Method, cloud server and readable storage medium for intelligent dialogue processing

Abstract

The application provides a method for intelligent dialogue processing, a cloud server and a readable storage medium. The method of the application can more comprehensively solve user problems by structuring knowledge information of a plurality of sources of a questioning object to form high-quality dialogue knowledge containing question-answer information pairs and question vectors, and establishing an intelligent dialogue knowledge base. In the knowledge retrieval stage, according to the question text input by the user and the text vector of the question text, multi-path retrieval of dialogue knowledge is performed in the intelligent dialogue knowledge base to obtain multi-path retrieval results. According to the multi-path retrieval results, dialogue knowledge matched with the question text is determined, and the answer information in the dialogue knowledge matched with the question text is taken as the answer information of the question text. The answer information is output. The multi-path retrieval mode based on the question text and the question vector can quickly retrieve high-quality answer information, greatly improving the efficiency of knowledge retrieval and the dialogue quality of intelligent customer service.

Description

TECHNICAL FIELD

[0001] The present application relates to the computer technology field, and in particular to an intelligent dialogue processing method, a cloud server and a readable storage medium. BACKGROUND

[0002] In a full-automatic dialogue process of a user and an intelligent customer service robot, there may be a case that the robot cannot answer a question or a reply cannot comprehensively solve a user problem.

[0003] In some intelligent customer service robot implementation schemes, based on knowledge information from multiple sources such as commodity comments, commodity instruction manuals and historical question and answer knowledge, in an online knowledge retrieval stage, a machine question and answer model is used to predict the relevance between a user input question and each piece of knowledge information, to select knowledge information with high relevance to the user input question as answer information. The knowledge retrieval is time-consuming and inefficient, and the same machine question and answer model is used to predict the relevance between questions and knowledge from different sources, which is inaccurate, resulting in low quality of the retrieved answer information and poor solution to user problems. SUMMARY

[0004] The present application provides an intelligent dialogue processing method, a cloud server and a readable storage medium, to solve the problem that the quality of answer information fed back by an existing intelligent customer service system to a user is low and the user problem cannot be well solved.

[0005] In one aspect, the present application provides an intelligent dialogue processing method, comprising:

[0006] obtaining an input question text and representing the question text as a text vector;

[0007] performing multi-path retrieval of dialogue knowledge in an intelligent dialogue knowledge base according to the question text and the text vector, to obtain a multi-path retrieval result, wherein the multi-path retrieval result contains at least one piece of dialogue knowledge, and the dialogue knowledge includes a question, answer information corresponding to the question and a question vector;

[0008] determining dialogue knowledge matching the question text according to the multi-path retrieval result;

[0009] outputting answer information of the question text as answer information in the dialogue knowledge matching the question text.

[0010] In another aspect, the present application provides a cloud server, comprising a processor and a memory in communication connection with the processor; the memory stores computer execution instructions; and the processor executes the computer execution instructions stored in the memory to implement the above-mentioned intelligent dialogue processing method.

[0011] In another aspect, the present application provides a computer readable storage medium, wherein computer execution instructions are stored in the computer readable storage medium, and the computer execution instructions are used to implement the intelligent dialogue processing method when executed by a processor.

[0012] The intelligent dialogue processing method, the cloud server and the computer readable storage medium provided by the present application can quickly retrieve high-quality answer information based on the multi-path retrieval of the question text and the question vector, greatly improve the efficiency of knowledge retrieval, and improve the quality of intelligent dialogue. BRIEF DESCRIPTION OF DRAWINGS

[0013] The accompanying drawings, which are incorporated into and form a part of the specification, illustrate an embodiment consistent with the present application and, together with the description, serve to explain the principles of the application.

[0014] Figure 1 An exemplary intelligent customer service system architecture applicable to the present application;

[0015] Figure 2 An intelligent dialogue processing method flowchart provided for an exemplary embodiment of the present application;

[0016] Figure 3 A method flowchart for establishing an intelligent dialogue knowledge base provided for an exemplary embodiment of the present application;

[0017] Figure 4 An intelligent dialogue processing method flowchart provided for another exemplary embodiment of the present application;

[0018] Figure 5 A framework schematic diagram of an intelligent dialogue processing method provided by the present application;

[0019] Figure 6 A structural schematic diagram of an intelligent dialogue processing device provided for an exemplary embodiment of the present application;

[0020] Figure 7 A structural schematic diagram of an intelligent dialogue processing device provided for another exemplary embodiment of the present application;

[0021] Figure 8 A structural schematic diagram of a cloud server provided for an embodiment of the present application.

[0022] The present application has been shown and described with reference to the preferred embodiments. Equivalent mechanisms and methods incorporating the spirit of the present application can be employed without departing from the scope of the application, which is defined in the following claims. DETAILED DESCRIPTION

[0023] The exemplary embodiments will be described in detail herein with reference to the attached drawings. The same reference numbers in different drawings represent the same or similar elements. The embodiments described in the following exemplary embodiments are not meant to represent all implementations consistent with the present application. Rather, they are merely examples of apparatuses and methods consistent with some aspects of the present application as detailed in the appended claims.

[0024] First, the terms involved in the present application are explained:

[0025] Knowledge mining: high-quality and valuable knowledge is mined from a large amount of raw text through mining algorithms.

[0026] Knowledge retrieval: knowledge is stored based on a certain organization form, and relevant knowledge is found according to the user's query problem through the corresponding retrieval algorithm.

[0027] In the full-automatic dialogue process between the user and the dialogue robot (such as the intelligent customer service robot), there may be cases where the robot cannot answer the question or the reply cannot fully solve the user's problem. In order to solve this problem, in some intelligent customer service robot implementation schemes, based on knowledge information from multiple sources such as product reviews, product manuals, and historical question and answer knowledge, the similarity between the user's question and the knowledge information from multiple sources is matched to obtain the answer information of the user's question.

[0028] In these schemes, although the retrieval library contains knowledge information from multiple sources, the coverage of knowledge information is expanded, but in the online knowledge retrieval stage, the relevance between the user's input question and each knowledge information is predicted by a machine question and answer model to select the knowledge information with high relevance to the user's input question as the answer information. The knowledge retrieval is time-consuming and inefficient, and the same machine question and answer model is used to predict the relevance between the questions and the knowledge from different sources, which is not accurate, resulting in low quality of the retrieved answer information and poor solution to the user's problem.

[0029] To address the issue that in fully automated dialogue processes between users and intelligent customer service, robots may be unable to answer questions or provide answers that do not fully resolve user problems, this application provides an intelligent dialogue processing method. This method uses a unified knowledge mining process to structure knowledge, unifying the structure of knowledge information from various sources such as attribute information from the questioning object (e.g., products), comment information, and historical dialogue information. This results in dialogue knowledge with a unified structure, containing question-answer pairs and question vectors, thus constructing an intelligent dialogue knowledge base. This knowledge base contains high-quality dialogue knowledge obtained through unified mining of knowledge information from multiple sources, enabling it to better and more comprehensively cover answers to user questions. In the online knowledge retrieval phase, the input question text is represented as a text vector. Based on the question text and the text vector, a multi-way retrieval of dialogue knowledge is performed in the intelligent dialogue knowledge base to obtain multi-way retrieval results. Each multi-way retrieval result contains at least one piece of dialogue knowledge. Further, based on the multi-way retrieval results, the dialogue knowledge that matches the question text is determined. The answer information in the dialogue knowledge that matches the question text is used as the answer information for the input question text, and the answer information is output. The multi-way retrieval method based on text and vectors greatly improves the efficiency of knowledge retrieval, can quickly retrieve high-quality answer information, can better and more comprehensively solve user problems, improve the quality of intelligent dialogue, and alleviate the need for human customer service.

[0030] The intelligent dialogue processing method provided in this application can be applied to intelligent customer service systems or intelligent customer service robots in various specific application scenarios, including but not limited to intelligent customer service robots applied to e-commerce platforms, intelligent customer service robots applied to financial institutions, and intelligent customer service robots applied to various network operators. These will not be listed one by one in this embodiment.

[0031] Figure 1 An exemplary intelligent customer service system architecture diagram applicable to this application is shown below. Figure 1 As shown, the system architecture may specifically include cloud servers and terminal devices.

[0032] Specifically, a cloud server can be a server cluster set up in the cloud. The cloud server has a communication link with each terminal device, which enables communication between the cloud server and each terminal device.

[0033] Terminal devices can specifically refer to hardware devices used by various users that have network communication, computing, and information display functions, including but not limited to smartphones, tablets, desktop computers, and Internet of Things (IoT) devices.

