A patent document indexing method, system, and electronic device

Abstract

The application provides a patent document indexing method, system and electronic equipment. The indexing method comprises the following steps: after obtaining structured text data of an original patent document, splitting an indexing task into multiple subtasks based on the structured text data; calling multiple processing agents to respectively execute the multiple subtasks in a parallel manner, and generating an initial indexing result set containing results of the subtasks; performing consistency evaluation on the initial indexing result set based on a semantic constraint relationship graph of the multiple subtask results; when the score is lower than a preset threshold, analyzing the contribution of each subtask result in the initial indexing result set to the overall semantic inconsistency, determining a target processing agent based on the contribution, selecting an adjustment strategy according to the type of the target processing agent, and performing adjustment; re-performing consistency evaluation based on a new indexing result set; and when the score is not lower than the preset threshold, outputting the current indexing result set, thereby realizing high-quality, high-efficiency and high-stability automatic patent indexing.

Description

Technical Field

[0001] This application relates to the field of patent information processing technology, and in particular to a patent document indexing method, system, and electronic device. Background Technology

[0002] With the iterative development of big data and artificial intelligence technologies, efficient, accurate, and intelligent indexing of massive patent documents has become a core foundation for patent information retrieval, technology mining, and trend analysis. Accurate and unified intelligent patent indexing results can effectively lower the threshold for obtaining patent information, significantly improve the retrieval efficiency and information utilization of massive patent documents, and are a key prerequisite for achieving structured and intelligent processing of patent data.

[0003] However, current patent indexing technologies generally employ a single-task, independent processing architecture, leading to significant semantic inconsistencies between the outputs of different indexing tasks. For example, patent CN102722495A relies solely on preset technology classifications and keyword matching rules to independently complete patent indexing, resulting in fragmented indexing dimensions and a high likelihood of semantic inconsistencies in the indexing results. Furthermore, the optimization of low-quality indexing results in related technologies primarily relies on manual correction, resulting in low optimization efficiency and poor indexing stability. For instance, patent CN101692240A manually revises indexing terms to generate the final indexing result.

[0004] No effective solutions have yet been proposed for the aforementioned technical problems. Summary of the Invention

[0005] This application provides a patent document indexing method, system, and electronic device to solve one or more of the above-mentioned technical problems.

[0006] In a first aspect, embodiments of this application provide a patent document indexing method, comprising: S1: after obtaining structured text data of the original patent document, dividing the indexing task into multiple sub-tasks based on the structured text data, wherein the multiple sub-tasks include at least: International Patent Classification (IPC) prediction, technical keyword extraction, and technical abstract generation; S2: calling multiple processing agents to execute the multiple sub-tasks in parallel, generating an initial indexing result set containing the results of each sub-task, wherein the initial indexing result set includes at least: IPC prediction results, a set of technical keywords, and technical abstract text; S3: based on the semantic constraint relationship graph relating the results of the multiple sub-tasks, performing a cross-task semantic consistency evaluation on the initial indexing result set to obtain a consistency score. In the process, the semantic constraint relationship graph defines at least the semantic matching relationship between the IPC prediction results and the set of technical keywords, and the semantic content coverage relationship between the set of technical keywords and the technical summary text; S4: When the consistency score is lower than a preset threshold, analyze the contribution of each subtask result in the initial indexing result set to the overall semantic inconsistency, and determine the target processing agent causing the inconsistency based on the contribution; S5: Select an adjustment strategy and perform an adjustment operation according to the type of subtask executed by the target processing agent; after the adjustment operation is triggered, return to S3 and re-evaluate the consistency based on the new indexing result set; S6: When the consistency score is not lower than the preset threshold, output the current indexing result set as the final indexing result.

[0007] Secondly, embodiments of this application provide a patent document indexing system, including: a task scheduling module, a multi-agent parallel processing module, a semantic consistency evaluation module, a decision module, and an output module; the output end of the task scheduling module is connected to the input end of the multi-agent parallel processing module, and is configured to, after obtaining the structured text data of the original patent document, divide the indexing task into multiple sub-tasks based on the structured text data, wherein the multiple sub-tasks include at least: International Patent Classification (IPC) prediction, technical keyword extraction, and technical abstract generation; the output end of the multi-agent parallel processing module is connected to the input end of the semantic consistency evaluation module, and is configured to execute the multiple sub-tasks in parallel to generate an initial indexing result set containing the results of each sub-task, wherein the initial indexing result set includes at least: IPC prediction results, a set of technical keywords, and technical abstract text; the output end of the semantic consistency evaluation module is connected to the input end of the decision module, and is configured to, based on the associated... A semantic constraint relationship graph of multiple sub-task results is used to perform cross-task semantic consistency evaluation on the initial indexing result set to obtain a consistency score. The semantic constraint relationship graph at least defines the semantic matching relationship between the IPC prediction results and the set of technical keywords, and the semantic content coverage relationship between the set of technical keywords and the technical abstract text. The output of the decision module is connected to the input of the semantic consistency evaluation module and the input of the output module. It is configured to analyze the contribution of each sub-task result in the initial indexing result set to the overall semantic inconsistency when the consistency score is lower than a preset threshold, and determine the target processing agent causing the inconsistency based on the contribution. An adjustment strategy is selected and executed according to the type of sub-task performed by the target processing agent. After the adjustment operation is triggered, a new consistency evaluation is performed based on the new indexing result set. When the consistency score is not lower than the preset threshold, the current indexing result set is output as the final indexing result.

