Method for enhancing summary reply ability of intelligent question-answering system based on graph clustering

Abstract

A method for enhancing summary reply capability of an intelligent question-answering system based on graph clustering comprises the following steps: step 1, acquiring text data of a document, and converting unstructured text data into a structured text information graph; step 2, performing hierarchical clustering on the text information atlas, and dividing the text information atlas into a plurality of different information communities; step 3, generating an information abstract for each information community, selecting an information abstract similar to a user question, and generating a candidate document set; and step 4, integrating the candidate document set to obtain a final summary reply. According to the method, firstly, the text information atlas and the graph clustering technology are creatively combined, and LLM optimization is supplemented, so that the summarization reply capability of the intelligent question-answering system is remarkably improved. Secondly, the deep semantic structure and the document content are organically combined, so that the recall quality, the information comprehensive efficiency and the reply continuity of the intelligent question-answering system are improved, and efficient and reliable technical support is provided for application of the intelligent question-answering system in a complex task scene.

Description

Technical Field

[0001] The present invention relates to the technical field of information retrieval and natural language processing, and more specifically, to a method for enhancing the summarizing and replying capabilities of an intelligent question-answering system by constructing a text information graph and applying graph clustering technology. Background Art

[0002] With the widespread adoption of large-scale pre-trained language models (such as BERT and GPT), intelligent question-answering systems have demonstrated powerful capabilities in natural language processing. Retrieval-Augmented Generation (RAG) technology, in particular, significantly improves question-answering performance by integrating with external knowledge bases.

[0003] However, existing technologies still face numerous challenges in generating high-quality, coherent, and accurate summary responses. For example, existing generative question-answering models primarily rely on embedded knowledge to generate answers. This knowledge is limited by the training data, making it difficult to dynamically update and incapable of timely acquiring and utilizing the latest information. This makes it difficult to meet the real-time and domain-diversity requirements of practical applications. Generative models are highly sensitive to the quality of the input context. Low-quality contextual information can lead to inaccurate or incoherent answers, reducing the system's practicality. In RAG systems, the context provided by the recall module directly impacts the output quality of the generative model. Traditional recall methods often directly retrieve the full text of the document, introducing a large amount of irrelevant information and reducing the effectiveness of the context. For long documents, existing methods struggle to accurately locate key information and often overlook content crucial to answering the question. Summarization content is often used to aid reading or facilitate quick comprehension, failing to fully integrate the recall strategy of RAG systems and failing to realize their potential for information concentration and relevance enhancement. Furthermore, existing question-answering systems suffer from information synthesis problems such as modal ambiguity, fragmented evidence, and hallucinatory reasoning. Automatic text summarization (ATS) systems often face challenges in generating summaries with respect to coherence, fluency, and factual accuracy. This is especially true in multi-document summarization (MDS) tasks, where the system needs to handle information duplication, integrate different viewpoints, and ensure the coherence and conciseness of the summary.

[0004] Therefore, the problems existing in the prior art need to be further improved and developed. Summary of the Invention

[0005] (1) Purpose of the Invention: To address the above-mentioned problems existing in the prior art, the purpose of the present invention is to provide a method for enhancing the summary and reply capabilities of an intelligent question-answering system based on graph clustering, thereby solving the problems of the coherence, accuracy, redundancy and insufficient information integration ability of the summary and reply generated by the question-answering system when processing complex and long documents, thereby significantly improving the overall performance of the question-answering system and the user experience.

[0006] (ii) Technical solution: In order to solve the above technical problems, the technical solution provides a method for enhancing the summarizing and replying capability of an intelligent question and answer system based on graph clustering, comprising the following steps: 1. obtaining text data of a document, and converting unstructured text data into a structured text information graph; 2. performing hierarchical clustering on the text information graph, and dividing the text information graph into a plurality of different information communities; 3. generating an information abstract for each information community, selecting an information abstract similar to a user question, and generating a candidate document set; 4. integrating the candidate document set to obtain a final summary reply.

[0007] Preferably, the step 1 specifically comprises: Step 101: performing text preprocessing and blocking on the original document; Step 102: defining nodes and edges according to different semantic blocks to obtain a text information graph; Step 103: converting the nodes into embedding vectors, assigning weights to the edges, storing the text information graph into a graph database and a vector database, and establishing an index; Step 104: identifying and merging the same entity to determine the accuracy of the text information graph.

[0008] Preferably, a sliding window is used to aggregate and calculate the semantic similarity of adjacent windows, and the cosine similarity of vectors is used to calculate the semantic similarity of adjacent windows. , wherein V A represents a vector of a text block A; V B represents a vector of a text block B; represents the dot product of the vector V A and the vector V B , represents the L2 norm of the vector V A and the vector V B . When the semantic similarity of adjacent windows is greater than or equal to a first similarity threshold, and the length of the text block is less than a first text block length, the same semantic block is merged; otherwise, the semantic block is divided into two semantic blocks.

