Man-machine dialogue device and method for implementing man-machine dialogue

Abstract

The application discloses a man-machine conversation device and a method for realizing man-machine conversation, and comprises the following steps: obtaining a user's inquiry in a conversation process for providing service for the user; putting forward a question to the user in a multi-round question and answer mode based on the user's inquiry and a question in a question database, and determining a user's intention according to the user's reply. Through the technical scheme provided by the application, when the user's intention is uncertain, the user's intention is clarified by actively putting forward a question to the user, the man-machine conversation process is intelligently realized, and the accuracy of the user's intention recognition is improved.

Description

[0001] This application is a divisional application of the patent application with application number 201611106354.X and titled "A man-machine dialogue device and a method for implementing man-machine dialogue thereof", which was filed on December 5, 2016. TECHNICAL FIELD

[0002] The present application relates to man-machine dialogue technology, and in particular to a man-machine dialogue device and a method for implementing man-machine dialogue thereof. BACKGROUND

[0003] An intelligent man-machine dialogue system uses natural language as an interactive medium to provide services for users. The intelligent man-machine dialogue system is a brand-new interactive mode between man and machine, and the interactive process is more natural and efficient. For example, the famous Eliza system was introduced in the 1960s, which can have a simple dialogue with users, making many users feel as if they are communicating with a real person. After Apple Siri, many dialogue products have appeared, including Google Now, Microsoft Cortana. Many chat robots have been released in China, such as Microsoft Xiaoice, which have attracted widespread attention.

[0004] Most of the existing intelligent man-machine dialogue systems only make single-turn dialogue. For the inquiry of the user, i.e. the service object, the intelligent man-machine dialogue system generally processes in two stages: in the first stage, the user intent is recognized, the user intent refers to the intention of the user to achieve a certain purpose, such as: hoping to solve a certain problem, hoping to complete a certain task, hoping to achieve a certain goal, etc.; in the second stage, the AGENT, i.e. the machine simulating customer service, provides corresponding answers or operations for the user intent. Here, accurately recognizing the user intent is the key to determining the service effect of the entire intelligent man-machine dialogue system. In the user intent recognition stage, the existing intelligent man-machine dialogue system generally directly recognizes the user intent according to the user's expression. The entire user intent recognition process is passive, that is, the AGENT does not ask any questions about the user's inquiry, but passively determines the user's intent based on a user's inquiry. In this way, if the user's expression is not clear, the user intent recognized by the intelligent man-machine system will also be inaccurate, and the answers or behaviors given by the AGENT based thereon will not achieve the user's intent, and even lead to errors and risks.

[0005] From the implementation of the above existing intelligent man-machine dialogue system, it cannot truly guarantee the accuracy of the user intent recognition, and the implementation of the existing intelligent man-machine dialogue system relies on a large number of human participation, which increases the cost and has a relatively long implementation period. SUMMARY

[0006] To solve the above technical problems, the application provides a man-machine dialogue device and a method for realizing man-machine dialogue, which can ensure the accuracy of user intention recognition, reduce the implementation cost, and shorten the implementation period.

[0007] To achieve the purpose of the application, the application provides a man-machine dialogue device, comprising an acquisition unit, a question and answer unit, and a question database for storing questions,

[0008] The acquisition unit is configured to acquire the user's inquiry during the dialogue process for providing services for the user.

[0009] The question and answer unit is configured to, when the user's intention is uncertain, actively ask questions to the user based on the user's inquiry and the questions in the question database in a multi-round question and answer mode, and determine the user's intention according to the user's reply.

[0010] In each round of question and answer, more than one question is acquired from the question database according to the user's inquiry or the user's reply, and the questions are asked to the user in the order of asking questions, and the reply from the user based on the asked questions is received.

[0011] Optionally, the man-machine dialogue device further comprises:

[0012] The learning unit is configured to determine the questions in the dialogue based on dialogue corpus samples and store the questions in the question database.

[0013] Optionally, the learning unit is specifically configured to:

[0014] extract the questions asked by the customer service to the user from the dialogue corpus samples, classify the questions asked to the user through text clustering, and store the questions in the question database.

[0015] Optionally, the question and answer unit comprises an intention prediction module and an intention decision module.

[0016] The intention prediction module is configured to, when the user's intention is uncertain, predict the probability of positive reply of each question corresponding to the inquiry in the question database based on the user's inquiry or the user's reply, and form representation information representing the user's intention.

[0017] The intention decision module is configured to, based on the representation information, update the probability of positive reply of the user predicted in the representation information by using the reply of the user to the questions asked to the user in the multi-round question and answer mode, and determine the user's intention according to the representation information obtained after the multi-round question and answer.

[0018] Optionally, the intention decision module comprises a question decision submodule and an intention decision submodule.

