The dialogue information acquisition unit 32 acquires dialogue information from a dialogue C between the user U and the chatbot B. The dialogue information may be text data or voice data. The dialogue information includes information regarding the intent of the communication service that the user U wants to use. The dialogue information acquisition unit 32 inputs the acquired dialogue information to the request information extraction unit 33.

[0051] The request information extraction unit 33 converts the input dialogue information into a natural language sentence format. The request information extraction unit 33 extracts a sentence corresponding to the request information from the converted natural language sentence. The request information extraction unit 33 inputs the extracted request information to the type determination unit 26.

[0052] The type determination unit 26 functions in the same way as the type determination unit 26 related to the training function, based on the trained type determination model 22. The type determination unit 26 inputs the request information extracted from the dialogue information to the trained type determination model 22, and calculates a data sequence of probability values ​​for each service type as an output. The type determination unit 26 inputs the calculated data sequence of probability values ​​for each service type to the explanation information acquisition unit 29.

[0053] The explanation information acquisition unit 29 selects the service type corresponding to the highest probability value from the data sequence of probability values ​​output from the type determination unit 26. The explanation information acquisition unit 29 acquires explanation information corresponding to the selected service type from the service DB 21. The explanation information acquisition unit 29 inputs the acquired explanation information to the consistency determination unit 30.

[0054] The consistency determination unit 30 functions in the same way as the consistency determination unit 30 related to the training function, based on the trained consistency determination model 23. The consistency determination unit 30 inputs the request information extracted by the request information extraction unit 33 and the explanation information acquired by the explanation information acquisition unit 29 to the trained consistency determination model 23, and calculates, as an output, a probability value indicating the consistency of the service types corresponding to each of these pieces of information. If the calculated probability values ​​indicate a mismatch, the consistency determination unit 30 inputs a request for a re-question together with the determination result to the question generation unit 34. If the calculated probability values ​​indicate a match, the consistency determination unit 30 inputs, together with the determination result, the service type corresponding to the highest probability value among the data strings of probability values ​​obtained from the type determination unit 26 to the output unit 35.

[0055] In response to a request for a re-question from the matching determination unit 30, the question generation unit 34 generates question information indicating a question to be input to the user U in order to obtain more accurate desired information. The question generation unit 34 inputs the generated question information to the chatbot B. As a result, the question information is input to the user U via the chatbot B. Then, by obtaining an answer to the question from the dialogue information, further desired information is extracted.

[0056] The output unit 35 outputs the service type obtained from the consistency determination unit 30 to the service providing device 200. With the above configuration, the intent extraction device 100 can extract an appropriate service type as an intent from the request information of the user U, using the trained type determination model 22 and the consistency determination model 23.

[0057] 2. Operation Next, the operation of the intent extraction device according to the embodiment will be described.

[0058] 2.1 Training Operation of Type Determination Model FIG. 6 is a flowchart showing an example of a training operation of a type determination model in the intent extraction device according to the embodiment.

[0059] When an instruction to perform training operations for the type determination model 22 is received from the service provider (start), the type determination label generation unit 24 selects a correct service type and generates a type determination label corresponding to the correct service type (S11).

[0060] The request information generating unit 25 generates request information relating to the correct service type based on the correct service type corresponding to the type determination label generated in the process of S11 (S12).

[0061] The type determination unit 26 inputs the request information generated in the process of S12 into the type determination model 22 and determines the service type corresponding to the request information (S13). Specifically, the type determination unit 26 calculates a probability value indicating the degree to which the service type corresponds to the request information as a data string for each service type.

[0062] The type determination model update unit 27 updates the parameters of the type determination model 22 based on the data sequence of probability values ​​for each service type calculated as the determination result by the processing of S13 and the type determination label generated by the processing of S11 (S14).

[0063] The control circuit 11 determines whether or not the update termination condition for the category determination model 22 is satisfied (S15).

[0064] If the update termination condition for the category determination model 22 is not satisfied (S15; no), the category determination label generation unit 24 selects a new correct service category and generates a category determination label corresponding to the new correct service category (S11). Then, the subsequent processes of S12 to S15 are executed. In this way, the processes of S11 to S15 are repeatedly executed until the update termination condition for the category determination model 22 is satisfied.

