A method for generating long text summaries based on evidence hierarchy graph-enhanced large language models

Abstract

This invention proposes a long text summarization method based on an evidence hierarchy graph-enhanced large language model. Addressing the shortcomings of existing large language models in processing long texts, such as attention dilution, easy disruption of global structure, and the introduction of illusions, this method first divides the long text into blocks, extracts verifiable atomic propositions and retrieves evidence, and constructs a basic proposition evidence graph through a two-stage filtering process. Topic communities are discovered, community summary nodes are generated, and a hierarchical graph structure is constructed. Subsequently, under the objective of a joint comparison and ranking loss function, an enhanced graph attention network is used for self-supervised training to learn node representations containing structural and semantic information. A cross-community Top-k filtering strategy is used for retrieval in a unified vector space, and the retrieval results are combined with the user query and input into the large language model to generate the final summary. This invention effectively restores the global structure of the document, significantly reduces illusions, and improves the quality and efficiency of long text summarization.

Description

Technical Field

[0001] This invention belongs to the field of computer science and technology, and in particular relates to a method for generating long text summaries based on an evidence hierarchy graph-enhanced large language model. Background Technology

[0002] With the rapid development of information technology and the acceleration of digitalization, massive amounts of long-text data have emerged in various application scenarios, such as hours-long meeting minutes, extensive government reports, complex legal documents, and large-scale literary works. Long-text summarization, as a key technology for extracting core insights from these lengthy documents and enabling rapid knowledge navigation, is becoming increasingly important. A robust long-text summarization system not only needs to efficiently answer diverse user queries, covering questions ranging from global topic overviews to specific factual details, but also must avoid computationally expensive full-text duplication.

[0003] In recent years, despite significant progress in long-context understanding using large language models (LLMs), they still face serious challenges when directly handling extremely long inputs. First, when the input length exceeds a certain threshold, large language models often experience attention dilution, causing them to easily overlook key details in long document paragraphs when generating summaries. Second, the inherent generative nature of large models makes them highly susceptible to illusions in the absence of strict factual constraints, leading to the fabrication of information not present in the source document and severely reducing system reliability. Existing retrieval-enhanced generation (RAG) methods mitigate the risk of illusions and improve factual accuracy to some extent by introducing external corpora for retrieval. However, traditional RAG techniques typically rely on shallow semantic similarity-based retrieval filtering mechanisms (i.e., segmenting documents into independent text blocks and calculating vector similarity), which is inherently short-sighted. It merely extracts fragmented text blocks, severing the deep logical connections between sentences and paragraphs, disrupting the global narrative structure and causal dependencies of the original document, resulting in a lack of coherence and a global perspective in the generated summaries.

[0004] To address the limitations of traditional RAG in structural modeling, graph-based RAG methods have recently emerged, attempting to model complex structural dependencies across dispersed information units. However, existing graph architecture methods still suffer from fundamental flaws: On the one hand, some advanced methods rely on unverified, hypothetical query-response pairs spontaneously generated by LLMs to construct graph nodes and edges. This "response generation" model directly uses unverified outputs, potentially containing model illusions, as the core skeleton of the retrieval, inevitably injecting errors into the underlying retrieval index, causing error cascading and propagation. On the other hand, most existing graph structures remain flat, lacking hierarchical abstraction from local details to high-level semantics. This flat design often traps the retrieval engine in local information neighborhoods, making it difficult to effectively reason about complex queries involving high-level topics, long-range dependencies, or global contexts, severely limiting the comprehensiveness and depth of long text summarization in capturing the global context. Summary of the Invention

[0005] The purpose of this invention is to address the problems mentioned in the background art. This invention proposes a long text summarization method based on an evidence hierarchy graph-enhanced large language model. This method constructs a strictly fact-based proposition-evidence hierarchy graph and combines it with an enhanced graph neural network trained with a joint comparison-ranking objective. This achieves structured information retrieval that considers both local details and global structure, and then, combined with a large language model, enables accurate and comprehensive long text summarization.

[0006] To achieve the objective of this invention, this invention provides a method for generating long text summaries based on an evidence hierarchy graph-enhanced large language model, comprising the following steps:

[0007] S1, Proposition Extraction and Underlying Graph Construction: The long input document is divided into blocks, verifiable atomic propositions are extracted, and corresponding supporting evidence is retrieved. After filtering and evaluation, an underlying proposition-evidence graph is constructed to ensure the semantic reliability and factual basis of the graph nodes.

[0008] S2, Hierarchical Structure Enhancement and Community Summarization: Based on the underlying proposition-evidence graph, a personalized PageRank (PPR) algorithm is used to calculate node influence and detect topic communities. For each community, core topics are extracted and community summary nodes are generated, thereby constructing a hierarchical graph structure that combines micro-facts and macro-semantics.