[0008] Thirdly, embodiments of this application provide an electronic device, including a memory, a processor, and a computer program stored in the memory, wherein the processor, when executing the computer program, implements the method described in any of the above-mentioned embodiments.

[0009] In the technical solutions provided in this application, the evaluation object is no longer the result of a simple single task, but rather the synergy between IPC prediction results, the set of technical keywords, and the technical summary text based on semantic matching and semantic content coverage relationships. This ensures a high degree of semantic consistency between the outputs of different indexing tasks, thereby solving the serious semantic inconsistency problem between the outputs of different indexing tasks in related technologies. Furthermore, when the consistency score is low, contribution analysis accurately identifies and adjusts the target processing agent causing the inconsistency. After a limited number of closed-loop iterations, the output consistency score is improved to above a preset threshold, greatly improving optimization efficiency and indexing stability. This solves the technical problems of low optimization efficiency and poor indexing stability in related technologies.

[0010] The above description is only an overview of the technical solution of this application. In order to better understand the technical means of this application, it can be implemented according to the contents of the specification. In order to make the above and other objects, features and advantages of this application more obvious and understandable, specific embodiments of this application are given below. Attached Figure Description

[0011] In the accompanying drawings, unless otherwise specified, the same reference numerals throughout the various drawings denote the same or similar parts or elements. These drawings are not necessarily drawn to scale. It should be understood that these drawings depict only some embodiments according to this application and should not be construed as limiting the scope of this application.

[0012] Figure 1 A flowchart illustrating a patent document indexing method provided in an embodiment of this application is shown;

[0013] Figure 2 This illustration shows a schematic diagram of the indexing results of a patent document provided in an embodiment of this application;

[0014] Figure 3 A flowchart illustrating another patent document indexing method provided in an embodiment of this application is shown;

[0015] Figure 4 This illustration shows a schematic diagram of a cross-task consistency constraint relationship provided in an embodiment of this application;

[0016] Figure 5 This paper shows a structural block diagram of a patent document indexing system provided in an embodiment of this application;

[0017] Figure 6 This paper shows a structural block diagram of another patent document indexing system provided in an embodiment of this application;

[0018] Figure 7A block diagram of an electronic device used to implement embodiments of this application is shown. Detailed Implementation

[0019] In the following description, only certain exemplary embodiments are briefly described. As those skilled in the art will recognize, the described embodiments can be modified in various ways without departing from the concept or scope of this application. Therefore, the drawings and description are considered to be exemplary in nature and not restrictive.

[0020] To facilitate understanding of the technical solutions of the embodiments of this application, the relevant technologies of the embodiments of this application are described below. The following relevant technologies are optional solutions and can be combined with the technical solutions of the embodiments of this application in any way, and all of them fall within the protection scope of the embodiments of this application.

[0021] Currently, several methods are commonly used to index patent documents. One approach involves independent processing by a single-task model, where each task, such as International Patent Classification (IPC) classification, keyword extraction, and abstract generation, is performed by a separate model with no correlation between them. Another approach uses a multi-task model without cross-evaluation; while multiple tasks may utilize the same set of models, there is a lack of assessment of the correlation between the results of each task. A third approach uses simple post-processing rules, such as filtering keywords based on classification results. However, existing methods have significant drawbacks. For example, the independent execution of sub-tasks, lacking interrelationships and constraints, leads to semantic inconsistencies between the final indexing results (IPC prediction results, sets of technical keywords, and technical abstract text), and the inability to automatically correct low-quality results. This significantly impacts the accuracy and reliability of the indexing results, thereby reducing the effectiveness and practicality of the entire indexing system and making it difficult to meet users' needs for high-quality patent information retrieval and utilization in real-world applications.

[0022] In view of this, embodiments of this application provide a patent document indexing method to solve or partially solve the technical problems existing in the prior art. This patent document indexing method is applicable to application scenarios involving the intelligent, refined, and standardized automatic indexing of massive amounts of patent documents, including but not limited to: structured extraction of core patent technology information, precise indexing of patent technical issues, standardized governance of patent databases, intelligent analysis of industry patent trends, precise empowerment of patent retrieval, discovery of R&D technical pain points, and patent compliance review and intelligent assessment.

[0023] like Figure 1 The diagram shown is a flowchart of a patent document indexing method according to an embodiment of this application, including:

[0024] S1: After obtaining the structured text data of the original patent documents, the indexing task is divided into multiple sub-tasks based on the structured text data. These sub-tasks include at least: International Patent Classification (IPC) prediction, technical keyword extraction, and technical abstract generation.