[0009] Preferably, the text information graph is a network comprising a node set and an edge set; the node represents a semantic block in the text, and comprises an entity node and a concept node; the entity node is a key information carrier in the text, and the concept node is a higher-level key concept or subject word abstracted from the text. Edges represent the relationships between nodes, including co-occurrence edges and reference / reference edges. The co-occurrence edges connect words, entities, or concepts that appear together in a specific window, and the reference / reference edges connect different expressions that refer to the same entity or concept.

[0010] Preferably, all nodes are converted into embedded vectors as node features. In terms of edge weight assignment, a frequency weighting method is used to assign edge weights. The edge weight reflects the strength of the relationship between nodes. Edge weights are assigned by co-occurrence frequency or semantic similarity. The text information map is stored in a graph database and a vector database and an optimized index is established. For short texts, we use co-occurrence frequency weighting. Local co-occurrence reflects direct correlation. Then, according to the two nodes N i and N j Appearing together in k windows, its edge weight is W ij， That is, the closeness between the two nodes is measured by the number of times they appear together: W ij =Count(Co-occurrence of N i ,N j ), Among them, W ij represents the edge weight between node i and node j; N i represents the i-th node that appears in k windows; N j Indicates the jth node that appears in the k window For long texts, semantic similarity weighting is used, which is based on the semantic similarity calculation of two node vectors: W ij =Sim(Node i ,Node j )= , Among them, W ij represents the edge weight between node i and node j; The vector representing the i-th node that appears in the k-th window; The vector representing the jth node that appears in the k windows; Represents vector V A and vector V B The L2 norm of .

[0011] Preferably, the text information graph is used as input, and an iterative calculation is performed based on the edge weights between nodes through a clustering algorithm to divide the text information graph into multiple information communities; The Louvain algorithm is selected to measure the strength of information community division by iteratively optimizing modularity. The larger the Q value, the better the division strength of the information community. , Among them, A ij is an element of the adjacency matrix, indicating the weight of the edge between nodes i and j or whether there is an edge; Represents the sum of the weights of all edges in the text information graph; represents the degree (or sum of weights) of node i; c i represents the information community to which node i belongs; δ(c i ,c j ) represents the Kronecker delta function. When ci=cj, δ(c i ,c j ) is 1, otherwise δ(c i ,c j ) is 0.

[0012] Preferably, the step 3 specifically includes: Step 301: stratify the information communities in the text information graph and generate information summaries; Step 302: extract the keywords of the user's query and search for significant information points in the information community; Step 303: vectorize the user question, calculate the semantic similarity between the user question vector and the information summary vector, and obtain a candidate document set.

[0013] Preferably, the hierarchical characteristics of the Louvain clustering algorithm are utilized to generate a corresponding information summary for each layer of clustering. The information summary includes a node-level summary and a cluster-level summary. The node-level summary is generated for each basic node in the text information graph; the cluster-level summary is generated for each level of information community identified by the Louvain algorithm, and the lower-level node summary or low-level cluster summary contained in the clustering algorithm is aggregated and abstracted to generate a cluster information summary.

[0014] Preferably, after the user question is vectorized, the user question vector v is calculated using cosine similarity with the information summary vector of each level cluster Q and the vector v of each cluster information summary C The similarity between them is selected, and the clustering information summary with a similarity greater than or equal to the first similarity is selected as the paragraph of the candidate document; The paragraphs of the candidate documents are rearranged through a cross encoder, and an intelligent truncation strategy based on statistical distribution is adopted to calculate the candidate score according to the formula: SCORE threshold =μ+γσ, Among them, μ represents the mean score of all candidate paragraphs; σ represents the standard deviation of all candidate paragraph scores; γ represents different paragraphs, γ∈[0.5,1.5], and paragraphs with scores greater than the first candidate score are retained to form the candidate document set.

[0015] Preferably, a high-quality, cluster-enhanced candidate document set is used as the context input for retrieval enhancement generation; the candidate document set is further optimized using the powerful generation and context learning capabilities of the large language model to output the final summary response.

[0016] (III) Beneficial Effects: This method, based on graph clustering, significantly enhances the summary and response capabilities of intelligent question-answering systems by innovatively combining text information graphs with graph clustering techniques, supplemented by LLM optimization. Secondly, by organically integrating deep semantic structure with document content, it improves the recall quality, information integration efficiency, and response coherence of intelligent question-answering systems, providing efficient and reliable technical support for the application of intelligent question-answering systems in complex task scenarios. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] Figure 1 This is a flowchart of the steps of a method for enhancing the summary and reply capability of an intelligent question-answering system based on graph clustering according to the present invention; Figure 2 It is a schematic diagram of a specific embodiment of the present invention. DETAILED DESCRIPTION

[0018] The present invention is further described in detail below in conjunction with preferred embodiments. More details are set forth in the following description to facilitate a full understanding of the present invention. However, the present invention can obviously be implemented in a variety of other ways different from the description. Those skilled in the art can make similar generalizations and deductions based on actual application situations without violating the connotation of the present invention. Therefore, the scope of protection of the present invention should not be limited by the content of this specific embodiment.

