A method and device for managing a customer service robot knowledge base

Abstract

The application provides a management method and device for a customer service robot knowledge base, which comprises the following steps: acquiring multi-source data; performing question semantic analysis on missed question data, and performing clustering processing on the missed question data based on the analysis result to obtain an unresolved question type set; constructing a knowledge association graph based on the unresolved question type set, artificial customer service answer data and the customer service robot knowledge base, and matching the knowledge association graph with the unresolved question type set to obtain question candidate knowledge items; for each knowledge item in the question candidate knowledge items, evaluating each knowledge item based on user session log data, artificial customer service answer data and the customer service robot knowledge base to obtain a confidence level of the knowledge item; and updating each knowledge item to the customer service robot knowledge base based on the confidence level of each knowledge item. The application solves the problem of knowledge base optimization delay under artificial interference, and improves the knowledge base updating efficiency.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of artificial intelligence, and in particular to a management method and device for a customer service robot knowledge base. BACKGROUND

[0002] With the continuous development of artificial intelligence technology, intelligent customer service robots are increasingly widely used in various fields. These robots can realize intelligent question and answer interaction through natural language processing technology and machine learning algorithms, and provide more efficient and accurate services for users. The core of the intelligent customer service robot is the knowledge base, and only effective management and optimization of the knowledge base can ensure high-quality question and answer services.

[0003] At present, the knowledge base of the customer service robot has certain limitations in actual application, and needs to be continuously expanded and improved during use. However, at present, the updating mode of the knowledge base mainly relies on manual operation, that is, after encountering new problems, manual updating and optimization of the knowledge base are carried out. This updating and optimization mode has a complicated operation process, which not only consumes a large amount of human and material resources, but also easily causes delay in optimization of the knowledge base, and cannot keep up with actual business needs in time. SUMMARY

[0004] The present application provides a management method and device for a customer service robot knowledge base, which solves the problem of a complicated operation process of the updating and optimization mode, which not only consumes a large amount of human and material resources, but also easily causes delay in optimization of the knowledge base under manual interference, improves the updating efficiency of the knowledge base, and enables timely response to business changes and user needs.

[0005] In a first aspect, the present application provides a management method for a customer service robot knowledge base, comprising:

[0006] obtaining multi-source data; the multi-source data includes user session log data, missed question data and artificial customer service answer data;

[0007] performing question semantic analysis on the missed question data, and performing clustering processing on the missed question data based on the analysis result to obtain an unresolved question type set;

[0008] based on the unresolved question type set, the artificial customer service answer data and the customer service robot knowledge base, constructing a knowledge association graph, and matching the knowledge association graph with the unresolved question type set to obtain a question candidate knowledge item;

[0009]

[0009] for each knowledge item in the question candidate knowledge item, based on the user session log data, the artificial customer service answer data and the customer service robot knowledge base, evaluating each knowledge item to obtain a confidence level of the knowledge item;

[0010] updating each of the knowledge items to the customer service robot knowledge base based on the confidence level of each of the knowledge items.

[0011] In a second aspect, the present application further provides a device for managing a customer service robot knowledge base, which is applied to the method for managing a customer service robot knowledge base as described in the first aspect; the device for managing a customer service robot knowledge base comprises:

[0012] a multi-source data collection module, configured to acquire multi-source data; the multi-source data comprises user session log data, missed question data and artificial customer service answer data;

[0013] a semantic clustering analysis module, configured to perform question semantic analysis on the missed question data, and perform clustering processing on the missed question data based on the analysis result, to obtain an unresolved question type set;

[0014] a knowledge association graph construction module, configured to construct a knowledge association graph based on the unresolved question type set, the artificial customer service answer data and the customer service robot knowledge base, and perform matching based on the knowledge association graph and the unresolved question type set, to obtain question candidate knowledge items;

[0015] a confidence review module, configured to, for each of the question candidate knowledge items, perform evaluation on each of the knowledge items based on the user session log data, the artificial customer service answer data and the customer service robot knowledge base, to obtain a confidence level of the knowledge item;

[0016] an updating management module, configured to update each of the knowledge items to the customer service robot knowledge base based on the confidence level of each of the knowledge items.

[0017] In a third aspect, the present application further provides an electronic device, comprising: a memory, configured to store a computer software program; and a processor, configured to read and execute the computer software program, to realize the method for managing a customer service robot knowledge base as described in any of the above aspects.

[0018] In a fourth aspect, the present application further provides a non-transitory computer readable storage medium, wherein the storage medium stores a computer software program, and the computer software program is executed by a processor to realize the method for managing a customer service robot knowledge base as described in any of the above aspects.

[0019] In a fifth aspect, the present application further provides a computer program product, comprising a computer program, and the computer program is executed by a processor to realize the method for managing a customer service robot knowledge base as described in any of the above aspects.

[0020] The management method of the customer service robot knowledge base provided by the embodiment of the application can quickly and accurately find out the unsolved problem type set through semantic analysis and clustering of automatically acquired multi-source data, and can further understand the nature of the user problem in depth, automatically mine more accurate knowledge association, reduce human and material resources, and provide accurate data support for subsequent updating and optimization. Finally, the problem candidate knowledge item association is evaluated according to the artificial customer service answer data, user session log data and knowledge base content, the knowledge added to the knowledge base is accurate and reliable from multiple dimensions, the knowledge base can be automatically updated and optimized according to the evaluation confidence level without a large amount of manual intervention, the problem of knowledge base optimization delay under artificial interference is solved, the knowledge base updating efficiency is improved, and the business changes and user demands can be responded in a timely manner. BRIEF DESCRIPTION OF DRAWINGS

[0021] Figure 1 is a flowchart of the management method of the customer service robot knowledge base provided by the embodiment of the application;

[0022] Figure 2 is a structural schematic diagram of the management device of the customer service robot knowledge base provided by the embodiment of the application;

[0023] Figure 3 is an embodiment diagram of an electronic device provided by the embodiment of the application;

[0024] Figure 4 is an embodiment diagram of a computer readable storage medium provided by the embodiment of the application. DETAILED DESCRIPTION

[0025] The technical solutions in the embodiments of the application will be clearly and completely described below with reference to the drawings in the embodiments of the application. Obviously, the described embodiments are only part of the embodiments of the application, rather than all the embodiments of the application. Based on the embodiments in the application, all other embodiments obtained by those skilled in the art without creative work fall within the scope of protection of the application.

[0026] In the description of the application, the terms "first", "second" are used only for the purpose of description, and cannot be understood as indicating or implying relative importance or implicitly indicating the number of the indicated technical features. Therefore, the features defined with "first", "second" can explicitly or implicitly include one or more of the features. In the description of the application, the meaning of "multiple" is two or more, unless otherwise specifically limited.

[0027] In the description of the present application, the term "for example" is used to mean "serving as an example, instance, or illustration." Any embodiment described as "for example" in the present application is not necessarily to be construed as preferred or advantageous over other embodiments. The following description is presented to enable any person skilled in the art to make and use the application. In the following description, for the purposes of explanation, numerous specific details are set forth in order to provide a thorough understanding of the present application. It will be apparent, however, to one skilled in the art that the present application can be practiced without using these specific details. In other instances, well-known structures and processes are not elaborated upon in order to avoid unnecessary detail, which can obscure the description of the present application. Thus, the present application is not intended to be limited by the embodiments shown, but is to be accorded the widest scope consistent with the principles and features disclosed herein.

[0028] Referring to Figure 1 , Figure 1 is a flowchart of the management method of the customer service robot knowledge base provided by the present application. In the embodiment of the present application, the execution subject of the management method of the customer service robot knowledge base is a knowledge base management device. Therefore, the management method of the customer service robot knowledge base comprises the following steps.