[0025] Optionally, in this embodiment, before executing S1 above, patent documents to be indexed can be obtained from a preset patent data source (including national or international patent office databases, internal enterprise document libraries, or third-party data platforms), and their file formats (such as PDF, XML, DOC, TXT, or scanned images) can be identified. For different formats, the corresponding parsing module is called to extract predetermined fields of text such as title, abstract, claims, and specification text. At the same time, text content is generated in the scanned or image format using Optical Character Recognition (OCR). The extracted text is cleaned and formatted, including removing invisible characters, redundant spaces or line breaks, and unifying symbol forms and encoding standards. The processed text content is then constructed into structured data objects (such as JSON or database records). This structured text data serves as input for automatic indexing, supporting subtasks such as IPC prediction, keyword extraction, and abstract generation.

[0026] S2: Invoke multiple processing agents to execute the multiple subtasks in parallel and generate an initial indexing result set containing the results of each subtask. The initial indexing result set includes at least: IPC prediction results, a set of technical keywords, and technical summary text.

[0027] It is understood that in the embodiments of this application, the aforementioned multiple processing agents each employ different task processing models to adapt to their respective sub-task requirements. For example, the agent for IPC prediction may employ a deep learning-based text classification model, the agent for technical keyword extraction may employ a keyword extraction model based on statistical algorithms and word vectors, and the agent for technical summary generation may employ a semantic embedding generation model based on a pre-trained language model. After receiving unified structured text data, each processing agent independently performs task processing according to its configured model and outputs the corresponding results.

[0028] S3: Based on the semantic constraint relationship graph that associates the results of the multiple sub-tasks, perform cross-task semantic consistency evaluation on the initial indexing result set to obtain a consistency score. The semantic constraint relationship graph at least defines the semantic matching relationship between the IPC prediction result and the set of technical keywords, as well as the semantic content coverage relationship between the set of technical keywords and the technical abstract text.

[0029] Understandably, in S3 above, the semantic coherence among IPCs, technical keywords, and technical abstracts is quantified using a semantic constraint relationship graph. During actual evaluation, these two relationships are used as constraints to perform a coherent analysis on the outputs of each subtask in the initial indexing result set (including IPC prediction results, technical keyword sets, and technical abstract text), and a comprehensive consistency score is calculated. This consistency score reflects the degree of semantic coherence among the three components under predefined constraints; a higher score indicates better overall consistency of the indexing results across multiple tasks. This semantic consistency evaluation effectively identifies information biases or insufficient coverage between different subtasks, allowing for targeted result correction and optimization in subsequent processing stages, thereby improving the accuracy and reliability of automatic indexing.

[0030] S4: When the consistency score is lower than the preset threshold, analyze the contribution of each subtask result in the initial indexing result set to the overall semantic inconsistency, and determine the target processing agent that causes the inconsistency based on the contribution.

[0031] In other words, when the consistency score is low, this application does not blindly make all processing agents run again, nor does it randomly pick one to adjust. Instead, it accurately identifies the target processing agent that causes inconsistency through contribution analysis. This attribution mechanism upgrades the adjustment behavior from "extensive" to "precise and targeted", which greatly improves the optimization efficiency and intelligence.

[0032] S5: Select an adjustment strategy and perform an adjustment operation based on the type of subtask performed by the target processing agent; after the adjustment operation is triggered, return to S3 and re-evaluate the consistency based on the new indexing result set.

[0033] S6: When the consistency score is not lower than the preset threshold, output the current indexing result set as the final indexing result.

[0034] In other words, this application proposes a negative feedback closed-loop control mechanism. As long as the trigger condition (consistency score is lower than a preset threshold) is met, this closed loop will raise the output consistency score to above the preset threshold after a finite number of iterations.

[0035] For example, assuming the initial indexing result set includes: the IPC preset result is "G06N3 / 08 (Neural Network)", the technical keyword set is "Support Vector Machine", and the technical abstract text does not mention neural networks at all, this application embodiment will give an extremely low consistency score. Then, through contribution analysis, it is found that the result of technical keyword extraction is furthest from the IPC prediction result and the semantic vector of the technical abstract, resulting in the highest contribution to inconsistency. Therefore, the processing agent corresponding to keyword extraction is precisely adjusted, and the result after adjustment becomes "Neural Network, Deep Learning, Backpropagation". Then, it is re-evaluated, and the semantic consistency between the new result and the IPC prediction result and the technical abstract text is significantly improved. After 1-2 iterations, the consistency score meets the standard.

[0036] It should be noted that, in this embodiment of the application, an upper limit for the number of iterations can be set (e.g., 5 times). If the consistency score still fails to meet the standard after reaching the upper limit, the current best result is output and marked as requiring manual review.

[0037] Through the aforementioned S1~S6, the evaluation object is no longer simply the result of a single task, but rather the synergy between IPC prediction results, the set of technical keywords, and the technical abstract text based on semantic matching and semantic content coverage relationships. This ensures a high degree of semantic consistency among the output results of different indexing tasks. Compared to the technical solution in patent CN1975723A, which independently completes indexing operations based on preset technical classifications and keyword matching rules, resulting in fragmented indexing dimensions and a lack of unified semantic constraints, this application's embodiment solves the serious semantic inconsistency problem among the output results of different indexing tasks in related technologies. Furthermore, when the consistency score is low, contribution analysis accurately identifies and adjusts the target processing agent causing the inconsistency. After a limited number of closed-loop iterations, the output consistency score is improved to above a preset threshold, significantly improving optimization efficiency and indexing stability. Compared to the single-forward extraction architecture of patent CN115640375A, which lacks an iterative feedback closed loop, this application's embodiment solves the technical problems of low optimization efficiency and poor indexing stability for low-quality indexing results in related technologies.

