Patent retrieval system and method based on dynamic weight adjustment and multi-modal semantics

Abstract

The application provides a patent retrieval system and method based on dynamic weight adjustment and multi-modal semantics. The system includes an input layer for receiving a natural language query of a user, an initial technical dimension weight adjustment instruction, and a display mode selection; a semantic retrieval layer for retrieving and sorting related patents from a patent document library based on the natural language query and the technical dimension weight to form a preliminary retrieval result; a processing layer for standardizing the preliminary retrieval result; a display layer for visualizing and rendering the standardized retrieval result in a corresponding multi-modal form according to the display mode selection, receiving a real-time weight adjustment instruction of the user, and obtaining an interactive feedback of the user; and a dynamic weight adjustment module for generating an initial technical dimension weight according to the initial technical dimension weight adjustment instruction or a preset retrieval strategy corresponding to the display mode selection, and dynamically optimizing the initial technical dimension weight based on the interactive feedback.

Description

Technical Field

[0001] This invention relates to the field of computer technology, and in particular to a patent retrieval system and method based on dynamic weight adjustment and multimodal semantics. Background Technology

[0002] The core flaw of existing patent search systems lies in their lack of dynamic adaptability. Their underlying design is based on a static rule system (fixed weights for technical dimensions, rigid semantic relationships, and fixed data processing procedures), which cannot respond to the dynamic needs and complex technical features of patent search scenarios. This is mainly reflected in the following three aspects: 1. Search bias caused by rigid strategies The system design failed to consider the dynamic differences in search scenarios from the outset, resulting in a fixed weight allocation strategy and the absence of a user feedback mechanism. Consequently, the system's weight allocation strategy lacks online adaptive optimization capabilities based on scenarios and user feedback. The system relies on a fixed rule engine for weight allocation; for example, regardless of whether a user is searching for "infringement evidence" (emphasizing technical means) or "technology trend research" (emphasizing technical effectiveness), the weights for dimensions such as "technical means" and "technical effectiveness" are pre-set to fixed values ​​(e.g., 0.5 each). Furthermore, the system architecture lacks a closed-loop system for collecting and analyzing user clicks, dwell time, and other behaviors. This prevents the search strategy from being dynamically adjusted according to specific scenarios (e.g., the weight of "technical means" cannot be reduced from 0.7 to 0.3), and weight updates lag significantly behind actual needs. The direct result is a systematic deviation between the obtained search results and the user's true intent.

[0003] 2. Semantic modeling failure The underlying semantic model used is a general-purpose model, which is not adapted to the technical, structural, and terminological specificities of patent texts. General-purpose semantic models struggle to accurately understand the deep logical connections within patent technologies, and semantic matching is disconnected from business weights. The system uses traditional bag-of-words models (such as TF-IDF) or general-domain natural language processing models (such as the basic version of BERT). Its modeling process cannot recognize the equivalence of heterogeneous terms commonly found in the patent field (e.g., it cannot understand that "lithium battery" and "lithium-ion battery" refer to the same technical concept). Simultaneously, the model ignores the strong contextual logical connections between the inherent "technical problem-technical solution-technical effect" triples in patent texts (e.g., it cannot accurately associate the technical effect of "reducing energy consumption" with the various specific technical means to achieve it). Ultimately, this results in superficial semantic understanding and an inability to accurately capture the core ideas of the patent. More importantly, due to the disconnect between semantic modeling and weight allocation, a deep semantic space capable of incorporating weight-oriented approaches has not been constructed.

[0004] 3. System architecture performance bottlenecks The data processing workflow is poorly designed, relying on simple pattern matching and lacking intelligent decision-making capabilities. Furthermore, the low level of intelligence in the underlying data processing workflow leads to overall system inefficiency and affects result quality. In the data preprocessing stage, complex tasks such as applicant name normalization and legal status determination rely solely on manually written regular expressions for rule matching. For example, it cannot intelligently identify "Company A" and "Company A Group Co., Ltd." as the same entity. This approach results in a high false positive rate (up to 18% in applicant alias scenarios), and due to the lack of legal status determination, different versions of the same patent (such as the published version and the authorized version) are repeatedly included, with a duplication rate exceeding 30%. This causes a huge waste of computing resources and storage space and reduces the cleanliness and reliability of the search results.

[0005] In summary, the fundamental problem facing existing patent search technologies is: how to build a patent search engine with dynamic response capabilities that can overcome the limitations of static rules, achieve adaptive optimization of weighting strategies, deep semantic understanding of patent technologies, and ensure the efficiency and accuracy of data processing. Summary of the Invention

[0006] The purpose of this invention is to provide a patent retrieval system and method based on dynamic weight adjustment and multimodal semantics, aiming to solve the above-mentioned problems in the prior art.

[0007] This invention provides a patent retrieval system based on dynamic weight adjustment and multimodal semantics, including an input layer, a semantic retrieval layer, a processing layer and a display layer that are communicatively connected to each other, and a dynamic weight adjustment module that is connected to the input layer, the semantic retrieval layer and the display layer respectively; The input layer is used to receive the user's natural language query, initial technical dimension weight adjustment instructions, and display mode selection; The semantic retrieval layer is used to retrieve and sort relevant patents from the patent document library based on the natural language query and the technical dimension weights provided by the dynamic weight adjustment module, forming preliminary retrieval results; wherein, the technical dimension weights include initial technical dimension weights and optimized technical dimension weights; The processing layer is used to standardize the preliminary search results; The display layer is used to select the standardized search results for visualization rendering in the corresponding multimodal form according to the display mode, and to receive the user's real-time weight adjustment instruction as a new initial technical dimension weight adjustment instruction through interactive controls, as well as to obtain the user's interactive feedback. The dynamic weight adjustment module is used to generate initial technical dimension weights according to the initial technical dimension weight adjustment instruction or the preset retrieval strategy corresponding to the display mode selection and output them to the semantic retrieval layer; and to receive the interactive feedback, dynamically optimize the initial technical dimension weights based on the interactive feedback, and output the optimized technical dimension weights to the semantic retrieval layer in real time.

[0008] This invention provides a patent retrieval method based on dynamic weight adjustment and multimodal semantics, comprising: The input layer receives users' natural language queries, initial technical dimension weight adjustment instructions, and display mode selections. The semantic retrieval layer retrieves and sorts relevant patents from the patent document library based on the natural language query and the technical dimension weights provided by the dynamic weight adjustment module, forming preliminary retrieval results; wherein, the technical dimension weights include initial technical dimension weights and optimized technical dimension weights; The preliminary search results are standardized through a processing layer. The display layer selects the standardized search results according to the display mode and renders them in a corresponding multimodal form. It also receives real-time weight adjustment instructions from users as new initial technical dimension weight adjustment instructions through interactive controls, and obtains user interaction feedback. The dynamic weight adjustment module generates initial technical dimension weights based on the initial technical dimension weight adjustment instruction or the preset retrieval strategy corresponding to the display mode selection, and outputs them to the semantic retrieval layer; it also receives the interactive feedback, dynamically optimizes the initial technical dimension weights based on the interactive feedback, and outputs the optimized technical dimension weights to the semantic retrieval layer in real time.