[0019] The accompanying drawings are schematic diagrams of embodiments of the present invention. It should be noted that the drawings are merely examples and are not drawn to scale, and should not be used to limit the actual scope of protection claimed in the present invention.

[0020] A method based on graph clustering to enhance the summary and reply ability of intelligent question answering system, such as Figure 1 、 Figure 2 As shown, the specific steps include: Step 1: Obtain the text data of the document and convert the unstructured text data into a structured text information graph.

[0021] Step 2: Perform hierarchical clustering on the text information graph to divide it into multiple different information communities.

[0022] Step 3: Generate information summaries for each information community, select information summaries similar to user questions, and generate candidate document sets.

[0023] Step 4: Integrate the candidate document set to obtain the final summary response.

[0024] More specifically, the step 1 specifically includes: The text data of the original document is obtained and converted into a structured text information graph, laying the foundation for subsequent graph clustering and knowledge discovery. It not only organizes scattered information points, but more importantly, it reveals the hidden semantic and structural relationships between information points, providing rich context for subsequent intelligent processing.

[0025] Step 101: Preprocess and segment the original document.

[0026] Text preprocessing and segmentation of original documents includes text cleaning, word segmentation, and syntactic analysis. To ensure the semantic integrity of each text chunk, the present invention adopts a semantic-driven segmentation strategy rather than a simple fixed-length segmentation. For example, the present invention uses a sliding window approach to aggregate related sentences and determines whether to merge them by calculating the semantic similarity of adjacent windows. The semantic similarity Sim(A, B) between text chunks A and B is usually represented by their vectors V. A and V B The cosine similarity is used to calculate, as shown in formula (1).

[0027] (1), Among them, V A A vector representing a text block A; V B A vector representing a text block B; Represents vector V A and vector V B The dot product of Represents vector V A and vector V B The L2 norm of .

[0028] When the semantic similarity between adjacent windows is greater than or equal to the first similarity threshold (e.g., 0.85), they are merged into the same semantic block; otherwise, a new semantic block is created. To avoid generating overly long paragraphs, the length of the text block is considered for control. When the length of the text block is less than the length of the first text block (e.g., 512 tokens), the text is merged into the same semantic block; otherwise, it is divided into two semantic blocks.

[0029] For paragraphs containing subheadings, the present invention also supplements the paragraph content with the subheadings, preserving the document's hierarchical structure and important structural information, ensuring the integrity of the semantic blocks and the coherence of the context. This refined block segmentation method provides a high-quality node foundation for the subsequent construction of the text information graph, enabling nodes in the text information graph to be directly associated with the document's chapter information, providing explicit structural clues for identifying information communities during subsequent clustering.

[0030] Step 102: Define nodes and edges according to different semantic blocks to obtain a text information graph.

[0031] The text information graph is a network consisting of a set of nodes and edges. By converting raw text into this structured graph, it effectively addresses the processing of long documents and the interference of redundant information. This approach eliminates the need to directly process the raw, potentially lengthy, and noisy text, but instead operates on a more manageable and semantically rich network representation. This allows the system to capture deeper semantic and contextual information that is difficult to detect using traditional methods, providing a solid foundation for subsequent knowledge discovery and summary generation.

[0032] Nodes represent semantic blocks in a text, while edges represent relationships between nodes, capturing the intrinsic structure and semantic connections of the text. The specific types of nodes and edges are defined in Table 1.

[0033] Nodes include entity nodes and concept nodes. Entity nodes identify named entities in text, such as names of people, places, organizations, and specific concepts, based on actual application scenarios, and treat them as independent nodes. Entity nodes are key information carriers in text, and their isolation facilitates subsequent entity-level analysis and relationship discovery. Concept nodes abstract higher-level key concepts or keywords from the text and serve as high-level nodes. Concept nodes help capture the macro-themes of a document and provide overview information for summarizing cross-documents or long documents.

[0034] Edges include co-occurrence edges and reference / coreference edges. Co-occurrence edges connect words, entities, or concepts that co-occur within a specific window (e.g., within the same paragraph or adjacent sentences). They reveal the local relevance of information within a text. Reference / coreference edges connect different expressions that refer to the same entity or concept, such as a pronoun and its antecedent. Reference / coreference edges ensure accurate understanding of entity references in a text and avoid information confusion.

[0035] Table 1 Definition of nodes and edge types in text information graph

[0036] Step 103: Convert the nodes into embedded vectors, assign weights to the edges, store the text information graph in the graph database and the vector database, and create indexes.

[0037] Specifically, a vector model adapted to the field (such as the BGE model) is used to convert all nodes into embedded vectors. As the feature representation of the nodes, it can capture the semantic information of the text and provide a numerical basis for subsequent similarity calculation and clustering.

[0038] When converting a node to an embedded vector representation, all attribute text of the node is converted to vectors. Finally, when calculating the node's embedding vector, different attributes are assigned different calculation weights based on their importance, and the corresponding embedding vector is finally obtained through weighted averaging. The importance of attributes is specified by the user when creating the node schema, ensuring that the calculation results are more closely aligned with application requirements.

