Dynamic content abstract generation method and system based on multi-dimensional intention perception

Abstract

The invention discloses a dynamic content abstract generation method and system based on multi-dimensional intention perception. Comprising the following steps: acquiring a document set, extracting a plurality of semantic unit tetrads through an algorithm based on dependency syntactic analysis, and generating a d-dimensional semantic vector; receiving a four-dimensional weight vector input by a user, and dynamically scheduling three basic strategies or a combination thereof based on the weight vector; selecting a representative unit set according to the multi-granularity hierarchical sequence; a visual abstract graph is constructed based on the semantic relation of representative units, graph visualization is achieved through node coding and edge screening, and multi-granularity smooth switching and down-drilling exploration based on global semantic indexes are supported. According to the method, the problems of single dimension, black box decision and interaction stiffness of a traditional abstract are solved, personalized, explainable and interactive abstract generation is realized through multi-dimensional intention quantification and dynamic strategy scheduling, the efficiency is remarkably improved, and the method is suitable for scenes such as academic literature review and legal document analysis.

Description

Technical Field

[0001] This invention relates to the fields of natural language processing and information retrieval technology, and in particular to a method and system for generating dynamic content summaries based on multidimensional intent awareness. Background Technology

[0002] In today's era of information explosion, rapidly extracting high-quality content summaries from massive amounts of text has become a critical need. Currently, content summarization in information retrieval and natural language processing mainly falls into two categories: traditional methods and deep learning methods. Traditional methods largely rely on shallow logic such as keyword extraction and key sentence ranking, while deep learning methods incorporate techniques like CNNs, attention mechanisms, and pre-trained models; however, both have significant limitations.

[0003] Traditional summarization methods generally suffer from a single objective dimension, often optimizing only a single metric such as relevance or coverage. This makes it difficult to balance users' core needs across multiple dimensions, including relevance, coverage, diversity, and semantic depth. Furthermore, the algorithm's decision-making process lacks transparency, creating a "black box" operation where users cannot understand the logic behind the summary generation, thus reducing the reliability of the results. In addition, existing methods often generate summaries in static text format with rigid interactive methods, failing to allow users to adjust the granularity of the summary or explore information details according to their needs, thus failing to meet the exploratory needs for in-depth information acquisition.

[0004] While some existing technologies attempt to introduce graph structures or multi-model fusion, they do not address the quantification and dynamic adaptation of user intent, and thus fail to solve the core pain points of "insufficient personalization, poor interpretability, and weak interactivity." Therefore, there is an urgent need for a technical solution that can perceive multi-dimensional user intent, dynamically optimize summary characteristics, and generate intuitive and interactive summaries to improve the efficiency and experience of information acquisition. Summary of the Invention

[0005] The technical problem to be solved by the present invention is to provide an improved dynamic content summarization method and system based on multi-dimensional intent awareness, so as to solve the technical problems of traditional summarization methods such as single dimension, black box decision-making and rigid interaction, and realize personalized, interpretable and interactive high-quality summarization generation.

[0006] A first aspect of this disclosure provides a method for generating dynamic content summaries based on multi-dimensional intent awareness, comprising: 1. Obtain a set of documents to be processed, extract multiple semantic units from the set of documents, and generate a corresponding d-dimensional semantic vector representation for each semantic unit, wherein each semantic unit is a quadruple (t,e,τ,s), where t is the text content, e is the semantic vector, τ is the type label, and s is the confidence score; 2. Receive a four-dimensional weight vector W = (α, β, γ, δ) input by the user, wherein the weight vector satisfies α, β, γ, δ ≥ 0 and α + β + γ + δ = 1, where α is the relevance weight, β is the coverage weight, γ is the diversity weight, and δ is the semantic depth weight. 3. Based on the four-dimensional weight vector, a target selection strategy or strategy combination is dynamically determined from three preset basic strategies. The three basic strategies include a clustering selection strategy for optimizing coverage, a ranking trade-off strategy for balancing relevance and diversity, and a semantic induction strategy for improving semantic depth. 4. Using the target selection strategy or a combination of strategies, select a representative unit set from all the semantic units, wherein the size of the representative unit set corresponds to a multi-granularity hierarchical sequence. ,in L is the maximum granularity level (1≤L≤10, default L=5); 5. Based on the semantic vectors and semantic relationships of each unit in the representative unit set, construct and generate an interactive multi-granularity visual summary map, which supports smooth granularity switching and drill-down exploration functions.