[0009] This invention also provides an electronic device, including: a memory, a processor, and a computer program stored in the memory and executable on the processor. When the computer program is executed by the processor, it implements the steps of the patent retrieval method based on dynamic weight adjustment and multimodal semantics described above.

[0010] This invention also provides a computer-readable storage medium storing an information transmission implementation program, which, when executed by a processor, implements the steps of the patent retrieval method based on dynamic weight adjustment and multimodal semantics described above.

[0011] The embodiments of the present invention can include the following beneficial effects: The embodiments of the present invention achieve adaptive optimization of patent search strategy through dynamic weight adjustment and user feedback closed loop, effectively overcoming the search bias caused by the fixed weight of traditional system; combined with the deep semantic understanding and term alignment of patent adaptation, it significantly improves the ability to accurately capture the core of patent technology; and through intelligent data processing and multimodal display, while ensuring the accuracy of results and high information density, it greatly improves the overall efficiency of the search system and the efficiency of user decision-making. Attached Figure Description

[0012] To more clearly illustrate the technical solutions in one or more embodiments of this specification or in the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in this specification. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0013] Figure 1 This is a schematic diagram of a patent retrieval system based on dynamic weight adjustment and multimodal semantics according to an embodiment of the present invention; Figure 2 This is a system architecture diagram of an embodiment of the present invention; Figure 3 This is a schematic diagram of the dynamic weight optimization interface according to an embodiment of the present invention; Figure 4 This is a visualization of the technical association map of an embodiment of the present invention; Figure 5 This is a system implementation flowchart of an embodiment of the present invention; Figure 6 This is a flowchart of a patent retrieval method based on dynamic weight adjustment and multimodal semantics according to an embodiment of the present invention. Detailed Implementation

[0014] To enable those skilled in the art to better understand the technical solutions in one or more embodiments of this specification, the technical solutions in one or more embodiments of this specification will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of this specification, and not all of the embodiments. Based on one or more embodiments of this specification, all other embodiments obtained by those skilled in the art without creative effort should fall within the protection scope of this document.

[0015] The terms used in the embodiments of this invention are explained as follows: Dynamic weight adjustment: A technology that optimizes technical means and technical parameter values ​​(0~1 range) in real time based on user needs or behavior characteristics; Semantic retrieval: An information retrieval method that uses natural language processing technology to match query intent with patent text at the semantic level; BERT-Patent: A BERT pre-trained model adapted for the patent text domain, which jointly optimizes semantic representation with the patent element classification task through a mask prediction module; SimHash: A locality-sensitive hashing algorithm used to generate text fingerprints and detect duplicate content or similar documents; Siamese neural network: A neural network architecture containing two encoders with shared weights, used to calculate semantic similarity between input samples; Conditional Random Field (CRF): A probabilistic graphical model used for sequence labeling tasks to identify entities such as technical terms and legal status in patent texts.

[0016] System Implementation Examples According to embodiments of the present invention, a patent retrieval system based on dynamic weight adjustment and multimodal semantics is provided. Figure 1 This is a schematic diagram of a patent retrieval system based on dynamic weight adjustment and multimodal semantics according to an embodiment of the present invention, as shown below. Figure 1 As shown, the patent retrieval system based on dynamic weight adjustment and multimodal semantics according to an embodiment of the present invention specifically includes: an input layer 10, a semantic retrieval layer 12, a processing layer 14 and a display layer 16 that are interconnected with each other, and a dynamic weight adjustment module 18 that is connected to the input layer, the semantic retrieval layer and the display layer respectively; The input layer 10 is used to receive the user's natural language query, initial technical dimension weight adjustment instruction, and display mode selection; The semantic retrieval layer 12 is used to retrieve and sort relevant patents from the patent document library based on the natural language query and the technical dimension weights provided by the dynamic weight adjustment module, forming preliminary retrieval results; wherein, the technical dimension weights include initial technical dimension weights and optimized technical dimension weights; The semantic retrieval layer specifically includes: a patent adaptive semantic encoder, a term alignment submodule, and a dynamic weight fusion unit; wherein, the patent adaptive semantic encoder is used to encode the natural language query and patent document into semantic vectors; The terminology alignment submodule is used to normalize heterogeneous technical terms in the natural language query and patent document before encoding, based on a pre-built patent terminology equivalence matrix. The dynamic weight fusion device has its weight input end connected to the output end of the dynamic weight adjustment module. It is used to receive the technical dimension weights and perform weighted fusion of the semantic vector and the technical dimension sub-vectors extracted from the patent document to generate the final search relevance score. The patent adaptive semantic encoder is a language model trained on patent text. Its training tasks include mask language modeling to enhance the technical terms in the patent text, and multi-label classification to identify technical means and technical effect elements.

[0017] The processing layer 14 is used to standardize the preliminary search results; The processing layer specifically includes: a data cleaning pipeline, a structured parsing engine, and a multimodal conversion engine; The data cleaning pipeline is used to deduplicate and filter low-quality data from the preliminary search results. The structured parsing engine is used to extract named entities from the filtered patent text and parse the citation relationships between claims to generate structured patent data. The multimodal conversion engine is used to convert the structured patent data into output data in various predetermined formats for use by the presentation layer.

[0018] The data cleaning pipeline includes: The duplicate patent identification unit is used to accurately match the preliminary search results based on the patent application number, and to generate text fingerprints for patents that cannot be accurately matched using the SimHash algorithm, to perform similarity comparison based on the text fingerprints, and to remove duplicates for patents with similarity greater than a preset threshold. The low-quality data filtering unit integrates a rule engine to filter patent data according to preset threshold conditions, which include at least one of the following: the number of claims, the number of words in a specific chapter of the specification, and the clarity score of the drawings. The structured parsing engine includes: Named entity recognition model for identifying and extracting technical terms, legal status, and timestamp entities from patent text; The claim tree construction module is used to construct a claim tree data structure that represents the hierarchical structure of claims based on the parsed claim reference relationships. The multimodal conversion engine supports generating output data in various formats, including XML, JSON, Markdown, and visual charts. The JSON format data encapsulates a correlation matrix between technical means and technical effects, and the visual charts include heatmaps for comparing the matching degree between input parameters and patented technical features. The processing layer further includes a feasibility scoring submodule, which calculates the feasibility score of the patent based on the number of embodiments, the clarity of the accompanying drawings, and the completeness of the experimental data, and outputs the feasibility score as an additional technical dimension parameter to the presentation layer.