[0009] S3, Self-Supervised Training and Node Representation Learning: Fine-grained self-supervised signals are constructed based on the similarity ranking of source text blocks and hierarchical graph nodes. Node feature learning is performed using an Enhanced Graph Attention Network (GAT) optimized with a joint contrast-ranking loss function, enabling node representations to integrate local factual semantics and global graph topology.

[0010] S4, Hierarchical Graph Joint Retrieval: Ordinary propositional nodes and community summary nodes are placed in a unified dense vector space. Given a summary query, its similarity to all graph nodes is calculated, and a cross-community filtering strategy is used to retrieve diverse and non-redundant relevant node data.

[0011] S5, Large Language Model Summarization: The retrieved node information is mapped back to natural language text and concatenated with the user's summary query in context. The result is then input into the large language model to generate a coherent, accurate, and globally-oriented long text summary.

[0012] Furthermore, the specific steps for proposition extraction and underlying graph construction in S1 include:

[0013] The original long document is segmented into semantically coherent text blocks. A large language model is used to extract atomic propositions from each text block. A dense retrieval model is used to retrieve the corresponding supporting text blocks as evidence for each proposition.

[0014] A two-stage filtering strategy is applied to the extracted propositions: the first stage filters out redundant or malformed propositions through rules; the second stage adopts LLM-as-a-Judge, that is, using a large language model as a fact judge to strictly evaluate the factual consistency, relevance and clarity of the propositions, and only retains propositions with high confidence.

[0015] Furthermore, the specific process of using the personalized PageRank algorithm to enhance the hierarchical structure in S2 is as follows:

[0016] For the source node in the graph, calculate its personalized PageRank (PPR) vector. The iterative calculation formula for the PPR vector is as follows: ;

[0017] in, It is a personalized PageRank steady-state vector obtained by convergence under the iterative formula, where each dimension represents the probability of reaching the corresponding node from a random walk starting from the source node; This is the probability parameter for stealth teleportation, and its value range is... ; For the source node The one-hot vector (i.e., a vector that is 1 only at the corresponding node position and 0 at all other positions); The matrix is ​​a columnar random state transition matrix; This represents the source node in the network graph. Initially, the PPR vector... Equal to the unique heat origin vector Iterate repeatedly until... The convergence stops, resulting in the final PPR vector.

[0018] Based on the calculated PPR scores, the nodes are ranked by row summation, and the nodes with the highest influence are selected as seed nodes. The remaining nodes are then assigned to the seed nodes that have the greatest influence on them, forming multiple densely connected topic communities.

[0019] For each community, the content of its internal nodes is aggregated, and a summary node representing the central theme of the community is generated using a large language model, thereby realizing the transformation from a flat graph to a hierarchical graph.

[0020] Furthermore, the specific process of self-supervised training and node representation learning in S3 includes:

[0021] The similarity between each source text block and all nodes in the hierarchical graph is calculated using pre-trained language evaluation metrics, generating a relative ranking list of nodes, which serves as a micro-supervision signal without manual annotation.

[0022] An enhanced graph attention network, incorporating multi-head attention, residual connections, layer normalization, and Jumping Knowledge pooling aggregation strategies, is used to encode graph nodes. The network training process employs a joint contrastive-ranking loss function for end-to-end optimization. The formula for this loss function is as follows: .

[0023] in, This represents the overall joint loss function; This represents the contrastive learning loss, used to distinguish between positive and negative samples; This represents the boundary-based triplet loss, used as a hard negative sample penalty term; This represents the focus ranking loss, used to preserve the relative order among negative samples; This represents the hyperparameter weighting coefficients used to balance the contrastive loss and the triplet loss; This represents a ranking weight function that dynamically increases with each training round. This indicates the current training epoch.

[0024] Furthermore, the hierarchical graph joint retrieval strategy in S4 is as follows:

[0025] The input query is encoded using the same dense feature extractor; the cosine similarity between the query vector and all nodes (including proposition nodes and community summary nodes) in the unified vector space is calculated;

[0026] When selecting the top-k most relevant nodes, a cross-community filtering mechanism is introduced: by checking the PPR community label of candidate retrieval nodes, a penalty score is applied to overrepresented nodes from the same community, forcing the retrieval system to extract relevant evidence across different topic clusters, so as to ensure that the retrieval results have global coverage and significantly reduce information redundancy.

[0027] Furthermore, the specific process of generating large language model summaries in S5 includes:

[0028] The retrieved graph nodes are mapped back to their original text form and then aggregated to build a context.