[0065] If the update termination condition for the type determination model 22 is met (S15; yes), the type determination model update unit 27 stores the parameters last updated in the processing of S14 in the storage 12 as the learned type determination model 22 (S16).

[0066] After the process of S16, the training operation of the type determination model 22 ends (end).

[0067] 2.2 Training Operation of the Consistency Determination Model FIG. 7 is a flowchart showing an example of the training operation of the consistency determination model in the intent extraction device according to the embodiment.

[0068] When an instruction is received from the service provider to perform training operations for the consistency judgment model 23 (start), the consistency judgment label generation unit 28 selects the service type to be judged for consistency and generates a consistency judgment label (S21).

[0069] The request information generating unit 25 generates request information relating to the service type to be subjected to the match determination based on the service type to be subjected to the match determination corresponding to the match determination label generated in the process of S21 (S22).

[0070] The explanation information acquisition unit 29 determines whether the match determination label generated in the process of S21 indicates a match (S23).

[0071] If the label for determining consistency indicates a match (S23; yes), the explanatory information acquisition unit 29 refers to the service DB 21 and acquires explanatory information regarding the service type that matches the service type corresponding to the request information generated in the processing of S22 (S24).

[0072] If the label for determining consistency indicates a mismatch (S23; no), the explanatory information acquisition unit 29 refers to the service DB 21 and randomly acquires explanatory information regarding a service type different from the service type corresponding to the request information generated in the processing of S22 (S25).

[0073] The consistency determination unit 30 inputs the request information generated in the process of S22 and the explanation information acquired in the process of S24 or S25 into the consistency determination model 23, and determines the consistency of the service types of the request information and the explanation information (S26). Specifically, the consistency determination unit 30 calculates a probability value indicating the consistency.

[0074] The consistency determination model update unit 31 updates the parameters of the consistency determination model 23 based on the probability value indicating consistency calculated as the determination result by the processing of S26 and the consistency determination label generated by the processing of S21 (S27).

[0075] The control circuit 11 determines whether or not the update termination condition for the consistency determination model 23 is satisfied (S28).

[0076] If the update termination condition for the consistency determination model 23 is not satisfied (S28; no), the consistency determination label generation unit 28 selects a new service type to be subjected to consistency determination and generates a corresponding consistency determination label (S21). Then, the subsequent processes of S22 to S28 are executed. In this manner, the processes of S21 to S28 are repeatedly executed until the update termination condition for the consistency determination model 23 is satisfied.

[0077] If the update termination condition for the consistency determination model 23 is met (S28; yes), the consistency determination model update unit 31 stores the parameters last updated in the processing of S27 in the storage 12 as the learned consistency determination model 23 (S29).

[0078] After the process of S29, the training operation of the consistency determination model 23 ends (END).

[0079] 2.3 Extraction Operation Next, the extraction operation of the intent extraction device according to the embodiment will be described.

[0080] FIG. 8 is a flowchart illustrating an example of an extraction operation in the intent extraction device according to the embodiment.

[0081] When a dialogue C occurs (starts) between a user U and a chatbot B, the dialogue information acquisition unit 32 acquires dialogue information (S31).

[0082] The desired information extraction unit 33 extracts desired information from the dialogue information acquired in the process of S31 (S32).

[0083] The type determination unit 26 inputs the request information extracted in the process of S32 into the trained type determination model 22, and determines the type of service requested by the user U (S33).

[0084] Based on the determination result of the process of S33, the explanation information acquisition unit 29 acquires, from the service DB 21, explanation information corresponding to the service type with the highest probability value of being the service type desired by the user U (S34).

[0085] The consistency determination unit 30 inputs the request information extracted in the processing of S32 and the explanatory information obtained in the processing of S34 into the trained consistency determination model 23, and determines the consistency of the service types corresponding to each of the request information and explanatory information (S35).

[0086] Based on the determination result of the process of S35, the agreement determination unit 30 determines whether or not the service types corresponding to the request information and the explanation information match (S36).

[0087] If the service types corresponding to the request information and the explanation information do not match (S36; no), the question generator 34 generates question information to elicit further request information from the user U (S37). The question information generated in the process of S37 is input to the user U via the chatbot B.

[0088] After the process of S37, the dialogue information acquisition unit 32 acquires the dialogue information (S31). Then, the subsequent processes of S32 to S37 are executed. In this way, the processes of S31 to S37 are repeatedly executed until the service types corresponding to the request information and the explanation information match.