[0029] In step 10, multi-source data is acquired. The multi-source data includes user session log data, missed question data, and artificial customer service answer data.

[0030] Optionally, the knowledge base management device is connected with a user interaction terminal, a customer service robot server, and an artificial customer service workstation, and establishes a data transmission channel. The knowledge base management device acquires real-time transmission data periodically (for example, every second or every minute) by using an API interface, a data interface protocol (for example, HTTP or WebSocket). The user session log data covers complete records of user interaction with the customer service robot, including a user ID, a question time, a question content, a reply content of the robot, and feedback information (for example, follow-up questions, approval, or denial) of the user to the reply. The missed question data refers to user question record data that cannot be matched with an answer or cannot be effectively answered by the customer service robot according to existing customer service robot knowledge base content after the user asks a question to the customer service robot. The missed question data is mainly used to accurately locate a knowledge blind spot of the customer service robot knowledge base. The artificial customer service answer data is mainly extracted from an artificial customer service system and is derived from records of the artificial customer service when processing user questions. The artificial customer service answer data contains information such as a time when a question is transferred to an artificial customer service, an artificial customer service answer text, and a user determination state, so as to provide a professional and effective question solution. Through integration of the source data, limitations of single data sources can be avoided, and problem scenarios that the user may encounter can be more comprehensively covered.

[0031] Further, after obtaining the user session log data, the missed question data and the artificial customer service answer data, in order to determine the accuracy and integrity of the data, the original data is preliminarily cleaned after being obtained, and the repeated records, garbled codes and irrelevant information are removed. For example, for the user session log data, if it is found that the same user repeatedly asks the same content in a short time, only one valid record is retained.

[0032] Further, in an embodiment, taking the e-commerce customer service system as an example, the user session log is obtained from the user App end or PC end every 5 minutes, and if the user asks: Does the product support 7-day no reason return? When the customer service robot retrieves the customer service robot knowledge base and no matching answer is found, the question is automatically marked as missed question data. At the same time, when the artificial customer service intervenes and answers: Except for customized products, all can be returned within 7 days without reason, the related answer content is collected into the artificial customer service answer data.

[0033] Step 20, the missed question data is analyzed for question semantics, and based on the analysis result, the missed question data is clustered to obtain a set of unsolved question types.

[0034] Optionally, after the knowledge base management device obtains the multi-source data, the missed question data in the multi-source data is processed through natural language processing technology to determine the lexical semantic features and semantic categories of each missed question in the missed question data, and the lexical semantic features and semantic categories are clustered and analyzed according to the association relationship between the lexical semantic features and semantic categories to obtain a set of unsolved question types, as described in steps 201-204.

[0035] Further, the set of unsolved question types obtained by the knowledge base management device includes a set of all unsolved question types. Similarly, taking the e-commerce customer service scene as an example, if the collected missed questions include Can I get a refund if I receive the goods halfway, Can I get a refund if I don't want it, etc., first, the two questions are processed by natural language processing to obtain corresponding lexical semantic features and semantic categories, and then they are classified as unsolved question types of product refund policy after clustering according to the association relationship between the corresponding lexical semantic features and semantic categories.

[0036] Step 30, based on the set of unsolved question types, the artificial customer service answer data and the customer service robot knowledge base, a knowledge association graph is constructed, and the knowledge association graph is matched with the set of unsolved question types to obtain a question candidate knowledge item.

[0037] Optionally, the knowledge base management device, after determining the set of unresolved problem types, constructs a knowledge association graph based on the set of unresolved problem types, the manual customer service answer data and the customer service robot knowledge base, taking the problem attributes of each unresolved problem in the set of unresolved problem types as nodes and the relationships between the problem attributes as edges. As described in steps 3011-3015, the knowledge association graph can quickly and intuitively show the relationships between the unresolved problem types and various knowledge, thereby quickly understanding the structure and connections of the knowledge system.

[0038] Further, after the knowledge base management device determines the knowledge association graph, it first performs natural language processing on each unresolved problem in the set of unresolved problem types, maps the semantic units obtained after the natural language processing to the nodes in the knowledge association graph, expands the association paths and supplements the paths, and finally performs matching screening and knowledge item fusion on each unresolved problem in the set of unresolved problem types, thereby obtaining the subsequent knowledge items of the problem. As described in steps 3021-3025, the candidate knowledge items related to the unresolved problem types can be accurately found, and the accuracy and efficiency of knowledge matching are improved.

[0039] Step 40, for each knowledge item in the problem candidate knowledge item, based on the user session log data, the manual customer service answer data and the customer service robot knowledge base, each knowledge item is evaluated to obtain the confidence level of the knowledge item.

[0040] Optionally, after the knowledge base management device determines the problem candidate knowledge item, for each knowledge item in the problem candidate knowledge item, it is evaluated from three aspects of the user session log data, the manual customer service answer data and the customer service robot knowledge base, thereby obtaining the confidence level of each knowledge item. As described in steps 401-405, the confidence level of the knowledge item is evaluated by considering multiple factors, which can comprehensively and objectively evaluate the quality and reliability of the knowledge item, avoid the one-sidedness of a single standard, and reduce the risk of incorrect knowledge storage.

[0041] Step 50, based on the confidence level of each knowledge item, each knowledge item is updated to the customer service robot knowledge base.

[0042] Optionally, after obtaining the confidence level of each knowledge item, the knowledge base management device adopts different updating strategies according to the confidence level of each knowledge item, wherein the confidence level contains at least three levels, such as the first confidence level, the second confidence level and the third confidence level, and the confidence standard of each confidence level can be adjusted according to actual conditions, as described in steps 501-503. It should be noted that the confidence level is not limited to three levels, and can be adjusted according to actual conditions and needs, so as to avoid the problem that the quality and reliability of knowledge items of different confidence levels are different, and the unified updating method may cause low-quality knowledge in the knowledge base or waste of audit resources. At the same time, the hierarchical updating strategy can reasonably allocate resources and improve the efficiency and quality of knowledge base updating by processing according to the actual situation of knowledge items.

[0043] The embodiment of the present application can quickly and accurately find out the set of unsolved problem types by semantic analysis and clustering of automatically obtained multi-source data, and can also deeply understand the nature of user problems by combining knowledge association graph and matching of unsolved problems, automatically mine more accurate knowledge association, reduce human and material resources, and provide accurate data for subsequent update optimization. Finally, according to the artificial customer service answer data, user session log data and knowledge base content, the problem candidate knowledge item association evaluation can be carried out from multiple dimensions to ensure that the knowledge added to the knowledge base is accurate and reliable, and the knowledge base can be automatically updated and optimized according to the evaluation confidence level without a large amount of manual intervention, which solves the problem of knowledge base optimization delay under manual interference, improves the knowledge base updating efficiency, and can respond to business changes and user needs in a timely manner.

[0044] In an embodiment, steps 201-204 are described as follows:

[0045] Step 201, performing word segmentation processing on the hit problem data to obtain a plurality of vocabulary data, and performing feature extraction on the vocabulary data based on a preset semantic knowledge base to obtain vocabulary semantic features.

[0046] Optionally, the knowledge base management apparatus adopts a word segmentation tool (such as Jieba word segmentation, HanLP) to perform word segmentation on the missed question data, to obtain, for example, the missed question in the e-commerce customer service scenario: Where can I see the size table of this dress? After Jieba word segmentation, the following vocabulary data is obtained: this, dress,, size table, where, and see. After word segmentation, the vocabulary data is subjected to feature extraction according to the preset semantic knowledge base. In addition to relying on a word vector model (such as Word2Vec), other semantic representation methods such as GloVe and FastText can also be used for analysis and processing. Taking the Word2Vec word vector model called in the preset semantic knowledge base as an example, the model has been pre-trained on a large-scale text corpus. Each word is input into the word vector model to generate a corresponding word vector. These word vectors are the semantic features of the words. For example, the word “dress” will obtain a vector of a specific dimension (such as 300 dimensions) in the Word2Vec model. The vector contains semantic information of the word “dress” in the semantic space.