[0007] A second aspect of this disclosure provides a dynamic content summarization system based on multi-dimensional intent awareness, comprising: 1. A semantic processing module, used to acquire a document set, extract semantic units, and generate a d-dimensional semantic vector representation for each semantic unit, wherein the semantic unit is a quadruple (t, e, τ, s), and simultaneously calculate the global semantic center. Construct a vector matrix ; 2. An intent parsing module is used to receive and parse the four-dimensional weight vector W=(α,β,γ,δ) input by the user, wherein the weight vector satisfies α,β,γ,δ≥0 and α+β+γ+δ=1, and output the weight parameters; 3. A strategy scheduling module, connected to the intent parsing module, is used to dynamically determine the target selection strategy or strategy combination from three preset basic strategies based on the four-dimensional weight vector, and supports the combined scheduling of two basic strategies (such as "clustering selection + ranking trade-off"). 4. A representative selection module, connected to the semantic processing module and the policy scheduling module, is used to execute the target selection policy or policy combination to select a set of representative units from the semantic units; 5. The summary generation and rendering module, connected to the representative selection module, is used to construct and render an interactive multi-granularity visual summary map based on the representative unit set, and provides granularity switching and drill-down exploration functions.

[0008] A third aspect of this disclosure provides an electronic device including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the steps of the above-described dynamic content summary generation method based on multidimensional intent awareness.

[0009] A fourth aspect of this disclosure provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the steps of the above-described dynamic content summary generation method based on multidimensional intent awareness.

[0010] The beneficial effects of the dynamic content summarization method and system based on multi-dimensional intent awareness provided in this disclosure are as follows: This disclosure achieves precise expression of summary requirements by quantifying user intent into a four-dimensional weight vector; it solves the multi-objective balancing problem of traditional methods by dynamically scheduling multiple strategies based on weights; and the generated multi-granularity visualization map supports interactive exploration, breaking the limitations of static summarization. The entire solution is logically clear and highly interpretable, capable of generating high-quality summaries according to users' personalized needs, significantly improving information acquisition efficiency and user experience, and is suitable for various scenarios such as academic literature reviews, news event tracking, and legal document analysis.

[0011] To more clearly illustrate the technical solutions in the embodiments of this disclosure, the accompanying drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of this disclosure. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort. Attached Figure Description

[0012] Figure 1 A system architecture diagram of a dynamic content summarization system based on multi-dimensional intent awareness provided in an embodiment of the present invention; The connection relationships and data flow of the semantic processing module, intent parsing module, strategy scheduling module, representative selection module, and summary generation and rendering module are shown (solid lines represent vector data flow, and dashed lines represent weight parameter flow). Figure 2 The flowchart of the method and system embodiment of the dynamic content summarization method and system based on multi-dimensional intent awareness provided by an embodiment of the present invention clearly presents the complete steps from document input to graph generation, and marks the correspondence between the four sub-steps of step S205 and the multi-granularity hierarchical sequence formula. Figure 3 The algorithm selection logic diagram of a specific embodiment of this disclosure is shown, with the balance parameter formula and the hybrid strategy quota allocation formula annotated. Figure 4This is a schematic diagram of the multi-granularity semantic graph structure of an embodiment of the present disclosure, showing the composition of nodes and edges and the visual attribute encoding rules, including coverage examples, similarity numerical annotations, and drill-down exploration entry points. Detailed Implementation

[0013] In the following description, specific details such as particular system architectures and techniques are set forth for illustrative purposes and not for limitation, so as to provide a thorough understanding of the embodiments of this disclosure. However, those skilled in the art will understand that this disclosure may also be implemented in other embodiments without these specific details. In other instances, detailed descriptions of well-known systems, apparatuses, circuits, and methods have been omitted so as not to obscure the description of this disclosure with unnecessary detail.

[0014] The following is an explanation of the terms used in the embodiments of this disclosure: Semantic unit: refers to the smallest text unit with independent semantics extracted from a document, which can be at the sentence level or clause level and can fully express a concept, fact, opinion, or method. This invention extracts semantic units using an algorithm based on dependency parsing, with the length controlled between 10-30 words, and units with a confidence level s≥0.7 are retained.

