Dynamic knowledge retrieval enhancement method based on large language model

Abstract

The invention discloses a method for enhancing dynamic knowledge retrieval based on a large language model, belongs to the field of knowledge retrieval, and aims to solve the problems of knowledge solidification, insufficient timeliness and illusion of a traditional LLM (Logistics Language Model). A multi-granularity knowledge base is dynamically constructed, and a rule and semantic partitioning technology is combined, so that a text is converted into a normalized vector, and a hybrid index is established; a two-channel retrieval triggering mechanism is adopted, keyword matching scores and BERT semantic probability analysis are fused, and retrieval requirements are intelligently judged; vectorization retrieval is realized through a BGE-M3 model, and candidate results are reordered in combination with a cross encoder to improve the precision. The system supports multi-language adaptive processing, dynamic switching of word segmentation strategies and cross-language retrieval, and introduces real-time knowledge updating and version control. According to the method, the answer timeliness and accuracy are remarkably improved, the context coherence of multiple rounds of dialogues is optimized, the method can be widely applied to the fields of intelligent customer service, professional questions and answers and the like, the LLM illusion risk is effectively reduced, and the knowledge traceability is enhanced.

Description

Technical Field

[0001] The present invention relates to the field of knowledge retrieval, and in particular to a method for dynamic knowledge retrieval enhancement based on a large language model. Background Art

[0002] In recent years, with the rapid development of large language models (LLMs), generative AI, represented by DeepSeek and GPT, has demonstrated powerful capabilities in natural language processing (NLP) tasks. However, traditional LLMs have the following key issues: Knowledge solidification and lack of timeliness LLM training relies on a static corpus and cannot dynamically update knowledge, resulting in limited performance when faced with the latest information or domain-specific data. For example, when answering news events or industry technology trends after 2024, an unenhanced LLM may generate outdated or incorrect answers.

[0003] Hallucination Problem Since LLM is essentially a probability-based text generation model, in the absence of reliable knowledge support, it may output seemingly reasonable but actually wrong answers (such as fictitious facts, incorrect citations, etc.).

[0004] Insufficient coverage of long-tail knowledge General LLMs have limited knowledge of low-frequency or professional fields (such as medicine, law, engineering, etc.) and are unable to meet the high-precision requirements of vertical scenarios.

[0005] To address these issues, Retrieval-Augmented Generation (RAG) technology has emerged. RAG dynamically retrieves external knowledge bases (such as structured databases, document collections, or real-time data sources) to inject relevant context into the LLM generation process, thereby improving the accuracy, timeliness, and interpretability of answers. Typical technical solutions include: Dense Retrieval: Use models such as BERT and DPR to encode queries and documents into vectors and match relevant paragraphs through approximate nearest neighbor search (ANN).

[0006] Hybrid Retrieval: Combines traditional keyword search (such as BM25) with vector search, taking into account both semantic matching and precise term recall.

[0007] Dynamic knowledge update (Hot-Reload): Supports real-time addition, deletion, modification, and query of the knowledge base, avoiding the retraining cost of traditional LLM.

[0008] Although the RAG technology has significantly improved the reliability of LLMs, its practice still faces challenges such as retrieval efficiency-accuracy trade-off, multimodal knowledge fusion, and construction of complex reasoning chains. Therefore, how to optimize the collaborative mechanism between dynamic knowledge retrieval and LLM generation has become a key direction in current research and applications. Summary of the Invention

[0009] In view of the above deficiencies in the prior art, the present invention provides a method for enhancing dynamic knowledge retrieval based on large language models.

[0010] To achieve the above invention objective, the technical solution adopted by the present invention is as follows: A method for enhancing dynamic knowledge retrieval based on large language models, comprising the following steps: S1. Construct a dynamic knowledge index system; S2. Based on the constructed dynamic knowledge index system, parse the user input intention and judge the knowledge retrieval requirements; S3. Based on the results obtained in S2, construct a decision-making model for multi-feature fusion, and perform intelligent routing selection between the retrieval channel and the generation channel; S4. Based on the decision-making model obtained in S3, construct a multi-modal fusion knowledge hybrid retrieval mechanism and output candidate knowledge fragments; S5. Organize the candidate knowledge fragments obtained in S4, the user's current input, and the context history into input pairs, and construct an automatic evaluation system that integrates semantic and structural information through multi-dimensional indicators such as semantic relevance calculation, context alignment analysis, and knowledge source credibility scoring. Evaluate the accuracy and reliability of the retrieved knowledge according to the source confidence weight of the knowledge fragment; S6. Encode the content of the knowledge fragments after accuracy and reliability evaluation into context vectors, use two-channel information fusion to control the information weight distribution between the knowledge fragments and the internal language modeling of the large prediction model, and monitor the user satisfaction and changes in the dialogue state in real time and update the dialogue state graph.