[0047] In step 202, the vocabulary semantic features are subjected to syntactic structure analysis to obtain syntactic structure information, and the vocabulary semantic features are subjected to semantic role labeling based on the syntactic structure information to obtain semantic role information.

[0048] Optionally, after determining the vocabulary data and vocabulary semantic features of the missed question data after word segmentation, the knowledge base management apparatus performs syntactic structure analysis on the vocabulary semantic features. In addition to determining the basic grammatical relationship according to the dependency syntactic analyzer, semantic dependency analysis can also be introduced to further mine the semantic dependency relationship between words in the sentence, such as cause-and-effect relationship and purpose relationship. It should be noted that for some long and difficult sentences, recursive syntactic structure analysis can be used to decompose the sentence into multiple clauses, analyze the syntactic structure of each clause, and then synthesize the syntactic structures of all clauses to obtain the syntactic structure information of the entire sentence.

[0049] Further, after determining the syntactic structure information, the knowledge base management apparatus performs semantic role labeling on the vocabulary semantic features according to the syntactic structure information. In addition to labeling common semantic roles (such as agent and patient), custom semantic role labeling can also be performed according to the specific business domain and question characteristics. For example, in the e-commerce customer service scenario, a commodity attribute role (used to describe the color, size, and other attributes of a commodity) and a transaction behavior role (such as purchase and return) can be added. In addition, for cases where the semantic role is ambiguous, more accurate labeling can be performed in combination with the semantic constraint rules in the semantic knowledge base. For example, in the case where a user purchases a dress, the dress is both a patient and a commodity attribute role. Through the semantic constraint rules, it can be determined that the commodity attribute role is more important in this sentence context, and the semantic role labeling is performed to determine the final semantic role information.

[0050] Further, the syntax structure analysis explicitly defines the grammatical relationship between words, which helps to understand the semantic combination logic of the sentence. Semantic role labeling further refines the component role from the semantic level, making the text semantic understanding more in-depth and comprehensive, and providing more accurate information for semantic reasoning.

[0051] In step 203, the semantic features of the words, the syntax structure information and the semantic role information are subjected to semantic reasoning based on the preset semantic reasoning rule library, and the semantic category is determined.

[0052] Optionally, when the preset semantic reasoning rule library is constructed in advance, the knowledge base management device not only includes simple rule matching based on the semantic features of the words, the syntax structure and the semantic role information, but also introduces reasoning rules based on probability statistics. For example, by statistically analyzing a large amount of historical missed question data, the probability of belonging to a certain semantic category under a certain combination of semantic features of the words, syntax structure and semantic role is calculated. For example, in the historical data, when the syntax structure of product name + cannot + function description appears, and the agent is the user and the patient is the product, there is an 80% probability of belonging to the product fault feedback semantic category. When performing semantic reasoning, in combination with the probability information, when encountering missed question data that meets the mode, not only rule matching is performed, but also the possibility of belonging to the product fault feedback semantic category is further judged according to the probability, so as to improve the accuracy of semantic reasoning.

[0053] In step 204, the semantic features of the words and the semantic categories are used to cluster the missed question data, and an unresolved problem type set is obtained.

[0054] Optionally, the knowledge base management device analyzes and clusters the missed question data according to the association relationship between the semantic features of the words and the semantic categories based on the obtained semantic features of the words and the semantic categories, so as to obtain the unresolved problem type set, which is specifically described in steps 2041-2045.

[0055] The embodiment of the application analyzes the text semantics from multiple levels by multiple steps such as word semantic feature extraction, syntax structure analysis, semantic role labeling and semantic reasoning, avoids the one-sidedness of simple text processing, makes the semantic understanding more accurate and comprehensive, and provides a reliable semantic basis for knowledge base management.

[0056] In an embodiment, steps 2041-2044 are described as follows:

[0057] In step 2041, a semantic hierarchy tree is constructed based on the association relationship between the semantic categories and the semantic features of the words.

[0058] Optionally, the knowledge base management device sorts the obtained semantic categories and lexical semantic features, first determines the top-level general semantic categories, such as product-related problems, and then gradually subdivides according to the inclusion and subordination relationship between the semantic categories, such as electronic product problems and household product problems under the product-related problems. For electronic product problems, it can be further subdivided into mobile phone problems and computer problems. In terms of lexical semantic features, if the lithium battery belongs to the battery category, and the battery is related to the mobile phone problem (such as mobile phone battery related problems), a semantic hierarchy tree is constructed according to this hierarchical relationship. And use graph data structure to store the semantic hierarchy tree, each node represents a semantic category or a set of words with specific semantics, and the edge between the nodes represents the semantic inclusion or subordination relationship.

[0059] Further, in an embodiment, taking the e-commerce customer service scenario as an example, the semantic categories include clothing problems, digital product problems, etc. The digital product problems have camera problems and mobile phone problems as subcategories. In the lexical semantic features, the lens belongs to the camera category, and the lithium battery belongs to the mobile phone battery category. In the constructed semantic hierarchy tree, the digital product problem is the upper node, the camera problem and the mobile phone problem are the lower sub-nodes, and the lexical semantic features such as lens and lithium battery are respectively attributed to the corresponding lower nodes.

[0060] Step 2042, locate the unresolved problem in the semantic hierarchy tree to obtain the unresolved problem node.

[0061] Optionally, for each unresolved problem in the hit problem data, the knowledge base management device locates it in the semantic hierarchy tree according to its semantic analysis results (including semantic categories and lexical semantic features). First, determine the semantic category to which the unresolved problem belongs, find the corresponding top-level or middle-level category node in the semantic hierarchy tree. Then, according to the specific lexical semantic features involved in the unresolved problem, further find more accurate sub-nodes under the category node. For example, the unresolved problem is how to deal with the short battery life of the mobile phone lithium battery, its semantic category belongs to the mobile phone problem, and the lexical semantic feature points to the lithium battery. First, locate the mobile phone problem node in the semantic hierarchy tree, and then further locate the lithium battery problem node under the mobile phone battery problem, which is the unresolved problem node.

[0062] Step 2043, perform path search on the unresolved problem node in the semantic hierarchy tree to obtain the unresolved problem node path, and perform depth processing on any two unresolved problem nodes based on a preset hierarchical correlation function to obtain the hierarchical correlation degree.

[0063] Optionally, the knowledge base management device performs a path search from the root node for each unresolved problem node and records all the nodes passed, forming an unresolved problem node path. The process of depth processing of any two unresolved problem nodes according to a preset hierarchical correlation function is as follows: let the unresolved problem node path of problem be , and the unresolved problem node path of problem be . The preset hierarchical correlation function is , where represents the maximum depth of the same node on the paths from the root node, and are the maximum depths of the same node on the paths of problem and problem , respectively, and and are the total lengths of the paths of problem and problem . For example, the path of problem is electronic product problem-mobile phone problem-mobile phone battery problem-lithium battery problem, and the depth is 4; the path of problem is electronic product problem-mobile phone problem-mobile phone charging problem, and the depth is 3. Problems and problem are the same before the mobile phone problem node, and the depth of the mobile phone problem node is 2, so , , , that is, the hierarchical correlation degree between problem A and problem B is

[0064] . The hierarchical correlation degree quantifies the correlation degree of unresolved problems in the semantic hierarchy, considers the hierarchical relationship of semantics, and is more in line with semantic logic than simple distance calculation, thereby providing a reasonable similarity measurement standard for subsequent clustering and helping to accurately aggregate semantically related problems.