[0019] The display layer 16 is used to select the standardized search results to be visualized and rendered in a corresponding multimodal form according to the display mode, and to receive the user's real-time weight adjustment instruction as a new initial technical dimension weight adjustment instruction through the interactive control, as well as to obtain the user's interactive feedback. The presentation layer includes a multimodal renderer, interactive controls, and a user feedback collector; wherein, the multimodal renderer is used to select and call the corresponding format data converted by the processing layer according to the presentation mode, and generate and render a view including a patent summary, claims, specification, and drawings; The interactive controls are integrated into the user interface generated by the multimodal renderer, and are used to provide real-time adjustment controls for the weight of technical dimensions, and respond to user operations. The user feedback collector is used to capture user interactions through the interactive controls or directly with the rendered view, and generate the interactive feedback. The multimodal renderer supports generating at least one of the following views: A summary view that presents the basic information of the patent in a structured card format; The claims view displays the dependency relationships of claims in a hierarchical layout and supports infringement comparison functionality; The instruction manual view is divided into sections based on technical issues, technical solutions, and technical effects, with each section analyzed and highlighted. The attached view provides an interactive viewing experience, supporting partial zoom and component annotation; The presentation layer further includes a role-adaptive presentation module and a retrieval, tracing, and evidence storage module; The adaptive role display module is used to dynamically adjust the interface layout and information focus of the multimodal renderer according to the current user's preset role; wherein, the adaptive role display module is used to professionally evaluate the legal status and claim tree of user role focus rendering, and to score the technical effect of enterprise user role focus rendering and provide patent drawings; The retrieval traceability and evidence storage module is used to perform hash processing on the user's key retrieval operations, weight adjustment records, and corresponding result sets, and then store them on the blockchain for evidence storage, supporting trusted verification and auditing of the retrieval process.

[0020] The dynamic weight adjustment module 18 is used to generate initial technical dimension weights according to the initial technical dimension weight adjustment instruction or the preset retrieval strategy corresponding to the display mode selection and output them to the semantic retrieval layer; and to receive the interactive feedback, and dynamically optimize the initial technical dimension weights based on the interactive feedback, and output the optimized technical dimension weights to the semantic retrieval layer in real time.

[0021] The following describes in detail the above-mentioned technical solutions of the present invention with reference to the specific circumstances of the patent retrieval system based on dynamic weight adjustment and multimodal semantics in the embodiments of the present invention.

[0022] This invention proposes a multimodal semantic retrieval system based on a Dynamic Weight Adjustment Mechanism (DWAM), suitable for solving the problems of fixed technical dimension weights and semantic understanding biases in patent retrieval. This system achieves high-precision patent retrieval and multimodal result display by integrating multi-dimensional parameter collaborative optimization, attention-enhanced semantic analysis, and adaptive data post-processing techniques. Its core components include: 1. Patent Semantic Intelligent Retrieval: Based on deep semantic matching and cross-dimensional parameter fusion algorithms using Natural Language Processing (NLP); 2. Dynamic weight adaptive adjustment mechanism: a hierarchical weight allocation and reinforcement learning optimization method oriented towards core retrieval dimensions such as technical means and technical effects; 3. Data processing and visualization technology system: Supports the entire backend processing and frontend presentation technology for structured analysis, visualization rendering and interactive display.

[0023] I. The technical architecture of the embodiments of the present invention includes: 1. Dynamic Strategy Layer: Solves the problem of strategy rigidity. This invention proposes a dynamic weight adaptive adjustment mechanism to reinforce learning-driven weight allocation: (1) Construct a dynamic weight adjustment module (DWAM): optimize the weight allocation strategy in real time through the Q-learning algorithm. The input layer receives user scenario tags (such as "infringement evidence retrieval"), retrieval history behavior and patent text features; the output layer generates dynamic weights for dimensions such as technical means, technical effects, and legal status (such as increasing the weight of technical means to 0.7 in the infringement scenario).

[0024] (2) Feedback closed-loop design: Input the user’s interactive behavior on the search results (such as clicking, marking related patents) into DWAM in real time, and adjust the strategy network parameters through the reward function (such as click rate ↑ → weight effectiveness ↑) to achieve end-to-end dynamic optimization.

[0025] Preferably, the dynamic weight adjustment module can also incorporate a federated learning framework, which supports multiple users or institutions to train weight models locally and only upload model gradient updates. Combined with differential privacy technology, this protects the privacy of user behavior data, thereby improving the availability and security of the system in cross-institutional collaboration scenarios.

[0026] 2. Patented Adaptive Semantic Layer: Solving the Problem of Semantic Modeling Failure (1) Patent adaptive semantic modeling: Through triplet-aware patent semantic encoding, a multimodal pre-trained model (such as a patent-adapted BERT variant) is adopted. The input layer integrates the patent title, abstract, claims and figure description, and captures the contextual association between technical problem-solution-effect through a multi-head attention mechanism.

[0027] Preferably, a cross-modal joint pre-training framework of patent text + figures + citation relationships can be adopted. In the pre-training stage, a patent figure description generation task and a figure-text alignment task are introduced to enhance the model's ability to visualize and understand technical means and realize joint image-text retrieval.

[0028] (2) Design term alignment module: Based on the patent term library (such as CPC classification number mapping table), construct a heterogeneous term equivalence matrix to realize the automatic normalization of terms such as "lithium battery → lithium-ion battery".

[0029] (3) Through weight-driven semantic space fusion, the dynamic weights output by DWAM are embedded into the semantic space to generate a joint vector, such as: ( , (These are dynamic weighting coefficients), supporting multi-dimensional weighted retrieval.

[0030] 3. Intelligent Architecture Layer: Solving architectural performance bottlenecks Intelligent architecture optimization: Through decision tree-driven data cleaning and deduplication, a legal status decision tree is constructed, and version relationships are automatically marked based on patent lifecycle rules (e.g., application → publication → grant → invalidation). This is combined with an applicant alias database (e.g., "XXX Technology Co., Ltd."). Entity alignment was performed on “Mouwei”, reducing the redundant patent duplication rate to <5%.

[0031] Furthermore, a feasibility scoring submodule can be added to the data processing flow. Based on features such as the number of embodiments, the clarity of the accompanying drawings, and the completeness of the experimental data, the actual feasibility of the patent can be automatically evaluated, and a feasibility score can be output for search, ranking, and display.

[0032] II. The system architecture of this invention adopts a four-layer heterogeneous collaborative architecture. like Figure 2 As shown, the diagram illustrates the input layer (weight adjustment UI, natural language query) → semantic retrieval layer (BERT model, dynamic weight fusion engine) → processing layer (cleaning / transformation pipeline) → presentation layer (multimodal rendering engine), and labels the key data streams: technical feature vectors, user feedback signals, and multi-format output streams.

[0033] 1. Input layer: Supports multi-source parameter input, including technical dimension weights, semantic search expressions, and display mode selection.

[0034] 2. Semantic Retrieval Layer: BERT-based deep semantic feature extraction and dynamic weight fusion module.

[0035] 3. Processing layer: Includes the data cleaning pipeline and the format conversion engine.

[0036] 4. Presentation layer: Multimodal Renderer and Feedback Loop.

[0037] III. Core Module Decomposition 1. Four-dimensional dynamic weight adjustment module (4D-DWAM) This module ensures that the system can resolve issues related to the inability to understand patent technology logic, terminological heterogeneity, and the correlation of technology triples.

[0038] A. Parameter system and technology mapping, as shown in Table 1. Table 1. Parameter System and Technology Mapping Relationship

[0039] The semantic extraction methods for the technical dimension include: extracting technical action-object pairs (such as "coating / electrode active material") from claims through dependency parsing; and identifying effect-indicator pairs (such as "increase / energy density") in the abstract based on semantic role labeling. +20%”).