[0015] Multidimensional Intent Weights: Used to quantify the strength of user preferences for summary features, including four orthogonal dimensions: relevance (α), coverage (β), diversity (γ), and semantic depth (δ), with a weight sum of 1. Users can input via slider or text; input is automatically normalized if it does not meet the constraints.

[0016] Multi-granularity summarization: This refers to support multi-level summary display from coarse to fine, with the number of representative units corresponding to each level increasing exponentially to meet users' different information needs from overview to detail. The maximum granularity level L satisfies 1≤L≤10, with L=5 by default.

[0017] Semantic graph: A visual network constructed with semantic units as nodes and semantic similarity as edges. The visual attributes (size, color) of nodes reflect their semantic features and support user interaction and exploration.

[0018] To make the purpose, technical solution, and advantages of this disclosure clearer, the following will be discussed in conjunction with the appendix. Figure 1-4 The following is an explanation using specific examples.

[0019] like Figure 1As shown, the system architecture of this invention mainly includes five modules: semantic processing module, intent parsing module, strategy scheduling module, representative selection module, and summary generation and rendering module. These modules work collaboratively to transform the original document into an interactive semantic graph summary. The hierarchical connections and data flow are displayed, clearly indicating the core output parameters of each module (e.g., the semantic processing module outputs a set of semantic units and a vector matrix, and the intent parsing module outputs a four-dimensional weight vector). The collaborative working logic of the system is intuitively presented, and the input-output logic of the core formulas (e.g., semantic vector matrix) is implicitly included in the diagram. ), which is consistent with the output parameters of the semantic processing module mentioned above.

[0020] Please refer to Figure 2 This is a flowchart illustrating a dynamic content summarization method based on multi-dimensional intent awareness provided in an embodiment of this disclosure. It fully presents the entire process from document set acquisition, semantic unit extraction, weight vector reception, strategy determination, representative unit selection to graph generation. The four sub-steps of step S205 are highlighted (S205-1 Node Definition, S205-2 Edge Construction, S205-3 Multi-granularity Switching, S205-4 Drill-down Exploration), clarifying the execution order and core functions of each sub-step, echoing the multi-granularity hierarchical sequence. Technical features.

[0021] The method includes: S201: Obtain the set of documents to be processed, extract semantic units, and generate corresponding d-dimensional semantic vector representations. 1. In this embodiment, the document set D to be processed can be represented as: The total number of documents in the document set D is N, where Let i be the i-th document. The document type can be any text format, such as academic papers, news reports, technical reports, or legal provisions.

[0022] 2. Semantic unit extraction is the process of decomposing each document into multiple independent semantic units. This is achieved using dependency parsing-based algorithms (such as the spaCy toolkit), segmenting according to subject-verb-object structure, with a length controlled between 10-30 words, forming a set of semantic units. , where n is the total number of semantic units, n=|U|, that is, the total number of elements in the semantic unit set U and n>N (a document corresponds to multiple semantic units). For example, a 1000-word academic paper can be decomposed into 15-25 semantic units, and 100 documents can be decomposed into 1500-2500 semantic units.

[0023] It can be represented as: in, This represents the set of semantic units extracted from all documents, where n is the total number of semantic units and n > N (where N is the total number of documents in the document set D). The units are retained.

[0024] 3. Each semantic unit u is represented by a quadruple (t, e, τ, s): t represents the text content of a semantic unit, such as "The Transformer architecture has excellent semantic capture capabilities in natural language processing tasks"; e is the d-dimensional semantic vector representation of t, e∈ℝᵈ, generated through a pre-trained embedding model. Optional models include BERT, BGE, etc., where d is the vector dimension. BERT... base The model corresponds to d=768, BERT large The model corresponds to d=1024, and the model training parameters are batch size 32 and learning rate 2. e -5. The training process consists of 3 rounds. The vectors are normalized to the [0,1] interval and then stored. τ is a type label used to identify the attributes of a semantic unit. Optional values ​​include concepts, facts, opinions, methods, conclusions, etc. For example, τ in the above example could be labeled "technical". achievement ”; s is the confidence score, s∈[0,1], which is calculated based on the performance of the semantic unit extraction algorithm and is used to reflect the reliability of the semantic integrity of the unit. Units with s≥0.7 are retained.

[0025] The vectors of all semantic units constitute a vector matrix.