[0011] Further, the specific steps of S1 are as follows: S11. Use data access and parsing to convert multi-source heterogeneous data into a unified intermediate format; S12. Perform knowledge extraction on the converted data, including entity extraction, relationship recognition, and semantic annotation, and map it to a unified knowledge body structure; S13. Construct a keyword index, a semantic vector index, and a cross-domain knowledge graph, where the semantic vector index encodes knowledge fragments based on the Sentence-BERT model and constructs approximate nearest neighbor retrieval through the HNSW algorithm.

[0012] Further, the specific steps of S2 are as follows: S21. Standardize the user input and construct a context embedding representation in combination with the conversation history; S22. Determine whether it is an information request statement through an intent classifier and calculate the knowledge demand score; S23. Dynamically trigger the knowledge retrieval process according to the knowledge demand score and the user behavior rules.

[0013] Further, the specific method for calculating the knowledge demand score in S22 is as follows:

[0014] where, is the knowledge demand score, is the wh-word density score, is the proper noun density score, is the knowledge entity matching score, are the weight parameters of the corresponding scores in sequence.

[0015] Further, S3 specifically includes the following steps: S31. Extract the word vector features, question structure features, and historical behavior features of the user input; S32. Calculate the probabilities of the retrieval path and the generation path through a lightweight neural network decision model. If the retrieval probability is higher than the threshold, trigger knowledge retrieval; S33. Write the decision result into the conversation state diagram for subsequent retrieval strategy optimization.

[0016] Further, S4 specifically includes the following steps: S41. Parse the user query, extract keywords, semantic vectors, and multi-modal demand markers; S42. Obtain candidate knowledge fragments through a keyword retrieval module, a semantic vector retrieval module, and a multi-modal completion module; S43. Perform several rounds of fusion sorting on the retrieval results, optimize the final sorting using the LambdaMART model, and eliminate redundant content through semantic similarity detection to obtain candidate knowledge fragments.

[0017] Further, the specific method for eliminating redundant content in step S43 is as follows: S431. Remove duplicate knowledge units, count the coverage and confidence of the results, and convert the text into a vector representation of a fixed dimension; S432. Perform word segmentation and stop word removal on the key texts of the remaining knowledge units, and generate embedding vectors for each knowledge unit; S433. Calculate the similarity between the embedding vectors of each knowledge unit. When the similarity exceeds the set threshold, judge the semantics as redundant and remove them to obtain candidate knowledge fragments.

[0018] Furthermore, the S5 specifically includes the following steps: S51, organize all candidate knowledge fragments returned in step 4 and the user's current input and context history into input pairs, and construct<query, document> Structure pair collection; S52. For each<query, document> Sentence pair matching is scored using the pre-trained BERT-based model to obtain semantic similarity; S53. Concatenate the semantic similarity, context adaptation semantics, and knowledge source credibility score into a vector and input it into the fully connected network FFN for confidence score prediction. Output the credibility score and set the confidence threshold. Filter out knowledge units that do not meet the standards and sort the remaining content in descending order of confidence.

[0019] The present invention has the following beneficial effects: Through the dynamic retrieval mechanism, the system can access the latest knowledge resources in real time, overcome the problem of outdated knowledge in traditional large language models, and ensure the timeliness and professionalism of the answer content.

[0020] The dual-channel decision module and fusion algorithm are used to enable efficient collaboration between the retrieval and generation processes, avoid invalid retrieval and information redundancy, and improve system response speed and resource utilization.

[0021] It supports cross-level retrieval from keywords to semantic fragments, meeting query requirements of different complexities and granularity, and improving the comprehensiveness and accuracy of knowledge recall.

[0022] Through the retrieval result credibility assessment model, relevant content is automatically screened and sorted, reducing the interference of misleading or irrelevant information, and improving the reliability of answers and user trust.

[0023] Combining context-aware search optimization and conversation state tracking, dynamic adjustment of search strategies is achieved, making the system's understanding of user intent more accurate, improving the interactive experience and conversation coherence.