[0038] In one possible implementation, S3 above includes:

[0039] S31: Map the IPC prediction result to a first semantic vector using a shared semantic coding model, map the set of technical keywords to a second semantic vector, and calculate the similarity between the first and second semantic vectors as the score of the semantic matching relationship; S32: Identify text segments in the technical abstract text that have semantic equivalence to the set of technical keywords, and calculate the score of the semantic content coverage relationship based on the proportion of the text segments to the total number of the set of technical keywords; S33: Obtain the consistency score based on the score of the semantic matching relationship and the score of the semantic content coverage relationship.

[0040] In this embodiment, a shared semantic coding model can be used to achieve unified semantic space mapping. Specifically, the IPC prediction result c is converted into a standard technical field description (e.g., G06N3 → "machine learning method based on neural network") through a lookup table and encoded as a vector Vc; the technical keyword set K is concatenated or pooled and encoded as a vector Vk; the technical summary text s is encoded as a vector Vs. Optionally, the shared semantic coding model uses a language model that is further pre-trained on the patent corpus (such as a Bidirectional Encoder Representations from Transformers (BERT) model or an equivalent model), and the mapping from IPC classification number to descriptive text uses the official IPC title library.

[0041] For example, the semantic matching relationship between the IPC prediction result c and the set of technical keywords K can be calculated using cosine similarity, as shown in Formula 1 below:

[0042] (Formula 1)

[0043] For each technical keyword Use Formula 2 to determine if it is covered. When match(K) i If )>θ, where θ∈[0.7,0.85], then the keyword is considered covered.

[0044] (Formula 2)

[0045] Furthermore, the semantic content coverage relationship between the set of technical keywords and the technical abstract text is determined by the following formula 3:

[0046] (Formula 3)

[0047] Where M represents the number of covered technical keywords, and N represents the total number of technical keywords.

[0048] The overall consistency score is defined by the following formula 4:

[0049] (Formula 4)

[0050] Where α and β are weighting parameters.

[0051] Through the above steps S31~S33, the consistency score is split into two quantitative indicators, Sim(c,k) and Coverage(k,s), which can accurately measure the semantic matching relationship and content coverage between the results of different sub-tasks, thus improving the reliability of the consistency score.

[0052] In one possible implementation, step S4 above may include:

[0053] S41: Apply a preset perturbation to the results of each subtask in the initial indexing result set; S42: Calculate the change in consistency score before and after the perturbation; S43: Use the change as the contribution value, and rank the agents performing each subtask according to the contribution value; S44: Select the agent with the highest contribution value as the target processing agent.

[0054] Optionally, the perturbation methods mentioned above include, but are not limited to: randomly replacing the output of a subtask, adding small numerical perturbations (e.g., adding or subtracting random noise to each component of the vector) to the semantic vector generated by the task, and calling a backup model to generate comparison results for similar tasks, thereby disrupting the original semantic performance of the subtask. Subsequently, in S42, the same cross-task semantic consistency evaluation method as in S3 is used to calculate the overall consistency score before and after the perturbation, and the difference is taken as the change amount ΔScore. This change amount reflects the degree of influence of the subtask's output quality on the overall semantic consistency. In S43, the above change amount is used as a contribution value to rank the executing agents of each subtask. The higher the contribution value, the greater the importance of the processing agent to the overall consistency, and its output anomalies will significantly reduce the consistency score.

[0055] For example, in an automatic indexing task, suppose three sub-task results are generated for a patent document: IPC prediction result C (e.g., predicted as G06N3, corresponding to "machine learning method based on neural network"), a set of technical keywords K (e.g., containing "convolutional neural network", "backpropagation", "image classification"), and technical abstract text S (e.g., describing the scheme as an image recognition method based on convolutional neural network). After cross-task semantic consistency evaluation in S3, the overall consistency score is Score = 0.68, while the preset threshold is set to 0.75, thus entering the contribution analysis mechanism. First, in S41, preset perturbations are applied sequentially to the three sub-tasks. For example, for the IPC classification result, it is randomly replaced with G06F17 ("information retrieval"); for the set of technical keywords, random noise is added to the corresponding semantic vectors, causing the representation of some technical keywords to deviate from the original semantic space; for the technical abstract text, a backup abstract generation model is called to output a different description. Subsequently, in S42, the same consistency evaluation method as in S3 was used to calculate the changes in Score before and after the perturbation: after perturbing the IPC classification, the Score dropped to 0.52, with a change of ΔScore = 0.16; after perturbing the keyword set, the Score dropped to 0.60, with a change of ΔScore = 0.08; and after perturbing the summary text, the Score dropped to 0.57, with a change of ΔScore = 0.11. In S43, these changes were used as contribution values ​​and ranked: the IPC classification agent had the highest contribution value (0.16), followed by the summary agent (0.11), and the keyword agent had the lowest (0.08). Therefore, the IPC classification agent with the highest contribution value was identified as the target agent causing the consistency score to fall below the preset threshold, and its results were prioritized for model adjustment and output correction in subsequent optimization processes.