[0034] The user interacts with the cloud server through the terminal device used to realize the function of the intelligent customer service system. Specifically, the user can submit the input question text to the cloud server through the terminal device used. The cloud server receives the input question text of the user sent by the terminal device, represents the question text as a text vector, and performs multi-path retrieval of the dialogue knowledge in the pre-constructed intelligent dialogue knowledge base according to the question text and the text vector, to obtain the retrieval results containing dialogue knowledge in multiple paths through the multi-path retrieval. According to the multi-path retrieval results, the dialogue knowledge with high matching degree with the question text is determined as the dialogue knowledge matched with the input question text, and the answer information in the dialogue knowledge is output to the terminal device as the answer information of the input question text. The terminal device displays the answer information returned by the cloud server to output the answer information to the user.

[0035] In addition, the cloud server can also perform knowledge mining based on the existing user questions in the intelligent customer service system and the knowledge information of multiple sources of the question object, such as attribute information, comment information, and historical dialogue information of the question object (such as an article), structure the knowledge information of multiple different sources, form high-quality dialogue knowledge with a unified structure, construct an intelligent dialogue knowledge base, and the intelligent dialogue knowledge base can better and more comprehensively cover the answers to user questions. In addition, the cloud server can also establish an index containing information such as the question object (such as an article), a merchant, a question, answer information, and a question vector to improve the efficiency of knowledge retrieval.

[0036] The technical solutions of the present application and how the technical solutions of the present application solve the above technical problems will be described in detail below with specific embodiments. The following specific embodiments can be combined with each other, and the same or similar concepts or processes can not be described in detail in some embodiments. The embodiments of the present application will be described below with reference to the accompanying drawings.

[0037] Figure 2 The intelligent dialogue processing method flowchart is provided for an exemplary embodiment of the present application. The execution subject of the present embodiment is the cloud server mentioned above. As shown in the figure, the specific steps of the method are as follows: Figure 2

[0038] Step S201, obtaining the input question text and representing the question text as a text vector.

[0039] The input question text refers to the text information of the question proposed by the user to the intelligent customer service system in the intelligent customer service system. The input question text can be the question text directly input by the user through the exchange interface provided by the terminal device, or the text information converted from the question voice input by the user.

[0040] ​After obtaining the input question text, the cloud server represents the question text as a text vector to perform vector retrieval of dialogue knowledge in the intelligent dialogue knowledge base based on the text vector, and obtains a first retrieval result.

[0041] In the embodiment, the question text is represented as a corresponding text vector. The input question text can be input into a pre-trained text representation model to obtain a corresponding text vector, or any other scheme in the prior art can be used to represent the question text as a text vector. The embodiment is not limited in detail.

[0042] In step S202, multi-path retrieval of dialogue knowledge is performed in the intelligent dialogue knowledge base according to the question text and the text vector, and a multi-path retrieval result is obtained, wherein the multi-path retrieval result contains at least one piece of dialogue knowledge, and the dialogue knowledge includes a question, answer information corresponding to the question, and a question vector.

[0043] The intelligent dialogue knowledge base refers to a pre-established retrieval library, which includes multiple pieces of dialogue knowledge, each piece of dialogue knowledge includes a question-answer information pair and a question vector, the question-answer information pair contains a question and answer information corresponding to the question, and the question vector is obtained by representing the question in the question-answer information pair. The intelligent dialogue knowledge base is described in detail in subsequent embodiments.

[0044] In the embodiment, based on the input question text and the text vector of the question text, multi-path retrieval of dialogue knowledge is performed in the pre-established intelligent dialogue knowledge base. Different paths correspond to different retrieval methods, and each path retrieval result contains at least one piece of dialogue knowledge retrieved based on different retrieval methods.

[0045] For example, the multi-path retrieval of dialogue knowledge in the intelligent dialogue knowledge base can include at least text retrieval based on the question text and the question in the dialogue knowledge, and vector retrieval based on the text vector and the question vector in the dialogue knowledge.

[0046] Optionally, the multi-path retrieval of dialogue knowledge in the intelligent dialogue knowledge base can also include text matching of the input question text and the answer information in the dialogue knowledge to achieve text retrieval based on the question text and the answer information in the dialogue knowledge, and obtain a path retrieval result.

[0047] Optionally, the dialogue knowledge in the intelligent dialogue knowledge base can also include other feature information corresponding to the question, such as intent information. The multi-path retrieval of dialogue knowledge in the intelligent dialogue knowledge base can also include matching of the feature information corresponding to the input question text and the corresponding feature information in the dialogue knowledge, and obtaining a path retrieval result.

[0048] Exemplarily, the dialogue knowledge in the intelligent dialogue knowledge base can also include intention information corresponding to the question, and the multi-path retrieval of the dialogue knowledge in the intelligent dialogue knowledge base can also include intention information corresponding to the input question text, and the intention information in the dialogue knowledge is matched to realize intention retrieval, and a retrieval result based on the intention information is obtained.

[0049] In step S203, the dialogue knowledge matched with the question text is determined according to the multi-path retrieval result.

[0050] After obtaining the multi-path retrieval result, if the multi-path retrieval result contains one piece of dialogue knowledge, the dialogue knowledge can be directly taken as the dialogue knowledge matched with the question text. If the multi-path retrieval result contains multiple pieces of dialogue knowledge, one piece of dialogue knowledge matched with the question text is selected from the dialogue knowledge contained in each path retrieval result.

[0051] It should be noted that not every path retrieval result contains dialogue knowledge, and the multi-path retrieval result recalls at least one piece of dialogue knowledge.

[0052] Exemplarily, a machine question and answer model with high accuracy can be used to predict the matching degree between each piece of dialogue knowledge in each path retrieval result and the input question text, and one piece of dialogue knowledge with high matching degree is selected as the dialogue knowledge matched with the question text according to the matching degree between the input question text and the dialogue knowledge. The machine question and answer model can be a pre-trained machine learning model capable of accurately predicting the matching degree between user questions and knowledge. For example, a pre-trained BERT model applied to question and answer (QA) matching, the backbone network of the BERT model has a large number of encoding layers, which can accurately predict the matching degree of question and answer. For example, the backbone network of the BERT model can include 12 encoding layers (Transformer layers).

[0053] Optionally, after obtaining the multi-path retrieval result, the answer information in the dialogue knowledge contained in each path retrieval result can also be spliced to obtain a knowledge paragraph, and an extractive machine reading comprehension model is used to extract the answer information of the input question text from the knowledge paragraph. The answer information is a text segment in the knowledge paragraph.

[0054] In step S204, the answer information in the dialogue knowledge matched with the question text is taken as the answer information of the question text, and the answer information is output.

[0055] After determining the dialogue knowledge matched with the question text, the cloud server takes the answer information in the dialogue knowledge matched with the question text as the answer information of the question text, and outputs the answer information to the terminal device.

[0056] In a possible scenario, the intelligent customer service system provides a visual interactive interface through which the user inputs the question text, and the terminal device outputs the answer information through the interactive interface.

[0057] In another possible scenario, the user performs voice interaction with the intelligent dialogue robot, and the cloud server converts the answer information of the question text into corresponding audio data after obtaining the answer information, sends the audio data to the terminal device, and the terminal device outputs the answer information to the user by playing the audio data. The specific output manner of the answer information in this embodiment is not limited.

[0058] In this embodiment, based on the pre-constructed intelligent dialogue knowledge base, the dialogue knowledge in the intelligent dialogue knowledge base includes questions, answer information corresponding to the questions, and question vectors. In the knowledge retrieval stage, the multi-path retrieval of dialogue knowledge is performed in the intelligent dialogue knowledge base according to the question text input by the user and the text vector of the question text, the multi-path retrieval result is obtained, the dialogue knowledge matched with the question text is determined according to the multi-path retrieval result, the answer information in the dialogue knowledge matched with the question text is taken as the answer information of the question text, and the answer information is output. Based on the multi-path retrieval of the question text and the question vector, high-quality answer information can be quickly retrieved, the efficiency of knowledge retrieval is greatly improved, and the quality of intelligent dialogue is improved.

[0059] Figure 3 A method flowchart for establishing an intelligent dialogue knowledge base is provided for an exemplary embodiment of the present application. The cloud server obtains the existing user questions in the intelligent customer service system and the knowledge information of the question object from various sources, structures the knowledge by unified knowledge mining according to the existing user questions and the knowledge information of the question object from various sources, unifies the structure of the knowledge information from various sources of the question object, forms high-quality dialogue knowledge with a unified structure containing question-answer information pairs and question vectors, and establishes an intelligent dialogue knowledge base containing dialogue knowledge. The dialogue knowledge in the intelligent dialogue knowledge base can better and more comprehensively cover the answers to user questions, and based on the intelligent dialogue knowledge base, user questions can be comprehensively solved, thereby providing more high-quality question and answer service capabilities for users.