[0024] The technical solution supports multi-source heterogeneous data interfaces, facilitates the expansion and maintenance of dynamic knowledge bases, adapts to different application scenarios such as enterprise customer service, professional Q&A, intelligent education, etc., and enhances the system's versatility and sustainable development capabilities. BRIEF DESCRIPTION OF THE DRAWINGS

[0025] Figure 1 It is a schematic diagram of the process of the present invention. DETAILED DESCRIPTION

[0026] The specific embodiments of the present invention will be described below to facilitate those skilled in the art of the present technology to understand the present invention. However, it should be clear that the present invention is not limited to the scope of the specific embodiments. For those of ordinary skill in the art of the present technology, as long as various changes are within the spirit and scope of the present invention defined and determined by the appended claims, these changes are obvious, and all inventions made using the concept of the present invention are within the scope of protection.

[0027] A method for enhancing dynamic knowledge retrieval based on large language models, as Figure 1 shown, includes the following steps: S1. Construct a dynamic knowledge index system; In this embodiment, this step aims to establish a knowledge base architecture that supports multi-granularity indexing to solve problems such as static knowledge aging and single data source. A unified embedding representation of structured, semi-structured, and unstructured data is introduced in the knowledge storage layer; keyword indexes, semantic indexes, etc. are constructed in the index layer. The specific steps are as follows: S11. Use data access and parsing to convert multi-source heterogeneous data into a unified intermediate format; Data uploaded by users or automatically obtained is converted into a unified intermediate format Markdown through API interfaces, HTML, PDF, Word documents, etc.

[0028] S12. Perform knowledge extraction on the converted data, including entity extraction, relationship recognition, and semantic annotation, and map it to a unified knowledge body structure; S13. Construct a keyword index, a semantic vector index, and a cross-domain knowledge graph. Among them, the semantic vector index encodes knowledge fragments based on the Sentence-BERT model and constructs approximate nearest neighbor retrieval through the HNSW algorithm.

[0029] Construct an inverted index for the processed knowledge units, and use the Analyzer tool to perform word frequency statistics, position recording, and fast retrieval index establishment for keywords, named entities, etc. Encode each knowledge fragment into a fixed-dimensional vector based on the Sentence-BERT model, and use the HNSW approximate nearest neighbor algorithm to construct a vector index to support efficient retrieval based on semantic similarity.

[0030] Extract entities and relationships to construct a heterogeneous graph spectrum, which is stored in a relational database to support graph structure queries and make up for the lack of logical reasoning in text indexes.

[0031] S2. Based on the constructed dynamic knowledge index system, perform user input intention parsing and knowledge retrieval requirement judgment; In this embodiment, this step mainly involves the intention recognition, context understanding, and knowledge dependence judgment of the user input prompt. The core mechanism is the integration of multi-modal semantic understanding + trigger scoring model. By analyzing the position, expression mode, and information density of the user input in the context, it is determined whether to call the knowledge base for retrieval enhancement. The specific steps are as follows: S21. Standardize the user input and construct a context embedding representation in combination with the conversation history; Standardize the user input, including word segmentation, removing blank lines, scripts, duplicate paragraphs, and stop word filtering, etc., to construct a structured input vector. Call the conversation history cache, perform window modeling on the recent several rounds of conversations, and form a context embedding representation in combination with positional encoding.

[0032] S22. Use an intention classifier to determine whether it is an information request statement and calculate the knowledge demand score; Based on the fine-tuned RoBERTa model, perform intention classification on the current input sentence, output whether it is an "information request type" sentence, count the number of interrogative words, proper nouns, and knowledge entities in the input sentence, and construct a knowledge demand score. The score is calculated using where QW_score: interrogative word density score, NNP_score: proper noun density score, NE_score: knowledge entity matching score. is a weight parameter, reflecting the contribution degree of each index to knowledge dependence, and the default value is .

[0033] S23. Dynamically trigger the knowledge retrieval process according to the knowledge demand score and user behavior rules.

[0034] Combine the user behavior (such as whether to repeat the question, whether to ask a rhetorical question, whether to express vaguely) and the score to judge whether to trigger the knowledge retrieval process and return a boolean value and confidence level for the subsequent dual-channel decision-making module to use.