[0040] B. Dynamic Weight Optimization Model Based on the aforementioned parameter system, this system adopts an end-to-end joint training framework. It optimizes weights in real time through user behavior feedback (click heatmaps, filtering operation sequences), eliminating retrieval bias caused by static weights and achieving accurate mapping of user intent. For example... Figure 3 As shown, a slider (continuously adjustable from 0 to 1) displays the weighting of technical methods / effects, providing a real-time radar chart of weight distribution and a preview of search results. A heatmap layer illustrates the impact of weight adjustments on the ranking of the Top 10 patents. Experiments show that 4D-DWAM improves the Top-5 search accuracy from 62% to 89% and reduces response time by 40%.

[0041] a. Weight Space Modeling: The weight allocation problem is transformed into an optimization problem in a high-dimensional space, with the objective function being: (1); In the formula, The technical means weight represents the relative importance of technical means in the semantic matching of patent retrieval; the model adjusts... This influences the ranking of patents related to technical methods and steps in the search results. For technical effect weighting, this parameter represents the relative importance of the technical effect dimension; adjustment It will affect the ranking of patents related to function and effectiveness in the search results. This is the accuracy balancing coefficient, used to assign weight to "accuracy" in the optimization objective. The default value of 0.7 indicates that the model focuses more on improving the accuracy of the top-N results, that is, ensuring that the most relevant patents retrieved are as relevant as possible. This is the recall balancing coefficient, used to assign weight to "recall" in the optimization objective. The default value of 0.3 indicates that while ensuring accuracy, the model also appropriately considers the ability to retrieve all relevant patents to avoid omissions.

[0042] In short, the algorithm automatically adjusts and The size of the target area, finding the optimal balance between technical means and technical effects, and the guiding principle of the optimization process ( and This ensures that the final result is both accurate and comprehensive.

[0043] b. Gradient Descent Optimization: Calculate the weight gradient based on user click behavior data (CTR): (2); In the formula, CTR is the click-through rate, which is the optimization objective of the model and reflects the quality of the search results; a higher CTR indicates that users are more accepting of the retrieved patents, indicating higher weight. and The more reasonable the settings, the better. N is the number of samples, referring to the total number of data pairs (query terms, search result patents) used in one iteration for model training. The technical means score for patent i represents the semantic relevance score between the content of patent i and the query term in the technical means dimension. The technical effect score for patent i represents the semantic relevance score between the content of patent i and the query term in the dimension of technical effect. The actual click-through rate of patent i is calculated based on real user click behavior data and represents the historical probability that patent i was clicked by users in the search results.

[0044] The core idea of ​​the gradient formula above is to adjust the model weights based on user feedback (click behavior).

[0045] c. Weight constraint conditions: satisfy + <1, supports manual adjustment by users via slider control (precision level 0.01) or automatic optimization via reinforcement learning model.

[0046] d. Initialization of the Analytic Hierarchy Process (AHP): Construct a judgment matrix and calculate initial weights, taking the comparison of the importance of technical means and technical effects as an example: (3); e. Reinforcement learning online updates: The Q-Learning algorithm is adopted, with user click behavior as the reward signal R (click-related patents (R=+1), click-irrelevant patents (R=-1)). The state space is defined as the current weight vector. , The action space is the weight adjustment step size (±0.05), and the weights are optimized iteratively using the Bellman equation: (4); In the formula, Let be the state-action value function, representing the expected cumulative reward obtained by performing action 'a' in state 's'; in patent retrieval, it is called the Q-value, which represents the expected long-term total return that the agent (i.e., the retrieval system) can obtain after performing a specific action (a) in the current state (s). 's' is the current state of the system, i.e., the current weight combination ( , ); 'a' represents the operation that the agent can perform, namely the weight adjustment action—making some kind of fine-tuning to the weights (e.g., adjusting the weights). (Increase by 0.05); therefore This represents how beneficial "making this fine-tuning action under the current weight settings" is in the long run. R represents the immediate reward: the direct return obtained after performing the action; in patent search, it represents user click behavior feedback: clicking on a relevant patent R=+1, clicking on an irrelevant patent R=-1. The learning rate controls the rate at which new information overwrites old information (0 ≤ 0). ≤1); it determines the update range of the weight adjustment strategy in patent search and affects the convergence speed of the algorithm. The discount factor measures the importance of future rewards (0 ≤ ≤1); its function is to balance current rewards and long-term benefits. The larger the scale, the more emphasis is placed on long-term optimization effects. The next state: the new state reached after performing an action; that is, the new state to which the environment transitions after performing action a—the new weight vector after adjusting the step size. , ). For the next action to be performed; The maximum expected value for the next state is the expected effect of choosing the optimal adjustment strategy under the new weight configuration.

[0047] f. Dynamic retrieval strategy recommendation engine: Based on user history and role tags, a lightweight strategy recommendation model is trained to recommend the optimal weight combination and semantic parameters during the retrieval initialization phase, and supports one-click adaptation functions such as infringement mode and survey mode.

[0048] 2. Attention-enhanced semantic retrieval module This module ensures that the system can deeply understand the patent content.

[0049] A. Input and Output Input: User Query: A piece of natural language text describing the technical requirements; Patent Document Collection: A vast collection of patent texts (such as titles, abstracts, and claims) to be searched.

[0050] Output: A list of patents sorted from highest to lowest relevance. This sorting combines semantic similarity, technical logic matching, and patent-specific business weights, aiming to place the most relevant patents at the top.

[0051] B. Technical Implementation Principles a. Domain-enhanced semantic representation model Pre-training framework: A BERT-Patent pre-trained model is built based on the BERT architecture, and domain adaptation is performed on 1 million patent texts. Training tasks include: Masked Language Model (MLM): Randomly mask 15% of the tokens (the masking ratio for technical terms is increased to 30%) to predict the original vocabulary; Patent element classification: Multi-label classification of technical means, technical effects, claim types, etc., with an output dimension of 128-dimensional semantic vector.

[0052] Siamese neural network matching: A Siamese network architecture is constructed, and two BERT-Patent encoders with shared weights process the query statement Q and the patent document D respectively to generate semantic vectors. , And calculate the cosine similarity: (5); Dynamic weight fusion: Combining the matching degree of technical means and technical effects to generate the final search score. (6); in, , These are the weights for technical means and technical effects, respectively. , These are the cosine similarity of the subvectors representing technical means and technical effects, respectively.

[0053] b. Search result ranking strategy The BM25+ semantic similarity hybrid ranking model is adopted, which incorporates semantic matching scores into the traditional information retrieval model: (7); in, This is the fusion coefficient (default 0.6), and the optimal value is determined through cross-validation. To calculate keyword matching scores based on traditional information retrieval.

[0054] Preferably, the specific implementation process of this technology is as follows: A. BERT pre-training and fine-tuning: Pre-training data: 5 million patent abstracts (from USPTO, EPO, JPO databases), with technical entity masking added in the Masked Language Modeling (MLM) task (e.g., [MASK] coating to improve conductivity).