[0065] In step 2044, each unresolved problem in the hit problem data is taken as an initial cluster, and the initial clusters are merged based on the hierarchical correlation degree, to obtain an unresolved problem type set. Optionally, the knowledge base management device initially takes each unresolved problem in the hit problem data as a separate cluster (that is, an initial cluster). Then, for all clusters , the average hierarchical correlation degree between them is calculated, where is the hierarchical correlation degree of problem and problem , and and are the clusters of problem

[0065] and problem , respectively.The number of problems is determined. And the two largest clusters are merged. This process is repeated until a stop condition is met, such as the number of clusters reaching a preset value , or the average hierarchical correlation of the clusters no longer significantly increases. The final cluster is the set of unresolved problem types.

[0066] Further, in an embodiment, assuming there are three initial clusters , containing problems ; ; The average hierarchical correlation between and is calculated , and is calculated , and is calculated . If is the largest, then and are merged into a new cluster, and the iteration of calculation and merging continues until the stop condition is met.

[0067] The embodiment of the application starts by constructing a semantic hierarchy tree, systematically combs the hierarchical relationship of semantic categories and lexical semantic features, and positions, analyzes and clusters the missed problem data based on the semantic hierarchy, comprehensively and deeply understands the problem semantics, and avoids fragmented processing. Through hierarchical correlation calculation and clustering based on this, the operation is closely around the semantic hierarchy structure, which conforms to the internal logic of semantics, compared with the traditional clustering method, can more accurately reflect the semantic correlation between problems, and the clustering result is more logical and reasonable.

[0068] In an embodiment, steps 3011-3015 are described as follows:

[0069] Step 3011, attribute extraction processing is performed on the unresolved problems in the set of unresolved problem types, and a set of problem attributes is obtained.

[0070] Optionally, the knowledge base management apparatus determines the key parts of speech such as nouns and adjectives for each unresolved problem in the problem type set by a part-of-speech tagging tool, and then identifies the entities and attribute-related words involved in the problem by a named entity recognition model. For example, for the question: which brand of notebook computer has good heat dissipation, the key words such as brand, notebook computer, and heat dissipation are determined by part-of-speech tagging, and the attributes of notebook computer (entity) such as brand and heat dissipation performance are extracted by the named entity recognition model. The extracted attributes are summarized to form a problem attribute set. By accurately extracting the problem attributes, the subsequent knowledge association and graph construction around the problem attributes are laid a foundation, the direction of the problem and the knowledge base knowledge association is clarified, and the knowledge association is more targeted.

[0071] Step 3012, classifying the problem attribute set to obtain problem attribute categories, and constructing a hierarchical structure among the problem attributes based on the hierarchical association among the problem attributes in the problem attribute categories.

[0072] Optionally, the knowledge base management apparatus classifies the extracted problem attributes into basic attributes, key attributes, and derived attributes according to the pre-set attribute classification standard. The basic attributes are usually inherent attributes directly related to the problem object, such as brand and model; the key attributes are attributes that play a key role in solving the problem, such as heat dissipation performance for notebook computer heat dissipation problem; and the derived attributes are more detailed attributes associated with the key attributes, such as heat dissipation design and heat dissipation material, which are derived attributes of heat dissipation performance. Then, according to the semantic association and hierarchical relationship among the attributes, a tree data structure is used to construct a hierarchical structure, with the root node being the problem object, the child nodes being attributes of different categories, and the hierarchical relationship among the attributes.

[0073] Step 3013, based on the problem attribute categories and the hierarchical structure among the problem attributes, matching with the comprehensive knowledge in the customer service robot knowledge base and the artificial customer service answer data to obtain problem attribute knowledge entries.

[0074] Optionally, the knowledge base management device traverses each attribute in the problem attribute category and hierarchy, and searches for matching in the customer service robot knowledge base and artificial customer solution data. Keyword matching, semantic matching, etc. can be used, wherein keyword matching is used to directly find knowledge entries containing attribute keywords; semantic matching is used to match the semantics of the attributes and the semantics of the knowledge entries in the knowledge base by using semantic similarity calculation (such as cosine similarity). For example, for the notebook computer heat dissipation performance attribute, in the computer related knowledge base, the knowledge entry containing the heat dissipation performance is found by keyword matching, and the knowledge entry with similar semantics (such as heat dissipation effect) to the heat dissipation performance is found by using semantic matching. The screened knowledge entry is the problem attribute knowledge entry. Through the matching of the attributes and the knowledge base and the artificial customer service solution data, the knowledge entries related to the problem can be accurately screened, and the knowledge in the knowledge base is associated according to the problem attributes.

[0075] Step 3014, analyzing the relationship between the problem attributes in the problem attribute knowledge entry, and constructing a problem attribute relationship model.

[0076] Optionally, the knowledge base management device performs in-depth analysis according to the obtained problem attribute knowledge entry, uses text mining and relationship extraction technology (such as a relationship extraction model based on deep learning), and identifies the logical relationships such as causality, association, and influence between the attributes. For example, in the knowledge entry about the heat dissipation of a notebook computer, it is found that the use of copper pipe heat dissipation design can improve the heat dissipation performance, and a positive influence relationship between the heat dissipation design and the heat dissipation performance is established. These identified attribute relationships are integrated to construct a problem attribute relationship model, which is represented by a graph structure, the nodes are problem attributes, and the edges are the relationships between the attributes. The weight of the edge can be set according to the strength of the relationship (such as the certainty degree of the causal relationship).

[0077] Step 3015, based on the problem attribute relationship model, taking the problem attributes as nodes and the relationships between the problem attributes as edges, a knowledge association graph is constructed.

[0078] Optionally, the knowledge base management device uses a problem attribute relationship model to construct a knowledge correlation graph using a graph database (such as Neo4j) or a graph computing framework (such as GraphX). The problem attributes are taken as nodes in the graph, and each node contains attribute names, attribute descriptions, and the like. The relationships between the attributes are taken as edges, and the types (such as cause and effect, correlation) and weights of the edges are set according to the problem attribute relationship model. At the same time, according to the relationships between the attributes, related knowledge is further expanded from the knowledge base to enrich the content of the graph. For example, the node of the heat dissipation material is expanded to the production manufacturers, characteristics, and the like of the material, and corresponding nodes and edges are added in the graph. By constructing the knowledge correlation graph with the problem attributes as the core, the correlation relationships between the related knowledge of the problem are intuitively presented, and the graph has strong pertinence and practicality. The graph is constantly enriched in details through attribute expansion, so that the graph can better support the customer service robot to perform knowledge retrieval and problem solving based on the graph, and improve the quality and efficiency of customer service.

[0079] The embodiment of the present application starts from problem attribute extraction, gradually constructs a knowledge correlation graph, and performs knowledge matching and correlation around the attributes of the problem itself, so that the knowledge correlation graph can closely match the problem requirements, accurately correlate the related knowledge in the knowledge base, avoid irrelevant knowledge interference, and clearly show the logical derivation relationships between the knowledge through attribute classification, hierarchical structure construction, and attribute relationship model construction, so that the knowledge correlation graph is not only a knowledge set, but also a knowledge network with logical connotations, facilitating user understanding and application of the knowledge.

[0080] In an embodiment, steps 3021-3025 are described as follows:

[0081] Step 3021: The semantic structure of the unresolved problem in the unresolved problem type set is decomposed to obtain a plurality of semantic units.