[0089] If the service types corresponding to the request information and the explanation information match (S36; yes), the output unit 35 outputs the service type determined in the process of S33 to the service providing device 200 as an intent (S38).

[0090] When the process of S38 is completed, the extraction operation is completed (end).

[0091] 3. Effects of the Embodiment According to the embodiment, the type determination unit 26 inputs request information that does not include a keyword for the correct service type to the type determination model 22 and determines whether the request information can be associated with the correct service type. The type determination model update unit 27 updates the parameters of the type determination model 22 based on the determination result of the type determination model 22. As a result, even if the user U does not know any words that can directly identify the service type that the user U wants to use, such as the name or synonyms of the service type, it is possible to extract a service type based on the request information from the user U.

[0092] Specifically, the type determination unit 26 inputs the request information into the type determination model 22 and calculates the probability that the request information will be associated with the correct service type and the probability that the request information will be associated with an incorrect service type. The type determination model update unit 27 updates the parameters of the type determination model 22 so that the probability that the request information will be associated with the correct service type approaches "1" and the probability that the request information will be associated with an incorrect service type approaches "0". This makes it possible to improve the prediction accuracy of the type determination model 22 and appropriately analyze the meaning of the request from the user U.

[0093] Furthermore, the consistency determination unit 30 inputs request information that does not include a keyword for a certain service type and description information that includes a keyword for a service type randomly selected from the service type into the consistency determination model 23, and determines whether the service types corresponding to the request information and the description information match. The consistency determination model update unit 31 updates the parameters of the consistency determination model 23 based on the determination result of the consistency determination model 23. As a result, even if an incorrect service type is determined in the type determination process, it is possible to further determine whether the result of the type determination process is correct by semantically comparing the request information and the description information. This reduces the possibility of misinterpreting a request from user U.

[0094] 4. Modifications, etc. Various modifications can be applied to the above-described embodiment.

[0095] In the above-described embodiment, the program for executing the training action and the extraction action is executed by the intent extraction device 100. However, this is not limiting. For example, the program for executing the training action and the extraction action may be executed by a computing resource on the cloud.

[0096] The present invention is not limited to the above-described embodiments, and various modifications can be made in the implementation stage without departing from the spirit of the invention. Furthermore, the embodiments may be implemented in appropriate combinations, in which case the combined effects can be obtained. Furthermore, the above-described embodiments include various inventions, and various inventions can be extracted by combining selected elements from the disclosed elements. For example, if the problem can be solved and the desired effect can be obtained even if some elements are deleted from all elements shown in the embodiments, the configuration from which these elements are deleted can be extracted as an invention.

[0097] DESCRIPTION OF SYMBOLS 1...Service providing system 11...Control circuit 12...Storage 13...Communication module 14...Interface 15...Drive 16...Storage medium 21...Service DB 22...Category determination model 23...Consistency determination model 24...Category determination label generation unit 25...Request information generation unit 26...Category determination unit 27...Category determination model update unit 28...Consistency determination label generation unit 29...Explanation information acquisition unit 30...Consistency determination unit 31...Consistency determination model update unit 32...Dialogue information acquisition unit 33...Request information extraction unit 34...Question generation unit 35...Output unit 100...Intent extraction device 200...Service providing device U...User B...Chatbot C...Dialogue

Claims

1. An intent extraction device comprising: a first judgment unit that inputs a first sentence in a natural language that does not include a keyword of a first intent into a first model and judges whether the first sentence can be associated with the first intent; and a first update unit that updates parameters of the first model based on a judgment result of the first judgment unit.

2. The intent extraction device as described in claim 1, wherein the first judgment unit is configured to input the first sentence into a first model to determine a first probability that the first sentence is associated with the first intent and a second probability that the first sentence is associated with a second intent, and the first update unit is configured to update parameters of the first model so that the first probability approaches 1 and the second probability approaches 0.

3. The intent extraction device as described in claim 1, further comprising: a second judgment unit that inputs the first sentence and a second sentence in natural language containing a keyword of the first intent or the second intent into a second model and judges whether the intents associated with the first sentence and the second sentence match; and a second update unit that updates parameters of the second model based on the judgment result of the second judgment unit.

4. A program for causing a computer to function as each unit of the intent extraction device according to any one of claims 1 to 3.

Images