[0055] Fine-tuning objective: Patent similarity calculation task (Siamese Network structure), loss function is: (8); In the formula, L is the loss function, which is the objective function that needs to be minimized during model training; the smaller the value, the better the model can distinguish between similar samples (positive sample pairs) and dissimilar samples (negative sample pairs). This is a similarity function used to calculate the similarity between two patent vectors; it usually refers to cosine similarity, with values ​​between [-1, 1], and the larger the value, the more similar they are. , The vector representation generated for the BERT model is typically the vector corresponding to the label obtained after the patent text has been processed by the model, which contains the semantic information of the entire patent. and This constitutes a positive sample pair (i.e., two patents with similar semantics). In other words, in the same batch, except Vectors of other patent samples, With each By forming negative sample pairs, the model learns to... and The similarity is much greater than with any other The similarity is calculated as follows: B is the batch size, which is the total number of samples processed by the model in one training session; the summation term in the denominator covers all samples within the batch, so this loss function utilizes all other samples within the batch as negative samples to achieve efficient contrastive learning. is the temperature coefficient, a hyperparameter greater than 0, used to adjust the smoothness of the softmax distribution. A smaller value will amplify the difference in similarity, making the model pay more attention to high-difficulty negative samples and making the learning objective more acute; Larger values ​​will smooth out differences in similarity, making the loss function smoother and easier to optimize, but the discriminative power may decrease.

[0056] B. Semantic-weighted fusion retrieval: a. Construct the Technical Feature Space, and the retrieval score formula is as follows: (9); In the formula, Score is the comprehensive retrieval score, which is used to determine the final relevance score between the query and the document. The higher the score, the more relevant the patent document is to the query intent, and the higher its ranking. For semantic fusion weights, there is an adjustable parameter between [0, 1] with an initial value of 0.6, which is used to balance the importance of semantic similarity and technical feature weight scores in the final score; It can be optimized according to different search scenarios (such as infringement search, technical research). Approaching 1: Retrieval relies more on deep semantic matching, similar to pure semantic retrieval; Close to 0: The search relies more on precise matching of technical dimensions (means / effects), similar to improved keyword weight search. Cosine similarity, i.e., user query vector With patent document vectors The semantic similarity between the two is the direct output of the deep semantic model (BERT), which is used to capture deep semantic associations beyond literal matching. , These represent the semantic vectors obtained after processing the user's query and the patent document to be retrieved using the BERT model, respectively. , These weights are dynamically set by the user or the system, representing the relative importance of two dimensions: technical means and technical effectiveness. For example, in infringement searches, more attention might be paid to technical means (…). (High setting); In patentability assessment, more attention may be paid to technical effects ( (Set high). , These are the scores for the technical means and technical effects extracted from patent documents using technologies such as entity recognition, and which match the query. They can also be understood as the matching score on that specific technical dimension.

[0057] b. Technical Implementation Path Entity recognition layer: Entities (such as metal structure optimization) are extracted based on BiLSTM-CRF model technology, with an F1 score of 92%.

[0058] Weighted embedding layer: Maps legal status weights (e.g., 0.8) to the word vector space, and calculates them using weighted cosine similarity. (10); In the formula, The weighted similarity score represents the overall relevance score between query q and document d after considering the specific weight of each word; the higher the score, the more similar the two are in a weighted sense. q and d represent the query vector and document vector, respectively, which are dense vector representations of the query statement and patent document after transformation by natural language processing models (such as the input layer of Word2Vec and BERT), and are the foundation of the entire calculation. , Let q and d be the components of vectors q and d in the i-th dimension, respectively. In traditional vector space models, this can be understood as the weight of the i-th word in the query and document (such as the TF-IDF value). In deep learning models, this is the numerical value of a high-dimensional dense vector in a certain dimension. For each term, the weight represents the importance weight assigned to the i-th word (or feature dimension); this weight does not come from the text itself, but from external knowledge or specific rules, for example: Legal status weight: If a term is strongly correlated with a high-value legal status (such as validity), its... It can be set to 0.9; Technical dimension weight: derived from the upper-level dynamic weight adjustment mechanism; for example, words highly related to technical means can receive higher weight. Domain knowledge weighting: In the patent field, words in the claims are usually more important than words in the abstract.

[0059] Attention mechanism: The retrieval focus is dynamically allocated through multi-head attention, and the attention weight of key technical dimensions (such as legal status = 0.9) is increased by 300%.

[0060] In summary, this module constructs a multi-layered, deep understanding framework that extends from perception to cognition: Precise perception (entity recognition layer) Solve the "what is" question; Domain knowledge fusion (weighted embedding layer) Solve the question of "who is more important"; Logical relevance focusing (attention mechanism) Solve the "why is it relevant" question; Deep Semantic Matching (BERT) Based on all the above information, perform end-to-end deep correlation calculation; Robust output results (mixed sorting) By combining the advantages of traditional methods, we can ensure the overall reliability of the system performance.

[0061] Ultimately, this process transforms the retrieval system from a simple keyword matcher into an intelligent patent analyst capable of understanding technical logic and respecting domain rules, thereby effectively overcoming the technical relevance bottleneck of shallow semantic models.

[0062] Preferably, this embodiment of the invention also proposes a dynamic patent family map and reordering mechanism: after sorting the search results, a dynamic patent family map is constructed, and the map is reordered based on the evolution of family, citation, and legal status, thereby improving the technical coherence and legal foresight of the results.

[0063] 3. Backend processing module This module aims to address the system architecture performance bottleneck, namely, the low level of intelligence in the data processing workflow, high error rate, and high duplication rate, which lead to low overall system performance and poor result quality.

[0064] A. Data Cleaning and Preprocessing Duplicate patent identification: The SimHash algorithm is used to generate patent text fingerprints (hash value length 128 bits), the similarity calculation threshold is set to 0.95, and combined with patent number uniqueness verification, the accuracy rate is >99.5%; Low-quality data filtering: Patents that do not meet the following criteria are removed using a rule engine: number of claims. 3 items; word count of the "Invention Content" section of the specification. 500 words; rating for the clarity of the attached image. 70 points (based on OCR text recognition accuracy).

[0065] B. Structured parsing and format conversion a. Patent element extraction: The Conditional Random Field (CRF) model is used to identify named entities (technical terms, legal statuses, timestamps, etc.), and regular expressions are used to parse the claim citation relationships to construct a claim tree data structure. class ClaimNode: def __init__(self, claim_id, text, parent=None): self.claim_id = claim_id self.text = text self.parent = parent self.children = [] b. Multimodal conversion engine output support, as shown in Table 2: Table 2. Schematic diagram of multimodal conversion

[0066] Structured data: Generate patent metadata JSON conforming to ISO 26304 standard, including fields such as technical means, technical effects, and claim hierarchy; Visualizing data: Technical feature matrix: Compare the matching degree between input parameters and patent technical features, and visualize using a heatmap (color depth mapping similarity score); Vectorization of Figures: Convert PDF figures into SVG format, supporting coordinate mapping and interactive annotation (such as clicking on a figure component to display the corresponding claim paragraph).

