Literature summary analysis method based on natural language processing

Abstract

The invention relates to the field of natural language processing, in particular to a literature summary analysis method based on natural language processing, which designs a multi-level context extension mechanism, constructs a rich, three-dimensional and multi-angle information environment for each user query, and lays a solid foundation for generating a comprehensive and deep abstract. According to the method, highly-customized abstract generation is realized by constructing intelligent processing of query perception, the actual information demand of a user is deduced through multi-level analysis, query dynamic adjustment is introduced to ensure that summarized abstract content closely surrounds a query core, and a granularity self-adaptive control mechanism is designed to meet abstract demands of different granularities; according to the method, a multi-engine mixed retrieval system and an intelligent information fusion framework are constructed, semantic related content and associated information are found by running three different types of retrieval engines in parallel, metadata such as quality indexes, timeliness and authority of documents are screened, the problem of incomplete information coverage is solved, and the quality of document summary generation is optimized.

Description

Technical Field

[0001] This invention belongs to the field of natural language processing, specifically a literature summary and analysis method based on natural language processing. Background Technology

[0002] With the exponential growth of scientific research, researchers face a massive amount of academic literature. It is estimated that in the biomedical field alone, more than one million new papers are published each year. Traditional manual reading and abstract writing methods are inefficient and cannot keep up with the speed of knowledge production. Therefore, there is an urgent need for automated and efficient information condensation tools to help scholars quickly filter, understand, and track dynamics in the field. However, existing literature summarization techniques often suffer from problems such as insufficient contextual information, incomplete information coverage, and insensitivity to queries, resulting in poor quality of generated literature summaries. Summary of the Invention

[0003] To address the aforementioned issues and overcome the shortcomings of existing technologies, this invention provides a literature summary and analysis method based on natural language processing. Addressing the problem of insufficient contextual information in existing literature summary techniques, this invention improves upon this by designing a multi-level context expansion mechanism, including core entity text retrieval, weak context expansion, and strong context expansion. Through these multi-level, intelligent context expansion mechanisms, a rich, comprehensive, and multi-faceted information environment is constructed for each user query, laying a solid foundation for generating comprehensive and in-depth summaries. Addressing the query insensitivity problem of existing literature summary techniques, this invention achieves highly customized summary generation by constructing query-aware intelligent processing. It infers the user's actual information needs through multi-level analysis and systematically divides query intent into multiple dimensions. This multi-dimensional intent analysis provides precise guidance for subsequent customized processing. To ensure accurate guidance, this invention introduces dynamic adjustments to query relevance, guaranteeing that the summary content closely revolves around the core query. A granularity adaptive control mechanism is designed to automatically determine the required summary granularity, meeting users' needs for summaries at different granularities. Addressing the issue of incomplete information coverage in existing literature summary technologies, this invention achieves complete information coverage by constructing a multi-engine hybrid retrieval system and an intelligent information fusion framework. It runs three different types of retrieval engines in parallel, finding semantically related content through vector similarity, identifying associated information through graph traversal algorithms, and using a metadata-based conditional retrieval engine to filter documents based on quality indicators, timeliness, authority, and other metadata. The results from the three engines are not simply merged but intelligently integrated through a learned fusion model, thereby solving the problem of incomplete information coverage and ultimately further optimizing the quality of generated literature summaries.

[0004] This invention provides a literature summary and analysis method based on natural language processing, which specifically includes the following steps:

[0005] Step S1: Multi-granularity document segmentation. An adaptive segmentation strategy is used to segment the document article into blocks, resulting in n text blocks. Different segmentation strategies correspond to different numbers of text blocks.

[0006] Step S2: Hierarchical text block relationship extraction. The importance score of a text block is obtained by weighting its salience, centrality, and novelty scores within the article. ;

[0007] In the formula, , and The representative weight parameter satisfies that the sum of the three equals 1. Represents a text block. The TF-IDF score represents the saliency of a text block within the article. The centrality score represents the text block. The novelty score of the text block is calculated based on the time decay function. This represents the importance score of the text block;

[0008] Step S3: Dynamic text block aggregation. A dynamic merging threshold is introduced to adaptively merge text blocks whose semantic similarity is within a preset threshold range. The dynamic merging threshold is determined by the maximum number of text blocks, the current total number of text blocks, and the semantic similarity of the text blocks. The semantic similarity is the similarity between the semantic features of the text blocks. The semantic features are extracted by a neural network. The formula for calculating the dynamic merging threshold is as follows: ;

[0009] In the formula, Represents the dynamic merging threshold. This represents the semantic similarity between text block i and text block j. To adjust the parameters, Represents the maximum number of text blocks. Represents the total number of current text blocks;

[0010] Step S4: Graph construction. For each merged text block, entities and relations are extracted using a large language model. All entities and their relations constitute a knowledge graph.

