Call data processing method and device, vehicle and storage medium
By building multiple knowledge bases and pushing relevant knowledge points based on relevance scores, the problem of artificial intelligence voice failing to effectively classify different conversation scenarios is solved, improving the quality of reply and user experience.
Patent Information
- Application Number
- CN202411152688.5
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-08-21
- Publication Date
- 2024-12-06
AI Technical Summary
The existing technology has failed to effectively classify different conversation scenarios, resulting in low quality of artificial intelligence voice reply and poor user experience.
Build multiple knowledge bases to obtain user data for different conversation scenarios (high-frequency Q&A, task flow, multi-limited language, chat, etc.), determine user questions, traverse the knowledge base, and push relevant knowledge points through correlation scores to realize the reply content and Relationship of scenes.
It has improved the quality of responses of artificial intelligence customer service, improved the relevance of reply content and dialogue scenarios, and improved the user experience.
Smart Images

Figure CN119088922A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the field of artificial intelligence voice data processing technology, and specifically to a call data processing method, a vehicle control device, a vehicle and a computer-readable storage medium. Background Art
[0002] With the rapid development of artificial intelligence and breakthroughs in technologies such as machine learning and natural language processing, more and more commercial and service websites have launched intelligent question-answering robots to provide users with real-time, automatic and convenient online question-answering services. The intelligent question-answering system is a new type of information retrieval system for natural language processing. Its emergence reflects people's enthusiasm for obtaining information quickly and accurately. The intelligent question-answering system is a research topic that integrates natural language processing, knowledge representation, and information retrieval. It is based on text retrieval, but is different from traditional search engines. Traditional search engines require users to enter a combination of keywords, and can only locate documents or web pages for queries submitted by users. Users must rely on themselves to filter out the useful information they need. The question-answering system allows users to enter a question in the form of natural language, and the final answer returned to the user is also a short and accurate answer in the form of natural language.
[0003] However, since the relevant technology does not classify different dialogue scenarios, the answers are rather stiff, which reduces the quality of the artificial intelligence voice responses and also results in a poor user experience. Summary of the invention
[0004] One of the purposes of the present invention is to provide a call data processing method to solve the problem in the prior art that different dialogue scenarios are not classified, resulting in relatively stiff answers, which reduces the quality of artificial intelligence voice responses; the second purpose is to provide a call data processing device; the third purpose is to provide a vehicle; and the fourth purpose is to provide a computer-readable storage medium.
[0005] In order to achieve the above object, the technical solution adopted by the present invention is as follows:
[0006] A call data processing method, characterized by comprising:
[0007] Build multiple knowledge bases for dialogue scenarios;
[0008] Get the user's call data;
[0009] determining a user problem based on the call data;
[0010] Traversing the knowledge base to determine knowledge points that are relevant to the user question and a relevance score for characterizing the strength of the relevance;
[0011] When the relevance score is greater than a preset score threshold, the knowledge point is pushed based on the relevance score.
[0012] Furthermore, the step of constructing multiple knowledge bases for dialogue scenarios includes:
[0013] When the dialogue scenario is a high-frequency question-and-answer scenario, a high-frequency question-and-answer scenario knowledge base is constructed;
[0014] When the dialogue scenario is a task flow dialogue scenario, constructing a task flow dialogue scenario knowledge base;
[0015] When the dialogue scenario is a multi-qualifier dialogue scenario, constructing a knowledge graph;
[0016] When the dialogue scenario is a small talk scenario, a small talk scenario knowledge base is constructed.
[0017] Furthermore, the step of constructing a high-frequency question-answering scenario knowledge base includes:
[0018] Obtaining high-frequency questions and answers corresponding to the high-frequency questions;
[0019] Creating a first knowledge item based on the high-frequency questions and the answers; the first knowledge item includes the high-frequency questions, the answers, and classification information to which the high-frequency questions and the answers belong;
[0020] The first knowledge items are used to construct the high-frequency question and answer scenario knowledge base.
[0021] Furthermore, it also includes:
[0022] Performing a generalization operation on the high-frequency problem to generate similar high-frequency problems;
[0023] Determine the answers corresponding to the similar high-frequency questions;
[0024] Creating a second knowledge item based on the similar high-frequency questions and the answers corresponding to the similar high-frequency questions; the second knowledge item includes the similar high-frequency questions, the answers corresponding to the similar high-frequency questions, and the classification information to which the similar high-frequency questions and the answers corresponding to the similar high-frequency questions belong;
[0025] The second knowledge item is used to update the high-frequency question and answer scenario knowledge base.
[0026] Further, the step of pushing the knowledge point based on the relevance score includes:
[0027] When the user question has an association relationship with a plurality of the knowledge points, the plurality of knowledge points are sorted based on the association scores, and the knowledge points whose association scores are higher than a preset threshold are pushed.
[0028] Furthermore, the step of constructing a task flow dialogue scenario knowledge base includes:
[0029] Determine related questions for expressing user intent, and determine a question order corresponding to the related questions;
[0030] Creating a conversation flow based on the sequence of questions;
[0031] A task flow dialogue scenario knowledge base is constructed based on the dialogue process.