[0060] As shown in Figure 3 , the specific steps of establishing an intelligent dialogue knowledge base are as follows:

[0061] Step S301, obtaining original data of a question object related to the intelligent customer service system from various sources.

[0062] In this embodiment, the question object in the intelligent customer service system refers to an object in the system that can be used as an object of a user question, for example, various types of goods, services, resources, etc. in the intelligent customer service system as commodities.

[0063] The original data of the multiple sources related to the questioning object can include at least one of the following: attribute information of the questioning object, comment data of the questioning object, and historical dialogue data.

[0064] The attribute information of the questioning object includes text information describing various attributes of the questioning object, such as the size, color, applicable crowd description, and usage method description of a product.

[0065] The comment data of the questioning object includes historical comment content of the user on the questioning object, such as the comment content of a user who has purchased a product on the function, effect, and applicable crowd of the product.

[0066] The historical dialogue data includes historical dialogue data generated by the intelligent customer service system in a historical period, which can be set and adjusted according to the actual application scene, and is not specifically limited here. In addition, in order to improve the quality of the knowledge information, the historical dialogue data of the artificial customer service can be used, so that the knowledge information obtained based on the historical dialogue data of the artificial customer service can better solve the user problems.

[0067] In step S302, the original data of the multiple sources is cleaned to retain the original data that does not contain preset invalid information.

[0068] The preset invalid information can include at least one of the following: information with garbled characters, special characters, link information, and defined dialogue templates. The preset invalid information can be set and adjusted according to the specific application scene, and is not specifically limited here.

[0069] In this embodiment, for the original data of the multiple sources in the intelligent customer service system, the data containing the preset invalid information (or noise data) is removed through data cleaning, and only the valid data is retained.

[0070] Specifically, the data containing at least one preset invalid information can be cleaned in this step, and only the original data that does not contain the preset invalid information is retained.

[0071] It should be noted that when data cleaning is performed, the cleaning granularity can be different for raw data of different sources. For example, for multiple pieces of comment data of the same commodity, each piece of comment is taken as a data cleaning object, and the comments of the commodity containing the preset invalid information are cleaned, and the comments of the commodity not containing the preset invalid information are retained. For multiple rounds of conversation content of the same conversation in the historical dialogue data of the artificial customer service, each question of the user and each reply of the artificial customer service are taken as a data cleaning object, and the replies containing the preset invalid information are cleaned, and the questions containing the preset invalid information are cleaned. Optionally, when a user question is cleaned, the reply information for the user question can also be cleaned. In addition, for raw data of different sources, different data cleaning rules can be used for data cleaning, and the specific data cleaning rules can be designed and set according to the content characteristics of the data of different sources, which are not limited here.

[0072] Step S303, for the retained raw data of different sources, the corresponding integration rules are used for integration, and multiple sources of knowledge information of the question object are obtained.

[0073] In actual application, various sources of raw data are usually unstructured text, for example, comment data of commodities, historical dialogue data and the like, without unified format and structure. In this embodiment, for raw data of different sources, different integration rules can be used for integration to obtain corresponding knowledge information. Since the format and structure of raw data of different sources are not unified, different integration rules can be set according to different sources, and the integration rules of various sources can be designed according to the format and structure characteristics of the data of each source in the actual application scene, which are not limited here.

[0074] Exemplarily, for the attribute information of the question object, the integration rule can be that each piece of attribute information of each question object is taken as a piece of knowledge information.

[0075] Exemplarily, for the comment data of the question object, the integration rule can be that the comment data of the same question object is spliced according to a preset rule to generate at least one piece of knowledge information of the question object, or each piece of comment data of each question object can be taken as a piece of knowledge information.

[0076] Exemplarily, for the historical dialogue data, the integration rule can be that at least one question-answer information pair is generated according to the user question and the reply information contained in the historical dialogue data, and each question-answer information pair is a piece of knowledge information.

[0077] For example, historical dialogue data can be segmented based on user questions to obtain multiple conversation fragments. Each fragment contains a user question and one or more corresponding responses. A question-answer pair is generated based on each user question and its corresponding responses. If a user question is followed by multiple responses, these responses can be concatenated to form the answer to the user question, generating a question-answer pair.

[0078] Through the above steps S301-S303, raw data from multiple sources related to the questioner in the intelligent customer service system is obtained, and the raw data from multiple sources is cleaned and integrated to obtain high-quality knowledge information about the questioner from multiple sources. This knowledge information can comprehensively cover various types of information about the questioner, thereby solving user problems more comprehensively.

[0079] The multi-source knowledge information obtained through the above steps includes some question-answer pairs and some that do not contain user questions. Steps S304-S305 then match each piece of knowledge information with a corresponding user question, thereby generating question-answer pairs based on each piece of knowledge information and its corresponding user question, thus unifying the multi-source knowledge information into a unified question-answer pair structure.

[0080] Step S304: Match the knowledge information from multiple sources of the questioner with the user question to determine the user question corresponding to knowledge information 5.

[0081] In this step, for knowledge information containing user questions, such as question-answer pairs obtained from historical dialogue data, the included user question is directly used as the corresponding user question for the knowledge information. For knowledge information not containing user questions, such as knowledge obtained from the attribute information of the questioner or comment data...

[0082] If the information does not contain a user question, then match a user question for the knowledge information that does not contain a user question among the known user questions.

[0083] Specifically, for any piece of knowledge information from multiple sources, if the knowledge information contains a user question, then the user question contained in the knowledge information is taken as the user question corresponding to the knowledge information. If the knowledge information does not contain a user question, then a second machine question answering model is used to predict the relevance information between the knowledge information and existing user questions, and the user question corresponding to the knowledge information is determined based on the relevance information.

[0084] 5. Among them, the second machine question answering model uses a pre-trained BERT model applied to question-answer (QA) matching.

[0085] The encoding layers of the backbone network of the BERT model have fewer layers, which can more accurately and quickly predict the relevance information of the question answer, but the prediction accuracy is not very high. For example, the backbone network of the BERT model can include 3 encoding layers (Transformer layers). In this way, the inference time of the BERT model can be shortened, so that the user questions with certain relevance (lower matching degree) can be quickly determined for a large amount of knowledge information.

[0086] 0Step S305, generating a question answer information pair according to the answer information of the user question corresponding to the knowledge information.

[0087] After determining the user question corresponding to each piece of knowledge information, a question answer information pair is generated according to each piece of knowledge information and the user question corresponding to the knowledge information, and a multi-source question answer information pair related to the questioning object is obtained.

[0088] Specifically, for any piece of knowledge information, if the knowledge information contains a user question, the knowledge information is taken as a question answer information pair.

[0089] The format of the question answer information pair is that the piece of knowledge information is a question answer information pair. If the knowledge information does not contain a user question, the user question corresponding to the knowledge information and the knowledge information form a question answer information pair.

[0090] Through the above steps S301-S305, the knowledge is structured by unified knowledge mining, and the knowledge information from multiple sources of the questioning object is uniformly structured to form structured question answer information pairs. These question answer information pairs can more comprehensively cover most of the user questions and answer information in the intelligent customer service scene.

[0091] Step S306, optimizing the question answer information pair.

[0092] This step is an optional step. In this embodiment, the obtained question answer information pair can be optimized to improve the quality of the question answer information pair.

[0093]

[0094] For example, at least one optimization process can be performed on the question answer information pair in this step: refining the answer information in the question answer information pair, filtering out question answer information pairs with lower matching degrees, and filtering out question answer information pairs that do not meet the healthy and civilized expression conditions.

[0095] ​5Specifically, in the implementation of refining the answer information in the question-answer information pair, the answer information in the question-answer information pair is optimized using a pre-trained extractive machine reading comprehension model. Specifically, for any question-answer information pair, the question-answer information pair is input into the extractive machine reading comprehension model, and through the machine reading comprehension model, the information segment more suitable as the answer to the question is extracted from the answer information in the question-answer information pair for the question in the question-answer information pair as the optimized answer information of the question, so as to refine the answer information of the question-answer information pair and obtain a more high-quality question-answer information pair.

[0096] wherein the extractive machine reading comprehension model can be a pre-trained machine reading comprehension (MRC) model for extractive question answering task, for extracting answer information of a given question from a given text.

[0097] Specifically, in the implementation of filtering out the question-answer information pair with lower matching degree, a third machine question answering model can be used to determine the matching degree of the question and the answer information in the question-answer information pair, and filter out the question-answer information pair with a matching degree less than or equal to a preset matching degree threshold. The third machine question answering model can be a pre-trained machine learning model that can accurately predict the matching degree between the user question and the knowledge. For example, a pre-trained BERT model applied to question-answer (QA) matching, the backbone network of the BERT model has a large number of encoding layers, which can accurately predict the matching degree of the question-answer. For example, the backbone network of the BERT model can include 12 or 24 encoding layers (Transformer layers). The preset matching degree threshold can be set and adjusted according to the actual application scenario and experience value, which is not limited here.