[0011] Step S5: Adaptive Context-Expanded Retrieval. Adaptive context-expanded retrieval processing is performed on the knowledge graph to obtain multimodal retrieval results. This specifically includes the following steps:

[0012] Step S51: Query generation. Based on the knowledge graph, initialize the query entity Q.

[0013] Step S52: Calculate query complexity. Based on the query entity density, query ambiguity score, and query length, calculate the query complexity. ;

[0014] In the formula, Represents the normalized query length. The entity density in the query is represented by the number of entities divided by the total number of words. The query ambiguity score represents the diversity of queries calculated based on entity type. =1, which is the preset weight parameter. This represents the query complexity;

[0015] Step S53: Adaptive context-based retrieval expansion, generating dynamic expansion strategies based on query complexity: ;

[0016] In the formula, and Represents the complexity threshold. Represents the number of dynamically expanded steps. This means retrieving only entities directly related to the query. This represents the first-level extension, from weak context to strong context. Represents multi-level expansion, with a depth of , This is the dynamic expansion strategy, where the dynamic expansion steps are calculated as follows: ;

[0017] In the formula, Represents the preset base extension steps. Represents the preset expansion coefficient;

[0018] Step S54: Hybrid strategy retrieval. Perform hybrid strategy retrieval on all entities using vector similarity, graph structure, and metadata enhancement. Optimize the union of the retrieval results from these three methods using Rerank technology to obtain the final retrieval results. ; ; ; ;

[0019] In the formula, This represents the number of entities returned by vector retrieval. The Top-k represents the entities with the highest similarity among their semantic features. This represents the search results based on vector similarity. This represents random walk retrieval based on a knowledge graph, where t represents the number of random walk steps and p represents the restart probability. This represents the search results based on graph structure retrieval. Metadata-enhanced retrieval, The threshold represents the importance of the filtering. This represents search results enhanced based on metadata. This is the final search result;

[0020] Step S55: Cross-modal information fusion, extracting multimodal features, aligning the multimodal features with the final retrieval results, and then performing information fusion to obtain multimodal retrieval results. The information fusion includes entity alignment and disambiguation, information complementarity detection, and conflict resolution.

[0021] Step S6: Entity importance adjustment. Calculate the similarity between the semantic features of the multimodal retrieval result entity v and the query Q, and use the similarity as a weight to update the importance score corresponding to the multimodal retrieval result entity to obtain the weighted importance.

[0022] Step S7: Entity clustering. Based on weighted importance, the entities in the multimodal retrieval results are clustered to obtain entity clusters. ;

[0023] In the formula, Represents multimodal search results. Represents weighted importance. Represents the number of clusters;

[0024] Step S8: Dynamic question generation, dynamically generating differentiated questions based on entity clusters and query complexity: ;

[0025] In the formula, This represents a dynamic question generation operation. This represents the depth of the problem, which is determined by weighted importance: ;

[0026] In the formula, 1, 2, and 3 represent the shallow, middle, and deep layers, respectively. and Represents the upper and lower limits of weighted importance;

[0027] Step S9: Layered generation of literature summaries. Based on differentiated questions and multimodal retrieval results, a lightweight LLM model is used to generate entity-level summaries. A medium-sized LLM model is used to integrate the entity-level summaries to obtain cluster-level summaries. A large LLM model is used to integrate the cluster-level summaries to obtain the literature summary. Summary quality assessment and iterative refinement are introduced to process the literature summary, and finally, an optimized literature summary is obtained. The literature summary is calculated as follows: ; ; ;

[0028] In the formula, This represents the prompt generator in the LLM model. , and These represent entity level, cluster level, and literature summary, respectively.

[0029] The beneficial results achieved by the present invention using the above solution are as follows:

[0030] (1) In view of the problem that existing literature summary technology has insufficient context information, this invention improves it by designing a multi-level context expansion mechanism, including core entity text retrieval, weak context expansion and strong context expansion. Through these multi-level and intelligent context expansion mechanisms, a rich, three-dimensional and multi-angle information environment is built for each user query, laying a solid foundation for generating comprehensive and in-depth abstracts.