[0067] C. Value Assessment of Sentiment Analysis-Assisted Technologies A sentiment analysis module is introduced into the technical effect dimension to extract sentiment tendencies from texts such as patent citations and examination opinions, serving as an auxiliary dimension for evaluating technical value.

[0068] 4. Page Display Module This module aims to address the problems of unintuitive result presentation, poor interactivity, and inability to support decision-making caused by rigid strategies and architectural bottlenecks.

[0069] A. Multimodal display architecture Summary view: Presents basic patent information (application number, publication date, applicant, IPC classification number) in card format, and supports batch preview; Claims view: The claims are displayed hierarchically (independent claims are bolded and highlighted in red, dependent claims are indented), and an "Infringement Comparison" button is provided to generate a feature comparison table with one click; Instruction manual view: It is divided into blocks according to the CPC (Cooperative Patent Classification) system, highlighting the "technical problem-solution-technical effect" triple, and supporting the display of term definitions when hovering the mouse over keywords; Attached image view: An interactive attached image view is implemented based on D3.js, supporting local zoom and feature highlighting.

[0070] B. Dynamic interactive functions a. Real-time weight adjustment: The weight is dynamically adjusted through the Ant Design Slider component, triggering real-time updates throughout the "forward propagation-retrieval-rendering" process (end-to-end latency < 800ms), and the matching score under the current weight is marked in the results list; b. Intelligent analysis tools Generate a weight-recall curve: to show the changes in search performance under different combinations; Provides a technology evolution roadmap: Based on IPC classification and application time, it visualizes the patent layout trends in the target technology field.

[0071] c. Interactive visualization Technology Association Graph: Dynamically lays out technology nodes and associated edges using the Force-Directed Layout algorithm; such as... Figure 4 As shown, the node classifications are: technical means, technical effects, and patents; edge weights represent the correlation strength between patents and technical features (represented by line width and transparency).

[0072] Legal Status Timeline: Based on D3.js Brush & Zoom interaction, it supports time range filtering and patent family aggregation display.

[0073] C. Blockchain-based evidence storage and retrieval traceability mechanism By storing key operations such as user search behavior, weight adjustment records, and result sets on the blockchain, it supports credible verification and traceability of the search process, and is suitable for judicial and auditing scenarios.

[0074] 5. Adaptive Character Display Module This module aims to enable different roles in the system to adaptively display different interfaces.

[0075] The strategy is generated dynamically, and the code example is as follows: def generate_display(user_role, patent_data): """ The core function of the character adaptive display module A customized patent information display interface is dynamically generated based on the user's role. To achieve a personalized display effect that is "one face for a thousand people". parameter: user_role: str - User role identifier, determines the display strategy. patent_data: dict - a dictionary of patent data containing complete structured information about patents. return: Corresponding view component instance - Patent information display interface customized according to role """ #Role Assessment Branch: Professional Evaluation of User Perspective #Professional assessment users focus on the stability of the legal status and the hierarchical structure of the claims, which are used for patent examination and validity determination. if user_role == "Professional Evaluation User": #Return to the legal professional view, focusing on the legal dimensions and scope of rights analysis of patents. return LegalView( # legal_status: Patent legal status (e.g., granted / substantive examination / invalidation), a key focus for professional assessment users. #Used for quickly assessing the validity and legal risks of patents legal_status=patent_data['status'], #claims_tree: A hierarchical layout of the claims. # Visualize the claims in a tree structure using the HierarchicalLayout component #To facilitate professional evaluation and help users clearly understand the logical relationship between independent and dependent claims. claims_tree=HierarchicalLayout(patent_data['claims']) ) #Role Assessment Branch: Corporate IP Department Perspective Enterprise users focus on the technological value and industrialization potential for use in technology planning and business decisions. elif user_role == "Enterprise IP Department": #Return to the technology value view to highlight the technical effects and implementation details of the patent. return TechView( #tech_effect: Technical effectiveness score (1-10 points), quantifying the technological innovation value of patents. #Help companies assess the competitive advantage of this patent in the technological arena. tech_effect=patent_data['effect_score'], #figures: Vector graphics rendering of patent illustrations #A technical illustration demonstrating high-definition interactive features using VectorGraphicsRenderer #Supports zooming, annotation, and other operations, facilitating a deeper understanding of the technical implementation plan by technical personnel. figures=VectorGraphicsRenderer(patent_data['diagrams']) ) #Expansion Points: Additional character branches can be added. #Customized display logic for different roles such as R&D engineers and intellectual property lawyers IV. Detailed Implementation 1. Hardware Deployment Plan A. Computing Cluster: Semantic retrieval node: NVIDIA DGX A100 (8×A100 80GB GPU), used for BERT inference.

[0076] Data processing node: Apache Spark cluster (32 cores / 256GB RAM / 10GbE network).

[0077] B. Storage Architecture: Patent original document storage: MongoDB sharded cluster (Shard Key is the hash value of the publication number).

[0078] Vector Index: Faiss GPU version (IVF-PQ index structure).

[0079] 2. Implementation of key algorithms A. BERT fine-tuning code (PyTorch): class PatentSiamese(nn.Module): def __init__(self, bert_model): super().__init__() self.bert = BertModel.from_pretrained(bert_model) self.cos = nn.CosineSimilarity(dim=1) def forward(self, input_ids1, attention_mask1, input_ids2, attention_mask2): vec1=self.bert(input_ids1, attention_mask1).last_hidden_state[:, 0,:] vec2=self.bert(input_ids2, attention_mask2).last_hidden_state[:, 0,:] return self.cos(vec1, vec2) The code above is a semantic similarity calculation model: it uses BERT to extract the semantic vectors of the two texts respectively, and then calculates their matching score through cosine similarity.

[0080] B. Dynamic weight gradient calculation (TensorFlow): optimizer=tf.keras.optimizers.Adam(learning_rate=0.001) def train_step(user_feedback): with tf.GradientTape() as tape: loss = tf.reduce_mean( tf.square(user_feedback['expected_score'] model(user_feedback['input'])) ) gradients=tape.gradient(loss, model.trainable_variables) optimizer.apply_gradients(zip(gradients, model.trainable_variables)) The code above is a TensorFlow training step function: it uses gradient strips to record the forward propagation loss (mean squared error), then calculates the gradients and updates the model parameters to dynamically optimize the model based on user feedback.

[0081] 3. Software implementation process, such as Figure 5 As shown A. Input parameter processing flow (Step 1) a. Data input validation Country of Origin: Verifies whether the input conforms to the ISO 3166-1 standard. Unmatched items are set to WO (World Intellectual Property Organization) by default. Legal status: Supports Boolean logic combinations (e.g., valid AND NOT rejected), and parses logical expressions using the Shunting-yard algorithm; Semantic parameters: SpaCy is used for word segmentation and part-of-speech tagging, stop words (such as "a kind of" and "as described") are filtered, and noun phrases are extracted as core search terms.

[0082] b. Weight initialization strategy When no user input is provided, the domain default weight is used. =0.6、 =0.4, automatically assigned if only a single semantic parameter is input. =1, and the other parameter has a weight of 0.