[0032] Furthermore, before the step of pushing the knowledge point based on the relevance score, the method further includes:
[0033] Set up clarifying words;
[0034] If the user's intention cannot be identified, the clarification words are pushed; the clarification words are used to guide the user to ask the question again.
[0035] Furthermore, before the step of traversing the knowledge base to determine the knowledge points associated with the user question, the method further includes:
[0036] Setting trigger conditions; the trigger conditions at least include intent matching, corpus matching and conforming variable verification;
[0037] extracting keywords from the call data;
[0038] The step of traversing the knowledge base to determine knowledge points associated with the user question includes:
[0039] When the call data contains the trigger condition, the task flow dialogue scenario knowledge base is traversed to determine the knowledge points associated with the keyword.
[0040] Furthermore, the step of constructing the knowledge graph includes:
[0041] Create an initial knowledge graph;
[0042] Acquire a triple; the triple includes a subject, a predicate and an object;
[0043] Determining object attributes, and adding the object attributes to the object;
[0044] The subject, the predicate, and the object to which the object attributes are added are added to the initial knowledge graph to construct a knowledge graph.
[0045] Furthermore, the step of traversing the knowledge base to determine knowledge points associated with the user question includes:
[0046] When it is determined that the user question does not hit a knowledge point in the high-frequency question-and-answer scenario knowledge base or the task flow dialogue scenario knowledge base, the knowledge graph is traversed.
[0047] Further, the chat scene knowledge base includes a generation library, and the step of constructing the chat scene knowledge base includes:
[0048] Collect text data in specific fields;
[0049] Performing word segmentation and part-of-speech tagging operations on the text data to generate target text data;
[0050] A feature vector is extracted from the target text data, and a generation library is constructed based on the feature vector.
[0051] A call data processing device, comprising:
[0052] A knowledge base construction module is used to build multiple knowledge bases for dialogue scenarios;
[0053] A call data acquisition module is used to acquire the user's call data;
[0054] A user problem determination module, used to determine user problems based on the call data;
[0055] A knowledge point determination module, used to traverse the knowledge base, determine knowledge points that are relevant to the user question, and a relevance score for characterizing the strength of the relevance;
[0056] The knowledge point pushing module is used to push the knowledge point based on the relevance score when the relevance score is greater than a preset score threshold.
[0057] A vehicle comprising:
[0058] one or more processors;
[0059] and one or more machine-readable media having instructions stored thereon, which, when executed by the one or more processors, cause the vehicle to perform one or more of the methods described above.
[0060] A computer-readable storage medium stores instructions, which, when executed by one or more processors, enable the processors to execute the method according to the embodiment of the present invention.
[0061] Beneficial effects of the present invention:
[0062] The embodiment of the present invention builds multiple knowledge bases for dialogue scenarios; obtains user call data; determines user problems based on the call data; traverses the knowledge base to determine knowledge points associated with the user problems, and a relevance score for characterizing the strength of the relevance; when the relevance score is greater than a preset score threshold, pushes the knowledge point based on the relevance score, thereby achieving the association between the reply content and the dialogue scenario, thereby improving the reply quality of the artificial intelligence customer service. BRIEF DESCRIPTION OF THE DRAWINGS
[0063] Figure 1 is a flowchart of a call data processing method provided in an embodiment of the present invention;
[0064] Figure 2 is a flow chart of a call data processing method provided in an embodiment of the present invention;
[0065] Figure 3 It is a schematic diagram of the construction process of a high-frequency question-answering scenario knowledge base provided in an embodiment of the present invention;
[0066] Figure 4 It is a schematic diagram of a process for constructing a task flow dialogue scenario knowledge base provided in an embodiment of the present invention;
[0067] Figure 5 is a specific dialogue flow chart provided in an embodiment of the present invention;
[0068] Figure 6 It is a schematic diagram of a specific dialogue flow triggering node provided in an embodiment of the present invention;
[0069] Figure 7 It is a schematic diagram of a construction process of a knowledge graph provided in an embodiment of the present invention;
[0070] Figure 8 It is a schematic diagram of a construction process of a chat scene knowledge base provided in an embodiment of the present invention;
[0071] Fig. 9 is a schematic diagram of a display interface of an intent corpus provided in an embodiment of the present invention;
[0072] Fig.10 is a flow chart of another call data processing method provided in an embodiment of the present invention;
[0073] Fig.11 It is a structural block diagram of a call data processing device provided in an embodiment of the present invention. DETAILED DESCRIPTION
[0074] The following will describe the embodiments of the present invention with reference to the accompanying drawings and preferred embodiments. Those skilled in the art can easily understand other advantages and effects of the present invention from the contents disclosed in this specification. The present invention can also be implemented or applied through other different specific embodiments, and the details in this specification can also be modified or changed in various ways based on different viewpoints and applications without departing from the spirit of the present invention. It should be understood that the preferred embodiments are only for illustrating the present invention, not for limiting the scope of protection of the present invention.
[0075] It should be noted that the illustrations provided in the following embodiments are only schematic illustrations of the basic concept of the present invention, and thus the drawings only show components related to the present invention rather than being drawn according to the number, shape and size of components in actual implementation. In actual implementation, the type, quantity and proportion of each component may be changed arbitrarily, and the component layout may also be more complicated.