[0031] (2) In response to the problem of insensitivity of existing literature summary technology, this invention realizes highly customized summary generation by constructing intelligent processing with query awareness. It infers the user's actual information needs through multi-level analysis and divides the query intent into multiple dimensions. The multi-dimensional intent analysis provides precise guidance for subsequent customized processing. It introduces dynamic adjustment of query relevance to ensure that the summary summary content closely revolves around the core of the query. It designs a granularity adaptive control mechanism to automatically determine the required summary granularity to meet the user's needs for summary summaries of different granularities.

[0032] (3) In response to the problem of incomplete information coverage in existing literature summaries, this invention achieves complete information coverage by constructing a multi-engine hybrid retrieval system and an intelligent information fusion framework. By running three different types of retrieval engines in parallel, semantically related content is found through vector similarity, and related information is found through graph traversal algorithm. The conditional retrieval engine based on metadata filters the document's quality indicators, timeliness, authority, and other metadata. The results of the three engines are not simply merged, but are intelligently integrated through a learned fusion model, thereby solving the problem of incomplete information coverage and ultimately further optimizing the quality of the generated literature summary. Attached Figure Description

[0033] Figure 1 A flowchart of a literature summary and analysis method based on natural language processing provided by the present invention.

[0034] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used together with the embodiments of the invention to explain the invention and do not constitute a limitation thereof. Detailed Implementation

[0035] 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. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0036] Example 1, see Figure 1 The present invention provides a literature summary and analysis method based on natural language processing, which specifically includes the following steps:

[0037] Step S1: Multi-granularity document segmentation. An adaptive segmentation strategy is used to segment the document article into blocks, resulting in n text blocks. Different segmentation strategies correspond to different numbers of text blocks.

[0038] Step S2: Hierarchical text block relationship extraction. The importance score of a text block is obtained by weighting its salience, centrality, and novelty scores within the article. ;

[0039] In the formula, , and The representative weight parameter satisfies that the sum of the three equals 1. Represents a text block. The TF-IDF score represents the saliency of a text block within the article. The centrality score represents the text block. The novelty score of the text block is calculated based on the time decay function. This represents the importance score of the text block;

[0040] Step S3: Dynamic text block aggregation. A dynamic merging threshold is introduced to adaptively merge text blocks whose semantic similarity is within a preset threshold range. The dynamic merging threshold is determined by the maximum number of text blocks, the current total number of text blocks, and the semantic similarity of the text blocks. The semantic similarity is the similarity between the semantic features of the text blocks. The semantic features are extracted by a neural network. The formula for calculating the dynamic merging threshold is as follows: ;

[0041] In the formula, Represents the dynamic merging threshold. This represents the semantic similarity between text block i and text block j. To adjust the parameters, Represents the maximum number of text blocks. Represents the total number of current text blocks;

[0042] Step S4: Graph construction. For each merged text block, entities and relations are extracted using a large language model. All entities and their relations constitute a knowledge graph.

[0043] Step S5: Adaptive context expansion retrieval. Perform adaptive context expansion retrieval on the knowledge graph to obtain multimodal retrieval results;

[0044] Step S6: Entity importance adjustment. Calculate the similarity between the semantic features of the multimodal retrieval result entity v and the query Q, and use the similarity as a weight to update the importance score corresponding to the multimodal retrieval result entity to obtain the weighted importance.

[0045] Step S7: Entity clustering. Based on weighted importance, the entities in the multimodal retrieval results are clustered to obtain entity clusters. ;

[0046] In the formula, Represents multimodal search results. Represents weighted importance. Represents the number of clusters;

[0047] Step S8: Dynamic question generation, dynamically generating differentiated questions based on entity clusters and query complexity: ;

[0048] In the formula, This represents a dynamic question generation operation. This represents the depth of the problem, which is determined by weighted importance: ;

[0049] In the formula, 1, 2, and 3 represent the shallow, middle, and deep layers, respectively. and Represents the upper and lower limits of weighted importance;

[0050] Step S9: Layered generation of literature summaries. Based on differentiated questions and multimodal retrieval results, a lightweight LLM model is used to generate entity-level summaries. A medium-sized LLM model is used to integrate the entity-level summaries to obtain cluster-level summaries. A large LLM model is used to integrate the cluster-level summaries to obtain the literature summary. Summary quality assessment and iterative refinement are introduced to process the literature summary, and finally, an optimized literature summary is obtained. The literature summary is calculated as follows: ; ; ;

[0051] In the formula, This represents the prompt generator in the LLM model. , and These represent entity level, cluster level, and literature summary, respectively.