[0019] a question decision sub-module configured to predict information gain of each question in the representation information in each round of the multi-round question-answering, take a question with the maximum information gain as a question to be asked to the user, and ask the question to the user;

[0020] an intention decision sub-module configured to update the predicted probability of the user's answer being positive in the representation information according to the user's reply to the question asked to the user, and determine the user's intention according to the updated representation information.

[0021] Optionally, the intention prediction module is specifically configured to:

[0022] learn, from a dialogue corpus, an association between each question and a system goal for representing a solution or behavior for achieving the user's intention based on the user's inquiry or the user's reply, predict a probability of the answer to each question corresponding to the inquiry in the question database being positive based on a distribution of the association, and form the representation information of the user's intention.

[0023] Optionally, the intention decision sub-module stores a preset mapping relationship between the representation information and the user's intention.

[0024] The intention decision sub-module is specifically configured to find the user's intention corresponding to the representation information currently formed according to the mapping relationship.

[0025] The application further provides a method for implementing human-computer dialogue, comprising:

[0026] acquiring a user's inquiry in a dialogue process for providing a service for the user;

[0027] when the user's intention is uncertain, asking questions to the user in a multi-round question-answering manner based on the user's inquiry and questions in a question database, and determining the user's intention according to the user's reply;

[0028] in each round of the question-answering, acquiring more than one question asked to the user from the question database according to the user's inquiry or the user's reply, asking the questions to the user in a question order, and receiving a reply from the user based on the questions asked.

[0029] Optionally, the method further comprises determining a question of the dialogue based on a dialogue corpus sample, and storing the question in the question database.

[0030] Optionally, the determination of the question of the dialogue comprises:

[0031] extracting a question asked to the user by a customer service from the dialogue corpus sample, classifying the question asked to the user by text clustering, and storing the question in the question database.

[0032] Optionally, the step of asking the user questions in the multi-turn question-answering mode based on the user's query and the questions in the question database comprises:

[0033] predicting the probability of each question in the question database corresponding to the query being answered positively based on the user's query or the user's reply, and forming the representation information representing the user's intent;

[0034] updating the predicted probability of the user's reply being positive in the representation information based on the representation information and the user's reply to the question asked in the multi-turn question-answering mode, and determining the user's intent according to the representation information obtained after the multi-turn question-answering.

[0035] Optionally, the step of updating the predicted probability of the user's reply being positive in the representation information based on the representation information and the user's reply to the question asked in the multi-turn question-answering mode, and determining the user's intent according to the representation information obtained after the multi-turn question-answering comprises:

[0036] in each turn of the multi-turn question-answering, predicting the information gain of each question in the representation information, taking the question with the largest information gain as the question asked in the next turn of the multi-turn question-answering, and asking the question to the user;

[0037] updating the predicted probability of the user's reply being positive in the representation information based on the user's reply to the question asked, and determining the user's intent according to the updated representation information.

[0038] Optionally, the step of predicting the probability of each question in the question database corresponding to the query being answered positively based on the user's query, and forming the representation information representing the user's intent comprises:

[0039] learning the association between each question and the system goal for representing a solution or action to achieve the user's intent from the dialogue corpus based on the user's query or the user's reply, predicting the probability of each question in the question database corresponding to the query being answered positively based on the question-answer distribution in the association, and forming the representation information representing the user's intent.

[0040] Optionally, the step of determining the user's intent according to the formed representation information representing the user's intent comprises:

[0041] finding the user's intent corresponding to the current formed representation information according to the pre-set mapping relationship between the representation information and the user's intent.

[0042] The application further provides a device for implementing human-computer dialogue, comprising at least a memory and a processor, wherein,

[0043] The executable instructions stored in the memory include: obtaining a query of a user in a conversation process of providing a service for the user; when the user intention is not determined, actively asking questions to the user based on the query of the user and questions in a question database in a multi-round question-answering manner, and determining the user intention according to a reply of the user; wherein in each round of question-answering, more than one question is obtained from the question database according to the query of the user or the reply of the user, and the questions are asked to the user in a question order, and a reply from the user based on the questions is received.

[0044] The scheme provided in the application includes: obtaining a query of a user in a conversation process of providing a service for the user; actively asking questions to the user based on the query of the user and questions in a question database in a multi-round question-answering manner, and determining the user intention according to a reply of the user. The technical scheme for realizing the human-computer conversation provided in the application intelligently realizes the human-computer conversation process by actively asking questions to the user to clarify the user intention when the user intention is not determined, and improves the accuracy of identifying the user intention.

[0045] Other features and advantages of the application will be set forth in the following description, and in part will become apparent to those skilled in the art from the description, or can be learned by practice of the application. The objects and other advantages of the application will be realized and attained by the structure particularly pointed out in the written description and claims. BRIEF DESCRIPTION OF DRAWINGS

[0046] The accompanying drawings, which are included to provide a further understanding of the technical scheme of the application and constitute a part of the specification, illustrate the technical scheme of the application and are used to explain the technical scheme of the application together with the embodiments of the application, and do not constitute a limitation on the technical scheme of the application.