[0076] Reference Figure 1 , shows a step flow chart of a call data processing method provided in an embodiment of the present invention.
[0077] Step 101, construct multiple knowledge bases for dialogue scenarios;
[0078] Step 102, obtaining the user's call data;
[0079] Step 103, determining the user problem based on the call data;
[0080] Step 104, traversing the knowledge base to determine knowledge points that are relevant to the user question and a relevance score for characterizing the strength of the relevance;
[0081] Step 105: When the relevance score is greater than a preset score threshold, the knowledge point is pushed based on the relevance score.
[0082] In practical applications, the embodiments of the present invention can be applied to an intelligent question-answering system equipped with artificial intelligence customer service. The intelligent question-answering system can be installed in a vehicle, and the conversation between the user and the artificial intelligence customer service can include different conversation scenarios, such as: the user has clear questions that need to be answered; the user has a series of related questions; the user's questions are relatively unpopular; the user wants to chat randomly with the artificial intelligence customer service. For the above scenarios, the embodiments of the present invention can establish different knowledge bases respectively.
[0083] Call data may be the content of conversations during communication with artificial intelligence customer service, and user questions may be questions contained in the conversation content, which may be raised by users.
[0084] refer to Figure 2 , Figure 2It is a flowchart of a call data processing method provided in an embodiment of the present invention.
[0085] In an embodiment of the present invention, a user can talk to an artificial intelligence customer service and ask questions or requests to the artificial intelligence customer service. Based on the questions input by the user, the system calls the corresponding artificial intelligence service to traverse all knowledge bases to calculate the similarity or correlation between the user's questions and all relevant knowledge points in the knowledge base, and determines the correlation score used to characterize the strength of the correlation. The system can preset a scoring threshold and compare the calculated correlation score with the preset scoring threshold. If the correlation score is greater than the preset scoring threshold and there are multiple knowledge points: the system will call the cloud arbitration model to sort the multiple knowledge points and select the knowledge point with the highest correlation as a reply. If there is a knowledge point: the system directly outputs the answer corresponding to the knowledge point. If the correlation score is less than or equal to the preset scoring threshold: the system will adopt a fallback mechanism, for example: provide relevant recommendations, the system will recommend some keywords or topics that may be related to the user's question, or prompt the user to ask the question again, and provide some question suggestions.
[0086] The embodiment of the present invention builds multiple knowledge bases for dialogue scenarios; obtains user call data; determines user problems based on the call data; traverses the knowledge base to determine knowledge points associated with the user problems, and a relevance score for characterizing the strength of the relevance; when the relevance score is greater than a preset score threshold, pushes the knowledge point based on the relevance score, thereby achieving the association between the reply content and the dialogue scenario, thereby improving the reply quality of the artificial intelligence customer service.
[0087] Based on the above embodiment, a variant embodiment of the above embodiment is proposed. It should be noted that in order to make the description concise, only the differences from the above embodiment are described in the variant embodiment.
[0088] In an optional embodiment of the present invention, the step of building multiple knowledge bases for dialogue scenarios includes:
[0089] When the dialogue scenario is a high-frequency question-and-answer scenario, a high-frequency question-and-answer scenario knowledge base is constructed;
[0090] When the dialogue scenario is a task flow dialogue scenario, constructing a task flow dialogue scenario knowledge base;
[0091] When the dialogue scenario is a multi-qualifier dialogue scenario, constructing a knowledge graph;
[0092] When the dialogue scenario is a small talk scenario, a small talk scenario knowledge base is constructed.
[0093] In a specific implementation, a high-frequency question-and-answer scenario may be a conversation scenario in which the frequency of questions asked is higher than a preset threshold. For example, if 10% of 100 users have asked the same question, then when the question is included in the call data of any user, the current conversation scenario may be determined to be a high-frequency question-and-answer scenario.
[0094] The task flow dialogue scenario may include multiple related questions in the call data, such as which mode of travel to use, the weather at the destination, the transportation at the destination, the accommodation environment at the destination, etc.
[0095] A multi-qualifier dialogue scenario may be a dialogue scenario for a specific field. When the call data contains multiple qualifying sentences, the system may determine the current dialogue scenario as a multi-qualifier dialogue scenario for a specific field.
[0096] A casual chat scenario may be one in which the user does not have specific questions and the answers do not require high accuracy and precision.
[0097] The embodiment of the present invention can establish different knowledge bases for the above-mentioned scenarios respectively, so that the answers provided subsequently are associated with different dialogue scenarios, thereby improving the quality and efficiency of the responses.
[0098] In an optional embodiment of the present invention, the step of building a high-frequency question-answering scenario knowledge base includes:
[0099] Obtaining high-frequency questions and answers corresponding to the high-frequency questions;
[0100] Creating a first knowledge item based on the high-frequency questions and the answers; the first knowledge item includes the high-frequency questions, the answers, and classification information to which the high-frequency questions and the answers belong;
[0101] Using the first knowledge item to construct the high-frequency question and answer scenario knowledge base;
[0102] Performing a generalization operation on the high-frequency problem to generate similar high-frequency problems;
[0103] Determine the answers corresponding to the similar high-frequency questions;
[0104] Creating a second knowledge item based on the similar high-frequency questions and the answers corresponding to the similar high-frequency questions; the second knowledge item includes the similar high-frequency questions, the answers corresponding to the similar high-frequency questions, and the classification information to which the similar high-frequency questions and the answers corresponding to the similar high-frequency questions belong;
[0105] The second knowledge item is used to update the high-frequency question and answer scenario knowledge base.
