An AI-driven, full-scenario automated question-answering interaction system and method

Abstract

This invention discloses an AI-driven, full-scene automatic question-answering interaction system and method, relating to the field of automatic question-answering interaction technology. The system includes: using an AI model to segment original documents in a knowledge base, generating logical document fragments, and analyzing the relationships between different knowledge entities within these fragments; analyzing the contextual matching degree between the user's question text and historical question-answering records; analyzing the relationship between the amount of question information and the question structure in the target question text, determining the complexity granularity of the target question text, adaptively expanding the knowledge subgraph of the target question text to obtain the target knowledge subgraph, and combining this with the original documents in the knowledge base to generate the knowledge text of the target knowledge subgraph; generating the answer text to the user's question text based on the knowledge text, and conducting automatic question-answering interaction with the user through the platform. This fundamentally ensures that the answer information is appropriate and logically complete, greatly improving the user's actual question-answering experience.

Description

Technical Field

[0001] This invention relates to the field of automatic question-and-answer interaction technology, specifically an AI-driven, full-scenario automatic question-and-answer interaction system and method. Background Technology

[0002] Currently, there are numerous chain brands on the market. As these chain brands continue to expand their businesses, they experience a surge in customer inquiries. Traditional human customer service models require people to answer customer questions, which not only necessitates hiring a large number of customer service personnel and incurring high operating costs, but also requires new customer service staff to be familiar with a wide range of knowledge. Once customer service staff leave, retraining is required, making knowledge transfer difficult. Furthermore, different customer service staff have varying skill levels, making it impossible to maintain consistent quality and easily leading to different answers to the same question. Therefore, traditional human customer service models cannot meet the ever-expanding business needs of chain brands.

[0003] Currently, to address the problems associated with traditional human customer service models, many chain brands have introduced automated question-and-answer systems to assist customer service staff. However, these systems are based on simple keyword matching, making it difficult to understand users' natural language and lacking contextual understanding. Furthermore, they cannot accurately grasp the level of detail in knowledge points during the matching process, resulting in either redundant information or difficulty for users to quickly extract answers, thus failing to accurately resolve user issues. This not only wastes users' time and negatively impacts user experience but can also affect the operation of chain brands. Summary of the Invention

[0004] The purpose of this invention is to provide an AI-driven, full-scene automatic question-answering interaction system and method to solve the problems raised in the prior art.

[0005] To achieve the above objectives, the present invention provides the following technical solution: an AI-driven, full-scene automatic question-answering interaction method, the method comprising:

[0006] Step S1: Use an AI model to segment the original documents in the knowledge base, generate logical document fragments, and analyze the relationships between different knowledge entities in the logical document fragments to obtain a knowledge association graph;

[0007] Step S2: Obtain the user's question text and historical question-and-answer records, analyze the degree of contextual matching between the user's question text and historical question-and-answer records, and obtain the target question text;

[0008] Step S3: Based on the knowledge association graph, analyze the relationship between the amount of question information and the question structure in the target question text, determine the complexity granularity of the target question text, adaptively expand the knowledge subgraph of the target question text to obtain the target knowledge subgraph, and combine it with the original documents in the knowledge base to generate the knowledge text of the target knowledge subgraph;

[0009] Step S4: Based on the knowledge text, generate the answer text to the user's question text, and conduct automatic question-and-answer interaction with the user through the platform.

[0010] Furthermore, step S3 includes:

[0011] Obtain the user's target question text within the current period, extract knowledge entities from the target question text, and record them as candidate knowledge entities;

[0012] Obtain the entity word set of candidate knowledge entities from the platform, obtain the relevant historical question and answer records of candidate knowledge entities, and calculate the context key value L of candidate knowledge entities;

[0013] Obtain the knowledge association graph, and obtain the total number of other nodes connected to the node corresponding to the candidate knowledge entity through directed edges and undirected edges. Obtain the edge weights between the node corresponding to the candidate knowledge entity and the other connected nodes. Calculate the global key value S of the node corresponding to the candidate knowledge entity in the knowledge association graph.

[0014] Calculate the comprehensive score U of the nodes corresponding to the candidate knowledge entities, and obtain the maximum comprehensive score U of the nodes corresponding to each candidate knowledge entity in the target question text. max , will U max The node to which it belongs is denoted as the main node of the target question text;

[0015] Obtain the word set of the target question text, obtain the knowledge base from the platform, obtain the total number of original documents containing words from the word set from the knowledge base, calculate the question information content I of the target question text, calculate the structural breadth Y of the target question text, and calculate the complexity granularity H´ of the target question text;

[0016] A knowledge subgraph is constructed based on the target question text, where the knowledge subgraph is an empty set. The main nodes of the target question text are incorporated into the knowledge subgraph. Several nodes connected to the main nodes are randomly selected from the knowledge graph and incorporated into the knowledge subgraph. The average degree *b* of each node in the knowledge subgraph and the sum of the edge weights *W* of the edges connecting each node are calculated. sum The complexity granularity H of the computational knowledge subgraph △ ;

[0017] Obtain the complexity granularity of each knowledge subgraph of the target question text, and select several knowledge subgraphs with a complexity granularity greater than H´. From these knowledge subgraphs, select the knowledge subgraph with the smallest total number of nodes as the target knowledge subgraph of the target question text.