[0029] In particular, when the search results contain community summary nodes, the pre-calculated community summary text is directly obtained and placed in the context of the generator, thereby highly compressing the massive amount of information from multiple underlying source text blocks into concise paragraphs that are easy to process.

[0030] Finally, the aggregated contextual text is concatenated with the user's summary query and input into the large language model for inference. This process fully leverages the local and global relevance provided by cross-community retrieval, guiding the large language model to generate a highly coherent, comprehensive, and well-structured final summary.

[0031] The significant advancement of this invention compared to existing technologies lies in:

[0032] (1) Significantly reduces the propagation of illusions and improves the reliability of underlying retrieval semantics. This invention abandons the inherent defect of existing graph RAG methods that rely on large models to generate unverified "hypothetical responses" to build graphs, and innovatively constructs a "proposition-evidence graph" with facts as the core. Through two-stage filtering (rules combined with large language model adjudication), rigorous atomic propositions are extracted, eliminating the gap in model fact verification and fundamentally preventing the cascading propagation of illusions in the graph index structure and subsequent retrieval mechanisms.

[0033] (2) Restoring the global narrative structure and overcoming the limitations of fragmented local perspectives. Addressing the contextual breakage problem caused by fragmented block retrieval in traditional RAG, this invention introduces personalized PageRank (PPR) clustering and community summary node mechanisms. This mechanism successfully transforms the flat information network into a high-level structure, enabling the large model's retrieval system to both "magnify" and accurately locate micro-level factual details, and "shrink" to grasp the overall thematic context of long documents, effectively alleviating the attention dilution phenomenon in ultra-long text processing.

[0034] (3) Deep structural alignment significantly improves the quality of long text summarization and retrieval inference efficiency. This invention designs an enhanced graph attention network optimized for long texts and proposes a self-supervised joint contrast-ranking training objective based on text block micro-ranking. The node feature representation deeply integrates local semantics and global topological structure, and together with the cross-community Top-k joint retrieval strategy, ensures that the contextual prompts provided to the large language model are both highly relevant and comprehensive. Compared with direct input of full text and existing retrieval enhancement baselines, it significantly improves the ROUGE index of generated summaries and online retrieval efficiency.

[0035] To more clearly illustrate the functional characteristics and structural parameters of the present invention, further explanation is provided below in conjunction with the accompanying drawings and specific embodiments. Attached Figure Description

[0036] Figure 1 This is an overall flowchart of a long text summarization method based on an evidence hierarchy graph-enhanced large language model, as described in this invention.

[0037] Figure 2 This is a schematic diagram of the proposition extraction and initial graph construction process described in this invention.

[0038] Figure 3 This is a schematic diagram illustrating the use of PPR clustering for hierarchical structure enhancement and community summary node generation as described in this invention.

[0039] Figure 4 This is a schematic diagram illustrating the calculation of BERTScore scores during the self-supervised signal generation process described in this invention.

[0040] Figure 5 This is a flowchart of the joint comparison-ranking loss calculation and self-supervised training described in this invention.

[0041] Figure 6 This is a schematic diagram illustrating node representation learning and graph feature enhancement using an improved graph attention network as described in this invention.

[0042] Figure 7 This is a schematic diagram of the hierarchical graph joint retrieval and large language model summary generation process described in this invention. Detailed Implementation

[0043] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. 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.

[0044] Implementation Environment Description: The algorithm in this embodiment of the invention is implemented based on the PyTorch deep learning framework and the Deep GraphLibrary (DGL) graph neural network library;

[0045] The model training and inference process is completed on a compute node equipped with a single NVIDIA GeForce RTX 4090 GPU (24GB of video memory).

[0046] In step S1, the specific implementation details of proposition extraction and underlying graph construction are as follows:

[0047] Step S11, perform semantic segmentation of long text: To adapt to the input constraints of dense retrieval models and preserve local context, the input long document is segmented using LangChain's TokenTextSplitter tool; specifically, each text block is set to 256 tokens, and the overlapping portion between adjacent text blocks is set to 32 tokens; the text block set constitutes the document. ,in This represents the total set of text blocks after full-text segmentation; to Representing the first to A text block, This indicates the total number of text blocks obtained after a long document is split.

[0048] Step S11 involves proposition extraction and evidence retrieval. A large language model (such as GPT-4) is used to extract atomic propositions from each text block. After extensive experiments and performance trade-off analysis, the optimal threshold for the number of propositions extracted per document is set to 30. Subsequently, a pre-trained dense retrieval model, Contriever, is used to retrieve the top-6 most relevant text blocks for each proposition as supporting evidence, thereby forming the initial proposition-evidence pairs.