[0106] refer to Figure 3 , Figure 3 It is a schematic diagram of the construction process of a high-frequency question-answering scenario knowledge base provided in an embodiment of the present invention;
[0107] The knowledge preparation stage includes:
[0108] Collect high-frequency questions and collect the most frequently asked questions by users through channels such as user feedback and customer service records.
[0109] Organize your responses to provide accurate, clear answers to each question.
[0110] Classify and summarize, classify questions according to topics, types, etc., for easy management and search.
[0111] The knowledge creation phase includes:
[0112] Question input, input the sorted questions into the system.
[0113] Answer input: Enter the corresponding answer for each question.
[0114] Category selection: select the category to which the knowledge item belongs, such as vehicle model knowledge, product function, etc.
[0115] Related information: If necessary, other related information can be associated, such as vehicle model, product model, etc.
[0116] The knowledge review and certification phase includes:
[0117] Expert review: Domain experts review the created knowledge to ensure the accuracy and professionalism of the content.
[0118] Once the certification is passed, the knowledge that has passed the review will be marked as certified and can be officially launched online.
[0119] Knowledge generalization stage:
[0120] Similar question generation: the system will generate a series of similar questions based on existing questions using artificial intelligence generative models.
[0121] Candidate set,For each similar question generated, the system provides multiple candidate answers.
[0122] Screening and optimization stage:
[0123] Sensitive word filtering, eliminating candidate answers containing sensitive words.
[0124] Semantic scoring: Perform semantic scoring on candidate answers and select the answer with the most coherent semantics.
[0125] Manual screening stage:
[0126] Manual review: similar questions generated by the system are screened one by one by humans.
[0127] Confirm similar questions, confirm which generated similar questions meet the requirements, and add them to the knowledge base to complete the construction of the knowledge base for high-frequency question-and-answer scenarios.
[0128] Building a knowledge base for high-frequency question-and-answer scenarios is a systematic project, involving multiple links such as knowledge collection, creation, review, generalization, and manual screening. Through this process, users' high-frequency questions can be effectively solved and user experience can be improved.
[0129] In an optional embodiment of the present invention, the step of pushing the knowledge point based on the relevance score includes:
[0130] When the user question has an association relationship with a plurality of the knowledge points, the plurality of knowledge points are sorted based on the association scores, and the knowledge points whose association scores are higher than a preset threshold are pushed.
[0131] In the embodiment of the present invention, if it is the only answer, the user is directly replied. If it is not the only answer, the relevance is sorted from high to low through the arbitration model in the cloud, and the answer with the highest relevance is selected to reply to the user.
[0132] In actual applications, business personnel first build a knowledge base for high-frequency question-and-answer scenarios, which is mainly aimed at specific scenarios where users ask and answer questions. Creating knowledge includes creating questions, as well as concise and detailed answers to the questions. Considering the deployment on the vehicle side, concise answers are convenient for voice TTS broadcasting, and users do not need to view every text content in detail. If a car model is involved, the corresponding vehicle model must also be selected. The question is selected and assigned to the corresponding knowledge base path.
[0133] In an optional embodiment of the present invention, the step of constructing a task flow dialogue scenario knowledge base includes:
[0134] Determine related questions for expressing user intent, and determine a question order corresponding to the related questions;
[0135] Creating a conversation flow based on the sequence of questions;
[0136] A task flow dialogue scenario knowledge base is constructed based on the dialogue process.
[0137] Optionally, before the step of pushing the knowledge point based on the relevance score, the method further includes:
[0138] Set up clarifying words;
[0139] If the user's intention cannot be identified, the clarification words are pushed; the clarification words are used to guide the user to ask the question again.
[0140] Optionally, before the step of traversing the knowledge base to determine knowledge points associated with the user question, the step further includes:
[0141] Setting trigger conditions; the trigger conditions at least include intent matching, corpus matching and conforming variable verification;
[0142] extracting keywords from the call data;
[0143] The step of traversing the knowledge base to determine knowledge points associated with the user question includes:
[0144] When the call data contains the trigger condition, the task flow dialogue scenario knowledge base is traversed to determine the knowledge points associated with the keyword.
[0145] refer to Figure 4 , Figure 4 It is a schematic diagram of the construction process of a task flow dialogue scenario knowledge base provided in an embodiment of the present invention.
[0146] The construction of the business flow dialogue scenario knowledge base can mainly include the following stages:
[0147] Intent identification and management stage: determine the purpose of the user's question and collect the corresponding question corpus.
[0148] Dialogue process design phase: Visually design the dialogue process, including setting trigger conditions, defining node actions, etc.
[0149] Node configuration phase: configure the specific actions of each node, such as system response, intent recognition, variable assignment, etc.
[0150] Trigger condition setting stage: Set the trigger conditions between nodes to achieve multi-round dialogues and conditional branches.