[0035] S3. Based on the results obtained in S2, construct a decision-making model with multi-feature fusion and perform intelligent routing selection between the retrieval channel and the generation channel; In this embodiment, this step introduces a "collaborative decision-making model" to perform intelligent routing selection between the retrieval channel and the generation channel. A context-aware path discrimination, by analyzing the current input, historical context, and previous rounds of system behavior, constructs a set of decision-making models with multi-feature fusion to achieve the dynamics, interpretability, and personalization of channel selection. The specific process is as follows: S31. Extract the word vector features, interrogative structure features, and historical behavior features of the user input; Extract word vector features, question structure, and sentence length from the user input. At the same time, reference the knowledge requirement score and intention discrimination result in S2. Reference the system behavior in the last three rounds in the state diagram, and extract features such as whether knowledge retrieval has been used in the past, retrieval effect score, and user satisfaction. S32. Calculate the probabilities of the retrieval path and the generation path through a lightweight neural network decision model. If the retrieval probability is higher than the threshold, trigger knowledge retrieval. Use a lightweight neural network with two branches (such as a two-tower structure). One branch processes the input features of the generation path, and the other branch processes the retrieval path features. Finally, judge the probability value through softmax fusion. If the probability of the retrieval channel is higher than the threshold (which can be dynamically set, such as 0.6), the system calls the knowledge base module for subsequent knowledge retrieval; otherwise, directly enter the generation channel.

[0036] S33. Write the decision result into the dialogue state diagram for subsequent retrieval strategy optimization.

[0037] Regardless of which path is selected, the decision basis and result of the current round are written into the dialogue state diagram. This step ensures the smooth switching between the retrieval and generation paths, has a high context awareness ability, and avoids useless retrieval or knowledge loss.

[0038] S4. Based on the decision model obtained in S3, construct a multi-modal fusion knowledge hybrid retrieval mechanism and output candidate knowledge fragments. This step aims to construct a hybrid knowledge retrieval mechanism that supports the collaborative work of keyword retrieval and semantic vector retrieval, and introduce multi-modal fusion technology to make the system have stronger adaptability and recall ability when processing multi-type knowledge units such as text, images, and tables. This mechanism adopts a three-level retrieval strategy of "keyword filtering + semantic matching + multi-modal extension", which significantly improves the coverage while ensuring the accuracy. The specific process is as follows: S41. Parse the user query textually, extract keywords, semantic vectors, and multi-modal requirement tags. Parse the user input textually, extract keyword entities, relationship words, hyponym and hypernym concepts, and construct a preliminary query vector. If the input involves multi-modal elements (such as mentioning keywords like "diagram", "table", etc.), mark it as a multi-modal requirement.

[0039] S42. Obtain candidate knowledge fragments through the keyword retrieval module, semantic vector retrieval module, and multi-modal completion module. Keyword retrieval module: Use the ElasticSearch inverted index structure to perform boolean logic matching retrieval on keywords, obtain the first batch of candidate knowledge fragments, and sort them according to the TF-IDF score.

[0040] Semantic Vector Retrieval Module: Call the Chinese version of Sentence-BERT to encode the user input into a vector, and use the Milvus engine to perform vector approximate nearest neighbor search (ANN) in the semantic index of the knowledge base, and return semantically relevant candidates.

[0041] Multimodal Completion Module: If the input involves non-text content such as images and charts, call the corresponding modality retrieval engine CLIP for cross-modal matching, and uniformly convert the results into the form of knowledge fragments.

[0042] S43. Perform several rounds of fusion sorting on the retrieval results, use the LambdaMART model to optimize the final sorting, and eliminate redundant content through semantic similarity detection to obtain candidate knowledge fragments.

[0043] Merge the above three types of results, fuse the scores (keyword retrieval score, semantic similarity, multimodal relevance) through a weighted sorting mechanism, use the learning-to-rank model LambdaMART for sorting, remove duplicate knowledge units, and count the coverage and confidence of the results to provide raw data for the credibility evaluation in the fifth step. The implementation is mainly based on semantic similarity detection. Use the semantic analysis technology Sentence-BERT to embed and vectorize the fragments, convert the text into a vector representation of a fixed dimension, and then judge whether the semantics is redundant by calculating the similarity between vectors greater than 0.8. This can find the situation where the statement expressions are different but the semantics are the same. Calculation steps: Perform text preprocessing (such as word segmentation, stop word removal, etc.) on each knowledge unit. Then use the model Sentence-BERT to generate the embedding vector of each unit. Finally, use the Cosine similarity or Euclidean distance to measure the similarity between knowledge units. If the similarity exceeds the preset threshold, mark it as repeated.