[0056] The contribution analysis mechanism in steps S41 to S44 above can quickly locate the key tasks and processing agents that affect the consistency score, providing a basis for subsequent targeted optimization, task correction or model replacement, thereby significantly improving the accuracy and semantic coherence of the automatic indexing results.

[0057] In one possible implementation, the above S5 includes:

[0058] S51: Select an adjustment strategy from a predefined set of priority adjustment strategies and execute the adjustment operation corresponding to that adjustment strategy.

[0059] Optionally, in this embodiment, for the IPC prediction processing agent, the adjustment items that can be adopted include model switching and multi-model fusion; for the processing agent for extracting technical keywords, the adjustment items that can be adopted include model switching, keyword quantity threshold adjustment, keyword ranking weight adjustment, and multi-model fusion; for the processing agent for extracting technical summaries, the adjustment items that can be adopted include model switching, temperature parameter adjustment, beam size parameter adjustment, and multi-model fusion. During execution, the adjustment items are not randomly selected, but are executed sequentially within the set of adjustment items corresponding to the target agent based on a preset priority order. For example, the first priority is model switching, used to replace the currently executing model to improve output quality; the second priority is parameter adjustment, including modification of the keyword quantity threshold, adjustment of ranking weight, and optimization of the Temperature and Beam Size parameters for summary generation; the third priority is multi-model fusion, which integrates the output results of multiple models to generate more consistent and stable indexing results.

[0060] The priority strategy in S51 above ensures that after locating the target agent that affects the consistency score, its output is corrected in the most efficient order, thereby improving the semantic coherence and accuracy of the overall automatic indexing results.

[0061] When the above adjustment strategy is model switching, S51 includes:

[0062] S511: Select an alternative model for the target processing agent from the model pool based on the model selection function, wherein the model selection function calculates a selection score based on the historical performance data of each candidate model, including accuracy, latency, and computational cost when processing similar or related patent documents.

[0063] Understandably, before completing the model pool deployment, offline benchmark tests are first performed on subsets of each technology field to obtain three-dimensional metrics data: accuracy, latency, and hardware overhead. These three-dimensional metrics data are then stored in metadata.

[0064] The aforementioned similar or related patent documents can be identified by matching major or minor IPC categories, or by determining that the vector distance between their texts in a unified semantic space is less than a preset threshold.

[0065] Optionally, in this embodiment, the candidate model with the higher score is selected as the replacement model to ensure that the output of the target processing agent can achieve an optimal balance between accuracy, efficiency and resource consumption after model switching.

[0066] For example, the selection score can be calculated using Formula 5:

[0067] (Formula 5)

[0068] Among them, W1, W2, and W3 are weight parameters used to balance the overall performance of the model in terms of accuracy, execution efficiency, and resource consumption.

[0069] In one possible implementation, the above S5 may also include:

[0070] S52: Obtain historical adjustment strategies and their corresponding adjustment effects; S53: Sort the historical adjustment strategies according to their adjustment effects, and select the historical adjustment strategy with the highest adjustment effect as the current adjustment strategy.

[0071] Optionally, the effects of the above adjustments include, but are not limited to, indicators such as the increase in consistency score, changes in task execution time, and changes in resource consumption, which are used to comprehensively evaluate the performance of the adjustment strategy in historical tasks.

[0072] For example, in an automatic patent indexing task, after performing a cross-task semantic consistency assessment, the overall consistency score is Score = 0.68, which is lower than the preset threshold of 0.75. The S4 contribution analysis identifies the processing agent for technical keyword extraction as the main source of the problem. Upon entering the S5 adjustment strategy stage, the historical adjustment strategy records and corresponding adjustment effects of the processing agent for technical keyword extraction are obtained, including: Strategy A: Switching the model to a more accurate keyword extraction model, with a historical average increase in consistency score of 0.12. Strategy B: Adjusting the keyword quantity threshold (from 15 to 20), with a historical average increase in consistency score of 0.08. Strategy C: Multi-model fusion (merging the results of two keyword extraction models and then sorting them after deduplication), with a historical average increase in consistency score of 0.10. The above historical adjustment strategies are sorted by the amount of consistency score improvement, and the results are: Strategy A (0.12) > Strategy C (0.10) > Strategy B (0.08). Since Strategy A has the largest improvement in historical tasks, Strategy A is directly selected as the adjustment strategy for the current task. This strategy has successfully improved consistency scores in many past tasks. Therefore, when encountering similar semantic consistency problems, this strategy can be quickly invoked without repeated verification, thereby speeding up the closed-loop processing efficiency.

[0073] In one possible implementation, the above method also includes:

[0074] S7: Map each subtask result in the initial indexing result set to a corresponding node; establish an edge between the nodes to represent the semantic relationship between tasks, and set a weight for consistency scoring for the edge; combine the node and the edge to construct the semantic constraint relationship graph.