[0047] Figure 1 A schematic diagram of a composition structure of the human-computer conversation device of the application;

[0048] FIG. 2(a) is a schematic diagram of a first embodiment of a distribution of a reply to a question of the application;

[0049] FIG. 2(b) is a schematic diagram of a second embodiment of the distribution of the reply to the question of the application;

[0050] FIG. 2(c) is a schematic diagram of a third embodiment of the distribution of the reply to the question of the application;

[0051] FIG. 2(d) is a schematic diagram of a fourth embodiment of the distribution of the reply to the question of the application;

[0052] Figure 3 A flowchart of a method for realizing the human-computer conversation of the human-computer conversation device of the application;

[0053] Figure 4This is a flowchart illustrating an embodiment of the method for implementing human-computer dialogue in this application. Detailed Implementation

[0054] To make the objectives, technical solutions, and advantages of this application clearer, the embodiments of this application will be described in detail below with reference to the accompanying drawings. It should be noted that, unless otherwise specified, the embodiments and features described in these embodiments can be arbitrarily combined with each other.

[0055] In a typical configuration of this application, the computing device includes one or more processors (CPU), input / output interfaces, network interfaces, and memory.

[0056] Memory may include non-persistent storage in computer-readable media, such as random access memory (RAM) and / or non-volatile memory, such as read-only memory (ROM) or flash RAM. Memory is an example of computer-readable media.

[0057] Computer-readable media includes both permanent and non-permanent, removable and non-removable media that can store information using any method or technology. Information can be computer-readable instructions, data structures, modules of programs, or other data. Examples of computer storage media include, but are not limited to, phase-change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other memory technologies, CD-ROM, digital versatile optical disc (DVD) or other optical storage, magnetic tape, magnetic magnetic disk storage or other magnetic storage devices, or any other non-transferable medium that can be used to store information accessible by a computing device. As defined herein, computer-readable media does not include non-transitory computer-readable media, such as modulated data signals and carrier waves.

[0058] The steps illustrated in the flowchart in the accompanying drawings can be executed in a computer system such as a set of computer-executable instructions. Furthermore, although a logical order is shown in the flowchart, in some cases the steps shown or described may be performed in a different order than that presented here.

[0059] Figure 1 This is a schematic diagram of the composition of the applicant's human-computer interaction device, as shown below. Figure 1 As shown, it includes an acquisition unit, a question-and-answer unit, and a question database for storing questions; wherein,

[0060] The acquisition unit is used to acquire user queries during the dialogue process of providing services to the user;

[0061] The question and answer unit is configured to ask questions to the user in a multi-round question and answer manner based on the user inquiry and the questions in the question database, and determine the user intent according to the user's reply.

[0062] The multi-round question and answer manner is a one-question-one-answer process for one or more times. Specifically, one round of question and answer in this embodiment can refer to the human-computer dialogue device asking questions to the user, and the user replying based on the questions asked.

[0063] The questions stored in the question database are questions learned for the user's inquiry based on dialogue corpus samples, through artificial learning, cluster analysis, or machine learning methods. The dialogue corpus samples can be, for example, natural language dialogue corpus about achieving user intent between real users and customer service, or historical dialogue data of human-computer dialogue, etc. The questions in the question database can be updated in real time. Specifically, the questions for the user's inquiry can be determined through machine learning or other methods based on dialogue data collected over a period of time. In this way, the questions in the question database can be updated in real time, and the performance of the human-computer dialogue device can be continuously improved through the accumulated question and answer corpus in the process of providing services for the user. The questions in the question database include but are not limited to yes / no questions, specific questions, and selection questions.

[0064] Optionally, the human-computer dialogue device of the present application can further include a learning unit configured to determine the questions of the dialogue based on dialogue corpus samples, such as learning the questions for the user's inquiry and storing them in the question database. Through the learning unit, the questions in the question database can be obtained based on historical corpus.

[0065] For the process of artificial learning, learning the questions for the user's inquiry can include: assuming that the user asks: "Can you tell me how to use my xx service?", through artificial learning, the customer service can ask questions such as: "Have you opened xx service?", "Can't you open it?", "Have you paid the relevant fees?", "How to open xx service?", etc. These questions belong to the same category and form a series of questions for the inquiry xx service in the question database. There is a corresponding order for asking these questions, for example, when the customer service asks "Have you opened xx service?", the user answers "No", then the customer service will ask "Can't you open it?", the user answers "No", then the customer service will ask "How to open xx service?", the user answers "Yes", and the customer service's questions are completed. For another example, when the customer service asks "Have you opened xx service?", the user answers "Yes", then the customer service will ask "Have you paid the relevant fees?", the user answers "No", and the customer service's questions are completed.