[0018] Retrieve logical document fragments from the original documents containing knowledge entities in the target knowledge subgraph from the knowledge base, and merge them to generate the knowledge text of the target knowledge subgraph;

[0019] The acquisition of knowledge text in the above steps starts from the actual questions asked by users, fundamentally solving the core pain points of "too coarse granularity (redundant answers)" and "too fine granularity (fragmented answers)" in traditional knowledge matching, realizing fine control over the granularity of knowledge, and significantly improving the quality of answers from automatic question answering systems for questions of different complexities.

[0020] Furthermore, step S2 includes:

[0021] Get the question text entered by the user from the platform, get the user's historical question and answer records on the platform, and get the historical question text from the historical question and answer records;

[0022] Retrieve similar historical question texts from the question text, and retrieve candidate historical question-and-answer records for the question text;

[0023] Obtain and aggregate all candidate historical question and answer records of the question text to obtain a candidate historical record set, extract the pronouns in the question text, and extract a candidate entity from the candidate historical question and answer records;

[0024] Retrieve candidate historical question and answer records containing a certain candidate entity from the candidate historical record set and mark them to obtain each marked candidate historical question and answer record of a certain candidate entity;

[0025] The sentence containing a certain pronoun is obtained from the question text, the sentence containing a certain candidate entity is obtained from the marked candidate historical question and answer records, the average semantic similarity between the sentence containing the certain pronoun and the sentences containing the certain candidate entity in each marked candidate historical question and answer record is calculated, and the semantic matching value E between the certain pronoun and the certain candidate entity is obtained.

[0026] Obtain the total number α of tagged candidate question-and-answer records containing a certain candidate entity and a certain pronoun from each tagged candidate question-and-answer record. Calculate the collinearity matching value G = α / α between a certain pronoun and a certain candidate entity. sum , where α sum The total number of candidate historical question and answer records in the candidate historical record set;

[0027] Calculate the context matching value ζ between a certain referring pronoun and a certain candidate entity, obtain each candidate entity from the candidate history set, and obtain the maximum value ζ of the context matching values between the certain referring pronoun and each candidate entity. max , and take the candidate entity corresponding to the maximum value ζ max as the target entity of the certain referring pronoun;

[0028] Obtain the target entities of each referring pronoun in the question text and optimize the intention of the question text. The specific optimization process is: replace the referring pronoun in the question text with the target entity to obtain the target question text.

[0029] Further, step S1 includes:

[0030] Obtain the preset original document and AI model from the platform, and use the AI model to cut the original document in the knowledge base to generate logical document fragments;

[0031] Collect the logical document fragments of each original text obtained from the platform to obtain a logical document fragment set. Randomly obtain a certain knowledge entity and another knowledge entity from the platform to form an entity pair A, and calculate the association probability value P between the certain knowledge entity and the other knowledge entity A ;

[0032] Set an association probability threshold P´. When P A > P´, it is determined that there is an association between a certain knowledge entity and the other knowledge entity in the entity pair A. Obtain the precedence value d and the postcedence value d´ of a certain knowledge entity with respect to the other knowledge entity, and calculate the sequence value C = (d - d´) / B A ;

[0033] Set a sequence absolute threshold c´. When |C| > c´ and C > 0, it is determined that a certain knowledge entity is a pre-knowledge entity of the other knowledge entity. When C < 0, it is determined that the other knowledge entity is a pre-knowledge entity of a certain knowledge entity. When |C| < c´, it is determined that there is no sequence relationship between a certain knowledge entity and the other knowledge entity;

[0034] Obtain each entity pair with an association. Obtain each knowledge entity from the platform, and use each knowledge entity as a node. Connect the two knowledge entities in each entity pair with an association by an edge, and use the association probability value between the two knowledge entities in the entity pair as the edge weight of the edge between the two knowledge entities;

[0035] When there is no order relationship between any two knowledge entities in a knowledge pair, the edge is an undirected edge; otherwise, the edge between any two knowledge entities is a directed edge. The directed edge points from the previous knowledge entity to other knowledge entities in the pair, thus constructing a knowledge association graph. When a node is not connected to any other node in the knowledge association graph by an edge, the node is removed from the knowledge association graph.

[0036] Furthermore, step S4 includes:

[0037] Obtain the knowledge text and target question text of the target knowledge subgraph, input the knowledge text and target question text into a pre-trained large language model, and generate the answer text of the question text entered by the user in the current period;

[0038] The platform outputs and displays the answer text to the user, and conducts automated question-and-answer interactions with the user, storing the user's question-and-answer records for the current period.

[0039] To better implement the above methods, a full-scene automatic question-answering interaction system is also proposed, which includes an association graph construction module, a context matching analysis module, a knowledge text generation module, and a question-answering interaction module.

[0040] The knowledge graph construction module is used to use AI models to segment the original documents in the knowledge base and analyze the relationships between different knowledge entities in the generated logical document fragments to construct a knowledge graph.

[0041] The context matching analysis module is used to acquire the user's question text and historical question and answer records, analyze the degree of context matching between the user's question text and historical question and answer records, and obtain the target question text;

[0042] The knowledge text generation module is used to acquire knowledge association graphs, determine the complexity granularity of the target question text, and generate knowledge text of the target knowledge subgraph by combining the original documents in the knowledge base.

[0043] The question-and-answer interaction module is used to acquire the knowledge text of the target knowledge subgraph, generate the answer text to the user's question text based on the knowledge text, and conduct automatic question-and-answer interaction with the user.