[0039] In terms of edge weight assignment, a frequency-weighted approach is used to assign edge weights, which reflect the strength or importance of the relationship between nodes. Edge weights are assigned based on co-occurrence frequency or semantic similarity. The text information graph is stored in a graph database and a vector database, and optimized indexes are established.

[0040] For short texts (such as sentence level), co-occurrence frequency weighting is preferred because local co-occurrence can better reflect direct correlation. As shown in formula (2), if the edge weight is calculated according to the co-occurrence frequency weighting, then according to the two nodes N i and N j Appearing together in k windows, its edge weight is W ij， That is, the closeness between the two nodes is measured by the number of times they appear together.

[0041] W ij =Count(Co-occurrence of N i ,N j )(2), Among them, W ij represents the edge weight between node i and node j; N i represents the i-th node that appears in k windows; N j represents the jth node that appears in the k windows.

[0042] For long texts (e.g., document level), semantic similarity weighting is used because global semantic association is more critical. As shown in formula (3), if the edge weight is calculated according to semantic similarity weighting, it is calculated based on the semantic similarity of the two node vectors.

[0043] W ij =Sim(Node i ,Node j )= (3), Among them, W ij represents the edge weight between node i and node j; The vector representing the i-th node that appears in the k-th window; The vector representing the jth node that appears in the k windows; Represents vector V A and vector V B The L2 norm of .

[0044] When existing graph clustering-based intelligent question-answering systems process dynamic text data (such as newly added documents and real-time information), the construction and update of text information graphs rely on static processing, and the information interaction between nodes lacks a dynamic propagation mechanism, resulting in incomplete generation of candidate document sets and insufficient real-time and accuracy of summary responses.

[0045] The present invention combines the self-organization, multi-hop transmission, dynamic node management and other characteristics of the Mesh network with the node interaction of the text information graph to construct a dynamic text information graph, realize dynamic node joining, adaptive update of edge weights, and hierarchical dissemination of information communities, and enhance the summary and reply capabilities of the intelligent question-answering system.

[0046] Specifically, when a new text block is added and a new node is generated, it automatically joins the existing links (no pre-networking is required) through the self-organizing nature of the mesh network, and establishes edge connections with adjacent nodes (nodes with high semantic similarity). For example, when the "2025 AI Ethics Regulations" document is added, a new node "AI Ethics Regulations" is generated and automatically connected to the node of the "Artificial Intelligence Ethics" community.

[0047] The edge weight is dynamically adjusted by combining semantic similarity and Mesh network transmission frequency, as shown in formula (4): W ij =α×Sim(Node i ,Node j )+(1-α)×Count(Co-occurrence of N i ,N j )(4), Among them, α represents the weight coefficient, Sim(Node i ,Node j ) represents the node vector cosine similarity, Count(Co-occurrence of N i ,N j ) represents the co-occurrence frequency between node i and node j.

[0048] Short texts focus on co-occurrence frequency (α is 0.3), and long texts focus on semantic similarity (α is 0.7).

[0049] The Mesh network is a decentralized network topology. Its core feature is that all nodes are interconnected and work together. Data can be dynamically transmitted through multiple paths without relying on a single central node, which significantly improves the stability, coverage and fault tolerance of the network.

[0050] In a mesh network, each node acts as both a terminal and a relay, forwarding data from other nodes. Nodes automatically discover neighbors and establish connections, forming dynamic paths. If a node fails, data automatically reroutes (e.g., A→B→C becomes A→D→C). Through dynamic collaboration, data transmission is relayed across multiple nodes, breaking through the limitations of single-point coverage and resolving the coverage and single-point failure issues of traditional networks.

[0051] Step 104: Identify and merge the same entity to determine the accuracy of the text information graph.

[0052] Because text segmentation and independent processing can result in the same entity appearing multiple times in the text information graph, this invention employs entity disambiguation technology. This technology utilizes a large language model (LLM) to identify and merge different expressions referring to the same entity, while also integrating their attributes. This ensures the accuracy and consistency of the text information graph, avoiding information redundancy and confusion. Through this comprehensive text information graph construction method, this invention provides intelligent question-answering systems with a highly structured and semantically rich knowledge base, significantly improving their ability to process complex information and generate high-quality summaries.

[0053] More specifically, step 2 is as follows: Using a text information graph as input, a clustering algorithm identifies information communities with high internal correlation and semantic consistency from the constructed text information graph. The clustering algorithm iteratively calculates edge weights (semantic similarity, co-occurrence frequency, etc.) between nodes, dividing the text information graph into multiple information communities. Each information community represents a highly semantically related "topic cluster" or "information unit," containing sentences, entities, and concepts that collectively describe a specific aspect or topic. For example, an information community includes all relevant sentences and entities regarding "artificial intelligence ethics," even if they are distributed across different locations in the original document.

[0054] This step organizes scattered textual information points into meaningful semantic units, thus resolving the fragmented evidence and context loss issues faced by existing systems when synthesizing information. By discovering these inherent semantic structures, the present invention provides more coherent and comprehensive information blocks for subsequent summary generation.