[0044] S5. Organize the candidate knowledge fragments obtained in S4, the user's current input, and the context history into input pairs, and construct an automatic evaluation system that integrates semantic and structural information through multi-dimensional indicators such as semantic relevance calculation, context alignment analysis, and knowledge source credibility scoring. Evaluate the accuracy and reliability of the retrieved knowledge according to the source confidence weight of the knowledge fragments; In this embodiment, this step constructs an automatic evaluation system that integrates semantic and structural information by introducing multi-dimensional indicators such as semantic relevance calculation, context alignment analysis, and knowledge source credibility scoring. The core algorithm is a context-based bidirectional attention scoring model, which combines the source confidence weight of the knowledge element to realize the comprehensive evaluation of the accuracy and reliability of the retrieved knowledge, specifically including: S51. Organize all the candidate knowledge fragments returned in step four, the user's current input, and the context history into input pairs, and construct a set of <query, document> structure pairs; S52. Use the pre-trained BERT-based model to perform sentence pair matching scoring on each <query, document> pair to obtain semantic similarity. S53. Concatenate the semantic similarity, context adaptation semantics, and knowledge source credibility score into a vector and input it into the fully connected network FFN for confidence score prediction. Output the credibility score and set the confidence threshold to filter out knowledge units that do not meet the criteria, and sort the remaining content in descending order of confidence.

[0045] S6. Encode the content of the knowledge fragments after precision and reliability evaluation into context vectors, use dual-channel information fusion to control the information weight distribution between the knowledge fragments and the internal language modeling of the large prediction model, and monitor the user satisfaction and the change of the dialogue state in real time and update the dialogue state graph.

[0046] In this embodiment, this step aims to achieve high-quality fusion between the output of the large language model and the knowledge retrieval results, ensuring that the generated text is both naturally fluent in language and accurate and reliable in facts. Improve the structure using the Retrieval-Augmented Generation (RAG) technique, design dual-channel information fusion, and control the information weight distribution between the knowledge fragments and the internal language modeling of the LLM. The specific steps are as follows: Knowledge fragment encoding: Encode the knowledge content after credibility evaluation into context vectors and add them to the encoder segment by segment (such as the Retriever encoder in RAG).

[0047] User input fusion: Concatenate the user's current input and the dialogue history as the input token sequence and interact and fuse with the knowledge support vector.

[0048] Fusion mechanism: By controlling the Attention fusion weight between the knowledge vector and the autoregressive language generation token, guide the generation to be more dependent on high-confidence knowledge points.

[0049] Enhanced annotation: Add a knowledge citation annotation mechanism to clearly cite the corresponding knowledge number in the output content, and at the same time enhance the knowledge fragment position embedding in the underlying Transformer.

[0050] Output filtering and post-processing: Conduct a round of knowledge alignment check on the generated response. If the key conclusion is not reflected in the knowledge, warn the generator to reduce the confidence of the relevant content or regenerate it to improve the accuracy of the response. This mechanism breaks through the disadvantages of the traditional LLM's "hallucinatory generation" and realizes natural language generation output under the guidance of knowledge, effectively integrating external fact support and the model's language advantages. In the present invention, specific embodiments are used to elaborate on the principles and implementation manners of the present invention. The description of the above embodiments is only for helping to understand the method of the present invention and its core idea; at the same time, for those of ordinary skill in the art, according to the idea of the present invention, there will be changes in the specific implementation manners and application scopes. In summary, the content of this specification should not be construed as a limitation to the present invention.

[0051] Those of ordinary skill in the art will realize that the embodiments described herein are for helping the reader understand the principles of the present invention, and it should be understood that the protection scope of the present invention is not limited to such specific statements and embodiments. Those of ordinary skill in the art can make various other specific deformations and combinations that do not depart from the essence of the present invention according to these technical revelations disclosed in the present invention, and these deformations and combinations are still within the protection scope of the present invention.