[0026] Where the j-th row is a semantic unit vector .

[0027] Simultaneously, calculate the global semantic center. This is used to measure the relevance of semantic units to the overall topic of the document set. Here, q represents the average vector of the entire semantic unit set, used to measure its relevance to the global topic.

[0028] : Receive the four-dimensional weight vector W = (α,β,γ,δ) input by the user. In this embodiment, the user inputs a four-dimensional weight vector through an interactive interface. The weights satisfy the constraints α,β,γ,δ≥0 and α+β+γ+δ=1. If the input does not satisfy these conditions, the system automatically normalizes the vector (normalization formula: The physical meaning of each weight is as follows: α (relevance weight): represents the degree of relevance between the summary and the core topic of the document collection. The larger the α, the more the summary focuses on the core topic. β (coverage weight): represents the summarization bias of the summary in terms of the comprehensiveness of the information range of the document set. The larger the β is, the more the summary can cover different semantic regions. γ (diversity weight): Represents the preference for the richness of the summary content. The larger the γ, the less repetitive or similar information is in the summary. δ (Semantic Depth Weight): Represents the preference for the degree of abstraction and generalization in the summary. The larger the δ is, the more the summary reflects higher-level semantic insights.

[0029] For example, if a user wants to quickly understand the core conclusions of academic literature while also ensuring comprehensiveness, they can set the weight W=(0.3,0.3,0.2,0.2); if they want to focus on in-depth inductive analysis of the abstract, they can set W=(0.1,0.1,0.1,0.7).

[0030] Dynamically determine target selection strategies or strategy combinations based on four-dimensional weight vectors. In this embodiment, three basic strategies are preset, each corresponding to a different optimization objective. The strategy scheduling logic is as follows: Figure 3 As shown: 1. Clustering Selection Strategy (AlgCluster): Prioritizes optimizing coverage (C). This strategy is selected when β is the weight with the highest proportion in the weight vector. The number of clusters is determined when used independently. ; 2. Ranking Trade-off Strategy (AlgMMR): This strategy balances the optimization of relevance (R) and diversity (D), and is chosen when the sum of the weights of α and γ has the highest proportion. 3. Semantic Induction Strategy (AlgLLM): Prioritizes optimizing semantic depth (S). This strategy is selected when δ is the weight with the highest proportion in the weight vector. When used independently, the number of topics... ; 4. Hybrid Selection: When there is no obvious dominant relationship between the proportions of α, β, γ, and δ, the number of representative units is allocated according to the weight ratio, and the three basic strategies are executed in combination. At the same time, it supports the combined scheduling of two basic strategies (such as cluster selection ranking trade-off).

[0031] like Figure 3 The diagram shown is a strategy scheduling logic diagram of a specific embodiment of the present invention. It intuitively illustrates the mapping relationship between the four-dimensional weight vector (α, β, γ, δ) and three basic strategies (clustering selection strategy, ranking trade-off strategy, and semantic induction strategy) and the hybrid strategy. It clarifies the strategy selection rules for four scenarios: "β accounts for at least 10% more than other weights", "the sum of α and γ weights accounts for the highest percentage", "δ accounts for at least 10% more than other weights", and "the difference in the percentage of each weight is less than 10%", clearly presenting the decision path of dynamic scheduling.

[0032] Strategy scheduling needs to determine the target strategy based on the proportion of the four-dimensional weights: when When the percentage exceeds 10% of any other weight, the clustering selection strategy is triggered; when and When the sum of weights accounts for the highest percentage, a ranking trade-off strategy is implemented, corresponding to the balance parameter formula graph marked in the figure. ;when When the proportion is dominant or the difference in the proportion of each weight is less than 10%, it corresponds to either a semantic induction strategy or a hybrid strategy, respectively. The hybrid strategy corresponds to the quota allocation formula diagram marked in the figure. , Figure 3 The complete path of this decision-making logic is clearly presented.

[0033] The specific execution steps of the hybrid strategy are as follows: Quota allocation: K l K represents the total number of representative units at level l. C K is the number of slots for the clustering selection strategy. R To determine the number of slots for the ranking strategy, K S The quota for semantic induction strategies. The function is an up-rounding function that ensures the total number of places is ≥ K. l (Extra slots will be sorted and discarded based on semantic similarity). Implementation of different strategies: ① Implement the cluster selection strategy and select from U. Each representative unit yields a set. ; ② Eliminate from U Then, a sorting trade-off strategy is selected. Each representative unit yields a set. Balance parameters ; ③ Exclude from U and After the union of the sets, a semantic induction strategy is executed to select K. S Each representative unit yields a set. ; Result merging: Representative unit set .