[0066] Optionally, for the process through machine learning, learning the proposed question for the user's inquiry can include:

[0067] Extracting the question asked by the customer service to the user from the dialogue corpus sample, classifying the question asked to the user by text clustering and storing in the question database. For example: the synonymous interrogative sentences are classified into the same category, and each category forms a question set for the inquiry of a certain business of the user in the question database. Wherein, how to extract the interrogative sentence of the customer service to the user, how to process and classify by text clustering can be realized by various existing ways, and the specific implementation is also easy to be thought of by those skilled in the art on the basis of the technical scheme provided in the present application, and does not serve to limit the protection scope of the present application. Here, it is emphasized that the dialogue corpus sample is used for the inquiry of the user, such as the inquiry of a certain business or a certain use, and a series of questions corresponding to the inquiry are obtained.

[0068] The question in the question database can be a question for different businesses, or a question for different aspects of the same business, etc., as long as it can be called a category.

[0069] Optionally, the question and answer unit can include: an intention prediction module, an intention decision module; wherein,

[0070] The intention prediction module is configured to predict the probability of each question in the question database corresponding to the inquiry answering yes or no based on the inquiry of the user, and form the representation information of the user's intention; wherein each piece of information of the representation information of the user's intention corresponds to a question, and the probability of yes or no for the question.

[0071] The intention decision module is configured to update the probability of the user answering yes (or no) predicted in the representation information based on the representation information in the form of the multi-round question and answer by using the reply of the user to the question asked to the user, and determine the user's intention according to the representation information obtained after the multi-round question and answer. After the intention prediction module obtains the probability of each question in the question database answering yes based on the inquiry of the user, and obtains the representation information of the user's intention, in the process of multi-round question and answer, the reply of the user to the proposed question based on each round of question and answer will update the probability of the corresponding question answering yes (or no) in the above obtained representation information in real time, so that the intention of the user can be determined based on the updated representation information.

[0072] Optionally, the intention decision module can include: a question decision submodule, and an intention decision submodule, wherein,

[0073] a question decision sub-module configured to predict information gain of each question in the representation information in each round of the multi-round question-answering, take a question with the maximum information gain as a question to be asked to the user, and ask the question to the user;

[0074] an intention decision sub-module configured to update the predicted probability of a positive (or negative) answer of the user in the representation information according to a reply of the user to the question asked to the user, and determine the intention of the user according to the updated representation information; in a specific implementation, the number of rounds of the multi-round question-answering can also be limited, that is, when the number of questions reaches a preset number of questions, the questioning is stopped even if the intention of the user cannot be determined, the closest intention of the user is determined based on the representation information, or the questioning is stopped and the user is informed that the intention of the user cannot be determined, the user is asked to re-ask the question, and the like.

[0075] The intention decision module can specifically ask corresponding questions (Q1, Q2, …, Qn) from the question database according to the information of the question and the reply of the user in the conversation process of providing services for the user, and predict the probability of a positive or negative answer of each question to form a vector (p1, p2, …, pn) as the representation information of the intention of the user.

[0076] Suppose that for a question of a user, if the answer of the user to each question in the question database is known (for example, for a yes-or-no question, the answer is known to be positive or negative), it is clear that the user needs to solve what problem or achieve what goal. That is, the series of questions in the question database and the corresponding answers constitute a representation information of the intention of the user.

[0077] When a question is asked for the question of the user, it is assumed that the user gives a positive answer, and according to the question and answer, the corresponding prediction of the question and answer is updated to be positive, as shown in FIG. 2(c); then, according to the questioning strategy in the question database, another question is asked, and it is assumed that the user gives a negative answer, and according to the question and answer, the corresponding prediction of the question and answer is updated to be negative, as shown in FIG. 2(d). In this way, the questioning is ended until the intention of the user is clear enough, such as reaching a preset probability threshold or the number of questions reaches a preset number of questions threshold. And finally, the representation information representing the intention of the user, that is, the distributed representation vector (p1, p2, …, pn) of the intention of the user, is formed.

[0078] More specifically, at the beginning of the dialogue, there is no information about the user's intention, and for all possible questions initiated by the user's inquiry, there is a prior distribution of the answers, as shown in Figure 2(a). After the user states his question or goal (i.e. gives an answer) in response to the proposed question, the answers to each question can be predicted by the question-answer prediction model, forming a distributed representation of the user's intention, as shown in Figure 2(b). Using the distributed representation of the user's intention, the association between each question (Q1, Q2, …, Qn) and the system goal T (which can be a solution or action to achieve the user's intention) can be learned from the dialogue corpus pt=F(p1, p2, …, pn). Based on the question-answer prediction distribution (p1, p2, …, pn), the information gain InfoGain(Qi) that each question Qi (i=1, 2, …n) can bring to the goal after getting a positive answer (i.e. pi=1) or a negative answer (i.e. pi=0) can be predicted, as shown in formula (1), i.e. after determining the question Qi, the prediction of the goal T changes from uncertainty to certainty.