[0049] In step S13, a two-stage filtering strategy is implemented:

[0050] The first stage is rule-based heuristic filtering, which aims to quickly eliminate low-quality propositions with low computational cost.

[0051] The execution rules include: 1. Length limit filtering: removing propositions whose text length, i.e., the number of tokens, is less than the set minimum threshold (e.g., less than 5 tokens, which usually lacks substantial information) or greater than the maximum threshold (e.g., more than 50 tokens, which violates the atomicity principle of propositions); 2. Syntactic structure filtering: using natural language processing syntactic analysis tools to remove incomplete sentences lacking a clear subject, as well as complex compound sentences containing multiple coordinating conjunctions (such as "and", "however"), ensuring that the retained propositions are "single declarative sentences"; 3. Literal deduplication filtering: calculating the Jaccard similarity between the newly extracted proposition and the set of extracted propositions. If the similarity is greater than the preset redundancy threshold (set to 0.85 in this embodiment), it is determined to be a redundant proposition and is removed.

[0052] The second stage introduces the "LLM-as-a-Judge" paradigm for in-depth fact-finding and quality verification. Since extensive experiments and sensitivity analyses have determined that retaining 30 high-quality propositions from a single long document is the optimal configuration for downstream generation, precise filtering in the second stage is crucial. The specific implementation steps are as follows: A structured evaluation prompt is constructed, using the retrieved original source text blocks as factual premises and the propositions retained in the first stage as hypotheses to be verified, both input into the large language model for evaluation, allowing the model to score. The prompt requires the large model to score within a discrete range of 1-5 points, according to the following five dimensions... Independent scoring is performed on the following dimensions: 1. Factual accuracy: Whether the proposition content is completely supported by the original text block without tampering; 2. Verifiability: Whether the proposition contains specific entities or clear events for retrieval; 3. Completeness: Whether the proposition is semantically independent and complete after being removed from the context; 4. Clarity: Whether there are ambiguous referents (such as containing isolated pronouns like "he" or "that"); 5. Specificity: Whether the description is sufficiently detailed rather than vague. The average score of these five dimensions output by the large language model is calculated to obtain a comprehensive score. Finally, only proposition nodes with a comprehensive score greater than or equal to the preset reliability threshold (set to 4.0 in this embodiment) are retained.

[0053] In step S2, the specific calculation and implementation process of hierarchical structure enhancement and community summarization is as follows:

[0054] Step S21, calculate Personalized PageRank (PPR): A flat proposition-evidence graph cannot capture high-level topics. Therefore, for any source node in the graph... ( (For the identifier of a single node in the graph), calculate the PPR vector. ,in Represents the set of real numbers. This represents the total number of nodes in the graph. That is, a dimension is A real-valued vector is used to diffuse node influence and identify densely connected topic communities; the iterative equation is defined as: ;in, To determine the probability of stealth teleportation, this embodiment sets... ; It is a node The unique heat vector; The dimension is The column random transition matrix, The maximum number of PPR iterations is set to 100 to form the personalized PageRank steady-state vector after convergence.

[0055] Step 22, Community segmentation and summary generation: Based on the calculated PPR score, calculate the row and influence of the nodes and sort them in descending order;

[0056] The top-ranked nodes are selected as core seeds, and the remaining nodes are assigned to the seed nodes that exert the strongest influence on them, thus completing the division of non-overlapping topic communities. Then, for each divided topic community, the text content of all nodes within the community is extracted and concatenated to construct a contextual prompt box containing instructions (e.g., "Based on the following interrelated propositional evidence, summarize the core concepts and main events of this topic community, with a word limit of 100 words"). This prompt is input into a large language model (such as GPT-4), and the model's output summary text is instantiated as a new "community summary node." This summary node is connected to all lower-level propositional nodes within the community, thereby aggregating information from lower-level concrete facts to higher-level abstract topics, successfully transforming the originally flat proposition-evidence graph into a graph with a hierarchical routing structure.

[0057] In step S3, the specific computation process of self-supervised training and node representation learning is as follows:

[0058] Step S31, generate self-supervised pseudo-tags: Since long documents lack fine-grained node-level annotations, a self-alignment strategy is adopted to treat propositions as micro-summaries of the text blocks in which they are located.

[0059] Calculate text blocks , Indicates the first A text block, with a hierarchy diagram The BERTScore similarity between all nodes is used to generate a similarity ranking list. ;in This represents the constructed graph structure. This represents the set of nodes in the graph. Represents the set of edges in the graph; Indicates a text block The generated list of nodes in relative order. This represents the positive sample node with the highest similarity score to the corresponding text block, ranking first. This indicates the second-ranked negative sample node. It means the last one, that is, the number. negative sample nodes of position 1 This represents the total number of candidate nodes participating in the ranking; the list provides an accurate ranking basis for subsequent training.