[0034] Among them, the granularity hierarchy sequence

[0035] L is the maximum granularity level (default). ),For example hour , hour , hour This enables multi-granularity control, from coarse to fine.

[0036] Select a representative set of units using a target selection strategy or a combination of strategies. The specific implementation methods of the three basic strategies are as follows: 1. Clustering selection strategy (AlgCluster): Performing the K-means clustering algorithm on the vector matrix E yields... Cluster and the corresponding cluster center The distance metric used is Euclidean distance, and the number of iterations is set to 10 to 100 (the iteration stops when the change in cluster center is less than 1e-6). Calculate cluster center ,in For clusters The number of semantic units; The semantic unit in each cluster that has the closest Euclidean distance to the cluster center is selected as the representative, i.e. Collect all The corresponding semantic units are obtained .

[0037] 2. Ranking Trade-off Strategy (AlgMMR): Initialize the representative set Global Semantic Center ; First round of selection: ,Will join in ; Subsequent rounds (k>1): ,in Cosine similarity is calculated as follows: ,Will join in until The scale reached .

[0038] Specifically: This step, used to balance correlation (R) and diversity (D), includes: Scheduling condition: When the relevance weight α and the diversity weight γ both dominate (e.g., α + γ > threshold).

[0039] Algorithm Objective: This strategy aims to generate a summary that is both highly relevant to the global topic and contains diverse information. It is based on the Maximum Boundary Relevance (MMR) framework, employing a greedy iterative approach to select, at each iteration, a unit that is relevant to the global topic and least similar to the selected set.

[0040] Algorithm formulation steps: ① Initialization ② Iterative selection ③ Output set Among them, the equilibrium parameters It is used to control the emphasis on correlation and diversity.

[0041] 3. Semantic Induction Strategy (AlgLLM): M is a large language model with semantic inductive capabilities, used to extract high-level topics from a set of semantic unit texts T. Optional models include GPT-4, Claude, and DeepSeek. The input is the prompt word, and the output is a set of high-level topics T. {themes} .

[0042] Extract the text set of all semantic units , where t i For semantic unit u i The text content t; Construct the prompt word prompt = "inductively derived" ⊕K S ⊕ "High-level topic:" ⊕ JSON(T), where ⊕ is a string concatenation operation; JSON is a general data exchange format, used here to transform the unstructured text set T into structured data that can be parsed by the large language model M, ensuring that the model accurately understands the semantic unit text to be summarized.

[0043] Input the prompt into a large language model M (GPT-4, Claude, DeepSeek, etc. are optional) to generate K. S High-level theme ; For each topic Vectors are generated by embedding models. Match the unit with the highest similarity among the original semantic units. Collected .

[0044] The specific steps are as follows: Scheduling condition: When the semantic depth weight δ is significantly high.

[0045] Algorithm Goal: This strategy aims to generate summaries with high-level abstraction and deep insights. It leverages the inductive power of Large Language Models (LLMs), first allowing the LLM to extract from all semantic units... We identify three core themes and then find the expressions from the original units that best match these themes as representatives.

[0046] Algorithm formulation steps: ① Extract text ② Construction hints ③ Model Induction ④ Vector matching ⑤ Output set Corresponding metric: The set generated by this strategy Expected in semantic depth measurement They performed well.

[0047] : Construct and generate interactive, multi-granularity visual summary maps based on representative unit sets. like Figure 2 As shown, step S205 (constructing and generating an interactive, multi-granularity visual summary map) further includes four sub-steps. The execution logic of each sub-step corresponds one-to-one with the annotations in the attached figures. Figure 4 The graph structure features are perfectly matched: S205-1 defines nodes and performs visual encoding; S205-2 completes edge selection and construction; and S205-3 supports smooth switching between multiple granularities, corresponding to the multi-granularity hierarchical sequence formula graph marked in the figure. S205-4 triggers the drill-down exploration function to ensure the feasibility of the map construction process. Figure 4 The core components of the semantic graph are shown in detail: node visual encoding size mapping coverage, with examples such as c(v)=2 and c(v)=5, consistent with the threshold verification of θcover=0.5; color mapping type labels, including categories such as facts, opinions, and conclusions, as well as edge filtering rules (only valid connections with semantic similarity >0.3 are retained, labeled with similarity values ​​such as 0.65 and 0.42, and weak connection filtering logic is marked with dashed lines), multi-granularity switching examples (K=4 when l=2, K=8 when l=3), and drill-down exploration entry (including source document fragment display formats such as "Document 1: 80% charge in 15 minutes"), fully demonstrating the visualization features and interactive functions of the graph.