[0075] For example, in the initial indexing results of the patent document, the IPC prediction result is G06N3 ("Machine Learning Method Based on Neural Networks"), the set of technical keywords includes "convolutional neural network", "backpropagation", and "image recognition", and the technical abstract text is "An Image Classification Method Based on Convolutional Neural Networks". In the semantic constraint relationship graph, node A represents the IPC classification result, node B represents the set of technical keywords, and node C represents the technical abstract text. An edge (weight 0.85) establishing a semantic matching relationship between node A and node B is used to reflect the consistency between the classification result and the keywords in the technical field; an edge (weight 0.78) establishing a content coverage relationship between node B and node C is used to reflect the degree to which the abstract text covers the technical points expressed by the keywords.

[0076] The semantic constraint relationship graph constructed through the above optional implementation methods enhances the expressive power of constraints between tasks, enabling precise modeling and quantification of semantic associations across different indexing dimensions. Furthermore, the structure of this semantic constraint relationship graph possesses good scalability, allowing for the easy addition of new indexing dimension nodes (such as applicant information, inventor information, and their association analysis) in the future. By creating or adjusting edges and their weights, the output results of the new dimensions and existing tasks can be incorporated into a consistency evaluation framework, thereby continuously improving the automatic indexing system's processing capabilities and analytical depth for multidimensional data.

[0077] Optionally, in the embodiments of this application, the above weights can be updated through the above consistency score to solve the problem that fixed weights cannot adapt to the indexing characteristics of different technical fields, such as the importance of keyword-abstract coverage in the chemical field and the importance of classification-keyword matching in the electronic field.

[0078] In one possible implementation, the above method further includes the following before executing S2:

[0079] S21: Perform complexity analysis on the structured text data to obtain the complexity analysis results; S22: Select a suitable processing agent based on the complexity analysis results.

[0080] Optionally, the complexity analysis described above includes, but is not limited to, a comprehensive assessment of the text length, sentence structure diversity, terminology density, semantic nesting level, and the number and type of charts or formulas to quantify the difficulty level of the structured text data during processing.

[0081] For example, in an automated patent indexing task, the structured text data to be processed is first analyzed for complexity. For document A, the analysis result is low complexity: the full text length is relatively short (less than 2000 words), the sentence structure is simple, the density of technical terms is low (less than 5%), and it does not contain complex diagrams or mathematical formulas. This type of task is judged as low difficulty, and a lightweight processing agent is selected for indexing to reduce computational resource consumption and speed up processing, with an execution time of only 0.5 seconds. For document B, the complexity analysis result is high complexity: the full text length exceeds 10000 words, the sentence nesting level is high, the density of technical terms is high (more than 20%), and it contains multiple technical diagrams and complex formulas. This type of task is judged as high difficulty, and a high-precision processing agent is selected, such as a complete BERT large model or a multi-model fusion scheme, to ensure the accuracy of indexing and the depth of semantic understanding, with an execution time of approximately 2.5 seconds, but the indexing accuracy is significantly improved.

[0082] Through the above S21~S22, the complexity is predicted and a suitable agent is matched before processing. Low-complexity tasks will not waste high-performance computing resources, while high-complexity tasks will receive sufficient processing power. The overall resource scheduling efficiency is improved, effectively avoiding the problems of resource waste in low-complexity tasks and insufficient processing of high-complexity tasks caused by "one-size-fits-all" resource allocation.

[0083] In one possible implementation, the above method also includes:

[0084] S11: Perform self-supervised consistency enhancement pre-training on the multiple processing agents, wherein the pre-training process includes: constructing positive and negative sample pairs using manually indexed historical patent documents, and optimizing the semantic consistency of the outputs of the multiple processing agents through joint training.

[0085] Understandably, the aforementioned positive sample pairs can be composed of samples with highly semantically matched sub-task outputs (such as IPC prediction results, sets of technical keywords, and technical abstract texts) within the same patent document, while negative sample pairs are composed of sub-task outputs from different patents or different technical fields, used to reflect low-consistency scenarios. During the pre-training phase, multiple processing agents are jointly trained. By optimizing the loss function, the outputs of different agents maintain a higher consistency score in a unified semantic space. This, to some extent, shifts the consistency requirement from "runtime correction" to "internalization during training," pre-establishing cross-task semantic collaboration mechanisms within the model to fundamentally improve the collaboration among agents, effectively reducing consistency adjustment overhead during operation, and improving the overall accuracy and stability of automatic indexing.

[0086] For example, in the pre-training phase of an automated patent indexing system, self-supervised training samples are constructed using historical patent documents indexed manually. For instance, a patent document is selected whose manually indexed results include: IPC classification result: G06N3 (Machine Learning Method Based on Neural Networks); technical keyword set: Convolutional Neural Network, Backpropagation, Image Recognition; and technical abstract text: Describes an image classification method based on convolutional neural networks, and details the training process and application scenarios. These three types of output content are highly consistent semantically, and therefore are added to the training set as positive sample pairs. To construct negative sample pairs, the IPC classification result of this patent document is replaced with a classification number that does not match its technical field, such as G06F17 (Information Retrieval), while retaining the technical keyword set and technical abstract text. This combination scores significantly lower in semantic consistency evaluation and is clearly identified as a negative sample pair. During training, both positive and negative sample pairs are input simultaneously. A shared semantic encoding model maps these outputs to a unified semantic space, and a consistency loss function is introduced to reduce the vector distance between positive samples and increase the vector distance between negative samples. After multiple rounds of iterative training, each processing agent (IPC prediction agent, technical keyword extraction agent, and technical summary generation agent) pre-solidifies cross-task semantic matching rules within the model, thereby shifting the consistency requirement from "runtime correction" to "internalization during training".