[0060] Step S32, Enhanced GAT network structure: Initialize all nodes using a fixed 768-dimensional Contriever vector;

[0061] To prevent excessive smoothing in deep networks, a two-layer enhanced graph attention network is constructed; the hidden layer dimension of the network is 768, and four attention heads are used. The dropout rate varies from 0.1 to 0.5 depending on the dataset configuration; in the... Layers, nodes The feature update process is as follows:

[0062] ;

[0063] in, Represents a node In the Hidden feature vectors of the layer; Indicates to The output of each attention head is used to perform feature concatenation. This represents the total number of heads receiving multi-head attention. Indicates the first One point of attention; Represents a non-linear activation function; This indicates a summation calculation; Represents a node The set of first-order neighbor nodes; Indicates the index of the neighboring node; Indicates the first Under each attention head, the node For neighboring nodes Attention weight coefficient; Indicates the first Each attention head corresponds to a learnable linear feature transformation weight matrix; Representing neighboring nodes In the Hidden feature vectors of the layer;

[0064] ;

[0065] This includes residual connectivity and layer normalization. The layer normalization function is represented. This represents the linear projection weight matrix used for dimension alignment; in the final layer, a max-pooling strategy using Jumping Knowledge is employed to fuse multi-scale features. ;in, This represents the feature representation vector that a node outputs after forward propagation through the network. This represents the element-wise max-pooling operation; This represents the input characteristics for node initialization; to These represent the layers 1 to 2 of the network, respectively. Intermediate features of the layer output; This represents the total number of layers in the graph attention network.

[0066] Step S33: Optimization is performed using joint contrast-ranking loss. The network is only fine-tuned for the GAT projection matrix and temperature parameters. The Adam optimizer is used in conjunction with a cosine annealing learning rate scheduler. The initial learning rate is 1e-3, the batch size is 32, and the training lasts for 300 epochs.

[0067] To fully utilize the sorted list signal, a joint loss function was designed. ;in, Represents the total loss function; This is an InfoNCE-style contrastive loss used to distinguish between positive and negative samples; Represents the weighting coefficients that balance the loss from hard negative samples; The boundary-based triplet loss represents the hard negative sample penalty term; This represents a ranking weight function that dynamically increases with each training round. This indicates the current training epoch. For pairwise focus sorting loss;

[0068] The contrastive learning loss is an InfoNCE-style loss used to distinguish between positive and negative samples; the method for determining the contrastive loss in a single training batch is as follows:

[0069] ;

[0070] in, This represents the contrastive learning loss function; the negative sign indicates that the goal of the model optimization is to minimize this loss value. This represents the total number of all propositions in a single training batch; This represents the set of propositional summaries that span all levels of the graph within a single training batch; for the set Each proposition in Perform a traversal and summation operation. Indicates the first Feature representation vectors for each proposition summary query; Represents the natural logarithm function; for sets All negative sample nodes Perform a traversal and summation operation. This indicates a correspondence to a proposition. The set consisting of all negative sample nodes; express A specific negative sample node in the set is represented by a vector; the goal of contrastive learning is to maximize the similarity between the summary query and its corresponding positive sample node, while minimizing its similarity with all negative sample nodes.

[0071] The basic similarity is based on a learnable temperature coefficient. dot product ;in, Indicates query With nodes The scaling index similarity score; Represents an exponential function with the natural constant as its base; Represents the cosine similarity function; Indicates the first Vector representation of each query; A vector representation of the target node; This represents a learnable temperature coefficient variable;

[0072] Triple boundary loss introduced for hard-negative mining strategies:

[0073] ;

[0074] in, This indicates the total number of query samples in the current training batch; Represents the query set for the current batch; Sum of all queries in the set. This represents the activation function of the Rectified Linear Unit; Represents the mined and queried The nearest and most easily confused hard negative sample node vector; Representation and Query The matched positive sample node vector; This represents a pre-defined safety interval boundary value. In this embodiment, the boundary value is set as follows: ;

[0075] The pairwise focus ranking loss is used to force the feature distance to highly restore the relative order of semantic similarity based on BERTScore pre-computed, thereby achieving deep semantic structure alignment.

[0076] ;

[0077] in, This represents the total number of sample pairs constructed. Represents a set of sample pairs; This represents a pair of node indexes with a ranking relationship; for all Pairwise summation For the focus adjustment factor parameter, in this embodiment ; This represents the Sigmoid activation function; and These represent the nodes predicted by the model. With nodes The relevance score; The focus attenuation index parameter is used in this embodiment. ; Represent the natural logarithm function; Here, it represents the base of the natural logarithm.