Claims

1. A method for enhancing dynamic knowledge retrieval based on large language models, characterized in that, It includes the following steps: S1. Construct a dynamic knowledge index system; S2. Based on the constructed dynamic knowledge index system, perform user input intention parsing and knowledge retrieval requirement judgment; S3. Based on the results obtained in S2, construct a decision model for multi-feature fusion, and perform intelligent routing selection between the retrieval channel and the generation channel; S4. Based on the decision model obtained in S3, construct a multi-modal fusion knowledge hybrid retrieval mechanism, and output candidate knowledge fragments; S5. Organize the candidate knowledge fragments obtained in S4 with the user's current input and context history into input pairs, and construct an automatic evaluation system that integrates semantic and structural information through multi-dimensional indicators such as semantic relevance calculation, context alignment degree analysis, and knowledge source credibility scoring. Evaluate the accuracy and reliability of the retrieved knowledge according to the source confidence weight of the knowledge fragments; S6. Encode the content of the knowledge fragments after accuracy and reliability evaluation into context vectors, use dual-channel information fusion to control the information weight distribution between the knowledge fragments and the internal language modeling of the large prediction model, and monitor the user satisfaction and the change of the dialogue state in real time and update the dialogue state graph.

2. The method for enhancing dynamic knowledge retrieval based on a large language model according to claim 1, wherein The specific steps of S1 are as follows: S11. Use data access and parsing to convert multi-source heterogeneous data into a unified intermediate format; S12. Perform knowledge extraction on the converted data, including entity extraction, relationship recognition, and semantic annotation, and map them to a unified knowledge body structure; S13. Construct a keyword index, a semantic vector index, and a cross-domain knowledge graph, where the semantic vector index encodes knowledge fragments based on the Sentence-BERT model and constructs approximate nearest neighbor retrieval through the HNSW algorithm.

3. The method for enhancing dynamic knowledge retrieval based on a large language model according to claim 1, wherein, The specific steps of S2 are as follows: S21. Standardize the user input and construct a context embedding representation in combination with the dialogue history; S22. Use an intention classifier to judge whether it is an information request statement and calculate the knowledge demand score; S23. Dynamically trigger the knowledge retrieval process according to the knowledge demand score and the user behavior rules.

4. The method for enhancing dynamic knowledge retrieval based on a large language model according to claim 3, wherein The specific method for calculating the knowledge demand score in S22 is: Among them, is the knowledge demand score, is the interrogative word density score, is the proper noun density score, is the knowledge entity matching score, are the weight parameters of the corresponding scores in sequence.

5. The method for enhancing dynamic knowledge retrieval based on large language models according to claim 1, wherein The specific steps of S3 are as follows: S31. Extract the word vector features, question structure features, and historical behavior features of the user input; S32. Calculate the probabilities of the retrieval path and the generation path through a lightweight neural network decision model. If the retrieval probability is higher than the threshold, trigger knowledge retrieval; S33. Write the decision result into the dialogue state graph for subsequent retrieval strategy optimization.

6. The method for enhancing dynamic knowledge retrieval based on a large language model according to claim 1, wherein The specific steps of S4 are as follows: S41. Parse the user query, extract keywords, semantic vectors, and multi-modal demand markers; S42. Obtain candidate knowledge fragments through the keyword retrieval module, the semantic vector retrieval module, and the multi-modal completion module; S43. Perform several rounds of fusion sorting on the retrieval results, optimize the final sorting using the LambdaMART model, and eliminate redundant content through semantic similarity detection to obtain candidate knowledge fragments.

7. The method for enhancing dynamic knowledge retrieval based on a large language model according to claim 1, characterized in that, The specific method for eliminating redundant content in step S43 is: S431. Remove duplicate knowledge units, count the coverage and confidence of the results, and convert the text into a vector representation of a fixed dimension; S432. Tokenize and remove stop words from the key texts of the remaining knowledge units, and generate embedding vectors for each knowledge unit; S433. Calculate the similarity between the embedding vectors of each knowledge unit. When the similarity exceeds the set threshold, judge the semantics as redundant and remove them to obtain candidate knowledge fragments.

8. The method for enhancing dynamic knowledge retrieval based on a large language model according to claim 1, characterized in that The specific steps of S5 are as follows: S51. Organize all the candidate knowledge fragments returned in step four, the user's current input, and the context history into input pairs, and construct a set of <query, document> structure pairs; S52. Use the pre-trained BERT-based model to score the sentence pairs for each <query, document> pair to obtain semantic similarity; S53. Concatenate the semantic similarity, context adaptation semantics, and knowledge source credibility score into a vector and input it into the fully connected network FFN for confidence score prediction. Output the credibility score, set the confidence threshold, filter out the knowledge units that do not meet the standards, and sort the remaining content in descending order of confidence.

Images