[0079] InfoGain(Qi) = Entropy(pt) - pi x Entropy(pt|pi=1) - (1-pi) x Entropy(pt|pi=0) (1)

[0080] Each time a question is asked, the question with the maximum predicted information gain is asked to the user. When the user gives an answer according to a proposed question, the corresponding question-answer prediction distribution (p1, p2, …, pn) is updated according to this question-answer; according to the updated distribution, the information gain of each question is re-predicted, i.e. InfoGain(Qi) is re-calculated, and the question with the maximum information gain is selected again to continue to ask the user; until the user's intention is clear enough, such as the predicted pt value > a pre-set probability threshold, such as 0.9, i.e. an accurate solution or action can be made, the questioning is ended, or when the questioning initiated by the user's inquiry reaches a pre-set questioning number threshold, it is considered that the user's intention is clear enough, and the questioning is ended.

[0081] The question-answer prediction model can be specifically obtained by machine learning based on dialogue corpus samples, and based on the question-answer prediction model, the user's inquiry can be taken as input, i.e. the probability of answering yes or no for each question in the question database can be obtained.

[0082] Optionally, the intention prediction module is specifically configured to:

[0083] learning, from the dialogue corpus, an association between each question and a system goal for representing a solution or an action to achieve the user's intention based on the user's query; predicting, based on a distribution of the association, a probability of a positive (or negative) answer of each question in the question database corresponding to the query, to form the representation information of the user's intention.

[0084] Optionally, the mapping relationship between the representation information and the user's intention is pre-stored in the intention decision sub-module, and the intention decision module only needs to find the user's intention corresponding to the current formed representation information according to the mapping relationship.

[0085] The distributed representation vector of the user's intention expresses all information related to the user's intention in the dialogue process, and according to the representation vector corresponding to the user's intention, a solution for the user's intention is finally selected and pushed to the user.

[0086] The question database can be a database formed based on a set of questions of one type of business, or can be a database formed based on a set of questions of multiple different types of business, and the questions corresponding to each type of business correspond to the business type. In this way, if the database is formed based on a set of questions of multiple different types of business, when obtaining a user's query, the business type can be determined based on the user's query, all questions corresponding to the business type are found, and the representation information of the user's intention is formed according to the answers to the questions.

[0087] FIG. 2 is a flowchart of a method for implementing human-computer dialogue by the human-computer dialogue device, as shown in FIG. 2, which includes:

[0088] Step 200: obtaining a user's query in a dialogue process for providing service to the user.

[0089] Step 201: based on the user's query and the questions in the question database, asking questions to the user in a multi-round question-answering manner, and determining the user's intention according to the user's answers.

[0090] The step of asking questions to the user in a multi-round question-answering manner based on the user's query and the questions in the question database includes:

[0091] predicting, based on the user's query, a probability of a positive or negative answer of each question in the question database corresponding to the query, to form the representation information of the user's intention;

[0092] updating, based on the representation information, the probability of the positive (or negative) answer of the user predicted in the representation information by using the answers of the user to the questions asked to the user in the multi-round question-answering manner; and determining the user's intention according to the representation information obtained after the multi-round question-answering.

[0093] wherein,

[0094] predicting, based on the user's query, a probability of a positive (or negative) answer to each question in the question database corresponding to the query, forms the representation information of the user's intent.

[0095] learning, based on the user's query, an association between each question and a system goal for representing a solution or an action to achieve the user's intent from a dialogue corpus; predicting, based on a distribution of the association, a probability of a positive (or negative) answer to each question in the question database corresponding to the query, forms the representation information of the user's intent.

[0096] If for a user's query, if the user's answer to each question in the question database is known (for example, for a yes-no question, the answer is known to be positive or negative), it is clear to know what problem the user needs to solve or what goal the user wants to achieve. That is, a series of questions in the question database and their corresponding answers constitute a representation information of the user's intent.

[0097] When a question is asked for the user's query, and a positive answer is given by the user, the corresponding answer prediction is updated to be positive according to the question and answer, as shown in Fig. 2(c); then, according to the questioning strategy in the question database, another question is asked, and a negative answer is given by the user, the corresponding answer prediction is updated to be negative according to the question and answer, as shown in Fig. 2(d). In this way, according to the questioning strategy, the questioning is ended until the user's intent is clear enough, or the questioning reaches a pre-set threshold of the number of questions. Finally, the representation information of the user's intent, i.e., the distributed representation vector of the user's intent (p1, p2, …, pn) is formed.

[0098] More specifically,

[0099] wherein,

[0100] updating, with the user's reply to the question asked to the user, the probability of the user's positive (or negative) answer predicted in the representation information, and determining the user's intent according to the representation information obtained after the multi-round question and answer.