[0055] Specifically, the present invention selects a clustering algorithm suitable for large-scale text graphs to effectively identify semantic communities. For example, the Louvain algorithm is used for graph clustering to identify information communities with high internal correlation as information units, and each information community corresponds to a topic cluster.

[0056] The Louvain algorithm is selected to measure the strength of information community division by iteratively optimizing modularity. Modularity Q is an indicator to measure the strength of network community structure. The goal of the Louvain algorithm is to maximize the Q value to find the optimal information community division method. That is, the larger the Q value, the better the information community division, as shown in formula (5): (5), Among them, A ij is an element of the adjacency matrix, indicating the weight of the edge between nodes i and j or whether there is an edge; Represents the sum of the weights of all edges in the text information graph; represents the degree (or sum of weights) of node i; c i represents the information community to which node i belongs; δ(c i ,c j ) represents the Kronecker delta function. When ci=cj, δ(c i ,c j ) is 1, otherwise δ(c i ,c j ) is 0.

[0057] The Louvain algorithm, as an efficient hierarchical clustering method, can quickly detect community structures in networks and is capable of handling large-scale networks. It measures the strength of community divisions by iteratively optimizing modularity, making it suitable for discovering subsets of nodes with dense internal connections and sparse external connections. The Louvain algorithm is characterized by high computational efficiency, high-quality results, and the ability to detect hierarchical community structures. This makes it effective in discovering semantic topics of varying granularity when processing complex text graphs.

[0058] When the Louvain algorithm is used to perform hierarchical clustering on the text information graph, each information community corresponds to a sub-link in the Mesh network, including but not limited to sub-trees, sub-networks, etc. Nodes within an information community share node-level summaries through Mesh multi-hop transmission, while information communities are connected through central nodes.

[0059] The central node refers to the node with the highest edge weight in the text information graph.

[0060] Through the graph clustering process, the present application realizes knowledge discovery and representation. The clustering results output the information community ID to which each node belongs. By analyzing the nodes (sentences, entities, concepts) within each information community and their relationships, hidden topics, viewpoints, and key information in the text can be discovered. The information community constitutes a condensed, structured representation of the original document, transforming scattered information into coherent, semantically consistent "information communities", which is the key to solving the "fragmented evidence" and "context loss" problems faced by existing systems. It enables the system to generate not only concise summaries, but also logically structured and contextually rich summaries. In addition, this structured knowledge discovery process also provides a natural solution for multi-document summarization, as similar content will naturally be grouped into the same cluster, thereby avoiding cross-document redundancy and providing high-quality, pre-organized information input for subsequent LLM generation.

[0061] More specifically, step 3 specifically includes: Using the results of graph clustering, key information is extracted from the identified information communities, and a preliminary candidate document set is generated. This process is an important bridge connecting structured knowledge discovery and final summary generation, ensuring that the most relevant and core content to the user query is accurately filtered from the vast amount of information.

[0062] Step 301, hierarchically clustering information communities in the text information graph and generating information summaries.

[0063] In the offline preprocessing stage, the key information extraction strategy and hierarchical summary generation, the present application fully utilizes the hierarchical characteristics of the Louvain clustering algorithm to generate information summaries for each layer of clustering, achieving multi-granularity information summarization. The information summaries include node-level summaries (core information of individual nodes) and cluster-level summaries (aggregated abstraction of node or lower-level cluster summaries within the cluster).

[0064] Specifically, the central node of each information community aggregates node-level summaries within the sub-link to generate cluster-level summaries, and propagates the cluster-level summaries to upper-level information communities or other related information communities through Mesh links. For example, the central node of the "AI ethics" community aggregates the summaries of "privacy protection" and "algorithm fairness" sub-nodes to generate a cluster-level summary of "core controversial points of AI ethics".

[0065] The node-level summary generation uses LLM to generate a concise and accurate node information summary for each basic node (such as a concept node, an entity node) in the text information graph. The node-level summary captures the core information of the node. The cluster-level summary generation uses LLM to aggregate and abstract the lower-level node summaries or low-level cluster summaries contained in the clustering algorithm for each level of clustering (information community) identified by the Louvain algorithm to generate a higher-level, more concise cluster information summary. This hierarchical summary mechanism ensures that the system can provide a coherent and information-rich overview at any granularity level. For multi-document scenarios, the graph clustering results naturally handle the problem of information duplication, because similar content will be classified into the same cluster, thereby avoiding redundancy across documents and ensuring the conciseness of the preliminary summary.

[0066] When a new document is added, only the new document is segmented, nodes are extracted, and edges are constructed. It is then merged with the existing graph, and entity disambiguation techniques are used to match new and old entities, avoiding a full graph rebuild. Cluster-level summaries are recalculated only for the information communities affected by the new node, while the remaining communities remain unchanged. For example, when a new document titled "2025 AI Ethics Regulations" is added, only the summary for the "Artificial Intelligence Ethics" community is updated, without recalculating other communities.

[0067] Step 302: extract the keywords of the user's query and search for significant information points in the information community.