[0044] Furthermore, the association graph construction module includes document segmentation units and association graph construction units;

[0045] The document segmentation unit is used to obtain the original document and AI model from the platform, and use the AI ​​model to segment the original document in the knowledge base to generate logical document fragments;

[0046] The association graph construction unit is used to extract knowledge entities from logical document fragments, calculate the association probability values ​​between different knowledge entities, and analyze the association between different knowledge entities in the logical document fragments based on the association probability values ​​to obtain a knowledge association graph.

[0047] Furthermore, the context matching analysis module includes a context matching analysis unit and a text intent optimization unit;

[0048] The context matching analysis unit is used to obtain similar historical question texts of the question text, obtain candidate historical question and answer records of the question text, calculate the context matching value between pronouns in the question text and candidate entities, and obtain the target entity of the pronouns in the question text based on the context matching value.

[0049] The text intent optimization unit is used to optimize the target entities of each pronoun in the question text and the intent of the question text to obtain the target question text.

[0050] Furthermore, the knowledge text generation module includes a complexity determination unit and a knowledge text generation unit;

[0051] The complexity determination unit is used to analyze the relationship between the amount of information and the structure of the problem in the target problem text based on the knowledge association graph, and to determine the complexity granularity of the target problem text.

[0052] The knowledge text generation unit is used to adaptively expand the knowledge subgraph of the target question text according to the complexity granularity of the target question text, obtain the target knowledge subgraph, and generate the knowledge text of the target knowledge subgraph by combining it with the original documents in the knowledge base.

[0053] Furthermore, the question-and-answer interaction module includes question-and-answer interaction units;

[0054] The question-and-answer interaction unit is used to acquire the knowledge text and target question text of the target knowledge subgraph, generate answer text, output and display the answer text to the user through the platform, and conduct automated question-and-answer interaction with the user through the platform.

[0055] Compared with existing technologies, the beneficial effects of this invention are as follows: By automatically segmenting documents into semantic boundaries using an AI model, the generated knowledge graph can explain the deep relationships between knowledge entities. Furthermore, by analyzing the contextual matching degree between user question text and historical question-and-answer records, it obtains pronouns in the question text, optimizes the intent of the question text, and significantly improves the semantic understanding accuracy in complex dialogue scenarios. Moreover, based on the question complexity of the target question text, it adaptively expands the knowledge subgraph of the target question text, so that the knowledge text of the target knowledge subgraph is dynamically adjusted according to the amount of question information and the breadth of structure. This fundamentally ensures that the answer information is appropriate and logically complete, quickly solves the user's question, saves the user's question-and-answer time, and greatly improves the user's actual question-and-answer experience. Attached Figure Description

[0056] Figure 1 This is a flowchart of the entity order determination process for an AI-driven, full-scene automatic question-and-answer interaction method according to the present invention.

[0057] Figure 2 This is a flowchart of the modules of an automatic question-and-answer interaction system for all scenarios according to the present invention. Detailed Implementation

[0058] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0059] Example: Figures 1-2 As shown, this invention provides a technical solution: an AI-driven, full-scene automatic question-answering interaction method, the method comprising:

[0060] Step S1: Use an AI model to segment the original documents in the knowledge base, generate logical document fragments, and analyze the relationships between different knowledge entities in the logical document fragments to obtain a knowledge association graph;

[0061] Step S1 includes:

[0062] The system retrieves pre-set original documents and AI models from the platform, and uses the AI ​​model to segment the original documents in the knowledge base to generate logical document fragments.

[0063] For example, using an AI model to segment original documents in a knowledge base and generate logical document fragments involves the following process:

[0064] The sentences in the original document are obtained and aggregated to obtain the sentence sequence set Q of the original document. The sentence sequence set Q is used as the input of the AI ​​model to obtain the segmentation probability between each pair of adjacent sentences in the sentence sequence set Q.

[0065] For example, the AI ​​model is a pre-trained model, specifically the BERT sequence labeling model, which predicts whether there are segmentation boundaries between sentences. The input data of the AI ​​model is manually labeled training documents, and the gaps between sentences in the training documents are labeled as logical boundaries. The output of the AI ​​model is the segmentation probability between adjacent sentences.

[0066] Set a segmentation probability threshold. When the segmentation probability between two adjacent sentences in the sentence sequence set Q is greater than the segmentation probability threshold, the adjacent sentences are segmented, and the pairs of sentences corresponding to the adjacent sentences are marked and recorded as marked sentences.

[0067] Get each tagged sentence in the sentence sequence set Q, and get the logical document fragments of the original document;

[0068] For example, the specific process of obtaining logical document fragments from the original document is as follows:

[0069] When there are no other marked sentences between a marked sentence and another marked sentence in the sentence sequence set Q, and the marked sentence and another marked sentence are not adjacent, the sentences containing the marked sentence and another marked sentence are cut out from the original document and aggregated to obtain the logical document fragment of the original document, wherein the logical document fragment contains the marked sentence and another marked sentence;

[0070] Logical document fragments from various original texts are retrieved from the platform and aggregated to obtain a logical document fragment set. A knowledge entity is randomly selected from the platform and paired with another knowledge entity to form an entity pair A. The association probability value P between the two knowledge entities is calculated. A ;

[0071] For example, calculating the probability value P of the association between one knowledge entity and another knowledge entity. A The specific calculation process is as follows:

[0072] Retrieve the total number B of logical document fragments containing entity pair A from the logical document fragment set. A Obtain the total number B of logical document fragments containing one or more knowledge entities from the logical document fragment set. sum ;

[0073] Calculate the probability value P of the association between one knowledge entity and another knowledge entity. A =B A / B sum;

[0074] For example, use a BERT-based named entity recognition (NER) model to obtain knowledge entities from logical document fragments;

[0075] For example, for the entity word set preset by the platform for knowledge entities, if there is a word in the logical text segment in the entity word set, it is determined that the logical document segment contains a knowledge entity. When a word in the logical document segment is in the entity word set and another word in the logical document segment is not in the entity word set, if the cosine similarity of the word vectors between the other word and the certain word is greater than the preset threshold, it is determined that the other word also belongs to a certain knowledge entity;

[0076] For example, knowledge entities have different meanings according to different scenarios. When the scenario is a chain platform scenario, knowledge entities include stores, printers, scheduling, ordering, returns, etc.;

[0077] Set the associated probability threshold P´. When P A > P´, it is determined that there is an association between a certain knowledge entity and another knowledge entity in the entity pair A. Obtain the prior value d and the posterior value d´ of a certain knowledge entity to another knowledge entity, and calculate the sequence value C = (d - d´) / B of a certain knowledge entity to another knowledge entity A ;

[0078] For example, the specific process of obtaining the prior value d and the posterior value d´ of a certain knowledge entity to another knowledge entity is as follows:

[0079] The total number of logical document segments in which the word corresponding to a certain knowledge entity in the logical document fragment set is before the word corresponding to another knowledge entity, and it is denoted as the prior value d of a certain knowledge entity to another knowledge entity;

[0080] The total number of logical document segments in which the word corresponding to a certain knowledge entity in the logical document fragment set is after the word corresponding to another knowledge entity, and it is denoted as the posterior value d´ of a certain knowledge entity to another knowledge entity;

[0081] Set the sequence absolute threshold c´. When |C| > c´ and C > 0, it is determined that a certain knowledge entity is the prior knowledge entity of another knowledge entity. When C < 0, it is determined that another knowledge entity is the prior knowledge entity of a certain knowledge entity. When |C| < c´, it is determined that there is no sequence relationship between a certain knowledge entity and another knowledge entity;

[0082] Obtain each entity pair with an association. Obtain each knowledge entity from the platform, and use each knowledge entity as a node. Connect the two knowledge entities in each entity pair with an association by an edge, and use the associated probability value between the two knowledge entities in the entity pair as the edge weight of the edge between the two knowledge entities;

[0083] When there is no order relationship between any two knowledge entities in a knowledge pair, the edge is an undirected edge; otherwise, the edge between any two knowledge entities is a directed edge. The directed edge points from the previous knowledge entity to other knowledge entities in the pair, thus constructing a knowledge association graph. When a node is not connected to any other node in the knowledge association graph by an edge, the node is removed from the knowledge association graph.

[0084] Step S2: Obtain the user's question text and historical question-and-answer records, analyze the degree of contextual matching between the user's question text and historical question-and-answer records, and obtain the target question text;

[0085] Step S2 includes:

[0086] Get the question text entered by the user from the platform, get the user's historical question and answer records on the platform, and get the historical question text from the historical question and answer records;

[0087] Retrieve similar historical question texts from the question text, and retrieve candidate historical question-and-answer records for the question text;

[0088] For example, the specific process for obtaining similar historical question texts is as follows:

[0089] The problem text is converted into a text feature vector using a preset text vector model. The text feature vectors of historical problem texts are obtained. The semantic similarity between the problem text and historical problem texts is calculated. A semantic similarity threshold is set. When the semantic similarity between the problem text and historical problem texts is greater than the semantic similarity threshold, it is determined that there is semantic similarity between the problem text and historical problem texts, and the historical problem text is recorded as a similar historical problem text of the problem text.

[0090] For example, the preset text vector models include BAAI / bge-large-zh-v1.5, etc.;

[0091] For example, the specific calculation process for semantic similarity between the question text and historical question text is as follows:

[0092] Calculate the cosine similarity of the text feature vectors between the question text and the historical question text, and denote it as the semantic similarity between the question text and the historical question text;

[0093] For example, the specific process for obtaining candidate historical question-and-answer records for a question text is as follows:

[0094] Set a feature value k to obtain the total number K of similar historical question texts for each question text. sum When k≥K sum, the historical Q&A records of each similar historical question text of the question text are recorded as candidate historical Q&A records. Otherwise, the historical Q&A records of the top k similar historical question texts are selected from each similar historical question text in descending order according to semantic similarity and are recorded as candidate historical Q&A records;

[0095] Obtain and gather each candidate historical Q&A record of the question text to obtain a candidate historical record set. Obtain the pronouns in the question text and obtain a certain candidate entity from the candidate historical Q&A records;

[0096] For example, the pronouns in the question text include "it", "that", "this", etc.;

[0097] For example, the specific process of obtaining a candidate entity from the candidate historical Q&A records is as follows:

[0098] Obtain the historical question text and the historical answer text from the candidate historical Q&A records in the candidate historical record set, perform word segmentation on the historical question text and the historical answer text, remove stop words, and obtain candidate entities from the historical question text and the historical answer text. Among them, candidate entities include personal names, place names, object names, method names, and noun phrases, etc.;

[0099] For example, stop words include auxiliary words (such as "de", "le", etc.), prepositions ("zai", "duiyu", "dui", etc.);