[0048] In this embodiment, the construction of the summary graph includes defining the node set, constructing the edge set, encoding the node visual attributes, and implementing interactive functions. The graph is formally defined as follows: Where l represents the granularity level; 1. Node set Each representative unit corresponds to one node; when coverage At that time, the node size is set to the minimum pixel size. ; 2. Edge set construction: Calculate the cosine similarity between the semantic unit vectors corresponding to any two nodes. ,in sim Cosine similarity calculation function ; Filter for similarity greater than a preset threshold The nodes form an edge, and the edge set is formally defined as follows: To avoid weak connections causing graph clutter ( (Determined based on the similarity statistics of 5000 sets of semantic units); 3. Node visual attribute encoding: Node size: calculated based on coverage. Node size Where S max Maximum pixel size (default 50px); Node color: based on type label τ encoding. "Concept" corresponds to 0° hue, "Fact" corresponds to 120°, "Opinion" corresponds to 240°, "Method" corresponds to 60°, and "Conclusion" corresponds to 180°. A hash function maps type labels to hues ranging from 0 to 359 degrees to ensure that nodes of the same type have consistent colors. 4. Implementation of interactive functions: Multi-granularity switching: Users can adjust the granularity level l via a slider (slider range [1, L], step size 1). The system pre-calculates and caches the graphs at each level (using an LRU caching strategy, with a cache capacity of L graph instances). LRU stands for "Least Recently Used" cache replacement strategy, used here to cache the pre-calculated graphs at each granularity level. When the cache capacity reaches L (maximum number of granularity levels), the least recently used graph instance is evicted, improving the response speed of multi-granularity switching. During switching, a linear interpolation formula is used. , ( A smooth transition in node position and size is achieved with a transition duration of 300ms. Drill-down exploration: In response to user clicks on a node, retrieve the k (10 by default) original semantic units and source document fragments with the highest semantic similarity to the node from the global semantic index and display them in the sidebar, enabling in-depth exploration from the abstract to the original text.

[0049] Please refer to Figure 1 Here is a structural block diagram of a dynamic content summarization system based on multi-dimensional intent awareness provided in an embodiment of this disclosure. The system 20 includes: 1. Semantic Processing Module 21: Used to obtain a document set, and through a dependency parsing-based algorithm (such as...) spaCy The toolkit performs semantic unit extraction (a 1000-word document is decomposed into 15-25 semantic units) and vector generation, outputting a semantic unit set U and a vector matrix E, as described in S201. 2. Intent parsing module 22: It is used to receive the four-dimensional weight vector W=(α,β,γ,δ) input by the user, verify the weight constraints (non-negative and sum to 1), and automatically normalize the input if it does not meet the requirements (normalization formula: a'=a / (a+β+γ+δ), β′, γ', δ' are similar); and pass the weight parameters to the policy scheduling module. 3. Strategy scheduling module 23: connected to intent parsing module 22, dynamically determines the target strategy or strategy combination based on weight vector, supports the combined scheduling of two basic strategies, and outputs strategy execution instructions. The specific scheduling logic is as described in S203 (supports the combined scheduling of two basic strategies, such as cluster selection + ranking trade-off). 4. Representative Selection Module 24: Connected to the semantic processing module 21 and the policy scheduling module 23, it performs clustering selection, ranking trade-offs, or semantic induction operations according to policy instructions, and outputs a set of representative units. ; 5. Summary generation and rendering module 25: Connected to the representative selection module 24, it constructs a multi-granular semantic graph, realizes the visual encoding and interactive functions of nodes and edges, and displays a visual summary to the user.

[0050] In this embodiment, the system may also include an information connection module for establishing an association index between the summary map and the source document to support drill-down exploration; and a storage module for caching map data, semantic unit vectors and global semantic indexes at each granularity level to improve interactive response speed.