[0098] Specifically, in the implementation of filtering out the question-answer information pair that does not meet the healthy and civilized expression condition, it is determined whether the question-answer information pair meets the healthy and civilized expression condition, and the question-answer information pair that does not meet the healthy and civilized expression condition is filtered out.

[0099] Exemplarily, the healthy and civilized expression condition includes at least one of the following: the corresponding emotion category does not belong to negative emotion, does not contain preset unhealthy information, and does not contain preset uncivilized information. Specifically, by performing emotion category identification, uncivilized information identification, and unhealthy information identification on the answer information in the question-answer pair, it is determined whether the question-answer information pair meets the preset healthy and civilized expression condition according to the identification result. The healthy and civilized expression condition and the preset information in each condition can be set and adjusted according to the historical data and experience in the actual application scenario, which is not limited here.

[0100] In step S307, the question in the question-answer information pair is represented as a question vector, and a dialogue knowledge containing the question-answer information pair and the question vector is generated, and an intelligent dialogue knowledge base containing the dialogue knowledge is established.

[0101] To support multi-path retrieval based on the question text and the question vector, in this embodiment, the question in the question-answer information pair is represented as a question vector, and a dialogue knowledge containing the question-answer information pair and the question vector is generated, and an intelligent dialogue knowledge base containing all dialogue knowledge is constructed.

[0102] Exemplarily, the dialogue knowledge in the form of "question-answer information-question vector" structure can be generated according to the question and the answer information contained in each question-answer information pair and the question vector of the question, and dialogue knowledge with a unified structure is obtained. In addition, dialogue knowledge in the form of "question vector-question-answer information" structure can also be generated, and dialogue knowledge with a unified structure is obtained. Each piece of dialogue knowledge contains a question, answer information corresponding to the question, and a question vector of the question, and the specific structure of the dialogue knowledge is not limited in this embodiment.

[0103] Through the above steps S304-S307, according to the existing user questions and the knowledge information of the multi-source knowledge of the question object, the knowledge is structured through unified knowledge mining, the multi-source knowledge information related to the question object is structured into dialogue knowledge containing question-answer information pairs and question vectors, and an intelligent dialogue knowledge base is established. The source of knowledge information for knowledge retrieval is expanded, so that the dialogue knowledge in the intelligent dialogue knowledge base can cover most of the user questions and answer information in the intelligent customer service scene, and can provide more comprehensive, high-quality and accurate knowledge information for users, better and more comprehensive solve user problems, improve the quality of intelligent customer service robot dialogue, alleviate the demand for manual customer service, and improve user purchase willingness and satisfaction.

[0104] In step S308, an index of the intelligent dialogue knowledge base is established.

[0105] In this embodiment, an index of the intelligent dialogue knowledge base can also be established according to the key information in the dialogue knowledge, so as to improve the efficiency of knowledge retrieval based on the intelligent dialogue knowledge base.

[0106] Optionally, in this step, an index of the intelligent dialogue knowledge base can be established based on the question in the dialogue knowledge and the question vector of the question.

[0107] Optionally, in this step, an index of the intelligent dialogue knowledge base can be established based on the question and the answer information in the dialogue knowledge, and the question vector of the question.

[0108] Optionally, at least one of the associated question object information, the merchant information, and the intent information of the question-answer information pair can also be obtained in this step; and the index of the intelligent dialogue knowledge base is established based on at least one of the associated question object information, the merchant information, and the intent information of the question-answer information pair, and the question and answer information in the question-answer information pair and the question vector of the question.

[0109] Exemplarily, the index can be constructed according to the associated question object information, the merchant information, and the intent information of the question-answer information pair, and the question and answer information in the question-answer information pair and the question vector of the question.

[0110] For example, after the intelligent dialogue knowledge base is constructed, ES (Elasticsearch) index construction is performed according to a preset data structure such as “merchant information-question object information-question-answer information-question vector-intent information”, and an ES index knowledge base is obtained. Based on the ES index knowledge base, for a question input by a user, multi-path retrieval is performed through text retrieval, vector retrieval (vector similarity matching), and the like based on the question text and the text vector of the question, and the answer information of the dialogue knowledge matched with the input question is determined according to the multi-path retrieval result, so that the user question can be replied to more quickly and accurately, and the dialogue quality of the intelligent customer service is improved.

[0111] In this embodiment, the cloud server structures the knowledge information of the question object from various sources into dialogue knowledge containing question-answer information pairs and question vectors through unified knowledge mining based on the existing user questions and the knowledge information of the question object from various sources, establishes an intelligent dialogue knowledge base, expands the sources of knowledge information for knowledge retrieval, and makes the dialogue knowledge in the intelligent dialogue knowledge base cover most of the user questions and answer information in the intelligent customer service scene, so that more comprehensive, high-quality, and accurate knowledge information can be provided for the user to better and more comprehensively solve the user questions; further, the index of the key information in the dialogue knowledge is established, the speed of knowledge retrieval is improved, the user questions can be replied to more quickly and accurately, the dialogue quality of the intelligent customer service robot is improved, the demand for manual customer service is alleviated, and the purchase willingness and satisfaction of the user are improved.

[0112] The method flow of knowledge retrieval for intelligent dialogue based on the pre-established intelligent dialogue knowledge base and index will be described in detail below. The intelligent dialogue processing method provided in this embodiment realizes knowledge retrieval for the input question text based on the intelligent dialogue knowledge base and index constructed in the foregoing embodiments, so that the answer information corresponding to the question text can be quickly and accurately retrieved.

[0113] In an optional embodiment, according to the question text and the text vector, multi-path retrieval of dialogue knowledge is performed in the intelligent dialogue knowledge base, at least including: vector retrieval based on the text vector of the question text and the question vector of the dialogue knowledge in the intelligent dialogue knowledge base, and text retrieval based on the question text and the questions in the intelligent dialogue knowledge base. Through two different retrieval methods, two paths of retrieval results are obtained, wherein one path of retrieval results contains at least one piece of dialogue knowledge retrieved by one retrieval method.

[0114] Figure 4 The intelligent dialogue processing method flowchart provided in the embodiments of the present application, Figure 4 The flow of knowledge retrieval of the intelligent dialogue is exemplarily described taking the multi-path retrieval including vector retrieval and text retrieval as an example. As shown in Figure 4 The specific steps of the intelligent dialogue processing method are as follows:

[0115] Step S401, receiving an input question text, and representing the question text as a text vector.

[0116] The input question text refers to the text information of the question proposed by the user to the intelligent customer service system in the intelligent customer service system. The input question text can be the question text directly input by the user through the exchange interface of the terminal device, or the text information converted from the question voice input by the user.

[0117] After obtaining the input question text, the cloud server represents the question text as a text vector, so as to perform vector retrieval of dialogue knowledge in the intelligent dialogue knowledge base based on the text vector, and obtain one path of retrieval results.

[0118] The question text is represented as a corresponding text vector, the input question text can be input into a pre-trained text representation model for representation to obtain a corresponding text vector, or any other scheme in the prior art can be used to represent the question text as a text vector, which is not limited here in the embodiments.

[0119] Step S402, performing vector similarity matching between the text vector and the question vector of the dialogue knowledge in the intelligent dialogue knowledge base, and obtaining a first path of retrieval results.

[0120] In the embodiments, one path of retrieval according to the text vector is specifically performed by calculating the similarity between the text vector of the input question text and the question vector of each dialogue knowledge in the intelligent dialogue knowledge base, and according to the calculated vector similarity, the dialogue knowledge containing the question with high similarity to the text vector of the question text is determined, and a first path of retrieval results is obtained.

[0121] Optionally, according to the calculated vector similarity, the dialogue knowledge with the vector similarity greater than or equal to a preset vector similarity threshold is recalled to obtain the first retrieval result. In this implementation, the first retrieval result can not include any dialogue knowledge, or can include one or more dialogue knowledge. The preset vector similarity threshold can be set and adjusted according to an actual application scenario and an experience value, which is not specifically limited here.

[0122] Optionally, according to the calculated vector similarity, the first preset number of dialogue knowledge is recalled to obtain the first retrieval result. The first preset number is a positive integer, for example, the first preset number can be 1, 3, 5, etc., and the value of the first preset number can be set and adjusted according to an actual application scenario and an experience value, which is not specifically limited here. In this implementation, the first retrieval result can include at least one dialogue knowledge.