[0087] The embodiments of this application will be illustrated below with specific examples.

[0088] Example 1

[0089] This example uses the title "An Image Classification Method and Related Device Based on a Deep Fuzzy Convolutional Neural Network Model" as an example. This solution relates to the field of image processing, where the deep fuzzy convolutional neural network model includes a feature extraction network, a feature fusion network, and a classification network. For the initial automatic indexing result of a certain patent document, the IPC prediction result is: G06N3 (Machine Learning Method Based on Neural Networks), the set of technical keywords is: "image classification," "convolutional neural network," "feature extraction," and "feature fusion," and the technical abstract text is: "This invention discloses an image classification method based on a deep fuzzy convolutional neural network, which extracts image features through a feature extraction network and fuses them through a feature fusion network to ultimately achieve image classification." Through cross-task semantic consistency evaluation, the semantic matching degree Sim=0.84 between the IPC prediction result and the set of technical keywords, the semantic coverage between the set of technical keywords and the technical abstract text is Coverage=0.70, and the overall consistency score is Score=0.77. Based on statistical verification of approximately 1000 manually annotated data points, when the threshold is set to 0.8, the accuracy of the indexing task is approximately 92% when the consistency score reaches or exceeds this value. Because the initial score was below the threshold, the process entered the adjustment phase. Two optimization measures were implemented on the technical summary generation agent: reducing the text generation temperature to decrease the randomness of the output content; and adding a hard condition of "must cover all keywords" to the technical summary generation constraints. After adjustment, the consistency index was recalculated, and the keyword coverage improved to 0.93, while the overall consistency score improved to 0.89, exceeding the preset threshold of 0.8. Simultaneously, the optimized summary covered all core technical keywords, significantly improving the accuracy and synergy of the automatic indexing results, such as... Figure 2 As shown.

[0090] Example 2

[0091] This example provides a method for indexing patent documents, such as... Figure 3 As shown, it includes:

[0092] S301, Patent Text Input;

[0093] S302, Text Parsing and Structuring;

[0094] S303, Task splitting and scheduling, and initial model selection;

[0095] S304, Parallel execution of subtasks (IPC prediction / technical keyword extraction / technical summary generation);

[0096] S305, Construct semantic constraints;

[0097] S306, Cross-task consistency assessment. The specific assessment method is as follows: Figure 4 As shown, the classification results, the set of technical keywords, and the abstract text are mapped to a unified semantic space, resulting in corresponding classification vector A, keyword vector B, and abstract vector C. Classification vector A and keyword vector B are bounded by a semantic similarity Sim(A,B), which measures the semantic matching degree between the classification results and the keyword set. Keyword vector B and abstract vector C are bounded by a content coverage Coverage(B,C), which measures the extent to which the abstract text covers the technical points expressed by the keyword set. Furthermore, an overall consistency score Score is calculated based on the semantic similarity and content coverage, and the calculation method is as follows: Here, α and β are weighting parameters used to balance the consistency between classification and keywords, as well as the contribution ratio of keyword and summary coverage to the overall score. In this embodiment, the classification vector A, keyword vector B, and summary vector C are all obtained through mapping using a shared semantic encoding model. The semantic encoding model can be a pre-trained language model or an equivalent implementation, such as a BERT-like model, to ensure the consistency and comparability of the representation of different task outputs in a unified semantic space.

[0098] S307, determine whether the consistency score is greater than or equal to the threshold. If the result is yes, proceed to S310. If the result is no, proceed to step S308.

[0099] S308, Problem-based task location based on contribution analysis;

[0100] S309, Select the adjustment strategy and execute the adjustment operation, then return to S304;

[0101] S310, output the final indexing result.

[0102] Experimental data shows that, on a test set consisting of 1,000 manually indexed patents, the indexing method provided in this application improved the average consistency score among IPC, keywords, and abstracts by 40% compared to single-task independent processing, and reduced the manual correction rate by 55%.

[0103] It should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, data stored, data displayed, etc.) involved in this application are all information and data authorized by the user or fully authorized by all parties. Furthermore, the collection, use and processing of the relevant data must comply with the relevant laws, regulations and standards of the relevant countries and regions, and corresponding operation entry points are provided for users to choose to authorize or refuse.

[0104] Corresponding to the application scenarios and methods provided in the embodiments of this application, the embodiments of this application also provide a patent document indexing system, such as... Figure 5 The diagram shown is a structural block diagram of a patent document indexing system according to an embodiment of this application. The indexing system may include: a task scheduling module 52, a multi-agent parallel processing module 54, a semantic consistency evaluation module 56, a decision module 58, and an output module 60.