[0082] Optionally, the knowledge base management device uses a syntax analysis and semantic analysis tool (such as the Stanford Parser, semantic role labeling tool) in natural language processing technology to analyze each unresolved problem in the unresolved problem type set. First, the syntax structure of the sentence is determined through syntax analysis, such as subject-predicate-object, state complement, and the like. Then, the sentence is split into a plurality of semantic units based on semantic analysis. For example, for the unresolved problem: how to optimize the voice wake-up function of the smart speaker, the syntax analysis determines that how to optimize is an adverbial, the smart speaker is the subject, and the voice wake-up function is the object. The semantic analysis further decomposes the semantic units of the smart speaker, the voice wake-up function, and the optimization method. The unresolved problem is refined into semantic units to provide a basis for subsequent accurate matching in the knowledge correlation graph.

[0083] Step 3022: The semantic units are mapped to nodes in the knowledge correlation graph to obtain semantic unit mapping nodes.

[0084] Optionally, the knowledge base management device traverses the nodes in the knowledge association graph, and compares each semantic unit with the graph nodes using a semantic similarity calculation method (such as cosine similarity, edit distance, etc.). Based on the text content and semantic features of the semantic unit, the matching nodes are found. If there is a node in the knowledge association graph that is similar or identical in semantics to the semantic unit, the node is taken as the mapping node of the semantic unit. For example, for the semantic unit of the smart speaker, the node containing the keyword of the smart speaker or having a related semantic description is found in the knowledge association graph. If the smart speaker product category node is found, the node is the mapping node of the smart speaker semantic unit.

[0085] Step 3023, starting from the semantic unit mapping node, the association path is expanded in the knowledge association graph to obtain an initial association path, and the initial association path is supplemented based on the customer service robot knowledge base and external knowledge sources to obtain a target association path.

[0086] Optionally, the knowledge base management device starts from the semantic unit mapping node and searches the path in the knowledge association graph according to the connection relationship between the nodes (such as attribute association, causal association, etc.). A depth-first search or breadth-first search algorithm can be used to explore other nodes associated with the mapping node to form an initial association path. For example, starting from the smart speaker mapping node, the related nodes such as the hardware configuration of the smart speaker and the software algorithm of the smart speaker are found through the association relationship to form an initial association path. Then, the customer service robot knowledge base and external knowledge sources (such as industry authoritative websites, professional forums, etc.) are queried to obtain supplementary knowledge related to the initial association path, and the initial association path is perfected to obtain a target association path. For example, knowledge about the optimization of the microphone array of the smart speaker is obtained from a professional audio technology website and added to the association path related to the voice wake-up function.

[0087] Step 3024, for each unresolved problem in the set of unresolved problems, the target association path is matched and screened based on the unresolved problem to obtain a knowledge item.

[0088] Optionally, for each unsolved problem, the knowledge base management device filters closely related knowledge entries along the target association path based on its semantic features and problem requirements. This filtering can be done by calculating the semantic similarity and relevance scores of knowledge entries along the target association path. For example, for the unsolved problem: "How to optimize the voice wake-up function of a smart speaker," the semantic similarity of each knowledge entry along the target association path is calculated, and a corresponding semantic similarity threshold is set. If the semantic similarity between "smart speaker microphone array optimization method" and this problem is greater than or equal to the semantic similarity threshold, while the semantic similarity between "smart speaker appearance design features" and the problem is less than the semantic similarity threshold, then knowledge entries with a semantic similarity greater than or equal to the threshold are filtered as knowledge entries related to this unsolved problem. It should be noted that each unsolved problem can have multiple corresponding knowledge entries.

[0089] Step 3025: Merge each unsolved problem and its corresponding knowledge entry in the unsolved problem type set to obtain candidate knowledge entries for the problem.

[0090] Optionally, the knowledge base management device integrates each unsolved problem with its selected knowledge entries to form candidate knowledge entries. This correspondence can be stored using data structures (such as dictionaries, lists, etc.), with each candidate knowledge entry containing the text content of the unsolved problem and a set of matching knowledge entries. For example, for the unsolved problem: "How to optimize the voice wake-up function of a smart speaker" and its selected knowledge entries such as "Smart speaker microphone array optimization methods" and "Smart speaker voice wake-up algorithm improvement strategies," these can be integrated into a single candidate knowledge entry, recorded as: {"How to optimize the voice wake-up function of a smart speaker: [Smart speaker microphone array optimization methods, Smart speaker voice wake-up algorithm improvement strategies]}.

[0091] This invention starts with semantic structure decomposition, and gradually matches the semantics and attributes of the problem with the knowledge graph through operations such as mapping, association path expansion, and filtering. This can accurately locate knowledge related to the unresolved problem, improve the accuracy of knowledge matching, and avoid interference from irrelevant knowledge. Furthermore, by expanding the association path and supplementing it with internal and external knowledge sources, it can mine potential knowledge in the knowledge graph and external related knowledge, making the acquired knowledge items more comprehensive and providing rich knowledge support for customer service robots to answer complex questions.

[0092] In one embodiment, steps 401-405 are described as follows:

[0093] Step 401: Search and match the paths related to knowledge items in the knowledge association graph to obtain knowledge item path information.

[0094] Optionally, the knowledge base management device adopts the query function of the graph database (such as the Cypher query language of Neo4j) to search the path from other nodes in the knowledge association graph to the target node of the knowledge item. For example, to check whether the printer cartridge is out of ink, the path from the printer unable to print node to the knowledge item node is searched in the knowledge association graph, and the path information of printer unable to print-printer hardware check-cartridge check-check whether the printer cartridge is out of ink is obtained, and the information of the nodes and edges on the path is recorded.

[0095] In step 402, the information related to the knowledge item in the artificial customer service answer data is searched and matched to obtain the context information of the knowledge item.

[0096] Optionally, the knowledge base management device adopts information retrieval technology (such as inverted index, keyword matching, etc.) to search the content related to the knowledge item in the artificial customer service answer data. The artificial customer service answer data is preprocessed, such as word segmentation, stop word removal, etc. Then, the data is searched and matched according to the keywords of the knowledge item. For example, to check whether the printer cartridge is out of ink, the dialogue records containing the keywords of printer cartridge and ink shortage are searched in the artificial customer service answer data to obtain the context information such as the description of printer failure, the previously given solution and user feedback, etc. The context information of the knowledge item obtained from the artificial customer service answer data can reflect the application scenario and user feedback of the knowledge item in the actual customer service dialogue.

[0097] In step 403, the knowledge logic of the knowledge item path information is analyzed, and the logic analysis result is matched with the context information of the knowledge item based on the fitting matching function to obtain the first fitting degree.

[0098] Optionally, the knowledge base management device analyzes the knowledge item path information to sort out the knowledge logic relationship therein, such as cause-effect relationship, step relationship, etc. For example, for the path of printer unable to print-printer hardware check-cartridge check-check whether the printer cartridge is out of ink, the logic is analyzed as deducing from the printer failure phenomenon to the possible hardware cause (cartridge problem). Then, according to the pre-constructed fitting matching function (such as the function based on semantic similarity), the semantic representation of the knowledge item path information is , the semantic representation of the context information of the knowledge item is , and the fitting matching function is , and the first fitting degree is obtained by calculating the semantic similarity between the two, etc. For example, the cosine similarity formula is used The cosine value of the two semantic vectors is calculated as the first fit degree score. Through knowledge logic analysis and fit matching, it can be preliminarily judged whether the logic of the knowledge item in the knowledge correlation graph and the application in the actual customer service dialogue are consistent.

[0099] In step 404, the user dialogue data related to the knowledge item in the user session log data is analyzed for emotion, and the first fit degree is initially optimized based on the emotion analysis to obtain the second fit degree.