[0068] Specifically, when a user performs a question-and-answer query, the focus keywords of the user's query are extracted through word embedding similarity or LLM prompts, and saliency detection is performed based on the semantic relevance between the focus keywords and the information community (cosine similarity calculation) to identify salient information points.

[0069] Query-focused keyword / phrase extraction uses word embedding similarity or LLM hinting to extract the most relevant keywords or phrases from highly relevant clusters when a user queries. These keywords and / or phrases serve as guiding points for summarization. These focused keywords and / or phrases further focus the summary content, ensuring a close match with user intent.

[0070] Saliency Detection, an online algorithm, identifies salient information points across an information community or that are most relevant to a user's query. This is achieved by combining the semantic relevance of salient information points with the user's query. For example, for a cluster about "artificial intelligence development," if the user's query is about "AI applications in the medical field," sentences or abstracts within that cluster that are most relevant to medical applications will be prioritized. This mechanism enables more targeted summaries to meet the user's specific information needs.

[0071] Step 303: vectorize the user question, calculate the semantic similarity between the user question vector and the information summary vector, and obtain a candidate document set.

[0072] During information retrieval, instead of directly retrieving original document fragments, the system prioritizes the most relevant information communities as reference content for answer generation based on the semantic similarity between the user question and the content of cluster summaries at each level. These recalled cluster summaries are combined with full-text fragments extracted from related clusters and ranked according to their relevance to the user question.

[0073] Specifically, after the user question is vectorized, the similarity is calculated with the information summary vector of each level cluster. In this invention, the cosine similarity is used to calculate the user question vector v Q and the vector v of each cluster information summary C The semantic similarity between them is used to select clustering information summaries with a similarity greater than or equal to the first similarity as paragraphs of candidate documents.

[0074] The conversion of natural language to vectors is achieved by mapping words or phrases into a continuous numerical vector space, where semantically similar words are closer in the vector space. Models such as BGE (BAAI General Embedding) are often used as tools to achieve the conversion of natural language to vectors. The BGE model takes input natural language text (such as a sentence, paragraph, or document) and converts it into a numerical vector of fixed dimension. This vector is called an embedding or representation. The BGE model is trained to capture the semantic information in the text, so that semantically similar texts have closer vector representations in the vector space.

[0075] The cluster information summary with a similarity greater than or equal to the first similarity is selected as the paragraph of the candidate document. The selected information community propagates the cluster-level summary to relevant nodes (such as upper-level communities and cross-domain association communities) through the Mesh link. After collecting multi-source summaries, the paragraphs of the candidate document are rearranged through a cross encoder (such as DeBERTa-v3), and an intelligent truncation strategy based on statistical distribution is adopted to calculate the mean μ and standard deviation σ of all candidate paragraph scores. The paragraphs with candidate scores greater than the first candidate score are retained to form a candidate document set to ensure coverage of multi-dimensional key information. The candidate score calculation formula is shown in formula (6): SCORE threshold =μ+γσ(6), Wherein, μ represents the mean score of all candidate paragraphs; σ represents the standard deviation of the scores of all candidate paragraphs; γ represents different paragraphs, γ∈[0.5,1.5], and γ=1 in the present invention.

[0076] This method simultaneously ensures that at least three candidate paragraphs are retained, ultimately forming a quality-controlled set of candidate documents for generating the final summary response. This not only ensures the completeness and accuracy of the candidate document set but also provides high-quality, refined input for large-scale language models, significantly improving their ability to generate high-quality summaries. By organizing information into coherent semantic units during the preprocessing phase, this method effectively resolves the trade-off between information coverage and length in traditional summary generation, ensuring concise coverage of core topics.

[0077] More specifically, the specific operations of step 4 are: The candidate document set generated in step 3 is input into the generative model as context for the Retrieval-Augmented Generation (RAG) system. An iterative refinement strategy is employed, combined with the LLM, to abstract and polish the preliminary summary. Hint engineering guides the LLM to enhance coherence through techniques such as parallel structure and discourse analysis. The LLM's reasoning capabilities are leveraged for fact-checking, with reference node IDs output to verify factual consistency. Ultimately, an optimized final summary response is generated. This is the key step in achieving "summary response capability" in this invention, combining the advantages of structured knowledge with the generative power of the LLM to address the shortcomings of a single approach.

[0078] First, it integrates with the question-answering system. A high-quality, cluster-enhanced set of candidate documents is provided as contextual input to the retrieval-augmented generation (RAG) system and fed into the generative model. This integration provides the LLM with a "pre-digested" and "pre-organized" set of information, rather than raw, redundant documents. This allows the LLM to generate answers based on more precise and focused context, significantly improving question-answering performance. Because the input received by the LLM has already been preprocessed with graph clustering and key information extraction, the burden on its ability to handle noise and synthesize fragmented information is greatly reduced, making the generation process more efficient and reliable.

[0079] Secondly, we conduct LLM-driven summary response optimization, leveraging the powerful generative and contextual learning capabilities of the Large Language Model (LLM) to further abstract, polish, and refine the candidate document set.