[0051] Application Examples Taking the generation of abstracts from 10 academic papers on the theme of "Latest Advances in Electric Vehicle Battery Technology" as an example, this demonstrates the specific application process of this disclosure: 1. The document set D contains 10 relevant academic papers. The semantic processing module 21 extracts 200 semantic units, generates a 768-dimensional vector through the BERT-base model, and calculates the global semantic center q. 2. The user wants to balance relevance and coverage, while taking into account a certain degree of diversity and depth. The input weight W = (0.3, 0.3, 0.2, 0.2) passes the verification of the intent parsing module 22. 3. The strategy scheduling module 23 determines to adopt a hybrid strategy, assuming... Allocate quotas ; 4. Representative selection module 24 performs cluster selection to obtain... (2 units), sorting and trade-offs yielded (2 units), semantic induction yielded (1 unit) was merged into 5 representative units; 5. The abstract generation and rendering module 25 constructs a three-level graph. Node size reflects coverage, and colors distinguish types such as "Technological Breakthrough," "Experimental Data," and "Application Prospects." Users can switch levels using a slider. Clicking the "Improving Solid-State Battery Energy Density" node allows drilling down to view the corresponding three source document fragments. Figure 4 The example of drilling down to explore the entry point is consistent.

[0052] The embodiments described above are merely preferred embodiments of the present invention and are not intended to limit the scope of the present invention. Various modifications and improvements made by those skilled in the art to the technical solutions of the present invention without departing from the spirit of the present invention should fall within the protection scope defined by the claims of the present invention.

Claims

1. A dynamic content summarization method based on multi-dimensional intent awareness, comprising: Obtain a set of documents to be processed, extract multiple semantic units from the document set, and generate a corresponding d-dimensional semantic vector representation for each semantic unit, wherein each semantic unit is a quadruple (t,e,τ,s), where t is the text content, e is the semantic vector, τ is the type label, and s is the confidence score; The system receives a four-dimensional weight vector W = (α, β, γ, δ) input from the user. The weight vector satisfies α, β, γ, δ ≥ 0 and α + β + γ + δ = 1, where α is the relevance weight, β is the coverage weight, γ is the diversity weight, and δ is the semantic depth weight. Based on the four-dimensional weight vector, a target selection strategy or strategy combination is dynamically determined from three preset basic strategies. The three basic strategies include a clustering selection strategy for optimizing coverage, a ranking trade-off strategy for balancing relevance and diversity, and a semantic induction strategy for improving semantic depth. selecting, by using the target selection strategy or the combination of strategies, a representative unit set from all the semantic units, a size of the representative unit set corresponding to a multi-granularity level sequence L is a maximum granularity level Based on the semantic vectors and semantic relationships of each unit in the representative unit set, an interactive multi-granularity visual summary map is constructed and generated, which supports smooth granularity switching and drill-down exploration functions. 2.The method of claim 1, wherein, Step 3, which dynamically determines the target selection strategy or strategy combination based on the four-dimensional weight vector, includes: If the proportion of β is at least 10% more than any other weight, the target selection strategy is the cluster selection strategy, and the number of clusters is independently adopted ; If the sum of the weights of a and y exceeds any other weight by at least 10%, the target selection strategy is the ranking trade-off strategy, and the number of independent applications is ; If the percentage of δ exceeds 10% of any other weight, the target selection strategy is a semantic induction strategy, and the number of topics is used independently. ; If the difference in the proportions of α, β, γ, and δ is less than 10%, a hybrid strategy is adopted, which involves allocating representative unit quotas according to their weights and then combining the three basic strategies.

3. The dynamic content summarization method based on multi-dimensional intent awareness according to claim 2, characterized in that, The execution of the hybrid strategy includes: The number of representative units shall be allocated according to the following formula: K l K represents the total number of representative units at level l. C K is the number of slots for the clustering selection strategy. R To determine the number of slots for the ranking trade-off strategy, K S The quota for semantic induction strategies. It is the floor function, and Excess slots will be discarded in descending order of semantic similarity; Execute a clustering selection strategy to select K from all semantic units. C Each representative unit yields a set. ; Excluded from the semantic unit set Then, a sorting trade-off strategy is implemented to select K. R Each representative unit yields a set. The balance parameter of the ranking trade-off strategy And 0 < λ < 1; Excluded from the semantic unit set and After the union of the sets, a semantic induction strategy is executed to select K. S Each representative unit yields a set. ; The representative unit set is , and The union of .