[0123] The similarity between the text vector of the input question text and the question vector of the dialogue knowledge can be the cosine similarity of the two vectors. In addition, the similarity between the text vector of the input question text and the question vector of the dialogue knowledge can also be other vector similarity or distance indicators commonly used to measure the semantic correlation between two text vectors, such as Euclidean distance, which is not specifically limited here.

[0124] In step S403, text similarity matching is performed between the question text and the answer information in the intelligent dialogue knowledge base to obtain a second retrieval result.

[0125] In this embodiment, one-way retrieval is performed according to the question text, specifically by performing text similarity matching between the question text and the answer information in the dialogue knowledge in the intelligent dialogue knowledge base, and according to the text similarity between the question text and the answer information in the dialogue knowledge, dialogue knowledge with high text similarity between the answer information and the question text is determined to obtain the second retrieval result.

[0126] Optionally, according to the text similarity between the question text and the answer information in the dialogue knowledge, dialogue knowledge with the text similarity greater than or equal to a preset first text similarity threshold is recalled to obtain the second retrieval result. In this implementation, the second retrieval result can not include any dialogue knowledge, or can include one or more dialogue knowledge. The preset first text similarity threshold can be set and adjusted according to an actual application scenario and an experience value, which is not specifically limited here.

[0127] Optionally, the second preset number of dialogue knowledge can also be recalled according to the text similarity between the question text and the answer information in the dialogue knowledge, to obtain a second retrieval result. The second preset number is a positive integer, and the second preset number can be equal to or different from the first preset number. For example, the second preset number can be 1, 3, 5, etc. The value of the first preset number can be set and adjusted according to the actual application scenario and experience value, which is not limited here. In this implementation, the second retrieval result can include at least one piece of dialogue knowledge.

[0128] This step can be implemented by using a text matching algorithm based on word frequency, such as a text matching algorithm based on TF-IDF (Term Frequency-Inverse Document Frequency) or a BM25 algorithm. The TF-IDF algorithm and the BM25 algorithm are relatively mainstream algorithms for calculating the similarity score between a user query and a document. The present embodiment is used to calculate the similarity between the question text input by the user and the answer information in the dialogue knowledge.

[0129] In an optional embodiment, the input question text can also be matched with the questions in the dialogue knowledge in the intelligent dialogue knowledge base for text similarity, to obtain another retrieval result. The TF-IDF algorithm, the BM25 algorithm, or other similar algorithms can be used to calculate the text similarity between the input question text and the questions in the dialogue knowledge.

[0130] In an optional embodiment, the dialogue knowledge in the constructed intelligent dialogue knowledge base also includes intent information corresponding to the questions. The multi-path retrieval of the dialogue knowledge according to the question text also includes:

[0131] The intent information corresponding to the question text is identified, and the intent information corresponding to the question text is matched with the intent information contained in the dialogue knowledge in the intelligent dialogue knowledge base, to obtain a third retrieval result.

[0132] Specifically, the input question text can be input into a pre-trained intent recognition model for intent recognition, to obtain the intent information corresponding to the question text. The intent recognition model is a machine learning model trained based on a large amount of training data, and is used to identify the intent information corresponding to the input text.

[0133] Further, according to the intent information corresponding to the question text, the dialogue knowledge containing the same intent information is retrieved from the intelligent dialogue knowledge base, to obtain the third retrieval result. The third retrieval result can not include any dialogue knowledge, or can include one or more pieces of dialogue knowledge.

[0134] In addition, if the dialogue knowledge in the constructed intelligent dialogue knowledge base also includes other feature information corresponding to the question, when the multi-path retrieval of the dialogue knowledge in the intelligent dialogue knowledge base is performed, the feature information corresponding to the input question text can be matched with the corresponding feature information in the dialogue knowledge to obtain another retrieval result.

[0135] In an optional embodiment, a limitation condition can also be set for the multi-path retrieval, and when each path retrieval is performed, the dialogue knowledge meeting the limitation condition is recalled to obtain the retrieval result.

[0136] Specifically, the dialogue knowledge in the constructed intelligent dialogue knowledge base can also include question-related asking object information and merchant information. Part of the dialogue knowledge does not have question-related asking object information and / or merchant information, and the asking object information and / or merchant information in the dialogue knowledge can be a null value or a preset value, and the preset value indicates the information. When the multi-path retrieval of the dialogue knowledge in the intelligent dialogue knowledge base is performed according to the question text and the text vector to obtain the multi-path retrieval result, the cloud server obtains the target object information and / or target merchant information related to the question text, and performs the multi-path retrieval of the dialogue knowledge in the intelligent dialogue knowledge base according to the question text and the text vector and the target object information and / or target merchant information to obtain the multi-path retrieval result associated with the target object information and / or target merchant.

[0137] For example, the dialogue knowledge in the constructed intelligent dialogue knowledge base includes question-related asking object information. When the multi-path retrieval of the dialogue knowledge in the intelligent dialogue knowledge base is performed according to the question text and the text vector to obtain the multi-path retrieval result, the cloud server obtains the target object information related to the question text, and performs the multi-path retrieval of the dialogue knowledge in the intelligent dialogue knowledge base according to the question text and the text vector and the target object information to recall the dialogue knowledge whose question-related asking object information is consistent with the target object information, and obtain the multi-path retrieval result associated with the target object information.

[0138] For example, the dialogue knowledge in the constructed intelligent dialogue knowledge base includes question-related merchant information. When the multi-path retrieval of the dialogue knowledge in the intelligent dialogue knowledge base is performed according to the question text and the text vector to obtain the multi-path retrieval result, the cloud server obtains the target merchant information related to the question text, and performs the multi-path retrieval of the dialogue knowledge in the intelligent dialogue knowledge base according to the question text and the text vector and the target merchant information to recall the dialogue knowledge whose question-related merchant information is consistent with the target merchant information, and obtain the multi-path retrieval result associated with the target merchant information.

[0139] Exemplarily, the dialogue knowledge in the constructed intelligent dialogue knowledge base comprises question-related questioning object information and merchant information. In the implementation of the multi-path retrieval of dialogue knowledge in the intelligent dialogue knowledge base according to the question text and the text vector, and the multi-path retrieval result is obtained, the cloud server obtains the target object information and the target merchant information associated with the question text, and according to the question text and the text vector, and the target object information and the target merchant information, the multi-path retrieval of dialogue knowledge in the intelligent dialogue knowledge base is performed, the dialogue knowledge in which the question-related questioning object information is consistent with the target object information and the question-related merchant information is consistent with the target merchant information is recalled, and the multi-path retrieval result associated with the target object information and the target merchant information is obtained.

[0140] The question-related questioning object information refers to the information of the questioning object to which the question is directed. For example, when a user raises a question about a certain product, the questioning object related to the question is the product, and the questioning object can be determined according to information such as the entry page (e.g., a product detail page) when the user enters the interactive page of the intelligent customer service, and the product link sent by the user.

[0141] The question-related merchant information refers to the information of the merchant to which the question is directed. In actual applications, the question raised by the user can be raised about the questioning object (e.g., a product), or can be raised about a certain merchant. If the question raised by the user is about a certain merchant, for example, the user enters the interactive page of the intelligent customer service from the page of the merchant (rather than the detail page of a certain product), the merchant related to the question is the merchant. If the question raised by the user is about a certain product, for example, the user enters the interactive page of the intelligent customer service through the detail page of a certain product, or sends a link of a certain product, the merchant related to the question is the merchant to which the product belongs. The merchant can be determined according to information such as the entry page (e.g., the store page of a certain merchant) when the user enters the interactive page of the intelligent customer service, and the product link sent by the user.

[0142] In step S404, the matching degree between the dialogue knowledge in each of the multi-path retrieval results and the question text is determined using the first machine question answering model according to the multi-path retrieval result.

[0143] After obtaining the multi-path retrieval result, the matching degree between the answer information in each of the dialogue knowledge contained in the multi-path retrieval result and the input question text is predicted using the first machine question answering model according to the answer information in each of the dialogue knowledge contained in the multi-path retrieval result.

[0144] Specifically, the answer information in each of the dialogue knowledge contained in the multi-path retrieval result and the question text are input into the first machine question answering model for matching degree calculation, and the matching degree between the answer information in each of the dialogue knowledge and the question text is obtained.

[0145] The first machine question and answer model can be a pre-trained machine learning model capable of accurately predicting the matching degree between a user question and knowledge. For example, a pre-trained BERT model applied to question and answer (QA) matching, which has a backbone network with a large number of encoding layers, can accurately predict the matching degree of question and answer. For example, the backbone network of the BERT model can include 12 or 24 encoding layers (Transformer layers).

[0146] In step S405, according to the matching degree with the question text, a dialogue knowledge from the dialogue knowledge in each search result is selected as the dialogue knowledge matched with the question text.