[0100] Optionally, the knowledge base management device uses sentiment analysis technology (such as a sentiment analysis model based on deep learning, such as BERT-for-Sentiment-Analysis) to analyze the user dialogue content related to the knowledge item in the user session log data. The user dialogue data is input into the sentiment analysis model, and the emotion tendency score of the user is output, which can be set in the range [-1, 1], -1 represents extreme negativity, 1 represents extreme positivity, and 0 represents neutrality. For example, if the user dialogue mentions: After operating according to this method, the printer still cannot print, which is too disappointing, the emotion tendency score obtained through sentiment analysis is -0.6. According to the emotion tendency score, the first fit degree is adjusted, if the emotion tendency is positive, the first fit degree score is increased, if the emotion tendency is negative, the score is decreased, and the second fit degree is obtained. For example, the first fit degree is 0.6, and the emotion tendency score is -0.6. The second fit degree can be obtained by adjusting the formula (such as second fit degree = first fit degree + emotion tendency score × adjustment coefficient, the adjustment coefficient can be set to 0.2 according to experience). The second fit degree is 0.6 + (-0.6) × 0.2 = 0.48. By considering the user emotion factor to optimize the fit degree, the effectiveness of the knowledge item in the actual experience of the user can be more comprehensively reflected, avoiding the one-sidedness of the evaluation from the knowledge logic and the context information, and making the evaluation result more close to the actual application scene.

[0101] In step 405, the historical update feedback data related to the knowledge item in the customer service robot knowledge base is analyzed for effectiveness, and the second fit degree is secondarily optimized based on the effectiveness analysis result to obtain the target fit degree. Based on the target fit degree, the confidence level corresponding to each knowledge item is determined.

[0102] Optionally, the knowledge base management apparatus queries historical update feedback data related to the knowledge item in the customer service robot knowledge base, such as the success rate of problem solving after updating the knowledge item, the frequency of users asking related questions again, and the like. The data is analyzed to evaluate the historical update effect of the knowledge item. For example, the success rate of problem solving before updating a certain knowledge item is 30%, and the success rate is increased to 50% after updating, which indicates that the update effect is good. According to the historical update effect, the second fit degree is secondarily optimized. If the update effect is good, the second fit degree is increased by a formula (such as target fit degree=second fit degree x (1+improvement coefficient), wherein the improvement coefficient is set according to the effect improvement amplitude, and herein the improvement coefficient is set to 0.1); if the update effect is not good, the second fit degree is reduced to obtain the final target fit degree. According to the pre-set confidence level division standard (such as target fit degree greater than 0.8 is high confidence level (i.e. first confidence level), 0.5-0.8 is medium confidence level (i.e. second confidence level), and less than 0.5 is low confidence level (i.e. third confidence level)), the confidence level of the knowledge item is determined. By secondarily optimizing the fit degree by combining the historical knowledge update effect, the evaluation is further improved from the long-term application and optimization effect of the knowledge item, the confidence level is determined by comprehensively considering multiple factors, the evaluation result of the knowledge item is more accurate and comprehensive, and the quality evaluation level of the knowledge item in the knowledge base is improved.

[0103] The embodiment of the application evaluates the knowledge item by comprehensively considering the knowledge correlation graph path information, the artificial customer service answer context information, the user emotion, and the historical update feedback, avoids the limitation of single information source evaluation, and comprehensively covers the performance of the knowledge item in the knowledge system, the actual application, the user experience, and the long-term effect. By considering the user emotion and the historical update effect, the evaluation result is closer to the real use of the knowledge item, and the effectiveness and reliability of the knowledge item in the actual application can be more accurately reflected.

[0104] In an embodiment, steps 501-503 are described as follows:

[0105] In step 501, if the confidence level corresponding to each knowledge item is the first confidence level, the knowledge item is directly updated to the customer service robot knowledge base.

[0106] Optionally, the knowledge base management device detects that each knowledge item has a confidence level of the first confidence level (high confidence level, target fit degree greater than or equal to 0.8) after step 405, and automatically triggers the updating operation. The knowledge item is added to the corresponding classification directory in the customer service robot knowledge base, and the index of the knowledge base is updated for subsequent quick retrieval. For example, in an e-commerce customer service knowledge base, the knowledge item of a certain brand of sports shoes adopts a breathable mesh material, and the confidence level of wearing comfort is 0.85, which belongs to the first confidence level. The system automatically updates it to the knowledge base under the product introduction classification of sports shoes. By directly updating the high-confidence-level knowledge item, the knowledge base updating efficiency is improved, and reliable knowledge that has been strictly evaluated can be quickly supplemented to the knowledge base, so that the customer service robot can quickly answer related questions and meet the user's demand for quickly obtaining accurate knowledge.

[0107] Step 502, if the confidence level of each knowledge item is the second confidence level, the knowledge item is updated to the customer service robot knowledge base after being audited and confirmed by a human.

[0108] Optionally, the knowledge base management device detects that each knowledge item has a confidence level of the second confidence level (medium confidence level, target fit degree between 0.5 and 0.8) after step 405, and automatically pushes the knowledge item to the manual review interface. The human auditors (such as customer service managers, field experts, etc.) audit the accuracy, completeness, applicability, etc. of the knowledge item. During the auditing process, the human auditors will refer to the logic of the knowledge item in the knowledge correlation graph, the application in the user session log, historical update feedback, etc. Information to determine whether the knowledge item is accurate and reasonable. If the audit is passed, the human auditors confirm and update the knowledge item to the customer service robot knowledge base; if not, go to step 503.

[0109] Step 503, if the confidence level of each knowledge item is the third confidence level, the feature information of the knowledge item is extracted, data mining is performed in multiple data sources according to the extracted feature information, and the relevance of the mined data is evaluated. According to the relevance evaluation result and the preset standard threshold, the data that meets the requirements is screened out, then the screened multi-source data is fused, and the knowledge item is updated and improved according to the fused data. The updated and improved knowledge item is subjected to secondary confidence evaluation.

[0110] Optionally, the knowledge base management device detects that each knowledge item has a third confidence level (medium confidence level, target matching degree less than 0.5) after step 405, and receives the audit failure knowledge item of step 502, and then uses a text analysis technique to analyze the knowledge item and extract characteristic information such as core concepts, key technologies, and main points. For example, for the knowledge item that a new type of environmental protection material can be used for building outer wall, the characteristic information such as new type of environmental protection material and building outer wall is extracted.

[0111] Then, according to the extracted characteristic information, data mining is performed in multiple data sources (such as professional academic databases, industry reports, authoritative websites, etc.). Using web crawler technology, database query interface and other tools, relevant data is obtained from different data sources. For example, for the characteristic information of the new type of environmental protection material, relevant research papers are searched in the academic database, performance indicators and application cases of the material are found in the industry report, and market dynamics and the latest information of the material are collected on the authoritative website.

[0112] Suppose the amount of data related to the knowledge item mined from the pth data source is , the total number of data sources is s, and the mined data is evaluated for relevance to determine the degree of relevance of the data to the knowledge item. The formula R is used, where represents the number of elements in the data mined from the pth data source that are associated with the knowledge item; represents the total number of elements in the data mined from the pth data source; represents a weight coefficient set according to the reliability of the data source, such as an academic database with a higher weight of 0.8, an industry report with a weight of 0.6, and an authoritative website with a weight of 0.4. The relevance score of each data source is calculated by the formula to filter out data with higher relevance.

[0113] The filtered multi-source data is fused to remove duplicate information and integrate complementary information to form supporting evidence for the knowledge item. The knowledge item is updated and improved based on the fused data, and the confidence level is evaluated again, i.e. the updated and improved knowledge item is evaluated for the second time. For low confidence level knowledge items, the knowledge item is improved by making full use of multi-channel information through a series of operations of multi-source data mining and fusion, and the accuracy and reliability of the knowledge item are improved. The quality of the final knowledge in the database is ensured through the second confidence evaluation. At the same time, the knowledge source channel of the knowledge base is widened, the knowledge content is enriched, and the practicality and credibility of the knowledge base as a whole are further improved.