[0100] Obtain and mark the candidate historical Q&A records containing a certain candidate entity from the candidate historical record set to obtain each marked candidate historical Q&A record of a certain candidate entity; [[ID=一十九]]

[0101] Obtain the sentence where a certain pronoun is located from the question text, obtain the sentence where a certain candidate entity is located from the marked candidate historical Q&A records, and calculate the average value of the semantic similarity between the sentence where a certain pronoun is located and the sentence where a certain candidate entity is located in each marked candidate historical Q&A record to obtain the semantic matching value E between a certain pronoun and a certain candidate entity;

[0102] Obtain the total number α of the marked candidate historical Q&A records containing a certain candidate entity and a certain pronoun from each marked candidate historical Q&A record, and calculate the collinear matching value G between a certain pronoun and a certain candidate entity = α / α sum where α sum is the total number of candidate historical Q&A records in the candidate historical record set;

[0103] For example, the keywords in the question text are, after word segmentation of the question text, the words that are personal names, place names, object names, method names, and noun phrases, etc. in the question text are used as the keywords of the question text;

[0104] Calculate the context matching value ζ between a given pronoun and a given candidate entity. Obtain each candidate entity from the candidate history set and find the maximum value ζ between the given pronoun and all candidate entities. max and the maximum value ζ max The corresponding candidate entity serves as the target entity for a certain pronoun;

[0105] For example, the specific calculation process of the context matching value ζ is as follows:

[0106] Normalize the collinearity matching value G and semantic matching value E between a candidate entity and the question text, and calculate the context matching value ζ=η. G ×G+η E ×E, where η G η is the preset collinearity weighting coefficient. E η is the preset semantic weight coefficient. G +η E =1, η G and η E All are greater than 0;

[0107] Obtain the target entities of each pronoun in the question text and optimize the intent of the question text. The specific optimization process is as follows: replace the pronouns in the question text with the target entities to obtain the target question text.

[0108] Step S3: Based on the knowledge association graph, analyze the relationship between the amount of question information and the question structure in the target question text, determine the complexity granularity of the target question text, adaptively expand the knowledge subgraph of the target question text to obtain the target knowledge subgraph, and combine it with the original documents in the knowledge base to generate the knowledge text of the target knowledge subgraph;

[0109] Step S3 includes:

[0110] Obtain the user's target question text within the current period, extract knowledge entities from the target question text, and record them as candidate knowledge entities;

[0111] Obtain the entity word set of candidate knowledge entities from the platform, obtain the relevant historical question and answer records of candidate knowledge entities, and calculate the context key value L of candidate knowledge entities;

[0112] For example, the specific process for obtaining the relevant historical question-and-answer records of candidate knowledge entities is as follows:

[0113] If a user's historical question and answer record contains any word from the entity word set, then that historical question and answer record will be recorded as a related historical question and answer record of the candidate knowledge entity.

[0114] For example, the specific formula for calculating the contextual key value L of a candidate knowledge entity is: L = γ / γsum γ sum γ represents the total number of historical question and answer records for each user, and γ represents the total number of related historical question and answer records for each candidate knowledge entity.

[0115] Obtain the knowledge association graph, and obtain the total number of other nodes connected to the node corresponding to the candidate knowledge entity through directed edges and undirected edges. Obtain the edge weights between the node corresponding to the candidate knowledge entity and the other connected nodes. Calculate the global key value S of the node corresponding to the candidate knowledge entity in the knowledge association graph.

[0116] For example, the specific formula for calculating the global key value S is:

[0117] ,

[0118] Where j is the total number of other nodes in the knowledge graph that are connected to the node corresponding to the candidate knowledge entity through undirected edges; w i is the edge weight between the node corresponding to the candidate knowledge entity and the i-th other node connected by an undirected edge; n is the total number of other nodes in the knowledge graph connected to the node corresponding to the candidate knowledge entity by directed edges; w (△,z) is the edge weight between the node corresponding to the candidate knowledge entity and the z-th other node connected by a directed edge; e is the preset node direction coefficient, 1>e>0.5;

[0119] Calculate the comprehensive score U of the nodes corresponding to the candidate knowledge entities, and obtain the maximum comprehensive score U of the nodes corresponding to each candidate knowledge entity in the target question text. max , will U max The node to which it belongs is denoted as the main node of the target question text;

[0120] For example, the calculation process for the comprehensive score U of the node corresponding to the candidate knowledge entity is as follows:

[0121] The global key value S and contextual key value L of the node corresponding to the candidate knowledge entity in the knowledge association graph are normalized, and the comprehensive score U=λ of the node corresponding to the candidate knowledge entity is calculated. S ×S+λ L ×L, where λ S λ is the preset global key weight. L λ is the preset contextual key weight. L +λ S =1, λ L >0, λ S >0;

[0122] Obtain the word set of the target question text, obtain the knowledge base from the platform, obtain the total number of original documents containing words from the word set from the knowledge base, calculate the question information content I of the target question text, calculate the structural breadth Y of the target question text, and calculate the complexity granularity H´ of the target question text;

[0123] For example, the specific process of obtaining the word set of the target question text is as follows:

[0124] The target question text is segmented into words, and the resulting words are aggregated to obtain the word set of the target question text.