[0147] After obtaining the multiple search results, if the multiple search results contain a dialogue knowledge, the dialogue knowledge can be directly selected as the dialogue knowledge matched with the question text. If the multiple search results contain multiple dialogue knowledge, according to the matching degree between the answer information in each dialogue knowledge contained in the search result and the question text, a dialogue knowledge with a higher matching degree is selected from the dialogue knowledge in each search result as the dialogue knowledge matched with the question text.

[0148] Through the above steps S404-S405, the dialogue knowledge matched with the question text is determined according to the multiple search results.

[0149] In step S406, the answer information in the dialogue knowledge matched with the question text is output as the answer information of the question text.

[0150] In an optional embodiment, if the multiple search results contain multiple dialogue knowledge, multiple dialogue knowledge can also be selected as reference dialogue knowledge from the dialogue knowledge in each search result according to the matching degree with the question text, and the answer information in at least one reference dialogue knowledge is output at the same time as the answer information, so as to provide more and more comprehensive high-quality knowledge to the user, so as to more comprehensively solve the user's problem, alleviate the demand for artificial customer service, and improve the user's purchase willingness and satisfaction.

[0151] In another optional embodiment, the answer information in the dialogue knowledge contained in each search result can also be spliced to obtain a text paragraph, and the text paragraph and the input question text are input into an extractive reading comprehension model. Through the reading comprehension model, a suitable text segment is extracted from the text paragraph as the answer information of the input question text.

[0152] After determining the dialogue knowledge matched with the question text, the cloud server outputs the answer information in the dialogue knowledge matched with the question text as the answer information of the question text to the terminal device.

[0153] In a possible scenario, the intelligent customer service system provides a visual interactive interface through which the user inputs the question text, and the terminal device outputs the answer information through the interactive interface.

[0154] In another possible scenario, the user performs voice interaction with the intelligent dialogue robot, and the cloud server converts the answer information of the question text into corresponding audio data after obtaining the answer information, sends the audio data to the terminal device, and the terminal device outputs the answer information to the user by playing the audio data. The specific output manner of the answer information in this embodiment is not limited.

[0155] The intelligent dialogue processing method provided in this embodiment can quickly and accurately retrieve the answer information corresponding to the question text based on the intelligent dialogue knowledge base and index constructed in the foregoing embodiments through multi-path retrieval, so that the user's question can be replied to more quickly and accurately, the quality of the dialogue of the intelligent customer service robot can be improved, the demand for manual customer service can be alleviated, and the purchase willingness and satisfaction of the user can be improved.

[0156] By combining the foregoing embodiments of establishing an intelligent dialogue knowledge base and the intelligent dialogue processing method of knowledge retrieval, an efficient and unified solution is provided for multi-source knowledge mining and retrieval of intelligent customer service, so as to realize efficient knowledge mining and retrieval of large-scale multi-source knowledge and provide better dialogue service capability for users.

[0157] Figure 5 A framework diagram of an intelligent dialogue processing method provided in this application is shown in FIG. 1. Figure 5 As shown in FIG. 1, for the obtained original data of multiple sources, data cleaning and integration are performed to obtain knowledge information of multiple sources after data cleaning and integration. For each piece of knowledge information, it is determined whether the knowledge information contains a question. For the knowledge information containing a question, a question-answer information pair is directly determined. For the knowledge information not containing a question, an existing user question is matched for the knowledge information, and a corresponding question-answer information pair is generated. For the obtained question-answer information pair, intent recognition, answer information refinement, unhealthy and uncivil information filtering, low matching filtering, question vector representation and the like are performed, and the dialogue knowledge obtained after the processing is stored in an ES and indexed, so that the establishment of the intelligent dialogue knowledge base and index is completed. In the online knowledge retrieval stage, according to the question raised by the user, ES text retrieval, vector retrieval and the like are performed in the intelligent dialogue knowledge base, multiple-path retrieval results containing multiple dialogue knowledge are obtained, the answer information of the question is determined according to the multiple-path retrieval results, and the answer information of the question is output to the user.

[0158] In this embodiment, the cloud server incorporates high-quality knowledge information from multiple sources, including the attribute text of the questioner, comment data, and historical dialogue data. This knowledge information from various sources is structured through a unified knowledge mining process, forming question-answer pairs with a unified structure. The question text is then represented using a text representation model to obtain an index vector. This constructs structured dialogue knowledge containing "question-answer-question vector," and an intelligent dialogue knowledge base and index are built based on this dialogue knowledge. During the knowledge retrieval phase, multi-path retrieval based on the question text and question vector is used to determine the answer information for the question text. By organically combining knowledge mining and retrieval from multiple sources, this method structures knowledge information from various sources into indexed knowledge containing question-answer-question vectors and performs unified retrieval, forming a unified and effective mining and retrieval scheme. While expanding the sources of knowledge retrieval information, it simplifies the complexity of the knowledge mining and retrieval process. Knowledge retrieval is performed using both question text and question vectors, based on vector representation and vector similarity calculation. Compared to existing knowledge retrieval technologies based on machine learning models, this significantly reduces the computational complexity of online real-time retrieval and reasoning, providing users with more and higher-quality, high-precision knowledge. It offers highly efficient knowledge retrieval services for user questions, improves the quality of chatbot dialogue, alleviates the need for human customer service, better and more comprehensively solves user problems, and enhances user purchase intention and satisfaction.

[0159] Figure 6 This is a schematic diagram of the structure of an intelligent dialogue processing apparatus provided in an exemplary embodiment of this application. The intelligent dialogue processing apparatus provided in this embodiment can execute the processing flow provided in the intelligent dialogue processing method embodiment. Figure 6 As shown, the intelligent dialogue processing device 60 includes: a question text processing module 61, a multi-path retrieval module 62, a dialogue knowledge determination module 63, and an answer information determination module 64.

[0160] The question text processing module 61 is used to obtain the input question text and represent the question text as a text vector.

[0161] The multi-way retrieval module 62 is used to perform multi-way retrieval of dialogue knowledge in the intelligent dialogue knowledge base based on the question text and text vector, and obtain multi-way retrieval results. The multi-way retrieval results contain at least one piece of dialogue knowledge, which includes the question, the corresponding answer information, and the question vector.

[0162] The dialogue knowledge determination module 63 is used to determine the dialogue knowledge that matches the question text based on the multi-path retrieval results.

[0163] The answer information determination module 64 is used to take the answer information from the dialogue knowledge that matches the question text and use it as the answer information for the question text, and output the answer information.

[0164] In an optional embodiment, when performing multi-way retrieval of dialogue knowledge in the intelligent dialogue knowledge base based on the question text and text vector to obtain multi-way retrieval results, the multi-way retrieval module 62 is further configured to:

[0165] The first retrieval result is obtained by performing vector similarity matching between the text vector and the question vector in the intelligent dialogue knowledge base; the second retrieval result is obtained by performing text similarity matching between the question text and the answer information in the intelligent dialogue knowledge base.

[0166] In an optional embodiment, the dialogue knowledge further includes intent information corresponding to the question. When performing multi-way retrieval of dialogue knowledge in the intelligent dialogue knowledge base based on the question text and text vector to obtain multi-way retrieval results, the multi-way retrieval module 62 is further configured to:

[0167] Identify the intent information corresponding to the question text; match the intent information corresponding to the question text with the intent information contained in the dialogue knowledge in the intelligent dialogue knowledge base to obtain the third retrieval result.

[0168] In an optional embodiment, the dialogue knowledge further includes: question-associated questioner information and merchant information. When implementing multi-way retrieval of dialogue knowledge in the intelligent dialogue knowledge base based on the question text and text vector to obtain multi-way retrieval results, the multi-way retrieval module 62 is further configured to:

[0169] Obtain the target object information and / or target merchant information associated with the question text; based on the question text and text vector, as well as the target object information and / or target merchant information, perform multi-way retrieval of dialogue knowledge in the intelligent dialogue knowledge base to obtain multi-way retrieval results associated with the target object information and / or target merchant.

[0170] In an optional embodiment, when determining the dialogue knowledge matching the question text based on the multi-path retrieval results, the dialogue knowledge determination module 63 is further configured to:

[0171] If the multi-way retrieval results contain multiple dialogue knowledge items, then based on the multi-way retrieval results, the first machine question answering model is used to determine the matching degree between the dialogue knowledge items in each retrieval result and the question text; based on the matching degree with the question text, one dialogue knowledge item is selected from the dialogue knowledge items in each retrieval result as the dialogue knowledge item that matches the question text.

[0172] In an alternative embodiment, such as Figure 7 As shown, the intelligent dialogue processing device 60 also includes:

[0173] Knowledge base building module 65 is used for:

[0174] Obtaining user questions in the intelligent customer service system and knowledge information of the questioning object from multiple sources; and establishing an intelligent dialogue knowledge base containing dialogue knowledge from multiple sources according to the user questions and the knowledge information of the questioning object from multiple sources.