[0078] In step S4, the specific process of hierarchical graph joint retrieval is as follows:

[0079] Step S41, Unified Vector Space: To ensure inference efficiency and preserve text mapping, the graph structure is maintained synchronously in two forms: a NetworkX instance for text storage mapping and a DGL graph instance for efficient tensor computation; both ordinary proposition nodes and community summary nodes are mapped and indexed in the same dense vector space.

[0080] Step S42, cross-community Top-k filtering retrieval: Given a natural language summary query, the query is encoded into a query vector using a frozen Contriever model. Then, the cosine similarity between the query vector and the embeddings of all learned nodes in the unified vector space is calculated. When extracting the top-k relevant nodes, a cross-community filtering strategy is enabled to avoid information homogenization.

[0081] Introducing a dynamic penalty function: Assuming candidate nodes Belongs to a themed community Then the final search ranking score for:

[0082] ;

[0083] in, For query With nodes The initial cosine similarity; The redundancy penalty coefficient is set in advance (0.05 in this embodiment). This indicates that they already belong to the same community in the current Top-k candidate list. The cumulative number of times a node has been selected; the retrieval system uses the final score as the basis. Perform dynamic descending sorting and node selection; if a topic cluster is over-extracted, the scores of its remaining candidate nodes will decrease accordingly. As the number of elements increases, increasingly heavier penalties are imposed, thereby forcing the retrieval branches to actively extend to different topic clusters, ensuring that the evidence input into the large model is neither redundant nor lacks macro-global coverage.

[0084] In step S5, the specific process for generating the large language model summary is as follows:

[0085] Step S51, Reverse mapping and compression of nodes to text: Map the selected nodes back to the original text content through NetworkX;

[0086] When a community summary node is retrieved, the system extracts the pre-generated text and places it at the front end of the context generation area of ​​the large language model. This mechanism effectively compresses the massive amount of scattered information distributed in multiple source text blocks into structured paragraphs that can be directly processed.

[0087] Step S52, joint context generation, combines the restored and compressed node text paragraphs with the user's original summary query command to construct the final prompt; this is then input into the inference big language model, which generates a coherent long text summary that is faithful to the original text and has a global structural perspective.

[0088] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method for generating long text summaries based on an evidence hierarchy graph-enhanced large language model, characterized in that, The method includes the following steps: S1, Proposition Extraction and Underlying Graph Construction: The long document is segmented into text blocks, verifiable atomic propositions are extracted using a large language model, and supporting evidence fragments are retrieved for each proposition. After a two-stage filtering strategy, the underlying proposition-evidence graph is constructed. S2, Hierarchical Structure Enhancement and Community Summary Formation: Based on the proposition-evidence graph, the personalized PageRank (PPR) algorithm is applied to calculate node influence, clustering to form topic communities, and a community summary node is generated for each community using a large language model to construct an evidence graph with a hierarchical structure. S3, Self-supervised training and node representation learning: Design an enhanced graph attention network (GAT) and perform self-supervised training with joint contrastive ranking loss as the optimization objective, so that node representations integrate local evidence features and global structural semantics; S4, Hierarchical Graph Joint Retrieval: In a unified vector space, calculate the similarity between the user query and the hierarchical graph nodes containing propositional nodes and summary nodes, combine the cross-community Top-k filtering strategy to retrieve relevant nodes, and map them back to the corresponding text content; the hierarchical graph nodes contain propositional nodes and summary nodes; S5, Large Language Model Summarization: The retrieved text content is concatenated with the user query and input into the large language model to generate the final long text summary.

2. The method of claim 1, wherein, Proposition extraction and underlying graph construction include: Step S11, perform long text semantic segmentation: segment the input long document; specifically configured as follows: the size of each text block is set to 256 tokens, and the overlapping part between adjacent text blocks is set to 32 tokens; the text block set that constitutes the document. ,in This represents the total set of text blocks after full-text segmentation; to Representing the first to A text block, This represents the total number of text blocks obtained after a long document is split. Step S12: Proposition extraction and evidence retrieval are performed. Atomic propositions are extracted from each text block using a large language model. The optimal threshold for the number of propositions extracted from each document is set to 30. Subsequently, the pre-trained dense retrieval model Contriever is used to retrieve the top-6 most relevant text blocks for each proposition as supporting evidence, thereby forming the initial proposition-evidence pair. In step S13, a two-stage filtering strategy is implemented: The first stage is rule-based heuristic filtering, which includes the following rules:

1. Length limit filtering: removing propositions whose text length (i.e., the number of tokens) is less than the set minimum threshold or greater than the set maximum threshold; 2. Syntactic structure filtering: using natural language processing syntactic analysis tools to remove incomplete sentences lacking a clear subject, as well as complex compound sentences containing multiple coordinating conjunctions; 3. Literal deduplication filtering: calculating the Jaccard similarity between newly extracted propositions and the set of extracted propositions. If the similarity is greater than the preset redundancy threshold, it is determined to be a redundant proposition and removed. The second stage introduces the "LLM-as-a-Judge" paradigm for in-depth fact and quality verification. A structured evaluation prompt is constructed, using the retrieved original source text blocks as factual premises and the propositions retained in the first stage as hypotheses to be verified. Both are input into the large language model for evaluation, where the model scores them. The prompt requires the large model to score independently within a discrete range of 1-5 points according to the following five dimensions:

1. Factual accuracy: Whether the proposition content is completely supported by the original text block without alteration; 2. Verifiability: Whether the proposition contains specific entities or clear events for retrieval; 3. Completeness: Whether the proposition is semantically independent and complete after being removed from its context; 4. Clarity: Whether there is any ambiguity in referents; 5. Specificity: Whether the description is sufficiently detailed rather than vague. The average score of these five dimensions output by the large language model is calculated to obtain a comprehensive score. Finally, only proposition nodes with a comprehensive score greater than or equal to a pre-set reliability threshold are retained.

3. The method of claim 2, wherein, The hierarchical structure is enhanced and forms a community summary, including: Step S21, Calculate Personalized PageRank (PPR): For any source node in the graph Calculate the PPR vector ,in Represents the set of real numbers. This represents the total number of nodes in the graph. That is, a dimension is A real-valued vector is used to diffuse node influence and identify densely connected topic communities; the iterative equation is defined as: ;in, The probability of stealth teleportation; It is a node The unique heat vector; The dimension is The column random transition matrix, The maximum number of PPR iterations is set to 100 to form the personalized PageRank steady-state vector after convergence. Step 22, Community segmentation and summary generation: Based on the calculated PPR score, calculate the row and influence of the nodes and sort them in descending order; The top-ranked nodes are selected as core seeds, and the remaining nodes are assigned to the seed nodes that exert the strongest influence on them, thus completing the division of non-overlapping topic communities. Then, for each divided topic community, the text content of all nodes in the community is extracted and the strings are concatenated to construct a context prompt box containing instructions. The prompt is input into the large language model, and the summary text output by the model is instantiated as a new "community summary node".