[0125] For example, the specific calculation process for calculating the question information content I of the target question text is as follows:

[0126] Retrieve the total number of original documents containing the a-th word in the word set from the knowledge base. sum Get the total number of original documents in the knowledge base, v, and calculate the information scarcity value q=log(a) of the a-th word. sum / v), calculate the sum of the information scarcity values ​​of each word in the word set of the target question text, and obtain the question information content I of the target question text;

[0127] For example, the specific formula for calculating the structural breadth Y of the target question text is:

[0128] ,

[0129] Where m is the total number of knowledge entities in the target question text;

[0130] For example, the specific formula for calculating the complexity granularity H´ of the target question text is:

[0131] ,

[0132] Wherein, β is a preset information scaling factor, used to scale the amount of information at the complex granularity; β´ is a structure breadth adjustment factor, used to control the amplification effect of the structure on the complex granularity of the target problem text.

[0133] A knowledge subgraph is constructed based on the target question text, where the knowledge subgraph is an empty set. The main nodes of the target question text are incorporated into the knowledge subgraph. Several nodes connected to the main nodes are randomly selected from the knowledge graph and incorporated into the knowledge subgraph. The average degree *b* of each node in the knowledge subgraph and the sum of the edge weights *W* of the edges connecting each node are calculated. sum The complexity granularity H of the computational knowledge subgraph △ ;

[0134] For example, the specific process of calculating the average degree of each node in a knowledge subgraph is as follows:

[0135] Obtain the edges connecting each node in the knowledge subgraph. When a node is connected to two nodes, if the edge connecting a node to a node is undirected, the degree of the edge connecting the node to a node is 1. If the edge connecting a node to a node is directed, the degree of the edge connecting the node to a node is 2. The degree of the node is then 3.

[0136] For example, computing the complexity granularity H of a knowledge subgraph. △ The specific formula is as follows:

[0137] ,

[0138] Obtain the complexity granularity of each knowledge subgraph of the target question text, and select several knowledge subgraphs with a complexity granularity greater than H´. From these knowledge subgraphs, select the knowledge subgraph with the smallest total number of nodes as the target knowledge subgraph of the target question text.

[0139] Retrieve logical document fragments from the original documents containing knowledge entities in the target knowledge subgraph from the knowledge base, and merge them to generate the knowledge text of the target knowledge subgraph;

[0140] For example, the process of retrieving logical document fragments containing knowledge entities in the target knowledge subgraph from the knowledge base and merging them to generate the knowledge text of the target knowledge subgraph is as follows:

[0141] The acquired logical document fragments are deduplicated and reordered, and then concatenated into a complete context to obtain the knowledge text of the target knowledge subgraph.

[0142] Step S4: Based on the knowledge text, generate the answer text to the user's question text, and conduct automatic question-and-answer interaction with the user through the platform;

[0143] Step S4 includes:

[0144] Obtain the knowledge text and target question text of the target knowledge subgraph, input the knowledge text and target question text into a pre-trained large language model, and generate the answer text of the question text entered by the user in the current period;

[0145] For example, pre-trained large language models include: Qwen2-7B-Instruct, Llama-3, etc.

[0146] The platform outputs and displays the answer text to the user, and conducts automated question-and-answer interactions with the user, storing the user's question-and-answer records for the current period.

[0147] To better implement the above methods, a full-scene automatic question-answering interaction system is also proposed, which includes an association graph construction module, a context matching analysis module, a knowledge text generation module, and a question-answering interaction module.

[0148] The knowledge graph construction module is used to use AI models to segment the original documents in the knowledge base and analyze the relationships between different knowledge entities in the generated logical document fragments to construct a knowledge graph.

[0149] The context matching analysis module is used to acquire the user's question text and historical question and answer records, analyze the degree of context matching between the user's question text and historical question and answer records, and obtain the target question text;

[0150] The knowledge text generation module is used to acquire knowledge association graphs, determine the complexity granularity of the target question text, and generate knowledge text of the target knowledge subgraph by combining the original documents in the knowledge base.

[0151] The question-and-answer interaction module is used to acquire the knowledge text of the target knowledge subgraph, generate the answer text to the user's question text based on the knowledge text, and conduct automatic question-and-answer interaction with the user.

[0152] The association graph construction module includes a document segmentation unit and an association graph construction unit.

[0153] The document segmentation unit is used to obtain the original document and AI model from the platform, and use the AI ​​model to segment the original document in the knowledge base to generate logical document fragments;

[0154] The association graph construction unit is used to extract knowledge entities from logical document fragments, calculate the association probability values ​​between different knowledge entities, and analyze the association between different knowledge entities in the logical document fragments based on the association probability values ​​to obtain a knowledge association graph.

[0155] The context matching analysis module includes a context matching analysis unit and a text intent optimization unit.

[0156] The context matching analysis unit is used to obtain similar historical question texts of the question text, obtain candidate historical question and answer records of the question text, calculate the context matching value between pronouns in the question text and candidate entities, and obtain the target entity of the pronouns in the question text based on the context matching value.

[0157] The text intent optimization unit is used to optimize the target entities of each pronoun in the question text and the intent of the question text to obtain the target question text.

[0158] The knowledge text generation module includes a complexity determination unit and a knowledge text generation unit.

[0159] The complexity determination unit is used to analyze the relationship between the amount of information and the structure of the problem in the target problem text based on the knowledge association graph, and to determine the complexity granularity of the target problem text.

[0160] The knowledge text generation unit is used to adaptively expand the knowledge subgraph of the target question text according to the complexity granularity of the target question text, obtain the target knowledge subgraph, and generate the knowledge text of the target knowledge subgraph by combining it with the original documents in the knowledge base.