4. The dynamic content summarization method based on multi-dimensional intent awareness according to claim 1, characterized in that, Step 5, which involves constructing and generating an interactive, multi-granularity visual summary map, includes: The set of representative units is used as the node set of the graph; Calculate the semantic similarity of corresponding vectors between nodes, construct edges based on a preset threshold, and form a network structure; The visual attributes of a node are determined and rendered based on the semantic coverage or type of the unit it represents. Interactive controls are provided to allow users to adjust the granularity of the summary and trigger a smooth transition of the graph between different pre-computed granularity levels.

5. The dynamic content summarization method based on multi-dimensional intent awareness according to claim 4, characterized in that, It also includes the drilling exploration step: A global semantic index is pre-built, which stores the association relationship and vector similarity data between all semantic units and source document fragments. An inverted index structure is used, and the retrieval time complexity is O(log n). In response to user interaction with a specific node in the summary graph, retrieve the k original semantic units with the highest semantic similarity to the selected node and their corresponding source document fragments from the global semantic index, where k ≥ 5, and k = 10 by default; The retrieved original semantic units and source document fragments are displayed in descending order of similarity. The displayed content includes text content, type tags, confidence scores, and source document page / paragraph numbers.

6. The dynamic content summarization method based on multi-dimensional intent awareness according to claim 1, characterized in that, The clustering selection strategy is implemented using the K-means algorithm, which includes performing 10-100 iterations of clustering on the vector matrix E of all semantic units. The iteration stops when the change in the cluster center is less than 1e-6. After calculating the cluster center, the semantic unit in each cluster that has the closest Euclidean distance to the cluster center is selected as the representative unit. The ranking trade-off strategy is implemented based on the maximum boundary relevance framework, prioritizing semantic units with "high relevance and low repetition" during iterative selection; The semantic induction strategy uses large language models, including GPT-4, Claude, and DeepSeek, to inductively summarize high-level topics and then matches the most similar expressions in the original semantic units as representative units, with a similarity threshold of ≥0.

6.

7. The dynamic content summarization method based on multi-dimensional intent awareness according to claim 1, characterized in that, In step 1, "generating a corresponding d-dimensional semantic vector representation for each semantic unit" is achieved using the BERT model or the BGE model, where d = 768 or 1024. The type label τ of the semantic unit includes at least three of the following: concept, fact, opinion, method, and conclusion; The confidence score s∈[0,1] is determined based on the performance evaluation results of the semantic unit extraction algorithm, and units with s≥0.7 are retained.

8. A dynamic content summarization generation system based on multi-dimensional intent awareness, characterized in that, include: The semantic processing module is used to acquire a document set, extract semantic units, and generate a d-dimensional semantic vector representation for each semantic unit. The semantic unit is a quadruple (t, e, τ, s), where t is the text content of the semantic unit, e is the d-dimensional semantic vector, τ is the type label, and s is the confidence score. Simultaneously, it calculates the global semantic center. n is the total number of semantic units. ; Constructing the semantic vector matrix: This provides foundational data for subsequent intent-aware computation. The intent parsing module receives and parses the user's input four-dimensional weight vector W = (α, β, γ, δ). It verifies the weight constraints: α, β, γ, δ ≥ 0 and α + β + γ + δ = 1. If the input does not meet these constraints, it automatically normalizes the data and outputs the weight parameters. The normalization formula is as follows: The same applies to β', γ', and δ'; The strategy scheduling module, connected to the intent parsing module, is used to dynamically determine the target selection strategy or strategy combination from three preset basic strategies based on the four-dimensional weight vector and according to the rules described in claim 2. The representative selection module, connected to the semantic processing module and the policy scheduling module, is used to execute the target selection policy or policy combination according to the method of claim 3 or 6, and select a set of representative units from the semantic units; The summary generation and rendering module, connected to the representative selection module, is used to construct and render an interactive multi-granularity visual summary map according to the method of claim 4 or 5, providing granularity switching and drill-down exploration functions.

9. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements the steps of the method according to any one of claims 1 to 7.

10. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is executed by a processor, it implements the steps of the method according to any one of claims 1 to 7.

Images