[0101] In each round of the multi-round question and answer, the information gain of each question in the representation information is predicted, the question with the largest information gain is taken as the question asked to the user in the next round of the multi-round question and answer, and the question is asked to the user.

[0102] updating, with the user's reply to the question asked to the user, the probability of the user's positive (or negative) answer predicted in the representation information; until the user's intent is determined according to the updated representation information.

[0103] More specifically, at the beginning of the dialogue, there is no information about the user's intention, and for all possible questions initiated by the user, there is a prior distribution of the answers, as shown in FIG. 2(a). After the user states his question or goal (i.e., gives an answer) in response to the proposed question, the answers to each question can be predicted by the question-answer prediction model, forming a distributed representation of the user's intention, as shown in FIG. 2(b). Using the distributed representation of the user's intention, the association between each question (Q1, Q2, …, Qn) and the system goal T (which can be a solution or an action to achieve the user's intention) can be learned from the dialogue corpus pt=F(p1, p2, …, pn). Based on the question-answer prediction distribution (p1, p2, …, pn), the information gain InfoGain(Qi) for the goal that each question Qi (i=1, 2, …n) can bring about after being answered positively (i.e., pi=1) or negatively (i.e., pi=0) can be predicted, as shown in formula (1), i.e., after determining the question Qi, the prediction of the goal T changes from uncertainty to certainty.

[0104] Each time a question is asked, the question with the maximum predicted information gain is asked to the user. When the user gives an answer in response to a proposed question, the corresponding question-answer prediction distribution (p1, p2, …, pn) is updated according to the question-answer; according to the updated distribution, the information gain of each question is re-predicted, i.e., InfoGain(Qi) is re-calculated, and the question with the maximum information gain is selected again to continue to ask the user; until the user's intention is clear enough, such as the predicted pt value > a pre-set probability threshold, e.g., 0.9, i.e., an accurate solution or action can be made, the questioning is ended, or when the questioning initiated in response to the user's inquiry reaches a pre-set questioning number threshold, it is considered that the user's intention is clear enough, and the questioning is ended.

[0105] More specifically,

[0106] wherein determining the user intention according to the formed representation information of the user intention comprises:

[0107] According to a pre-set mapping relationship between the representation information and the user intention, the user intention corresponding to the current formed representation information is found. The distributed representation vector of the user intention expresses all information related to the user intention in the dialogue process, and according to the representation vector corresponding to the user intention, a solution for the user intention is finally selected and pushed to the user.

[0108] The technical scheme for realizing human-computer dialogue provided in the application, in terms of working mode, clarifies the user's intention by actively asking the user questions when the user's intention is uncertain, intelligently realizes the human-computer dialogue process, and improves the accuracy of identifying the user's intention.

[0109] The method of the application further comprises:

[0110] Determining the questions of the dialogue based on the dialogue corpus samples, such as learning the proposed questions of the inquiries to the user and storing in the question database.

[0111] The questions in the question database include, but are not limited to, yes-or-no questions and other types of questions such as specific questions and selection questions.

[0112] Optionally, for the process of artificial learning, the proposed questions of the inquiries to the user include: assuming that the user inquires, "May I ask how to use my xx service?", through artificial learning, the questions that the customer service may ask include: "Have you opened the xx service?", "Can't you open it?", "Have you paid the relevant fees?", "How to open the xx service?", and the like, which are classified into the same category to form a series of proposed questions of the inquiry of the xx service in the question database. The questions are in a corresponding order of inquiry, such as when the customer service asks, "Have you opened the xx service?", the user answers, "No", then the customer service asks, "Can't you open it?", the user answers, "No", then the customer service asks, "How to open the xx service?", the user answers, "Yes", and the questions of the customer service end. For another example, when the customer service asks, "Have you opened the xx service?", the user answers, "Yes", then the customer service asks, "Have you paid the relevant fees?", the user answers, "No", and the questions of the customer service end.

[0113] Optionally, for the process of machine learning, the proposed questions of the inquiries to the user include:

[0114] Extracting the questions asked by the customer service to the user from the dialogue corpus samples, classifying the questions asked to the user through text clustering, and storing in the question database. For example: classifying the synonymous questions into the same category, and each category forms a question set of the inquiry of a certain service of the user in the question database. The specific implementation of how to extract the questions asked by the customer service to the user and how to process and classify through the text clustering method can be implemented in various ways, and the specific implementation is easily thought of by those skilled in the art based on the technical solutions provided in the application, and does not limit the protection scope of the application. Here, it is emphasized that the inquiries to the user are obtained from the dialogue corpus samples, such as the inquiry of a certain service or the inquiry of a certain use, and a series of questions corresponding to the inquiry are obtained.

[0115] Further, in the evolution mode of the man-machine dialogue device, the dialogue record in the man-machine dialogue process is put into the dialogue corpus sample to correct or supplement the learned question, which continuously improves the performance of the man-machine dialogue device, thereby reducing the updating cost and improving the updating speed of the man-machine dialogue device.