[0080] An iterative refinement approach uses a candidate document set as the initial input. It then gradually introduces more relevant information (e.g., supplementary details from other related clusters) and uses the LLM to iteratively update and optimize the summary. Each iteration uses refinement prompts to guide the LLM in improving the summary, ensuring the quality of the final output. This iterative process gradually improves the detail and comprehensiveness of the summary.

[0081] To overcome the blandness, abstractness, repetition, and lack of logical structure that can occur in LLM-generated text, this invention uses prompting to guide LLMs in adopting specific writing techniques to enhance coherence and fluency. This approach encourages LLMs to use parallel sentence structures, enhancing the rhythm and balance of their final summary responses, emphasizing key ideas and making complex issues easier to understand. For example, when describing multiple characteristics, use "X has characteristics A, B, and C" rather than "X has characteristics A, B, and also has characteristic C."

[0082] Through sophisticated prompt engineering, the LLM is encouraged to consider the text's discourse structure (such as causality, chronology, and transitional phrases) during generation, ensuring logical connections and smooth transitions between ideas. This directly addresses the common AI problem of combining unrelated facts without smooth transitions. Leveraging the LLM's reasoning capabilities for fact-checking reduces generative hallucinations. This invention achieves this through the design of specific verification prompts, requiring the LLM to output the IDs of reference nodes in the graph when generating results, making it easier to identify the source of the answer.

[0083] This paper presents a method for enhancing the summary-response capabilities of intelligent question-answering systems based on graph clustering. First, by combining structured graph information with the generative capabilities of LLM (Learning Language Model), this method addresses the issues of hallucinations and incoherence that can arise when processing fragmented input. It also leverages the powerful generative capabilities of LLM to transform structured information into fluent, natural summary responses. This hybrid approach represents a significant advancement in the summary-response capabilities of intelligent question-answering systems, ensuring that the final output is not only accurate and concise, but also highly readable and effective. Secondly, by converting raw text into a structured graph and using graph clustering to identify core information communities, it can more accurately locate key information relevant to the user's question, effectively reducing the interference and redundancy of irrelevant information, significantly improving recall accuracy and context effectiveness. Furthermore, through graph construction and clustering, a higher-level semantic representation of the document is created, enabling the efficient extraction and organization of key information from long documents. This ensures that core content is accurately retrieved and delivered, significantly improving the performance of question-answering systems in long document scenarios. Finally, through graph clustering, the inherent semantic structure and thematic connections within the text are discovered, guiding the extraction and integration of summary information, making the information synthesis process more transparent and controllable. Combined with the refining ability of LLM, the generated summary response is ensured to be logically coherent and semantically rich, and fragmented evidence and illusions are effectively avoided. Information positioning is prioritized at the structured graph and clustering levels, which greatly reduces the computational complexity. By elevating information retrieval from the original text level to the semantic graph level, relevant information can be screened and located more efficiently, achieving a balance between efficiency and effectiveness. The present invention is not limited to specific fields or data types. In complex tasks that require processing large-scale document libraries or long documents and require high-precision summary responses, it has wide applicability and can provide users with more accurate and efficient information services.

[0084] In summary, the present invention, a method for enhancing the summary and response capabilities of intelligent question-answering systems based on graph clustering, firstly significantly improves the summary and response capabilities of intelligent question-answering systems by innovatively combining text information graphs with graph clustering technology, supplemented by LLM optimization. Secondly, by organically integrating deep semantic structure with document content, the recall quality, information integration efficiency, and response coherence of intelligent question-answering systems are improved, providing efficient and reliable technical support for the application of intelligent question-answering systems in complex task scenarios.

[0085] The above content is an explanation of the preferred embodiments of the present invention, which can help those skilled in the art to more fully understand the technical solutions of the present invention. However, these embodiments are merely illustrative, and it cannot be determined that the specific implementation methods of the present invention are limited to the description of these embodiments. For those skilled in the art of the present invention, without departing from the concept of the present invention, several simple deductions and transformations can be made, which should be considered to fall within the scope of protection of the present invention.

Claims

1. A method for enhancing the summarization and replying ability of an intelligent question-answering system based on graph clustering, characterized in that: The method includes the steps of:

1. obtaining text data of a document and converting the unstructured text data into a structured text information graph; Step 2: performing hierarchical clustering on the text information graph to divide it into multiple different information communities; Step 3: Generate information summaries for each information community, select information summaries similar to the user's questions, and generate candidate document sets; Step 4: Integrate the candidate document set to obtain the final summary response.

2. The method for enhancing the summarizing and replying ability of an intelligent question-answering system based on graph clustering according to claim 1, characterized in that: The step 1 specifically includes: Step 101: pre-process and segment the original document. Step 102: Define nodes and edges according to different semantic blocks to obtain a text information graph; Step 103: Convert the nodes into embedded vectors, assign weights to the edges, store the text information graph in the graph database and the vector database, and create indexes; Step 104: Identify and merge the same entity to determine the accuracy of the text information graph.