[0151] The details are as follows:
[0152] Intent identification and management stage:
[0153] Identify related questions to express user intent and clarify the purpose of each user conversation, such as "check the car condition", "book a flight", etc.
[0154] Collect corpus and various expressions that users may ask, such as "How much gas is left in my car tank?" and "I want to book a flight to Beijing tomorrow."
[0155] Corpus generalization: The system can generate more similar expressions based on the existing corpus to enrich the coverage of intent.
[0156] Dialogue process design phase:
[0157] refer to Figure 5 , Figure 5 is a specific dialogue flow chart provided in an embodiment of the present invention;
[0158] Determine the order of questions corresponding to related questions. Users may ask AI customer service questions continuously about vehicle control instructions. There is a certain order between the instructions, such as controlling the temperature of the air conditioner, adjusting the air volume, adjusting the position of the seat, etc.
[0159] Use visualization tools to connect the nodes in the conversation process to form a clear flowchart.
[0160] Design multi-round dialogues based on the order of questions. For example, after the user asks about the weather, the system can further ask the user whether he needs to check the weather in a specific city.
[0161] Design clarification words. When the system cannot understand the user's intention, you can set clarification words to guide the user to express it again.
[0162] Design follow-up questions. During the conversation, the system can ask users follow-up questions based on the context to obtain more information.
[0163] Node configuration phase:
[0164] refer to Figure 6 , Figure 6 It is a schematic diagram of a specific dialogue flow triggering node provided in an embodiment of the present invention;
[0165] The connection between nodes is a verification condition. If it is met, an action of the corresponding node will be executed, thus splitting the complete business flow into actions executed under different conditional branches. The system's reply content at each node can be text, voice, or other forms of output. When configuring node triggering, the recognized intent can be used as the trigger condition. By setting variable assignments, context information in the conversation can be stored for use in subsequent conversations.
[0166] Trigger condition settings:
[0167] Intent matching: When the statement entered by the user matches a certain intent, the corresponding node is triggered.
[0168] Corpus matching: You can set specific keywords or phrases, and when the user input contains these keywords, the corresponding node is triggered.
[0169] Variable validation: Determine which branch of the dialog flow to execute based on the value of the variable.
[0170] When the call data contains a trigger condition, the task flow dialogue scenario knowledge base is traversed to determine the knowledge points associated with the keywords.
[0171] Building a task flow dialogue scenario knowledge base is essentially to establish an efficient and natural communication bridge between users and artificial intelligence systems. Through visual design, flexible configuration and support for multiple rounds of dialogue, an intelligent and personalized dialogue experience can be created.
[0172] In an optional embodiment of the present invention, the step of constructing a knowledge graph includes:
[0173] Create an initial knowledge graph;
[0174] Acquire a triple; the triple includes a subject, a predicate and an object;
[0175] Determining object attributes, and adding the object attributes to the object;
[0176] The subject, the predicate, and the object to which the object attributes are added are added to the initial knowledge graph to construct a knowledge graph.
[0177] refer to Figure 7 , Figure 7 It is a schematic diagram of a construction process of a knowledge graph provided in an embodiment of the present invention;
[0178] A knowledge graph is a semantic network that links together information from all different sources (structured, semi-structured, and unstructured) and represents the relationships between possible events or concepts in a symbolic way.
[0179] 1. The starting point for building a knowledge graph.
[0180] 2. Create a new knowledge graph: Create a new knowledge graph database or file to store all knowledge triples.
[0181] 3. Add triples:
[0182] Subject: The thing or concept described in the knowledge graph, for example: Apple, Beijing, artificial intelligence.
[0183] Predicate: describes the relationship between the subject and the object, for example: is, is located in, belongs to.
[0184] Object: A thing or concept that has a relationship with a subject, for example: fruit, city, computer science.
[0185] 4. Add attributes and attribute values to the object:
[0186] Attribute: A characteristic or quality of a subject, e.g., color, population, founder.
[0187] Attribute value: the specific value of the attribute, for example: red, 20 million, Turing.
[0188] 5. Establish association relationships for knowledge points: connect different triples through shared entities to form a complex network structure.
[0189] 6. Forming a graph: All connected triplets together form a knowledge graph, which can be regarded as a huge semantic network.
[0190] 7. Construction completed: The knowledge graph is constructed and can be used for subsequent queries and reasoning.
[0191] In an optional embodiment of the present invention, the step of determining knowledge points associated with the user question includes:
[0192] When it is determined that the user question does not hit a knowledge point in the high-frequency question-and-answer scenario knowledge base or the task flow dialogue scenario knowledge base, the knowledge graph is traversed.
[0193] In the specific implementation, if the question raised by the user cannot hit the high-frequency question and answer knowledge base and the task flow dialogue scenario knowledge base, the long-tail unpopular knowledge can be obtained by traversing the knowledge graph, avoiding the unexpected results given directly by the generative model.
[0194] In an optional embodiment of the present invention, the chat scene knowledge base includes a generation library, and the step of constructing the chat scene knowledge base includes:
[0195] Collect text data in specific fields;
[0196] Performing word segmentation and part-of-speech tagging operations on the text data to generate target text data;
[0197] A feature vector is extracted from the target text data, and a generation library is constructed based on the feature vector.