[0116] The method for realizing man-machine dialogue of the present application will be described in detail below in combination with a specific embodiment. Figure 4 The flowchart of the embodiment of the method for realizing man-machine dialogue of the present application is shown in FIG. 1, which includes the following parts: Figure 4

[0117] Firstly, the single-turn question recognition is performed on the user's inquiry. The input of the single-turn interactive question recognition model is the description text of a sentence of the user's inquiry, such as "my account is stolen", and the output is the business classification corresponding to the description, such as "how to limit". Taking the user's inquiry, i.e., the initial question description "my alipay account cannot be logged in" as an example, after performing the classification task, the classification target is 1000 business questions, then it is considered that the single-turn question recognition is successful, and the corresponding solution result is outputted; if the classification result cannot be obtained, it is considered that the single-turn question recognition fails, and the multi-turn interactive process is entered, i.e., further questions are asked to the user and the answers are obtained to help determine the intention of the user's inquiry.

[0118] Then, the question to be further asked is predicted, and further recognition is performed according to the reply of the user to the predicted question. The goal of question prediction is to select a question, and if the question obtains a positive answer, it helps the classification to a certain business most. Assuming that the number of questions in the question database is N, the total number of business categories is K, and P i i (i=1~N) is the probability of positive answer, T j =P (business classification is j | P1P2…P N ) is the conditional probability of business type j (j=1~K), according to formula (1), the information gain when question i becomes a positive answer is: InfoGain(i)=Entropy(T)–P i ×Entropy(T|P i =1)–(1-P i )×Entropy(T|P i ​= 0)), the question to be selected is i = argmax(InfoGain(i)). A multi-layer neural network FFNN can be used to model the mapping from the distribution of answers to the N questions to the distribution of business types, and the training data of the model is also derived from the sample data, such as each telephone call as a training data. Similar to the establishment of the question database, the questions asked by the customer service and the answers of the user can be extracted from the telephone recording text, converted into (P1P2…P N ) vector, and the business type label of the telephone call is added to form a supervised training corpus of data-label.

[0119] The interaction process of the multi-round process is that the model predicts and proposes a question, the user gives an answer, and the corresponding question and answer distribution (P1P2…P N ) is updated according to the question and answer. According to the updated input, the information gain of each question answer is recalculated, the question with the maximum information gain is selected, and the user is further asked. The multi-round question recognition is successful until the user's intention is clear enough. The LSTM network of multi-round interaction can give a classification target higher than a threshold. Multi-round question recognition can be realized by training a multi-round interactive question recognition model. The multi-round interactive question recognition model takes all the dialogue data with the user as input, takes the question "I can't open my Huabei" raised by the user as an example, and executes a classification task with the AGENT further asking "Are you a seller?" and the user answering "Yes". The classification target and the single-round interactive question recognition model are the same, which is 1000 business questions.

[0120] Among them, the input of the question prediction model is the entire dialogue content of the user, and the output is a question in the pre-defined question library. When the confidence score of the question recognition result is not higher than the set threshold, it is considered that the amount of information described by the user is insufficient, and the question prediction model will select a question from the question library that is most helpful for classification to ask the user, and collect the user's answer to use the multi-round interactive question recognition model to identify again.

[0121] The single-round interactive question recognition model, the multi-round interactive question recognition model and the question prediction model mentioned above can be pre-trained, such as a deep neural network (DNN) model.

[0122] The application also provides a device for realizing human-computer dialogue, at least comprising a memory and a processor, wherein the memory stores the following executable instructions: in the dialogue process of providing services for the user, obtaining the inquiry of the user; based on the inquiry of the user and the questions in the question database, using a multi-round question and answer mode to ask the user questions, and determining the user's intention according to the reply of the user.

[0123] Although the present application has been described with reference to the above embodiments, the contents described are merely employed embodiments for facilitating the understanding of the present application, and are not intended to limit the present application. Any modification and change in the form and details can be made by any person skilled in the art without departing from the spirit and scope of the present application, and the patent protection scope of the present application shall be subject to the scope defined by the appended claims.

Claims

1. A human-computer dialogue device, characterized in that, include: The system includes an acquisition unit, a question-and-answer unit, and a question database for storing questions; among which, The acquisition unit is used to acquire user queries during the dialogue process of providing services to the user; The question-and-answer unit is used to proactively ask questions to the user in a multi-round question-and-answer manner when the user's intent is uncertain, based on the user's inquiry and the questions in the question database, and to determine the user's intent based on the user's response. In each round of question and answer, based on the user's inquiry and the user's reply, one or more questions are retrieved from the question database and asked to the user in the order of questioning, and the user's reply based on the question is received; The question-answering unit includes: an intent prediction module and an intent decision module; wherein... The intent prediction module is used to predict the probability of an affirmative answer to each question in the question database corresponding to the question when the user's intent is uncertain, based on the user's question, to form representation information representing the user's intent; The intent decision module includes a question decision submodule and an intent decision submodule. The question decision submodule is used to update the question-and-answer prediction distribution based on the user's response in each round of the multi-round question-and-answer process. The question-and-answer prediction distribution represents the association information between the response and the question. Based on the updated question-and-answer prediction distribution, the information gain of each question in the representation information is re-predicted. The question with the largest information gain is selected as the next question to be posed to the user in the multi-round question-and-answer process. The intent decision submodule is used to update the predicted probability of a positive user answer in the representation information based on the user's response to the question posed to the user, obtaining the updated probability, until the updated probability reaches a probability threshold, thereby determining the user's intent.