[0052] Example 2, this example is based on the above example, step S5 specifically includes the following steps:

[0053] Step S51: Query generation. Based on the knowledge graph, initialize the query entity Q.

[0054] Step S52: Calculate query complexity. Based on the entity density, query ambiguity score, and query length, calculate the query complexity. ;

[0055] In the formula, Represents the normalized query length. The entity density in the query is represented by the number of entities divided by the total number of words. The query ambiguity score represents the diversity of queries calculated based on entity type. =1, which is the preset weight parameter;

[0056] Step S53: Adaptive context-based retrieval expansion, generating dynamic expansion strategies based on query complexity: ;

[0057] In the formula, and Represents the complexity threshold. Represents the number of dynamically expanded steps. This means retrieving only entities directly related to the query. This represents the first-level extension, from weak context to strong context. Represents multi-level expansion, with a depth of , This is the dynamic expansion strategy, where the dynamic expansion steps are calculated as follows: ;

[0058] In the formula, Represents the preset base extension steps. Represents the preset expansion coefficient;

[0059] Step S54: Hybrid strategy retrieval. Perform hybrid strategy retrieval on all entities using vector similarity, graph structure, and metadata enhancement. Optimize the union of the retrieval results from these three methods using Rerank technology to obtain the final retrieval results. ; ; ; ;

[0060] In the formula, This represents the number of entities returned by vector retrieval. The Top-k represents the entities with the highest similarity among their semantic features. This represents the search results based on vector similarity. This represents random walk retrieval based on a knowledge graph, where t represents the number of random walk steps and p represents the restart probability. This represents the search results based on graph structure retrieval. Metadata-enhanced retrieval, The threshold represents the importance of the filtering. This represents search results enhanced based on metadata. This is the final search result;

[0061] Step S55: Cross-modal information fusion, extracting multimodal features, aligning the multimodal features with the final retrieval results, and then performing information fusion to obtain multimodal retrieval results. The information fusion includes entity alignment and disambiguation, information complementarity detection, and conflict resolution.

[0062] Example 3: Applying the above scheme to the generation of abstracts for review articles in the medical field:

[0063] Application scenario: A scholar needs to quickly understand 100 relevant articles on "the latest research progress of mRNA vaccines in the prevention and control of COVID-19". These articles include different types of documents such as clinical trial reports, mechanism studies, and epidemiological analyses.

[0064] Implementation steps

[0065] Step S1: Multi-granularity document segmentation

[0066] Input data: 100 medical literature PDFs, totaling approximately 500,000 words;

[0067] Block processing:

[0068] Use a PDF parsing tool to extract text and chart data;

[0069] Apply an adaptive block partitioning strategy:

[0070] Small-grained blocks (200-400 words): used to extract specific experimental data and results;

[0071] Medium-grained blocks (400-800 words): used to extract descriptions of research methods;

[0072] Large-grained blocks (800-1200 words): used to extract the review and discussion sections;

[0073] The final result was approximately 800 text blocks, each labeled with the source literature, chapter location, and granularity level;

[0074] Step S2: Extracting Hierarchical Text Block Relationships

[0075] Calculate an importance score for each text block:

[0076] TF-IDF score: Identifies technical terms in each block (such as "spike protein", "neutralizing antibody");

[0077] Centrality score: calculated based on inter-block references and semantic associations;

[0078] Novelty score: based on publication date (2020-2023), with a time decay coefficient λ=0.01;

[0079] Weight parameters set: α=0.4, β=0.3, γ=0.3;

[0080] Example calculation:

[0081] A text block from the 2022 NEJM regarding Moderna's vaccine clinical trials:

[0082] TF-IDF(v) = 0.75 (includes several highly specific terms);

[0083] Centrality(v) = 0.82 (cited in multiple papers);

[0084] Novelty(v) = 0.60 (relatively recent publication date);

[0085] s(v)=0.4×0.75+0.3×0.82+0.3×0.60=0.726;

[0086] Step S3: Dynamic aggregation of text blocks

[0087] Semantic similarity calculation:

[0088] The semantic features of each text block are extracted using the BioBERT model;

[0089] Calculate the cosine similarity between blocks;

[0090] Dynamic merging threshold setting:

[0091] Nmax=1000 (maximum expected number of blocks);

[0092] Nnow=800 (total number of blocks currently);

[0093] φ=0.05 (adjustment parameter);

[0094] For the two text blocks discussing "vaccine safety":

[0095] ρ(i,j)=0.85 (semantically highly similar);

[0096] τ=0.85+0.05×log(800) / log(1000)=0.85+0.05×0.903=0.895;

[0097] Since 0.85 < 0.895, no merging is performed (two independent but different perspectives are retained);

[0098] After aggregation, the number of text blocks was reduced from 800 to 650;

[0099] Step S4: Graph Construction

[0100] Entity extraction: Identifying medical entities using the PubMedBERT model;

[0101] Entity types: drugs (mRNA-1273, BNT162b2), proteins (Spike protein), diseases (COVID-19), biological processes (immune response);

[0102] Approximately 5,000 entities were extracted;

[0103] Relationship extraction:

[0104] "Moderna vaccine" - "targets" - "SARS-CoV-2"

[0105] "mRNA vaccines"-"induces"-"neutralizing antibodies"

[0106] "booster dose"-"enhances"-"immune memory"

[0107] Construct a knowledge graph containing 5000 nodes and 12000 edges;

[0108] Step S5: Adaptive Context-Extended Retrieval

[0109] User query: "Evaluate the effectiveness of mRNA vaccines in preventing severe COVID-19 illness and death, and analyze differences in immune responses among different populations";

[0110] Step S52: Calculate query complexity;

[0111] Entity recognition: ["mRNA vaccine", "COVID-19", "severe illness", "death", "immune response", "population differences"];

[0112] Lnorm(Q) = 18 / 50 = 0.36 (query length 18 words, maximum reference 50 words);

[0113] Edensity(Q) = 6 / 18 = 0.333;

[0114] Aambiguity(Q) = 1 - 4 / 6 = 0.333 (6 entities, 4 types);

[0115] δ1=0.3, δ2=0.4, δ3=0.3;

[0116] C(Q)=0.3×0.36+0.4×0.333+0.3×0.333=0.340;

[0117] Complexity classification: Medium (0.3 ≤ 0.340 < 0.7);

[0118] Step S53: Adaptive Context-Extended Retrieval

[0119] The OneStepExpansion strategy (first-level expansion) is adopted.

[0120] k=⌈2×(1+1.5×0.340)⌉=⌈2×1.51⌉=4;

[0121] Search extension:

[0122] Weak context: directly related entities (vaccine name, virus strain);

[0123] Strong context: expanded entities (clinical trial design, statistical methods, side effect data);

[0124] Step S54: Hybrid Strategy Retrieval

[0125] Vector retrieval (k=15): Finds semantically similar text blocks;

[0126] Graph structure retrieval (t=10, p=0.15): Finding related information through random walk;

[0127] Metadata filtering (τ=0.3): Filter out documents with low importance;

[0128] Re-ranking: The top 50 results are ranked based on their overall relevance score;

[0129] Step S55: Cross-modal information fusion

[0130] Extract tabular data (efficacy rate, confidence interval) from clinical trials;

[0131] Extract trend information (antibody titer changes over time) from the chart;

[0132] Align text and numerical data;

[0133] Conflict resolution: Prioritize the use of data from large-sample randomized controlled trials;

[0134] Step S6: Entity Importance Adjustment

[0135] For the entity "neutralizing antibody":

[0136] The original importance s(v) = 0.72;

[0137] Relevance to query: 0.85;

[0138] σ=0.5;

[0139] s'(v)=0.72×exp(0.85 / 0.5)=0.72×exp(1.7)=0.72×5.47=3.94;

[0140] After normalization: Adjusted importance = 0.92;

[0141] Step S7: Entity Clustering

[0142] Cluster the 200 retrieved entities based on adjusted importance;

[0143] The number of clusters n = min(max(2, ⌊√200⌋), 10) = min(max(2,14),10) = 10;

[0144] Clustering results:

[0145] Cluster 1: Vaccine efficacy-related entities (efficacy, protective efficacy, breakthrough infection);

[0146] Cluster 2: Safety-related entities (side effects, adverse reactions, contraindications);

[0147] Cluster 3: Entities related to population differences (elderly, immunocompromised individuals, pregnant women);

[0148] Cluster 4: Entities related to immune mechanisms (antibody response, cell-mediated immunity, memory B cells);

[0149] Step S8: Dynamic Problem Generation

[0150] For the core entity "vaccine efficacy" (s'(v)=0.95):

[0151] L_dep=3 (Deep problem)

[0152] Question generation:

[0153] What are the differences in the effectiveness of different mRNA vaccines in preventing severe COVID-19?