[0198] In actual applications, the chat scene knowledge base can be used to cover the scenarios of users chatting with AI customer service, which is divided into a matching library and a generation library. The matching library is answered by business personnel based on specific user questions, and it relies more on the generation library. The generative big model generates corresponding questions and answers based on the user's semantics.
[0199] refer to Figure 8 , Figure 8 It is a schematic diagram of a construction process of a chat scene knowledge base provided in an embodiment of the present invention;
[0200] Exemplarily, the chat scene knowledge base may be determined in the following manner.
[0201] Phase 1: Data preparation and feature extraction
[0202] Construction of domain-specific knowledge base: Collect text data in specific fields, such as technology news, entertainment news, etc.
[0203] Perform preprocessing on the text, such as word segmentation and part-of-speech tagging.
[0204] Extract information such as entities and relationships to build a knowledge graph or knowledge base.
[0205] Feature extraction:
[0206] Extract features from preprocessed text data, such as word vectors, sentence vectors, etc.
[0207] These features will serve as the basis for training the generative model.
[0208] Phase 2: Model Training
[0209] Generative model training:
[0210] Use the extracted features to train generative models (such as Seq2Seq, Transformer, etc.).
[0211] The goal is to enable the model to generate contextual responses based on a given input (question).
[0212] Classification model training:
[0213] Train a classification model to determine the intent of user input, whether it is small talk or task-based conversation.
[0214] Similarity model training:
[0215] Train a similarity model to measure the similarity between user input and questions in the knowledge base.
[0216] Phase 3: Building a chat knowledge base
[0217] Matching library build:
[0218] Add manually written question-answer pairs to the matching library.
[0219] For some specific scenarios or sensitive questions, answers in the matching library can be used first.
[0220] Generate the library build:
[0221] Integrate the trained generative model into the system.
[0222] When the user input cannot find an exact match in the matching library, the generation model is called to generate a response.
[0223] Phase 4: Optimization and Iteration
[0224] Model optimization:
[0225] As data increases and user feedback increases, the model is continuously optimized to improve the quality of generated responses.
[0226] Optimization can be achieved by adjusting hyperparameters, introducing new features, etc.
[0227] Knowledge Base Updates:
[0228] The knowledge base is updated regularly to add new knowledge points and question-answer pairs.
[0229] User feedback collection:
[0230] Collect user feedback on system responses to improve the model and chat scenario knowledge base.
[0231] In order to enable those skilled in the art to better understand the embodiments of the present invention, a complete example is used below to illustrate the embodiments of the present invention.
[0232] First, build a knowledge base for high-frequency question-and-answer scenarios, which is mainly aimed at specific scenarios where users ask and answer questions. Create knowledge, including creating questions, and concise and detailed answers to the questions. Considering the deployment on the vehicle side, concise answers are convenient for voice TTS (Text-to-Speech) broadcasting, without the need for users to view every text content in detail. If a car model is involved, the vehicle model must also be selected accordingly. Assign the question to the corresponding knowledge base path.
[0233] Select expert certification review. Each knowledge point created needs to be true and accurate. The expert evaluation method is used for secondary confirmation. After certification, the knowledge point can be viewed under the corresponding knowledge base path. Faced with the wide variety of ways of expressing user questions, it is necessary to generalize the knowledge question sentences. For the questioning method, the business personnel can first enter the question corpus, and then the artificial intelligence model can generate other similar question corpora based on the existing corpus. The principle is that for each entered text, the NLU (Natural Language Understanding) algorithm calculates a candidate text set in the optimal solution space. If these sets hit the sensitive word library, they are directly discarded. If they do not hit, they are formed into sentences. Then, based on the sentence scoring model, each produced question corpus is scored. The generalized question corpus that exceeds a certain score will be recommended to the business personnel to screen whether it is a similar question, thereby improving the hit rate of user questions. In the case of limited human resources, recommending similar questions can greatly reduce the pressure of business personnel to enter corpus.
[0234] Build a conversation flow knowledge base. This is mainly aimed at multi-round conversation scenarios with clear user purposes. For example, in the vehicle cabin, the user may continuously ask the AI customer service about vehicle control instructions. There is a certain order between the instructions, such as controlling the temperature of the air conditioner, adjusting the size of the air volume, adjusting the position of the seat, etc. By building a conversation flow, it can help the AI customer service accurately identify a series of commonly used instructions. Business personnel create an intent, which is a series of user questions. Fig. 9 As shown, Fig. 9 is a schematic diagram of a display interface of an intention corpus provided in an embodiment of the present invention, which expresses a user intention corpus set about the vehicle control range. In this way, the user's conventional vehicle use instructions are enumerated.
[0235] After the intent is created, the dialogue effects can be freely combined in the form of nodes and edges, with edges as trigger conditions and nodes as execution actions (currently supporting start nodes, extraction nodes, service nodes, assignment nodes, and reply nodes).
[0236] Trigger conditions: You can configure the trigger conditions for entering the node, supporting three trigger modes: intent corpus, corpus matching, and variable verification;
[0237] Start node: You can configure the exit words of the dialog flow. For example, you can configure the dialog flow to exit when the user says "exit".