2. The human-computer interaction device according to claim 1, characterized in that, Also includes: The learning unit is used to extract questions posed by customer service representatives to the user from the dialogue corpus sample, classify the questions posed to the user through text clustering, and store the classified questions in the question database.

3. The human-computer dialogue device according to claim 1, characterized in that, The intent prediction module is specifically used for: Based on the user's queries and responses, the association between each question and a system goal representing an answer or action to achieve the user's intent is learned from the dialogue corpus; based on the question-answer distribution in the association, the probability of answering affirmatively to each question corresponding to the query in the question database is predicted, forming the representation information representing the user's intent.

4. The human-computer dialogue device according to claim 1, characterized in that, The intent decision submodule stores a pre-set mapping relationship between the representation information and the user intent; The intent decision submodule is specifically used to: find the user intent corresponding to the currently formed representation information based on the mapping relationship.

5. A method for realizing human-computer dialogue, characterized in that, include: During the conversations where services are provided to users, obtain their inquiries; When the user's intent is uncertain, based on the user's inquiries and questions in the question database, a multi-round question-and-answer approach is used to proactively ask the user questions, and the user's intent is determined based on the user's responses. In each round of question and answer, based on the user's inquiry and the user's reply, one or more questions are retrieved from the question database and asked to the user in the order of questioning, and the user's reply based on the question is received; Specifically, when the user's intent is uncertain, a multi-round question-and-answer approach is used to proactively pose questions to the user based on the user's inquiry and questions in the question database. The user's intent is then determined based on their responses. This includes: predicting the probability of an affirmative answer to each question in the question database corresponding to the user's inquiry, forming representation information indicating the user's intent; updating the question-and-answer prediction distribution based on the user's response in each round of the multi-round question-and-answer process, where the prediction distribution represents the association between the response and the question; re-predicting the information gain of each question in the representation information based on the updated prediction distribution; selecting the question with the largest information gain as the next question to be posed to the user in the multi-round question-and-answer process; updating the predicted probability of an affirmative answer in the representation information based on the user's response to the question posed to the user, obtaining the updated probability, until the updated probability reaches a probability threshold, thus determining the user's intent.

6. The method according to claim 5, characterized in that, The step of predicting the probability of an affirmative answer to each question in the question database corresponding to the user's query, and forming representational information indicating the user's intent, includes: Based on the user's queries and responses, the association between each question and a system goal representing an answer or action to achieve the user's intent is learned from the dialogue corpus; based on the question-answer distribution in the association, the probability of answering affirmatively to each question corresponding to the query in the question database is predicted, forming representational information representing the user's intent.

7. The method according to claim 6, characterized in that, Based on the formed representation information indicating the user's intent, the user intent is determined to include: Based on the pre-set mapping relationship between the representation information and the user intent, the user intent corresponding to the currently formed representation information is identified.

8. An apparatus for realizing human-computer dialogue, comprising at least a memory and a processor, wherein, The memory stores the following executable instructions: during a dialogue to provide services to a user, obtain the user's inquiry; when the user's intention is uncertain, proactively ask the user questions in a multi-round question-and-answer manner based on the user's inquiry and questions in the question database, and determine the user's intention based on the user's response; wherein, in each round of question-and-answer, based on the user's inquiry and the user's response, obtain one or more questions from the question database, ask the user questions in the order of questioning, and receive the user's response based on the questions asked; Specifically, when the user's intent is uncertain, a multi-round question-and-answer approach is used to proactively pose questions to the user based on the user's inquiry and questions in the question database. The user's intent is then determined based on their responses. This includes: predicting the probability of an affirmative answer to each question in the question database corresponding to the user's inquiry, forming representation information indicating the user's intent; updating the question-and-answer prediction distribution based on the user's response in each round of the multi-round question-and-answer process, where the prediction distribution represents the association between the response and the question; re-predicting the information gain of each question in the representation information based on the updated prediction distribution; selecting the question with the largest information gain as the next question to be posed to the user in the multi-round question-and-answer process; updating the predicted probability of an affirmative answer in the representation information based on the user's response to the question posed to the user, obtaining the updated probability, until the updated probability reaches a probability threshold, thus determining the user's intent.

Images