4. The method of claim 3, wherein, Self-supervised training and node representation learning include: Step S31, generate self-supervised pseudo-labels: adopt a self-alignment strategy to treat propositions as micro-summaries of the text blocks in which they are located; Calculate text blocks , Indicates the first A text block, with a hierarchy diagram The BERTScore similarity between all nodes is used to generate a similarity ranking list. ;in This represents the constructed graph structure. This represents the set of nodes in the graph. Represents the set of edges in the graph; Indicates a text block The generated list of nodes in relative order. This represents the positive sample node with the highest similarity score to the corresponding text block, ranking first. This indicates the second-ranked negative sample node. It means the last one, that is, the number. negative sample nodes of position 1 This represents the total number of candidate nodes participating in the ranking; the list provides an accurate ranking basis for subsequent training. Step S32, Enhanced GAT network structure: Initialize all nodes using a fixed 768-dimensional Contriever vector; To prevent excessive smoothing in deep networks, a two-layer enhanced graph attention network is constructed; the hidden layer dimension of the network is 768, and four attention heads are used. The dropout rate varies from 0.1 to 0.5 depending on the dataset configuration; in the... Layers, nodes The feature update process is as follows: ； in, Represents a node In the Hidden feature vectors of the layer; Indicates to The output of each attention head is used to perform feature concatenation. This represents the total number of heads receiving multi-head attention. Indicates the first One point of attention; Represents a nonlinear activation function; This indicates a summation calculation; Represents a node The set of first-order neighbor nodes; Indicates the index of the neighboring node; Indicates the first Under each attention head, the node For neighboring nodes Attention weight coefficient; Indicates the first Each attention head corresponds to a learnable linear feature transformation weight matrix; Representing neighboring nodes In the Hidden feature vectors of the layer; ； This includes residual connectivity and layer normalization. Indicates the layer normalization function; This represents the linear projection weight matrix used for dimension alignment; in the final layer, a max-pooling strategy using Jumping Knowledge is employed to fuse multi-scale features. ;in, This represents the feature representation vector that a node outputs after forward propagation through the network. This represents the element-wise max pooling operation; This represents the input characteristics for node initialization; to These represent the layers 1 to 2 of the network, respectively. Intermediate features of the layer output; This represents the total number of layers in the graph attention network; Step S33: Optimization is performed using joint contrast-ranking loss. The network is only fine-tuned for the GAT projection matrix and temperature parameters. The Adam optimizer is used in conjunction with a cosine annealing learning rate scheduler. The initial learning rate is 1e-3, the batch size is 32, and the training lasts for 300 epochs. A joint loss function was designed. ;in, Represents the total loss function; This is an InfoNCE-style contrastive loss used to distinguish between positive and negative samples; Represents the weighting coefficients that balance the loss from hard negative samples; Represents the boundary-based triplet loss for the hard negative sample penalty term; This represents a ranking weight function that dynamically increases with each training round. This indicates the current training round; For pairwise focus sorting loss; InfoNCE style contrastive learning loss is used to distinguish positive and negative samples. The determination method of the contrastive loss on a single training batch is as follows: ； in, This represents the contrastive learning loss function; the negative sign indicates that the goal of the model optimization is to minimize this loss value. This represents the total number of all propositions in a single training batch; This represents the set of propositional summaries that span all levels of the graph within a single training batch; for the set Each proposition in Perform a traversal and summation operation. Indicates the first Feature representation vectors for each proposition summary query; Represents the natural logarithm function; for sets All negative sample nodes Perform a traversal and summation operation. This indicates a correspondence to a proposition. The set consisting of all negative sample nodes; express A specific negative sample node in the set is represented by a vector; The basic similarity is based on a learnable temperature coefficient. dot product ;in, Indicates query With nodes The scaling index similarity score; Represents an exponential function with the natural constant as its base; Represents the cosine similarity function; Indicates the first Vector representation of each query; A vector representation of the target node; This represents a learnable temperature coefficient variable; Triplet margin loss introduced for hard negative mining strategy: ； in, This indicates the total number of query samples in the current training batch; Represents the query set for the current batch; for Sum of all queries in the set. This represents the activation function of the Rectified Linear Unit; Represents the mined and queried The nearest and most easily confused hard negative sample node vector; Representation and Query The matched positive sample node vector; This represents the pre-defined safety interval boundary value; To sort the pair of focus loss, force the feature distance to highly restore the semantic similarity based on BERTScore pre-computed relative order relationship, realize the alignment of deep semantic structure: ； in, This represents the total number of sample pairs constructed. Represents a set of sample pairs; This represents a pair of node indexes with a ranking relationship; for all Pairwise summation This refers to the focus adjustment factor parameter; This represents the Sigmoid activation function; and These represent the nodes predicted by the model. With nodes The relevance score; This refers to the focal decay index parameter; Represent the natural logarithm function; Here, it represents the base of the natural logarithm.

5. The method of claim 4, wherein, Hierarchical graph joint retrieval includes: Step S41, Unified Vector Space: The graph structure is synchronously maintained in two forms: a NetworkX instance for text storage mapping and a DGL graph instance for efficient tensor computation; both ordinary proposition nodes and community summary nodes are mapped and indexed in the same dense vector space; Step S42, cross-community Top-k filtering retrieval, given a natural language summary query, use the frozen Contriever model to encode into a query vector, then calculate the cosine similarity between the query vector and all learned node embeddings in the unified vector space ; In extracting Top-k relevant nodes, to avoid information homogenization, a cross-community filtering strategy is enabled; Introducing a dynamic penalty function: Assuming candidate nodes Belongs to a themed community Then the final search ranking score for: ； in, For query With nodes The initial cosine similarity; The pre-set redundancy penalty coefficient; This indicates that they already belong to the same community in the current Top-k candidate list. The cumulative number of times a node has been selected; the retrieval system uses the final score as the basis. Perform dynamic descending sorting and node selection; if a topic cluster is over-extracted, the scores of its remaining candidate nodes will decrease accordingly. As the number of elements increases, increasingly heavier penalties are imposed, thereby forcing the retrieval branches to actively extend to different topic clusters, ensuring that the evidence input into the large model is neither redundant nor lacks macro-global coverage.

6. The method of claim 5, wherein, Large language model summarization generation, including: Step S51, Reverse mapping and compression of nodes to text: Map the selected nodes back to the original text content through NetworkX; When a community summary node is retrieved, the system extracts the pre-generated text and places it at the front end of the large language model context generation area; Step S52: Joint context generation, which combines the restored and compressed node text paragraphs with the user's original summary query command to construct the final prompt; this is then input into the inference big language model, which generates a long text summary based on it.

Images