[0238] Extraction node: You can define the keywords to be extracted from the user text and save them in variables to facilitate subsequent logical judgment or request for external services. For example, you can configure the extraction of time and location keywords in the user text for subsequent service nodes to request weather query services.
[0239] Service node: The interface service of the open platform. When the flow is transferred to this node, the relevant services can be directly called. For example, the interface service of requesting weather query can be requested when the dialogue flow is transferred to this node.
[0240] Assignment node: In this node, you can reassign the variable, for example, you can configure the variable vehicle type to be assigned "UNIT";
[0241] Reply node: You can configure the reply script in this node, and support configuring multiple scripts.
[0242] At this point, a dialogue flow can be created by the business personnel, and the trigger conditions and outputs can be customized. The reply method can support TTS voice broadcast or text reply.
[0243] When the user's question involves relatively niche and unpopular knowledge, but very clearly involves relevant information about the vehicle model, the AI customer service cannot directly answer using the knowledge base because it has not yet been included. However, if a large generative model is used to provide a backup reply, the accuracy of the vehicle information involved cannot be guaranteed. For car companies, the solution to this problem focuses on building a huge knowledge graph.
[0244] To build a knowledge graph, we first create triples, in short: entity, attribute, attribute value. Through a series of triples, we can build a knowledge graph in the way of "entity-relationship-entity". Relationship refers to subordinate relationships such as containment and location. Based on this relationship, we gradually enrich the knowledge graph.
[0245] After the knowledge graph is constructed, the questions raised by users cannot hit the high-frequency question-and-answer knowledge base and the dialogue flow knowledge base. By traversing the knowledge graph, long-tail and unpopular knowledge can be obtained, thus avoiding the generative model directly giving results beyond the expectations of the car brand.
[0246] When users chat with AI customer service without any clear purpose, there are two situations. Based on the car company's own strategy, a matching library needs to be set for certain questions. After the user asks a question, the AI customer service will reply to the user according to the answer in the matching library to ensure that it meets the relevant requirements of the company; questions that do not match the library are answered by the generative AI model. The current AI model integrates self-developed NLP algorithm models, including Chinese word segmentation, named entity recognition, entity relationship extraction, keyword extraction, syntactic analysis, semantic vector model, text classification, text similarity, grammatical correction, text generation, intent classification, sentiment analysis, etc., which can ensure that the recognition accuracy of user questions is above 80%.
[0247] refer to Fig.10 , Fig.10 is a flow chart of another call data processing method provided in an embodiment of the present invention;
[0248] After building all the above knowledge bases, the artificial intelligence customer service will be associated with the corresponding knowledge base based on the needs of the scenario. After the association is completed, it is necessary to set the hit threshold of each knowledge base. For example, in order to improve the recognition accuracy and avoid related risks, the threshold can be set to the highest value "1"; in high-frequency question-and-answer scenarios, the knowledge base based on the input of business personnel is relatively controllable and has a lower risk. Under a certain sample space, refer to the curve of precision and recall rate and select a certain threshold so that the accuracy can meet the relevant requirements.
[0249] After associating with the knowledge base, the artificial intelligence customer service can be applied to different terminals such as mobile phones, car computers, and WEB terminals (applications or services accessed through a web browser) by passing parameters through the application management interface API.
[0250] When users ask questions on different terminals, the AI customer service will first traverse all related knowledge bases based on the user's questions, calculate the relevance of the matched answers, and compare the relevance with the threshold set for the knowledge base. If it does not exceed the threshold, it will be discarded directly, and if it exceeds the threshold, it will be retained. If all matched answers do not exceed the threshold, the AI model will generate an answer as a fallback answer. If some of all matched answers exceed the threshold, they will be retained and then judged whether they are unique. If it is the only answer, the user will be directly replied. If it is not the only answer, the relevance will be sorted from high to low through the arbitration model in the cloud, and the answer with the highest relevance will be selected to reply to the user.
[0251] Reference Fig.11 , shows a structural block diagram of a call data processing device provided in an embodiment of the present invention, which may specifically include the following modules:
[0252] A knowledge base construction module 1101 is used to construct multiple knowledge bases for dialogue scenarios;
[0253] A call data acquisition module 1102 is used to acquire the user's call data;
[0254] A user problem determination module 1103, configured to determine a user problem based on the call data;
[0255] A knowledge point determination module 1104 is used to traverse the knowledge base and determine knowledge points that are relevant to the user question, as well as a relevance score for representing the strength of the relevance;
[0256] The knowledge point pushing module 1105 is configured to push the knowledge point based on the relevance score when the relevance score is greater than a preset score threshold.
[0257] As for the device embodiment, since it is basically similar to the method embodiment, the description is relatively simple, and the relevant parts can be referred to the partial description of the method embodiment.
[0258] The embodiment of the present invention further discloses a vehicle, comprising:
[0259] one or more processors;
[0260] and one or more machine-readable media having instructions stored thereon, which, when executed by the one or more processors, cause the vehicle to perform one or more of the methods described above.
[0261] The embodiment of the present invention further discloses a computer-readable storage medium having instructions stored thereon, which, when executed by one or more processors, enables the processors to execute the method described in the embodiment of the present invention.