[0083] B. Semantic Retrieval Implementation Details (Step 2) a. Model training optimization Pre-training phase: Train the BERT-Patent model on a TPU server with a batch size of 512, a training period of 100 epochs, and a loss function convergence threshold of 0.3. Fine-tuning phase: Training data is generated using a triplet mining strategy. Positive samples are patent pairs with the same technical theme, while negative samples are randomly selected unrelated patents. 100,000 triplet pairs are generated in each training round.

[0084] b. Search performance optimization The Faiss index is built to store patent semantic vectors, supporting near nearest neighbor search (ANN), and the retrieval speed is increased to 1000 articles / second.

[0085] C. Backend data processing implementation (Step 3) a. Data cleaning project Deduplication of duplicate patents: First, precise matching is performed using application numbers. Then, SimHash fingerprint comparison is performed on patents without application numbers. Similarity calculation uses Hamming distance (threshold). 5); Noise data filtering: The DBSCAN algorithm is used to cluster features such as the number of claims and the number of words in the specification, and outliers (low-density data points) are identified and removed.

[0086] b. Format conversion technology Structured output: Extend the patent metadata model according to the LOM (Learning Object Metadata) standard, and define the RDF triple mapping rules for core elements such as technical means and technical effects; Visualization generation: Matplotlib is used to draw a heatmap of technical features. The horizontal axis represents the input parameter keywords, the vertical axis represents the patent technical features, and the color values ​​correspond to the weighted sum of TF-IDF weights and semantic similarity.

[0087] D. Page display technology implementation (Step 4) a. Front-end technical architecture Single-page application (SPA) routing management is implemented using React Router, and component state is centrally managed through Redux; The weight adjustment component uses React Beautiful DnD to implement drag-and-drop weight assignment, and supports precise adjustment via keyboard input (accuracy level 0.001).

[0088] b. Dynamic rendering mechanism When a user switches display tabs, a format conversion request is sent to the backend via WebSocket, and structured data is received and rendered in a streaming manner. Supports progressive rendering: text content is loaded first, followed by asynchronous loading of attached images and visualizations, improving the user experience.

[0089] Furthermore, in the software implementation process, each newly added module has corresponding sub-processes and interfaces. For example: The cross-modal pre-training module synchronously trains the image-text alignment network during the BERT fine-tuning phase; The feasibility assessment module is executed after data cleaning, and the output score is written into the patent metadata. The blockchain evidence storage module triggers on-chain operations at key user interaction nodes, and the log structure conforms to the IPFS+Ethereum dual-chain storage specification.

[0090] In summary, the core of the embodiments of the present invention lies in: 1. This invention proposes a weight adaptive algorithm based on reinforcement learning to achieve dynamic optimization of retrieval strategies. Unlike traditional fixed weight or heuristic adjustment methods, this invention achieves dynamic coordination of four dimensions: legal status, technical means, technical effects, and technical field. Moreover, the weight adjustment algorithm has online learning capabilities. 2. By combining the Analytic Hierarchy Process (AHP) with reinforcement learning, a dynamic weight adjustment framework adapted to patent retrieval scenarios was constructed, solving the technical challenge of prioritizing multi-dimensional parameters. 3. An end-to-end patent retrieval pipeline was designed, including preprocessing, semantic matching, weight fusion, and data processing, to achieve full automation from natural language input to multimodal output; 4. A BERT-Patent pre-trained model specifically for the patent field was developed, which significantly improves the semantic parsing accuracy of technical and legal terms (F1 score is 18% higher than that of the general BERT model). 5. By combining reinforcement learning (PPO) with multi-head attention mechanism, the contradiction between "retrieval accuracy and user demand adaptation" is resolved, and the response time is reduced by 40%.

[0091] Method Implementation Examples According to embodiments of the present invention, a patent retrieval method based on dynamic weight adjustment and multimodal semantics is provided. Figure 6 This is a flowchart of a patent retrieval method based on dynamic weight adjustment and multimodal semantics according to an embodiment of the present invention, as shown below. Figure 6 As shown, the patent retrieval method based on dynamic weight adjustment and multimodal semantics according to an embodiment of the present invention specifically includes: Step S601: Receive the user's natural language query, initial technical dimension weight adjustment instruction, and display mode selection through the input layer; Step S602: Based on the natural language query and the technical dimension weights provided by the dynamic weight adjustment module, the semantic retrieval layer retrieves and sorts relevant patents from the patent document library to form preliminary retrieval results; wherein, the technical dimension weights include initial technical dimension weights and optimized technical dimension weights; Step S603: The preliminary search results are standardized by the processing layer; Step S604: The display layer selects the standardized search results in the corresponding multimodal form for visualization rendering according to the display mode, and receives the user's real-time weight adjustment instruction as a new initial technical dimension weight adjustment instruction through the interactive control, and obtains the user's interactive feedback. Step S605: The dynamic weight adjustment module generates initial technical dimension weights according to the initial technical dimension weight adjustment instruction or the preset retrieval strategy corresponding to the display mode selection and outputs them to the semantic retrieval layer; and receives the interactive feedback, dynamically optimizes the initial technical dimension weights based on the interactive feedback, and outputs the optimized technical dimension weights to the semantic retrieval layer in real time.

[0092] The embodiments of the present invention are method embodiments corresponding to the system embodiments described above. The specific operations of each step can be understood by referring to the description of the system embodiments, and will not be repeated here.

[0093] In summary, compared with the prior art, the beneficial effects of the embodiments of the present invention include: 1. Retrieval Accuracy: Under the same retrieval conditions, the relevant patent recall rate of this invention reaches 92%, which is 45% higher than traditional keyword retrieval and 22% higher than existing semantic retrieval systems; and in the infringement analysis scenario, the patent relevance ranking accuracy rate increases from 0.52 to 0.84, the applicant alias misjudgment rate decreases to 5%, and the patent version deduplication accuracy rate increases to 95%. 2. Processing efficiency: The average time for a single batch search (100,000 patents) is 1.2 seconds; data cleaning and format conversion time is controlled within 800ms; page interaction latency is <500ms; weight update latency has been reduced from hours to seconds; and the closed loop of user behavior feedback has improved the efficiency of weight strategy iteration by 50%. 3. Improved accuracy through scene adaptation: The accuracy of weight allocation for technical means in infringement evidence retrieval scenarios is improved by 40%, and the relevance of technical effect weights in technical investigation scenarios is improved by 35%. 4. Supports cross-language retrieval (e.g., matching Chinese queries with English patents), with an accuracy rate exceeding that of traditional systems by 35%; 5. Breakthrough in interactive experience: Multimodal visualization first screen loading in less than 1 second, supporting real-time interactive analysis of 100,000+ patent data.

[0094] Device Example 1 This invention provides an electronic device, including: a memory, a processor, and a computer program stored in the memory and executable on the processor. When the computer program is executed by the processor, it performs the steps described in the method embodiment.

[0095] Device Example 2 This invention provides a computer-readable storage medium storing an information transmission implementation program, which, when executed by a processor, performs the steps described in the method embodiment.

[0096] The computer-readable storage media described in this embodiment include, but are not limited to, ROM, RAM, disk, or optical disk.