[0161] The question-and-answer interaction module includes question-and-answer interaction units;

[0162] The question-and-answer interaction unit is used to acquire the knowledge text and target question text of the target knowledge subgraph, generate answer text, output and display the answer text to the user through the platform, and conduct automated question-and-answer interaction with the user through the platform.

[0163] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the invention can be implemented in other specific forms without departing from its spirit or essential characteristics. Therefore, the embodiments should be considered in all respects as exemplary and non-limiting, and the scope of the invention is defined by the appended claims rather than the foregoing description. Thus, all variations falling within the meaning and scope of equivalents of the claims are intended to be included within the present invention. No reference numerals in the claims should be construed as limiting the scope of the claims.

Claims

1. An AI-driven, full-scenario automatic question-answering interaction method, characterized in that, The method includes: Step S1: Use an AI model to segment the original documents in the knowledge base, generate logical document fragments, and analyze the relationships between different knowledge entities in the logical document fragments to obtain a knowledge association graph; Step S2: Obtain the user's question text and historical question-and-answer records, analyze the degree of contextual matching between the user's question text and historical question-and-answer records, and obtain the target question text; Step S3: Based on the knowledge association graph, analyze the relationship between the amount of question information and the question structure in the target question text, determine the complexity granularity of the target question text, adaptively expand the knowledge subgraph of the target question text to obtain the target knowledge subgraph, and combine it with the original documents in the knowledge base to generate the knowledge text of the target knowledge subgraph; Step S4: Based on the knowledge text, generate the answer text to the user's question text, and conduct automatic question-and-answer interaction with the user through the platform.

2. The AI-driven, full-scene automatic question-answering interaction method according to claim 1, characterized in that, Step S3 includes: Obtain the user's target question text within the current period, extract knowledge entities from the target question text, and record them as candidate knowledge entities; Obtain the entity word set of candidate knowledge entities from the platform, obtain the relevant historical question and answer records of candidate knowledge entities, and calculate the context key value L of candidate knowledge entities; Obtain the knowledge association graph, and obtain the total number of other nodes connected to the node corresponding to the candidate knowledge entity through directed edges and undirected edges. Obtain the edge weights between the node corresponding to the candidate knowledge entity and the other connected nodes. Calculate the global key value S of the node corresponding to the candidate knowledge entity in the knowledge association graph. Calculate the comprehensive score U of the nodes corresponding to the candidate knowledge entities, and obtain the maximum comprehensive score U of the nodes corresponding to each candidate knowledge entity in the target question text. max , will U max The node to which it belongs is denoted as the main node of the target question text; Obtain the word set of the target question text, obtain the knowledge base from the platform, obtain the total number of original documents containing words from the word set from the knowledge base, calculate the question information content I of the target question text, calculate the structural breadth Y of the target question text, and calculate the complexity granularity H´ of the target question text; A knowledge subgraph is constructed based on the target question text, where the knowledge subgraph is an empty set. The main nodes of the target question text are incorporated into the knowledge subgraph. Several nodes connected to the main nodes are randomly selected from the knowledge graph and incorporated into the knowledge subgraph. The average degree *b* of each node in the knowledge subgraph and the sum of the edge weights *W* of the edges connecting each node are calculated. sum The complexity granularity H of the computational knowledge subgraph △ ; Obtain the complexity granularity of each knowledge subgraph of the target question text, and select several knowledge subgraphs with a complexity granularity greater than H´. From these knowledge subgraphs, select the knowledge subgraph with the smallest total number of nodes as the target knowledge subgraph of the target question text. Extract logical document fragments from the original documents containing knowledge entities in the target knowledge subgraph from the knowledge base, and merge them to generate the knowledge text of the target knowledge subgraph.

3. The AI-driven, full-scene automatic question-answering interaction method according to claim 1, characterized in that, Step S2 includes: Get the question text entered by the user from the platform, get the user's historical question and answer records on the platform, and get the historical question text from the historical question and answer records; Retrieve similar historical question texts from the question text, and retrieve candidate historical question-and-answer records for the question text; Obtain and aggregate all candidate historical question and answer records of the question text to obtain a candidate historical record set, extract the pronouns in the question text, and extract a candidate entity from the candidate historical question and answer records; Retrieve candidate historical question and answer records containing a certain candidate entity from the candidate historical record set and mark them to obtain each marked candidate historical question and answer record of a certain candidate entity; Obtain the sentence where a certain pronoun is located from the question text, obtain the sentence where a certain candidate entity is located from the marked candidate historical Q&A records, calculate the average value of the semantic similarity between the sentence where the pronoun is located and the sentence where a certain candidate entity is located in each of the marked candidate historical Q&A records, and obtain the semantic matching value E between the pronoun and the candidate entity; Obtain the total number α of tagged candidate historical question-and-answer records containing a certain candidate entity and a certain pronoun from each of the aforementioned candidate historical question-and-answer records, and calculate the collinearity matching value G = α / α between a certain pronoun and a certain candidate entity. sum , where α sum The total number of candidate historical question and answer records in the candidate historical record set; Calculate the context matching value ζ between a given pronoun and a given candidate entity. Obtain each candidate entity from the candidate history set and find the maximum value ζ between the given pronoun and all candidate entities. max and the maximum value ζ max The corresponding candidate entity serves as the target entity for a certain pronoun; Obtain the target entities of each pronoun in the question text and optimize the intention of the question text. The specific optimization process is as follows: Replace the pronouns in the question text with the target entities to obtain the target question text.