[0175] In an optional embodiment, when the intelligent dialogue knowledge base containing dialogue knowledge from multiple sources is established according to the user questions and the knowledge information of the questioning object from multiple sources, the knowledge base establishing module 65 is further configured to:

[0176] matching the knowledge information of the questioning object from multiple sources with the user questions to determine user questions corresponding to the knowledge information; generating a question-answer information pair according to answer information of the user questions corresponding to the knowledge information; representing a question in the question-answer information pair as a question vector to generate dialogue knowledge containing the question-answer information pair and the question vector, and establish an intelligent dialogue knowledge base containing the dialogue knowledge.

[0177] In an optional embodiment, when the knowledge information of the questioning object from multiple sources is obtained, the knowledge base establishing module 65 is further configured to:

[0178] obtaining original data of the questioning object from multiple sources in the intelligent customer service system; performing data cleaning on the original data from multiple sources to retain original data not containing preset invalid information; and integrating the retained original data from different sources according to corresponding integration rules to obtain the knowledge information of the questioning object from multiple sources.

[0179] In an optional embodiment, the original data of the questioning object from multiple sources includes at least one of the following:

[0180] attribute information of the questioning object, comment data of the questioning object, and historical dialogue data.

[0181] In an optional embodiment, when the original data from different sources is integrated according to the corresponding integration rules to obtain the knowledge information of the questioning object from multiple sources, the knowledge base establishing module 65 is further configured to:

[0182] taking each piece of attribute information of each questioning object as a piece of knowledge information; splicing the comment data of the same questioning object according to a preset rule to generate at least one piece of knowledge information of the questioning object; and generating at least one question-answer information pair according to user questions and reply information contained in the historical dialogue data, each question-answer information pair being a piece of knowledge information.

[0183] In an optional embodiment, when the knowledge information of the questioning object from multiple sources is matched with the user questions to determine user questions corresponding to the knowledge information, the knowledge base establishing module 65 is further configured to:

[0184] For any of the knowledge information of the knowledge information of multiple sources, if the knowledge information contains a user question, the user question contained in the knowledge information is taken as the user question corresponding to the knowledge information; if the knowledge information does not contain a user question, a second machine question answering model is used to predict the relevance information of the knowledge information and the existing user question, and the user question corresponding to the knowledge information is determined according to the relevance information.

[0185] In an optional embodiment, after realizing that the knowledge information is taken as the answer information of the user question corresponding to the knowledge information, the knowledge base establishing module 65 is further used to:

[0186] The question and answer information pair is subjected to at least one optimization processing as follows:

[0187] The answer information in the question and answer information pair is optimized using an extractive machine reading comprehension model;

[0188] A third machine question answering model is used to determine the matching degree of the question and the answer information in the question and answer information pair, and the question and answer information pair with a matching degree less than or equal to a preset matching degree threshold is filtered out;

[0189] It is judged whether the question and answer information pair meets the healthy and civilized expression condition, and the question and answer information pair that does not meet the healthy and civilized expression condition is filtered out.

[0190] In an optional embodiment, after realizing the establishment of the intelligent dialogue knowledge base, the knowledge base establishing module 65 is further used to:

[0191] At least one of the question and answer information pair associated question object information, merchant information and intent information is obtained; based on at least one of the question and answer information pair associated question object information, merchant information and intent information, and the question and answer information in the question and answer information pair and the question vector of the question, an index of the intelligent dialogue knowledge base is established.

[0192] The device provided by the embodiments of the present application can be specifically used to execute the method provided by any of the above method embodiments, and the specific functions realized and the technical effects that can be achieved will not be repeated here.

[0193] Figure 8 A structural schematic diagram of a cloud server provided by the embodiments of the present application is shown in FIG. 8. As shown in FIG. 8, the cloud server includes a memory 801 and a processor 802. The memory 801 is used to store computer execution instructions and can be configured to store other various data to support operations on the cloud server. The processor 802 is in communication connection with the memory 801 and is used to execute the computer execution instructions stored in the memory 801 to realize the technical solutions provided by any of the above method embodiments, and the specific functions and the technical effects that can be achieved are similar, and will not be repeated here. Figure 8 The memory 801 is used to store computer execution instructions and can be configured to store other various data to support operations on the cloud server. The processor 802 is in communication connection with the memory 801 and is used to execute the computer execution instructions stored in the memory 801 to realize the technical solutions provided by any of the above method embodiments, and the specific functions and the technical effects that can be achieved are similar, and will not be repeated here.

[0194] Optionally, as shown in Figure 8 the cloud server further includes a firewall 803, a load balancer 804, a communication component 805, a power supply component 806, and other components. Figure 8 Some components are only schematically shown in the cloud server, and it does not mean that the cloud server only includes Figure 8 the components shown in the figure.

[0195] The embodiments of the present application also provide a computer readable storage medium, the computer readable storage medium stores computer execution instructions, and the computer execution instructions are executed by a processor to implement the technical solutions provided by any of the above method embodiments, and the specific functions and the technical effects that can be achieved are not repeated here.

[0196] The embodiments of the present application also provide a computer program product, the computer program product includes a computer program, the computer program is stored in a readable storage medium, at least one processor of an end-side device can read the computer program from the readable storage medium, and the at least one processor executes the computer program to make the end-side device execute the technical solutions provided by any of the above method embodiments, and the specific functions and the technical effects that can be achieved are not repeated here.

[0197] The embodiments of the present application provide a chip, including a processing module and a communication interface, the processing module can execute the technical solutions of the cloud server in the above method embodiments. Optionally, the chip further includes a storage module (such as a memory), the storage module is used to store instructions, the processing module is used to execute the instructions stored in the storage module, and the execution of the instructions stored in the storage module makes the processing module execute the technical solutions provided by any of the above method embodiments.

[0198] The above storage can be object storage (Object Storage Service, OSS).

[0199] The above storage can be realized by any type of volatile or non-volatile storage device or a combination thereof, such as static random access memory (SRAM), electrically erasable programmable read-only memory (EEPROM), erasable programmable read-only memory (EPROM), programmable read-only memory (PROM), read-only memory (ROM), magnetic storage, flash memory, magnetic disk or optical disk.

[0200] The communication component is configured to facilitate wired or wireless communication of device to device. The device can access a wireless network based on a communication standard, such as WiFi, a 2G, 3G, 4G / LTE, 5G, or the like cellular communication network, or a combination thereof. In an example embodiment, the communication component receives broadcast signals or broadcast related information from an external broadcast management system via a broadcast channel. In an example embodiment, the communication component further includes a Near Field Communication (NFC) module to facilitate short-range communication. For example, the NFC module can be implemented based on Radio Frequency Identification (RFID) techniques, infrared data association (IrDA) techniques, ultra-wideband (UWB) techniques, Bluetooth (BT) techniques, and other techniques.

[0201] The power component provides power to various components of the device in which the power component is located. The power component can include a power management system, one or more power supplies, and other components associated with generating, managing, and distributing power for the device in which the power component is located.

[0202] Those skilled in the art will appreciate that embodiments of the present application can be readily used as software, hardware, or a combination of software and hardware. In one example, the present application can be implemented as a computer program product, which can include one or more computer program elements. These computer program elements can be stored on a computer program distribution medium, which can be, for example, a compact disc read-only memory (CD-ROM), a floppy disk, a hard disk, or a solid state memory device. The computer program elements can also be stored in the memory of a computer system, which can be, for example, a personal computer, a server, a mobile device, or a computer system that includes one or more computers.

[0203] The present application is described in relation to flowcharts and / or block diagrams of methods, devices (systems), and computer program products according to embodiments of the present application. It is understood that each flow and / or block in the flowcharts and / or block diagrams, and combinations of flows and / or blocks in the flowcharts and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general purpose computer, special purpose computer, embedded processing system, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, create means for implementing the functions specified in the flowcharts and / or block diagrams block or blocks. Figure 1 The flowchart and / or block diagram in which functions are specified in the flowchart and / or block diagram. Figure 1 The flowchart and / or block diagram in which functions are specified in the flowchart and / or block diagram.

[0204] These computer program instructions can also be stored in a computer-readable memory that can direct a computer or other programmable data processing apparatus to function in a particular manner, such that the instructions stored in the computer-readable memory produce an article of manufacture including instructions which implement the functions specified in the flowcharts and / or block diagrams block or blocks. Figure 1 The flowchart and / or block diagram in which functions are specified in the flowchart and / or block diagram. Figure 1 The flowchart and / or block diagram in which functions are specified in the flowchart and / or block diagram.