[0114] The embodiment of the application adopts different processing modes according to different confidence levels of knowledge items, and designs direct updating, manual review, and multi-source data mining and improvement processes for high, medium and low confidence levels, so that the processing is more targeted, the high confidence level knowledge is quickly stored in the database, the medium and low confidence level knowledge is reasonably optimized, and the overall efficiency and quality of the knowledge base updating are improved.

[0115] Further, the management device of the customer service robot knowledge base provided by the application is described below, and the management device of the customer service robot knowledge base described below can be correspondingly referred to the management method of the customer service robot knowledge base described above.

[0116] Optionally, referring to Figure 2 , Figure 2 is a structural schematic diagram of the management device of the customer service robot knowledge base provided by the application, and the management device of the customer service robot knowledge base comprises.

[0117] The multi-source data acquisition module 210 is configured to acquire multi-source data, and the multi-source data includes user session log data, missed question data and artificial customer service answer data.

[0118] The semantic clustering analysis 220 is configured to perform question semantic analysis on the missed question data, and perform clustering processing on the missed question data based on the analysis result, to obtain an unresolved question type set.

[0119] The knowledge association graph construction module 230 is configured to construct a knowledge association graph based on the unresolved question type set, the artificial customer service answer data and the customer service robot knowledge base, and match the knowledge association graph with the unresolved question type set, to obtain a question candidate knowledge item.

[0120] The confidence review module 240 is configured to evaluate each knowledge item in the question candidate knowledge item based on the user session log data, the artificial customer service answer data and the customer service robot knowledge base, to obtain a confidence level of the knowledge item.

[0121] The update management module 250 is configured to update each knowledge item to the customer service robot knowledge base based on the confidence level of each knowledge item.

[0122] The embodiment of the present application can quickly and accurately find out the unsolved problem type set through semantic analysis and clustering of automatically acquired multi-source data, and can further understand the nature of the user's problem, automatically mine more accurate knowledge association, reduce human and material resources, and provide accurate data for subsequent updating and optimization, finally, according to the artificial customer service answer data, the user session log data and the knowledge base content, the problem candidate knowledge item association evaluation can be carried out from multiple dimensions to ensure that the knowledge added to the knowledge base is accurate and reliable, and the knowledge base can be updated and optimized automatically according to the evaluation confidence level without a large amount of manual intervention, solving the problem of knowledge base optimization delay under artificial interference, improving the knowledge base updating efficiency, and responding to business changes and user needs in a timely manner.

[0123] Please refer to Figure 3 , Figure 3 The embodiment of the electronic device provided by the embodiment of the present application is shown in the figure. As shown in Figure 3 , the present embodiment provides an electronic device 300, which comprises a memory 310, a processor 320 and a computer program 311 stored in the memory 310 and executable on the processor 320, and the processor 320 executes the computer program 311 to realize the following steps:

[0124] Obtain multi-source data; the multi-source data comprises user session log data, missed problem data and artificial customer service answer data;

[0125] Perform problem semantic analysis on the missed problem data, and perform clustering processing on the missed problem data based on the analysis result to obtain an unsolved problem type set;

[0126] Based on the unsolved problem type set, the artificial customer service answer data and the customer service robot knowledge base, a knowledge association graph is constructed, and matching is performed based on the knowledge association graph and the unsolved problem type set to obtain problem candidate knowledge items;

[0127] For each knowledge item in the problem candidate knowledge item, each knowledge item is evaluated based on the user session log data, the artificial customer service answer data and the customer service robot knowledge base to obtain a confidence level of the knowledge item;

[0128] Each knowledge item is updated to the customer service robot knowledge base based on the confidence level of each knowledge item.

[0129] Please refer to Figure 4 , Figure 4 The embodiment of the computer readable storage medium provided by the embodiment of the present application is shown in the figure. As shown in Figure 4 , the present embodiment provides a computer readable storage medium 400, which stores a computer program 311, and the computer program 311 is executed by a processor to realize the following steps:

[0130] obtaining multi-source data; the multi-source data comprises user session log data, missed question data and artificial customer service answer data;

[0131] performing question semantic analysis on the missed question data, and performing clustering processing on the missed question data based on an analysis result, to obtain an unresolved question type set;

[0132] constructing a knowledge association graph based on the unresolved question type set, the artificial customer service answer data and a customer service robot knowledge base, and performing matching based on the knowledge association graph and the unresolved question type set, to obtain question candidate knowledge items;

[0133] performing evaluation on each knowledge item in the question candidate knowledge items based on the user session log data, the artificial customer service answer data and the customer service robot knowledge base, to obtain a confidence level of the knowledge item;

[0134] updating each knowledge item to the customer service robot knowledge base based on the confidence level of the knowledge item.

[0135] In another aspect, the present application also provides a computer program product, which comprises a computer program, the computer program being stored on a non-transitory computer readable storage medium, and the computer program being executable by a processor to enable a computer to perform the management method of the customer service robot knowledge base provided by the above-mentioned methods, the method comprising:

[0136] obtaining multi-source data; the multi-source data comprises user session log data, missed question data and artificial customer service answer data;

[0137] performing question semantic analysis on the missed question data, and performing clustering processing on the missed question data based on an analysis result, to obtain an unresolved question type set;

[0138] constructing a knowledge association graph based on the unresolved question type set, the artificial customer service answer data and a customer service robot knowledge base, and performing matching based on the knowledge association graph and the unresolved question type set, to obtain question candidate knowledge items;

[0139] performing evaluation on each knowledge item in the question candidate knowledge items based on the user session log data, the artificial customer service answer data and the customer service robot knowledge base, to obtain a confidence level of the knowledge item;

[0140] updating each knowledge item to the customer service robot knowledge base based on the confidence level of the knowledge item.

[0141] The device embodiments described above are merely illustrative, wherein the units illustrated as separate components can or can not be physically separated, and the components illustrated as units can or can not be physical units, i.e., can be located in one place, or can be distributed to multiple network units. Part or all of the modules can be selected to achieve the purposes of the embodiments according to actual needs. Those skilled in the art can understand and implement without creative labor.

[0142] Through the description of the above embodiments, those skilled in the art can clearly understand that the embodiments can be realized by means of software and the necessary general hardware platform, and of course can also be realized by hardware. Based on such understanding, the above technical solutions can be embodied in the form of software products, and the computer software products can be stored in a computer readable storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and include a plurality of instructions for making a computer device (which can be a personal computer, a server, or a network device, etc.) execute the methods of the embodiments or some parts of the embodiments.

[0143] Finally, it should be noted that: the above embodiments are only used to illustrate the technical solutions of the present application, and not to limit them; although the present application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that: it can still modify the technical solutions recorded in the foregoing embodiments, or make equivalent replacement for part of the technical features; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the spirit and scope of the technical solutions of the embodiments of the present application.