[0097] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A patent retrieval system based on dynamic weight adjustment and multimodal semantics, characterized in that, It includes an input layer, a semantic retrieval layer, a processing layer, and a presentation layer that are interconnected with each other, as well as a dynamic weight adjustment module that is connected to the input layer, the semantic retrieval layer, and the presentation layer respectively; The input layer is used to receive the user's natural language query, initial technical dimension weight adjustment instructions, and display mode selection; The semantic retrieval layer is used to retrieve and sort relevant patents from the patent document library based on the natural language query and the technical dimension weights provided by the dynamic weight adjustment module, forming preliminary retrieval results; wherein, the technical dimension weights include initial technical dimension weights and optimized technical dimension weights; The processing layer is used to standardize the preliminary search results; The display layer is used to select the standardized search results for visualization rendering in the corresponding multimodal form according to the display mode, and to receive the user's real-time weight adjustment instruction as a new initial technical dimension weight adjustment instruction through interactive controls, as well as to obtain the user's interactive feedback. The dynamic weight adjustment module is used to generate initial technical dimension weights according to the initial technical dimension weight adjustment instruction or the preset retrieval strategy corresponding to the display mode selection and output them to the semantic retrieval layer; and to receive the interactive feedback, dynamically optimize the initial technical dimension weights based on the interactive feedback, and output the optimized technical dimension weights to the semantic retrieval layer in real time.

2. The system according to claim 1, characterized in that, The semantic retrieval layer specifically includes: a patent adaptive semantic encoder, a term alignment submodule, and a dynamic weight fusion unit; The patent adaptive semantic encoder is used to encode the natural language query and patent document into a semantic vector. The terminology alignment submodule is used to normalize heterogeneous technical terms in the natural language query and patent document before encoding, based on a pre-built patent terminology equivalence matrix. The dynamic weight fusion device has its weight input end connected to the output end of the dynamic weight adjustment module. It is used to receive the technical dimension weights and perform weighted fusion of the semantic vector and the technical dimension sub-vectors extracted from the patent document to generate the final search relevance score. The patent adaptive semantic encoder is a language model trained on patent text. Its training tasks include mask language modeling to enhance the technical terms in the patent text, and multi-label classification to identify technical means and technical effect elements.

3. The system according to claim 1, characterized in that, The processing layer specifically includes: a data cleaning pipeline, a structured parsing engine, and a multimodal conversion engine; The data cleaning pipeline is used to deduplicatize and filter low-quality data from the preliminary search results. The structured parsing engine is used to extract named entities from the filtered patent text and parse the citation relationships between claims to generate structured patent data. The multimodal conversion engine is used to convert the structured patent data into output data in various predetermined formats for use by the presentation layer.

4. The system according to claim 3, characterized in that, The data cleaning pipeline includes: The duplicate patent identification unit is used to accurately match the preliminary search results based on the patent application number, and to generate text fingerprints for patents that cannot be accurately matched using the SimHash algorithm, to perform similarity comparison based on the text fingerprints, and to remove duplicates for patents with similarity greater than a preset threshold. The low-quality data filtering unit integrates a rule engine to filter patent data according to preset threshold conditions, which include at least one of the following: the number of claims, the number of words in a specific chapter of the specification, and the clarity score of the drawings. The structured parsing engine includes: Named entity recognition model for identifying and extracting technical terms, legal status, and timestamp entities from patent text; The claim tree construction module is used to construct a claim tree data structure that represents the hierarchical structure of claims based on the parsed claim reference relationships. The multimodal conversion engine supports generating output data in various formats, including XML, JSON, Markdown, and visual charts. The JSON format data encapsulates a correlation matrix between technical means and technical effects, and the visual charts include heatmaps used to compare the matching degree between input parameters and patented technical features.

5. The system according to claim 3, characterized in that, The processing layer further includes a feasibility scoring submodule, which calculates the feasibility score of the patent based on the number of embodiments, the clarity of the accompanying drawings, and the completeness of the experimental data, and outputs the feasibility score as an additional technical dimension parameter to the presentation layer.

6. The system according to claim 1, characterized in that, The presentation layer includes a multimodal renderer, interactive controls, and a user feedback collector; The multimodal renderer is used to select and call the corresponding format data after being converted by the processing layer according to the display mode, and generate and render a view including a patent summary, claims, specification and drawings; The interactive controls are integrated into the user interface generated by the multimodal renderer, and are used to provide real-time adjustment controls for the weight of technical dimensions, and respond to user operations. The user feedback collector is used to capture user interactions through the interactive controls or directly with the rendered view, and generate the interactive feedback. The multimodal renderer supports generating at least one of the following views: A summary view that presents the basic information of the patent in a structured card format; The claims view displays the dependency relationships of claims in a hierarchical layout and supports infringement comparison functionality; The instruction manual view is divided into sections based on technical issues, technical solutions, and technical effects, with each section analyzed and highlighted. The attached view provides an interactive viewing experience, supporting zoom-in and component annotation.

7. The system according to claim 6, characterized in that, The presentation layer further includes a role-adaptive presentation module and a retrieval, tracing, and evidence storage module; The adaptive role display module is used to dynamically adjust the interface layout and information focus of the multimodal renderer according to the current user's preset role; wherein, the adaptive role display module is used to professionally evaluate the legal status and claim tree of user role focus rendering, and to score the technical effect of enterprise user role focus rendering and provide patent drawings; The retrieval traceability and evidence storage module is used to perform hash processing on the user's key retrieval operations, weight adjustment records, and corresponding result sets, and then store them on the blockchain for evidence storage, supporting trusted verification and auditing of the retrieval process.

8. A patent retrieval method based on dynamic weight adjustment and multimodal semantics, used in the system of any one of claims 1 to 7, the method comprising: The input layer receives users' natural language queries, initial technical dimension weight adjustment instructions, and display mode selections. The semantic retrieval layer retrieves and sorts relevant patents from the patent document library based on the natural language query and the technical dimension weights provided by the dynamic weight adjustment module, forming preliminary retrieval results; wherein, the technical dimension weights include initial technical dimension weights and optimized technical dimension weights; The preliminary search results are standardized through a processing layer. The display layer selects the standardized search results according to the display mode and renders them in a corresponding multimodal form. It also receives real-time weight adjustment instructions from users as new initial technical dimension weight adjustment instructions through interactive controls, and obtains user interaction feedback. The dynamic weight adjustment module generates initial technical dimension weights based on the initial technical dimension weight adjustment instruction or the preset retrieval strategy corresponding to the display mode selection, and outputs them to the semantic retrieval layer; it also receives the interactive feedback, dynamically optimizes the initial technical dimension weights based on the interactive feedback, and outputs the optimized technical dimension weights to the semantic retrieval layer in real time.

9. An electronic device, characterized in that, include: The memory, the processor, and a computer program stored in the memory and executable on the processor, wherein the computer program, when executed by the processor, implements the steps of the patent retrieval method based on dynamic weight adjustment and multimodal semantics as described in claim 8.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores an information transmission implementation program, which, when executed by a processor, implements the steps of the patent retrieval method based on dynamic weight adjustment and multimodal semantics as described in claim 8.

Images