[0205] These computer program instructions can also be loaded into computer or other programmable data processing devices, so that a series of operational steps are performed on the computer or other programmable devices to generate a computer-implemented process, so that the instructions executed on the computer or other programmable devices provide a process for implementing the flowchart Figure 1 one flowchart or multiple flowcharts and / or blocks Figure 1 one flowchart or multiple flowcharts and / or blocks

[0206] In a typical configuration, a computing device includes one or more processors (CPUs), input / output interfaces, network interfaces, and memory.

[0207] The memory can include non-persistent memory and / or volatile memory, such as random access memory (RAM) and / or read-only memory (ROM) 615, which can be used to temporarily store information during the execution of instructions. The memory 615 is an example of computer-readable media.

[0208] Computer-readable media includes permanent and non-permanent, removable and non-removable media implemented by any method or technology for storing information. Information can be computer-readable instructions, data structures, program modules 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 technology, compact disc read-only memory (CD-ROM), digital versatile disc (DVD), or other optical storage, magnetic cassette, magnetic tape, magnetic disk storage or other magnetic storage devices, or any other non-transmission medium that can be used to store information that can be accessed by a computing device. According to the definition in this paper, computer-readable media does not include transitory computer-readable media, such as modulated data signals and carriers.

[0209] It should be noted that the user information (including but not limited to user equipment information, user attribute information, etc.) and data (including but not limited to data for analysis, stored data, displayed data, etc.) involved in this application are all information and data authorized by the user or authorized by all parties, and the collection, use and processing of related data need to comply with relevant laws, regulations and standards, and provide corresponding operation portal for user to choose authorization or refusal.

[0210] In addition, some of the flowcharts described in the embodiments and the accompanying drawings contain operations that appear in a particular, sequential order. However, it should be understood that the implementation of some of these operations can not occur in the same chronological order as shown in the figures. The sequences of operations should be considered in the context of the previous description of the embodiments. Also, some of the operations can occur in different orders or be expanded upon or combined together or broken apart. Moreover, some of the operations can be performed concurrently or with priority over some of the other operations described. It is intended that each step, operation, or block include every and all permutations of other steps, operations, or blocks that can be similarly grouped. It is intended that the steps, operations, or blocks can be performed in any order, or not performed at all, unless otherwise specified or limited by context. Furthermore, many, if not all, of the functions described in the specification can be performed in the manner described above or other known, equivalent or other functionally similar manner.

[0211] Other embodiments of the application will be apparent to those skilled in the art from consideration of the specification and practice of the application disclosed herein. It is intended that the specification and examples be considered as exemplary only, with the true scope and spirit of the application being indicated by the following claims.

[0212] It is to be understood that the application is not limited to the precise details of construction and the embodiments described above and illustrated in the drawings. The scope of the application is to be determined by the appended claims only.

Claims

1. A method of intelligent dialog processing, characterized by, include: Obtain the input question text and represent the question text as a text vector; Based on the question text and the text vector, a multi-way retrieval of dialogue knowledge is performed in the intelligent dialogue knowledge base to obtain a multi-way retrieval result, wherein the multi-way retrieval result contains at least one piece of dialogue knowledge, and the dialogue knowledge includes a question, the corresponding answer information of the question, and a question vector. Based on the multi-path retrieval results, determine the dialogue knowledge that matches the question text; The answer information from the dialogue knowledge that matches the question text is used as the answer information for the question text, and the answer information is output. The method further includes: The system acquires existing user questions from the intelligent customer service system, as well as knowledge information from multiple sources about the questioner, including the questioner's attribute information, comment information, and historical dialogue information. The knowledge information from multiple sources of the questioner is matched with the user question to determine the user question corresponding to the knowledge information; Based on the answer information of the user's question corresponding to the knowledge information, generate question-answer information pairs; The questions in the question-answer information pairs are represented as question vectors, and dialogue knowledge containing the question-answer information pairs and question vectors is generated to establish an intelligent dialogue knowledge base containing the dialogue knowledge.

2. The method of claim 1, wherein, The step of performing multi-way retrieval of dialogue knowledge in the intelligent dialogue knowledge base based on the question text and the text vector to obtain multi-way retrieval results includes: The text vector is matched with the question vector of the dialogue knowledge in the intelligent dialogue knowledge base to obtain the first retrieval result; The second retrieval result is obtained by matching the text of the question with the answer information in the intelligent dialogue knowledge base based on text similarity.

3. The method of claim 2, wherein, The dialogue knowledge also includes intent information corresponding to the question. The step of performing multi-way retrieval of dialogue knowledge in the intelligent dialogue knowledge base based on the question text and the text vector to obtain multi-way retrieval results further includes: Identify the intent information corresponding to the question text; The intent information corresponding to the question text is matched with the intent information contained in the dialogue knowledge in the intelligent dialogue knowledge base to obtain the third retrieval result.

4. The method according to any one of claims 1 to 3, characterized in that, The dialogue knowledge also includes: information about the questioner and the merchant associated with the question. The step of performing multi-way retrieval of dialogue knowledge in the intelligent dialogue knowledge base based on the question text and the text vector to obtain multi-way retrieval results includes: Obtain the target object information and / or target merchant information associated with the question text; Based on the question text and the text vector, as well as the target object information and / or target merchant information, a multi-way retrieval of dialogue knowledge is performed in the intelligent dialogue knowledge base to obtain multi-way retrieval results associated with the target object information and / or target merchant.

5. The method of claim 1, wherein, The step of determining the dialogue knowledge matching the question text based on the multi-path retrieval results includes: If the multi-way retrieval results contain multiple dialogue knowledge, then based on the multi-way retrieval results, the first machine question answering model is used to determine the matching degree between the dialogue knowledge in each retrieval result and the question text; According to the matching degree with the question text, a dialogue knowledge is selected from dialogue knowledge in each path retrieval result as the dialogue knowledge matched with the question text.

6. The method of claim 5, wherein, Obtaining knowledge information of the questioning object from multiple sources, including: Obtaining original data of the questioning object from multiple sources in the intelligent customer service system; Data cleaning is performed on the original data from multiple sources to retain original data not containing preset invalid information; For the retained original data from different sources, integration is performed according to corresponding integration rules to obtain the knowledge information of the questioning object from multiple sources.

7. The method of claim 6, wherein, The original data of the questioning object from multiple sources includes at least one of the following: Attribute information of the questioning object, comment data of the questioning object, and historical dialogue data; The integration of the retained original data from different sources according to the corresponding integration rules to obtain the knowledge information of the questioning object from multiple sources includes: Each piece of attribute information of each questioning object is taken as a piece of knowledge information; The comment data of the same questioning object is spliced according to a preset rule to generate at least one piece of knowledge information of the questioning object; According to the user questions and reply information contained in the historical dialogue data, at least one question-answer information pair is generated, and each question-answer information pair is a piece of knowledge information.

8. The method of claim 6, wherein, The matching of the knowledge information of the questioning object from multiple sources with the user question to determine the user question corresponding to the knowledge information includes: For any knowledge information in the knowledge information from multiple sources, if the knowledge information contains a user question, the user question contained in the knowledge information is taken as the user question corresponding to the knowledge information; If the knowledge information does not contain a user question, a second machine question answering model is used to predict the relevance information of the knowledge information and the existing user question, and the user question corresponding to the knowledge information is determined according to the relevance information.

9. The method of claim 6, wherein, After generating the question-answer information pair according to the answer information of the user question corresponding to the knowledge information, the following at least one optimization processing is further included: An extractive machine reading comprehension model is used to optimize the answer information in the question-answer information pair; A third machine question answering model is used to determine the matching degree of the question and the answer information in the question-answer information pair, and the question-answer information pair with a matching degree less than or equal to a preset matching degree threshold is filtered out; It is judged whether the question-answer information pair satisfies the healthy and civilized expression condition, and the question-answer information pair not satisfying the healthy and civilized expression condition is filtered out. After establishing the intelligent dialogue knowledge base, the following is further included:

10. The method of claim 6, wherein, Obtaining at least one of the questioning object information, the merchant information, and the intent information associated with the question-answer information pair; Based on at least one of the questioning object information, the merchant information, and the intent information associated with the question-answer information pair, and the question and answer information in the question-answer information pair, and the question vector of the question, an index of the intelligent dialogue knowledge base is established. It includes:

11. A cloud server, characterized by, A processor and a memory in communication connection with the processor; The memory stores computer execution instructions; ​ The processor executes computer-executed instructions stored in the memory to implement the method of any one of claims 1-10.

12. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores computer-executed instructions which, when executed by a processor, implement the method of any one of claims 1-10.

13. A computer program product, characterised in that, The computer program product stores computer-executed instructions which, when executed by a processor, implement the method of any one of claims 1-10.

Images