[0154] What are the mechanisms by which vaccine effectiveness declines over time?

[0155] How do booster shots affect the long-term protective effect of vaccines?

[0156] For the minor entity "injection site pain" (s'(v)=0.35):

[0157] L_dep=1 (Shallow problem)

[0158] Question: "What is the incidence of local adverse reactions to mRNA vaccines?"

[0159] Step S9: Generate a hierarchical summary of the literature.

[0160] Entity-level summary (using LLaMA-7B):

[0161] Generate a 2-3 sentence summary of "vaccine efficacy," citing key data;

[0162] Cluster-level summary (using LLaMA-13B):

[0163] Integrate all entity-level summaries of Cluster 1 (Vaccine Effectiveness) and generate a structured summary;

[0164] Overall Summary (using GPT-4):

[0165] Integrate all cluster-level summaries and generate a complete literature review.

[0166] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0167] 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.

[0168] The present invention and its embodiments have been described above. This description is not restrictive, and the accompanying drawings are only one embodiment of the present invention; the actual structure is not limited thereto. In conclusion, if those skilled in the art are inspired by this description and design similar structures and embodiments without departing from the spirit of the invention, such designs should fall within the protection scope of the present invention.

Claims

1. A literature summary and analysis method based on natural language processing, characterized in that: Specifically, the following steps are included: Step S1: Use an adaptive segmentation strategy to segment the literature article into blocks, resulting in n text blocks; Step S2: The importance score of the text block is obtained by weighting the salience, centrality and novelty scores of the text block in the article; Step S3: Introduce a dynamic merging threshold to adaptively merge text blocks whose semantic similarity is within a preset threshold range. The dynamic merging threshold is determined by the maximum number of text blocks, the current total number of text blocks, and the semantic similarity of the text blocks. The semantic similarity is the similarity between the semantic features of the text blocks, and the semantic features are extracted by a neural network. Step S4: For each merged text block, use a large language model to extract entities and relations. All entities and their relations constitute a knowledge graph. Step S5: Perform adaptive context expansion retrieval processing on the knowledge graph to obtain query and multimodal retrieval results. Step S6: Calculate the similarity between the entities in the multimodal retrieval results and the semantic features of the query, and use the similarity as a weight to update the importance score corresponding to the entities in the multimodal retrieval results to obtain the weighted importance. Step S7: Cluster the entities in the multimodal retrieval results according to weighted importance to obtain entity clusters; Step S8: Dynamically generate differentiated questions based on entity clusters and query complexity; Step S9: Based on the differentiated questions and multimodal retrieval results, a lightweight LLM model is used to generate entity-level summaries. A medium-sized LLM model is used to integrate the entity-level summaries to obtain cluster-level summaries. A large LLM model is used to integrate the cluster-level summaries to obtain literature summaries. Summary quality assessment and iterative refinement are introduced to process the literature summaries, and finally, an optimized literature summary is obtained.

2. The document summary and analysis method based on natural language processing according to claim 1, characterized in that: Step S5 specifically includes the following steps: Step S51: Query generation. Based on the knowledge graph, initialize the query entity Q. Step S52: Calculate query complexity. Calculate the query complexity based on the entity density, query ambiguity score, and query length in the query. Step S53: Adaptive context-based retrieval expansion, generating dynamic expansion strategies based on query complexity; Step S54: Hybrid strategy retrieval. Perform hybrid strategy retrieval on all entities based on vector similarity, graph structure, and metadata enhancement. Optimize the union of the retrieval results from the three retrieval methods using Rerank technology to obtain the final retrieval results. Step S55: Cross-modal information fusion, extracting multimodal features, aligning the multimodal features with the final retrieval results, and then performing information fusion to obtain multimodal retrieval results. The information fusion includes entity alignment and disambiguation, information complementarity detection, and conflict resolution.

Images