Claims

1. A method for managing a customer service robot knowledge base, characterized in that, include: Acquire multi-source data; the multi-source data includes user session log data, missed question data, and human customer service answer data. The missing question data is subjected to semantic analysis, and the missing question data is clustered based on the analysis results to obtain a set of unresolved question types; Based on the set of unresolved problem types, the data of human customer service answers, and the knowledge base of customer service robots, a knowledge association graph is constructed, and the knowledge association graph is matched with the set of unresolved problem types to obtain candidate knowledge entries for problems. For each knowledge item in the candidate knowledge items of the question, the confidence level of the knowledge item is obtained by evaluating each knowledge item based on the user session log data, the human customer service answer data and the customer service robot knowledge base. Each knowledge entry is updated to the customer service robot knowledge base based on its confidence level. The construction of a knowledge association graph based on the set of unresolved problem types, the human customer service answer data, and the customer service robot knowledge base includes: The unresolved problems in the unresolved problem type set are subjected to attribute extraction processing to obtain the problem attribute set; The problem attribute set is classified to obtain problem attribute categories, and a hierarchical structure between problem attributes is constructed based on the hierarchical relationship between problem attributes in the problem attribute categories. Based on the hierarchical structure between the question attribute categories and the question attributes, the knowledge entries of the question attributes are obtained by matching them with the comprehensive knowledge in the customer service robot knowledge base and the human customer service answer data. Analyze the relationships between problem attributes in the problem attribute knowledge entries and construct a problem attribute relationship model; Based on the aforementioned problem attribute relationship model, the knowledge association graph is constructed using problem attributes as nodes and the relationships between problem attributes as edges.

2. The management method for the customer service robot knowledge base according to claim 1, characterized in that, The process involves performing semantic analysis on the missed question data and clustering the missed question data based on the analysis results to obtain a set of unresolved question types, including: The missing problem data is segmented to obtain multiple word data, and the word data is feature extracted based on a preset semantic knowledge base to obtain word semantic features; Syntactic structure analysis is performed on the semantic features of the words to obtain syntactic structure information, and semantic role labeling is performed on the semantic features of the words based on the syntactic structure information to obtain semantic role information; Based on a preset semantic reasoning rule base, semantic reasoning is performed on the semantic features of the words, the syntactic structure information, and the semantic role information to determine the semantic category; The lexical semantic features and semantic categories are used to cluster the missing question data to obtain a set of unresolved question types.

3. The management method for the customer service robot knowledge base according to claim 2, characterized in that, The step of clustering the missed question data using the lexical semantic features and the semantic categories to obtain a set of unresolved question types includes: Based on the association between the semantic category and the semantic features of the words, a semantic hierarchy tree is constructed; The unresolved issues in the missed issue data are located by matching them with the nodes in the semantic hierarchy tree to obtain the unresolved issue nodes; The unresolved problem nodes are searched within the semantic hierarchy tree to obtain the unresolved problem node paths. Then, based on a preset hierarchical association function, any two unresolved problem nodes are subjected to in-depth processing to obtain the hierarchical association degree. Each unresolved issue in the missed issue data is used as an initial cluster. The initial clusters are then merged based on the hierarchical correlation to obtain the set of unresolved issue types.

4. The management method for the customer service robot knowledge base according to claim 1, characterized in that, The process of matching the knowledge association graph with the set of unresolved problem types to obtain candidate knowledge entries for problems includes: The unresolved problems in the set of unresolved problem types are semantically decomposed to obtain multiple semantic units; The semantic unit is mapped to a node in the knowledge association graph to obtain the semantic unit mapping node; Starting from the semantic unit mapping node, the association path is expanded in the knowledge association graph to obtain the initial association path, and the initial association path is supplemented based on the customer service robot knowledge base and external knowledge sources to obtain the target association path; For each unresolved problem in the set of unresolved problem types, a matching and filtering process is performed based on the target association path to obtain knowledge entries; By fusing each unsolved problem and its corresponding knowledge entry in the set of unsolved problem types, the candidate knowledge entries for the problem are obtained.

5. The management method for the customer service robot knowledge base according to claim 1, characterized in that, For each knowledge item in the candidate knowledge items for the problem, the confidence level is evaluated based on the user session log data, the human customer service answer data, and the customer service robot knowledge base to obtain the confidence level corresponding to each knowledge item, including: Search and match the paths related to the knowledge entries in the knowledge association graph to obtain the knowledge entry path information; Search and match the knowledge entry-related information in the human customer service response data to obtain the knowledge entry context information; The path information of the knowledge entries is subjected to knowledge logic analysis, and the results of the logic analysis are matched with the context information of the knowledge entries based on the matching function to obtain the first degree of matching. Sentiment analysis is performed on user dialogue data related to the knowledge item in the user conversation log data, and the first fit is initially optimized based on the sentiment analysis to obtain the second fit. The historical update feedback data related to the knowledge item in the customer service robot knowledge base is analyzed for effectiveness, and the second fit is optimized a second time based on the effect analysis results to obtain the target fit. The confidence level corresponding to each knowledge item is determined based on the target fit.

6. The management method for the customer service robot knowledge base according to claim 5, characterized in that, The confidence levels include a first confidence level, a second confidence level, and a third confidence level. Based on the confidence level corresponding to each knowledge entry, the knowledge entry is updated to the customer service robot knowledge base, including: If the confidence level corresponding to each knowledge entry is the first confidence level, then the knowledge entry is directly updated to the customer service robot knowledge base; If the confidence level corresponding to each knowledge entry is the second confidence level, then the knowledge entry will be updated to the customer service robot knowledge base after being manually reviewed and confirmed. If the confidence level corresponding to each knowledge entry is the third confidence level, then the feature information of the knowledge entry is extracted, data mining is performed on multiple data sources based on the extracted feature information, and the relevance of the mined data is evaluated. Based on the relevance evaluation results and the preset standard threshold, the data that meets the requirements is selected. Then, the selected multi-source data is fused, and the knowledge entries are updated and improved based on the fused data. A second confidence level evaluation is performed on the updated and improved knowledge entries.

7. A management device for a customer service robot knowledge base, characterized in that, A management method for a customer service robot knowledge base as described in any one of claims 1 to 6; the management device for the customer service robot knowledge base includes: The multi-source data acquisition module is used to acquire multi-source data, including user session log data, missed question data, and human customer service answer data. The semantic clustering analysis module is used to perform semantic analysis on the missing question data and cluster the missing question data based on the analysis results to obtain a set of unresolved question types. The knowledge association graph construction module is used to construct a knowledge association graph based on the unresolved problem type set, the human customer service answer data, and the customer service robot knowledge base, and to match the knowledge association graph with the unresolved problem type set to obtain candidate knowledge entries for the problem. The confidence assessment module is used to evaluate each knowledge item in the candidate knowledge items of the question based on the user session log data, the human customer service answer data and the customer service robot knowledge base, and obtain the confidence level of the knowledge item. The update management module is used to update each knowledge entry to the customer service robot knowledge base based on the confidence level of each knowledge entry; The construction of a knowledge association graph based on the set of unresolved problem types, the human customer service answer data, and the customer service robot knowledge base includes: The unresolved problems in the unresolved problem type set are subjected to attribute extraction processing to obtain the problem attribute set; The problem attribute set is classified to obtain problem attribute categories, and a hierarchical structure between problem attributes is constructed based on the hierarchical relationship between problem attributes in the problem attribute categories. Based on the hierarchical structure between the question attribute categories and the question attributes, the knowledge entries of the question attributes are obtained by matching them with the comprehensive knowledge in the customer service robot knowledge base and the human customer service answer data. Analyze the relationships between problem attributes in the problem attribute knowledge entries and construct a problem attribute relationship model; Based on the aforementioned problem attribute relationship model, the knowledge association graph is constructed using problem attributes as nodes and the relationships between problem attributes as edges.

8. An electronic device, comprising: Memory, used to store computer software programs; A processor for reading and executing the computer software program, characterized in that, when the processor executes the computer software program, it implements the management method of the customer service robot knowledge base as described in any one of claims 1 to 6.

9. A non-transitory computer-readable storage medium, wherein a computer software program is stored therein, characterized in that, When the computer software program is executed by the processor, it implements the management method of the customer service robot knowledge base as described in any one of claims 1 to 6.

Images