4. The AI-driven, full-scene automatic question-answering interaction method according to claim 1, characterized in that, The step S1 includes: Obtain the preset original document and the AI model from the platform, and use the AI model to cut the original document in the knowledge base to generate logical document fragments; Logical document fragments from various original texts are retrieved from the platform and aggregated to obtain a logical document fragment set. A knowledge entity is randomly selected from the platform and paired with another knowledge entity to form an entity pair A. The association probability value P between the two knowledge entities is calculated. A ; Set an association probability threshold P', when P' A When P' >, ​​it is determined that there is a relationship between a knowledge entity and another knowledge entity in entity pair A. The preceding value d and the following value d' of a knowledge entity to another knowledge entity are obtained, and the order value C=(dd') / B of a knowledge entity to another knowledge entity is calculated. A ; Set the sequential absolute threshold c´. When |C|>c´ and C>0, it is determined that a certain knowledge entity is the pre-knowledge entity of another knowledge entity. When C<0, it is determined that the other knowledge entity is the pre-knowledge entity of a certain knowledge entity. When |C|<c´, it is determined that there is no sequential relationship between a certain knowledge entity and another knowledge entity; Obtain each entity pair with an association, obtain each knowledge entity from the platform, and use each knowledge entity as a node. Connect the two knowledge entities in each entity pair with an association by an edge, and use the association probability value between the two knowledge entities in the entity pair as the edge weight of the edge between the two knowledge entities; When there is no sequential relationship between the two knowledge entities in a certain knowledge pair, the edge is an undirected edge. Otherwise, the edge between the two knowledge entities is a directed edge, and the directed edge points from the pre-knowledge entity to the other knowledge entity of the two knowledge entities, constructing a knowledge association graph. When a certain node is not connected to other nodes by an edge in the knowledge association graph, remove the certain node from the knowledge association graph.

5. The AI-driven, full-scene automatic question-answering interaction method according to claim 1, characterized in that, The step S4 includes: Obtain the knowledge text of the target knowledge subgraph and the target question text, input the knowledge text and the target question text into the pre-trained large language model, and generate the answer text of the question text input by the user in the current cycle; Output and display the answer text to the user through the platform, and conduct an automated Q&A interaction with the user through the platform, and store the Q&A records of the user in the current cycle.

6. A full-scene automatic question-answering interaction system, used to execute the AI-driven full-scene automatic question-answering interaction method according to any one of claims 1-5, characterized in that, The system includes an association graph construction module, a context matching analysis module, a knowledge text generation module, and a Q&A interaction module; The association graph construction module is used to cut the original document in the knowledge base using the AI model, analyze the relevance between different knowledge entities in the generated logical document fragments, and construct a knowledge association graph; The context matching analysis module is used to obtain the user's question text and historical Q&A records, analyze the context matching degree between the user's question text and the historical Q&A records, and obtain the target question text; The knowledge text generation module is used to obtain the knowledge association graph, determine the complexity granularity of the target question text, and combine the original document in the knowledge base to generate the knowledge text of the target knowledge subgraph; The question-and-answer interaction module is used to acquire the knowledge text of the target knowledge subgraph, generate the answer text to the user's question text based on the knowledge text, and conduct automatic question-and-answer interaction with the user.

7. The full-scene automatic question-and-answer interaction system according to claim 6, characterized in that, The association graph construction module includes a document segmentation unit and an association graph construction unit; The document segmentation unit is used to obtain the original document and AI model from the platform, and use the AI ​​model to segment the original document in the knowledge base to generate logical document fragments. The association graph construction unit is used to obtain knowledge entities from logical document fragments, calculate the association probability values ​​between different knowledge entities, and analyze the association between different knowledge entities in the logical document fragments based on the association probability values ​​to obtain a knowledge association graph.

8. The full-scene automatic question-and-answer interaction system according to claim 6, characterized in that, The context matching analysis module includes a context matching analysis unit and a text intent optimization unit; The context matching analysis unit is used to obtain similar historical question texts of the question text, obtain candidate historical question and answer records of the question text, calculate the context matching value between the pronouns in the question text and the candidate entities, and obtain the target entity of the pronouns in the question text based on the context matching value. The text intent optimization unit is used to optimize the target entities of each pronoun in the question text and the intent of the question text to obtain the target question text.

9. A full-scene automatic question-and-answer interaction system according to claim 6, characterized in that, The knowledge text generation module includes a complexity determination unit and a knowledge text generation unit; The complexity determination unit is used to analyze the relationship between the amount of information about the problem and the structure of the problem in the target problem text based on the knowledge association graph, and to determine the complexity granularity of the target problem text. The knowledge text generation unit is used to adaptively expand the knowledge subgraph of the target question text according to the complexity granularity of the target question text, obtain the target knowledge subgraph, and generate the knowledge text of the target knowledge subgraph by combining it with the original documents in the knowledge base.

10. A full-scene automatic question-and-answer interaction system according to claim 6, characterized in that, The question-and-answer interaction module includes a question-and-answer interaction unit; The question-and-answer interaction unit is used to acquire the knowledge text and target question text of the target knowledge subgraph, generate answer text, output and display the answer text to the user through the platform, and conduct automated question-and-answer interaction with the user through the platform.

Images