[0262] The above embodiments are only preferred embodiments for fully illustrating the present invention, and the protection scope of the present invention is not limited thereto. Any equivalent substitution or change made by a person skilled in the art based on the present invention is within the protection scope of the present invention.
Claims
1. A call data processing method, characterized in that: include: Build multiple knowledge bases for dialogue scenarios; Get the user's call data; determining a user problem based on the call data; Traversing the knowledge base to determine knowledge points that are relevant to the user question and a relevance score for characterizing the strength of the relevance; When the relevance score is greater than a preset score threshold, the knowledge point is pushed based on the relevance score.
2. The method according to claim 1, characterized in that: The step of constructing multiple knowledge bases for dialogue scenarios includes: When the dialogue scenario is a high-frequency question-and-answer scenario, a high-frequency question-and-answer scenario knowledge base is constructed; When the dialogue scenario is a task flow dialogue scenario, constructing a task flow dialogue scenario knowledge base; When the dialogue scenario is a multi-qualifier dialogue scenario, constructing a knowledge graph; When the dialogue scenario is a small talk scenario, a small talk scenario knowledge base is constructed.
3. The method according to claim 2, characterized in that The steps of constructing a high-frequency question-answering scenario knowledge base include: Obtaining high-frequency questions and answers corresponding to the high-frequency questions; Creating a first knowledge item based on the high-frequency questions and the answers; the first knowledge item includes the high-frequency questions, the answers, and classification information to which the high-frequency questions and the answers belong; The first knowledge items are used to construct the high-frequency question and answer scenario knowledge base.
4. The method according to claim 3, characterized in that Also includes: Performing a generalization operation on the high-frequency problem to generate similar high-frequency problems; Determine the answers corresponding to the similar high-frequency questions; Creating a second knowledge item based on the similar high-frequency questions and the answers corresponding to the similar high-frequency questions; the second knowledge item includes the similar high-frequency questions, the answers corresponding to the similar high-frequency questions, and the classification information to which the similar high-frequency questions and the answers corresponding to the similar high-frequency questions belong; The second knowledge item is used to update the high-frequency question and answer scenario knowledge base.
5. The method according to claim 3, characterized in that: The step of pushing the knowledge point based on the relevance score includes: When the user question has an association relationship with a plurality of the knowledge points, the plurality of knowledge points are sorted based on the association scores, and the knowledge points whose association scores are higher than a preset threshold are pushed.
6. The method according to claim 2, characterized in that The step of constructing a task flow dialogue scenario knowledge base includes: Determine related questions for expressing user intent, and determine a question order corresponding to the related questions; Creating a conversation flow based on the sequence of questions; A task flow dialogue scenario knowledge base is constructed based on the dialogue process.
7. The method according to claim 6, characterized in that Before the step of pushing the knowledge point based on the relevance score, the method further includes: Set up clarifying words; If the user's intention cannot be identified, the clarification words are pushed; the clarification words are used to guide the user to ask the question again.
8. The method according to claim 6 or 7, characterized in that: Before the step of traversing the knowledge base to determine the knowledge points associated with the user question, the method further includes: Setting trigger conditions; the trigger conditions at least include intent matching, corpus matching and conforming variable verification; extracting keywords from the call data; The step of traversing the knowledge base to determine knowledge points associated with the user question includes: When the call data contains the trigger condition, the task flow dialogue scenario knowledge base is traversed to determine the knowledge points associated with the keyword.
9. The method according to claim 2, characterized in that: The steps of constructing the knowledge graph include: Create an initial knowledge graph; Acquire a triple; the triple includes a subject, a predicate and an object; Determining object attributes, and adding the object attributes to the object; The subject, the predicate, and the object to which the object attributes are added are added to the initial knowledge graph to construct a knowledge graph.
10. The method according to claim 9, characterized in that The step of traversing the knowledge base to determine knowledge points associated with the user question includes: When it is determined that the user question does not hit a knowledge point in the high-frequency question-and-answer scenario knowledge base or the task flow dialogue scenario knowledge base, the knowledge graph is traversed.
11. The method according to claim 2, characterized in that The chat scene knowledge base includes a generation library, and the steps of constructing the chat scene knowledge base include: Collect text data in specific fields; Performing word segmentation and part-of-speech tagging operations on the text data to generate target text data; A feature vector is extracted from the target text data, and a generation library is constructed based on the feature vector.
12. A call data processing device, characterized in that: include: A knowledge base construction module is used to build multiple knowledge bases for dialogue scenarios; A call data acquisition module is used to acquire the user's call data; A user problem determination module, used to determine user problems based on the call data; A knowledge point determination module, used to traverse the knowledge base, determine knowledge points that are relevant to the user question, and a relevance score for characterizing the strength of the relevance; The knowledge point pushing module is used to push the knowledge point based on the relevance score when the relevance score is greater than a preset score threshold.
13. A vehicle, characterized in that: include: one or more processors; and One or more machine-readable media having instructions stored thereon, when executed by the one or more processors, cause the vehicle to perform the method of any one of claims 1-11.
14. A computer-readable storage medium, characterized in that: Instructions are stored thereon, which, when executed by one or more processors, cause the processors to perform the method according to any one of claims 1 to 11.