3. The method for enhancing the summarizing and replying ability of an intelligent question-answering system based on graph clustering according to claim 2, characterized in that: Aggregate using a sliding window and calculate the semantic similarity of adjacent windows using the cosine similarity of vectors. ， Among them, V A A vector representing a text block A; V B A vector representing a text block B; Represents vector V A and vector V B The dot product of Represents vector V A and vector V B The L2 norm of When the semantic similarity between adjacent windows is greater than or equal to the first similarity threshold and when the text block length is less than the first text block length, they are merged into the same semantic block; otherwise, they are divided into two semantic blocks.

4. The method for enhancing the summarization and replying ability of an intelligent question-answering system based on graph clustering according to claim 2, characterized in that: The text information graph is a network consisting of a set of nodes and a set of edges; the nodes represent semantic blocks in the text, including entity nodes and concept nodes. The entity nodes are the key information carriers of the text, and the concept nodes are higher-level key concepts or keywords abstracted from the text. Edges represent the relationships between nodes, including co-occurrence edges and reference / reference edges. The co-occurrence edges connect words, entities, or concepts that appear together in a specific window, and the reference / reference edges connect different expressions that refer to the same entity or concept.

5. The method for enhancing the summarizing and replying ability of an intelligent question-answering system based on graph clustering according to claim 2, characterized in that: All nodes are converted into embedding vectors as node features. A frequency-weighted method is used to assign edge weights. Edge weights reflect the strength of the relationship between nodes. Edge weights are assigned based on co-occurrence frequency or semantic similarity. The text information graph is stored in a graph database and a vector database, and an optimized index is established. For short texts, we use co-occurrence frequency weighting to measure the closeness between two nodes based on the number of co-occurrences: W ij =Count(Co-occurrence of N i ,N j ), Among them, W ij represents the edge weight between node i and node j; N i represents the i-th node that appears in k windows; N j Indicates the jth node that appears in the k window For long texts, semantic similarity weighting is used, which is based on the semantic similarity calculation of two node vectors: W ij =Sim(Node i ,Node j )= , in, The vector representing the i-th node that appears in the k-th window; The vector representing the jth node that appears in the k windows; Represents vector V A and vector V B The L2 norm of .

6. The method for enhancing the summarization and replying ability of an intelligent question-answering system based on graph clustering according to claim 1, characterized in that: Taking the text information graph as input, the clustering algorithm performs iterative calculations based on the edge weights between nodes to divide the text information graph into multiple information communities. The Louvain algorithm is selected to measure the strength of information community division by iteratively optimizing modularity. The larger the Q value, the better the division strength of the information community. ， Among them, A ij is an element of the adjacency matrix, indicating the weight of the edge between nodes i and j or whether there is an edge; Represents the sum of the weights of all edges in the text information graph; represents the weight of node i; c i represents the information community to which node i belongs; δ(c i ,c j ) represents the Kronecker delta function. When ci=cj, δ(c i ,c j ) is 1, otherwise δ(c i ,c j ) is 0.

7. The method for enhancing the summarization and replying ability of an intelligent question-answering system based on graph clustering according to claim 1, characterized in that: The step 3 specifically includes: Step 301: stratify the information communities in the text information graph and generate information summaries; Step 302: extract the keywords of the user's query and search for significant information points in the information community; Step 303: vectorize the user question, calculate the semantic similarity between the user question vector and the information summary vector, and obtain a candidate document set.

8. The method for enhancing the summarizing and replying capability of an intelligent question-answering system based on graph clustering according to claim 7, characterized in that: Utilizing the hierarchical characteristics of the Louvain clustering algorithm, a corresponding information summary is generated for each layer of clustering. The information summary includes a node-level summary and a cluster-level summary. The node-level summary is generated for each basic node in the text information graph; the cluster-level summary is generated for each level of information community identified by the Louvain algorithm, and the lower-level node summary or low-level cluster summary contained in the clustering algorithm is aggregated and abstracted to generate a cluster information summary.

9. The method for enhancing the summarizing and replying capability of an intelligent question-answering system based on graph clustering according to claim 7, characterized in that: After vectorizing the user question, the user question vector v is calculated using cosine similarity with the information summary vector of each level clustering Q and the vector v of each cluster information summary C The similarity between them is selected, and the clustering information summary with a similarity greater than or equal to the first similarity is selected as the paragraph of the candidate document; The paragraphs of the candidate documents are rearranged through a cross encoder, and an intelligent truncation strategy based on statistical distribution is adopted to calculate the candidate score according to the formula: SCORE threshold =μ+γσ, Among them, μ represents the mean score of all candidate paragraphs; σ represents the standard deviation of all candidate paragraph scores; γ represents different paragraphs, γ∈[0.5,1.5], and paragraphs with scores greater than the first candidate score are retained to form the candidate document set.

10. The method for enhancing the summarizing and replying capability of an intelligent question-answering system based on graph clustering according to claim 1, characterized in that: The high-quality, clustering-enhanced candidate document set is used as the context input for retrieval enhancement generation; the powerful generation and context learning capabilities of the large language model are used to further optimize the candidate document set and output the final summary response.

Images