[0105] The output of the task scheduling module 52 is connected to the input of the multi-agent parallel processing module 54. It is configured to, after obtaining the structured text data of the original patent document, split the indexing task into multiple sub-tasks based on the structured text data. The multiple sub-tasks include at least: International Patent Classification (IPC) prediction, technical keyword extraction, and technical abstract generation.

[0106] The output of the multi-agent parallel processing module 54 is connected to the input of the semantic consistency evaluation module 56 and is configured to execute the multiple subtasks in parallel to generate an initial indexing result set containing the results of each subtask. The initial indexing result set includes at least: IPC prediction results, a set of technical keywords, and technical summary text.

[0107] The output of the semantic consistency evaluation module 56 is connected to the input of the decision module 58. It is configured to perform cross-task semantic consistency evaluation on the initial indexing result set based on the semantic constraint relationship graph that associates the results of the multiple sub-tasks, and obtain a consistency score. The semantic constraint relationship graph at least defines the semantic matching relationship between the IPC prediction result and the set of technical keywords, as well as the semantic content coverage relationship between the set of technical keywords and the technical abstract text.

[0108] The output of the decision module 58 is connected to the input of the semantic consistency evaluation module 56 and the input of the output module 60. It is configured to analyze the contribution of each subtask result in the initial indexing result set to the overall semantic inconsistency when the consistency score is lower than a preset threshold, determine the target processing agent that caused the inconsistency based on the contribution, select an adjustment strategy and perform the adjustment operation according to the type of subtask performed by the target processing agent, re-evaluate the consistency based on the new indexing result set after the adjustment operation is triggered, and output the current indexing result set as the final indexing result when the consistency score is not lower than the preset threshold.

[0109] pass Figure 5The system described here evaluates not just the results of a single task, but the synergy between IPC prediction results, a set of technical keywords, and technical abstract text based on semantic matching and semantic content coverage. This results in a high degree of semantic consistency between the outputs of different indexing tasks. Compared to the technical solution in patent CN1975723A, which independently performs indexing operations based on preset technical classifications and keyword matching rules, leading to fragmented indexing dimensions and a lack of unified semantic constraints, this application's embodiment solves the serious semantic inconsistency problem between the outputs of different indexing tasks in related technologies. Furthermore, when the consistency score is low, contribution analysis accurately identifies and adjusts the target processing agent causing the inconsistency. After a limited number of closed-loop iterations, the consistency score is improved to above a preset threshold, significantly enhancing optimization efficiency and indexing stability. Compared to the single-forward extraction architecture in patent CN115640375A, which lacks an iterative feedback loop, this application's embodiment solves the technical problems of low optimization efficiency and poor indexing stability for low-quality indexing results in related technologies.

[0110] In one possible implementation, the semantic consistency evaluation module 56 is further configured to: map the IPC prediction result to a first semantic vector and the set of technical keywords to a second semantic vector using a shared semantic coding model, and calculate the similarity between the first and second semantic vectors as a score for the semantic matching relationship; identify text segments in the technical abstract text that have semantic equivalence to the set of technical keywords, and calculate a score for the semantic content coverage relationship based on the proportion of the text segments to the total number of the set of technical keywords; and obtain the consistency score based on the score for the semantic matching relationship and the score for the semantic content coverage relationship.

[0111] In one possible implementation, the decision module 58 is further configured to: apply a preset perturbation to the results of each subtask in the initial indexing result set; calculate the change in consistency score before and after the perturbation; use the change as a contribution value, and rank the agents performing each subtask according to the contribution value; and select the agent with the highest contribution value as the target processing agent.

[0112] In one possible implementation, the decision module 58 is further configured to: select an adjustment strategy from a predefined set of priority adjustment strategies and execute an adjustment operation corresponding to the adjustment strategy. When the adjustment strategy is model switching, executing the adjustment operation corresponding to the adjustment strategy includes: selecting an alternative model for the target processing agent from a model pool based on a model selection function. The model selection function calculates a selection score based on historical performance data of each candidate model, including accuracy, response time, and computational cost when processing similar or related patent documents. Alternatively, the decision module 58 is further configured to: obtain historical adjustment operations and their corresponding adjustment effects; sort the historical adjustment operations according to their adjustment effects; and select the historical adjustment operation with the highest adjustment effect as the current adjustment operation.

[0113] In one possible implementation, the indexing system further includes: a first processing module configured to map each subtask result in the initial indexing result set to a corresponding node; establish an edge between the nodes to represent the semantic association between tasks, and set a weight for consistency scoring for the edge; and combine the node and the edge to construct the semantic constraint relationship graph.

[0114] Optionally, the above weights can be updated using the aforementioned consistency score.

[0115] In one possible implementation, the indexing system further includes: a second processing module configured to perform complexity analysis on the structured text data to obtain the complexity analysis result; and to select an appropriate processing agent based on the complexity analysis result.

[0116] In one possible implementation, the indexing system further includes a third processing module configured to perform self-supervised consistency enhancement pre-training on the plurality of processing agents, wherein the pre-training process includes: constructing positive and negative sample pairs using historical patent documents indexed manually, and optimizing the semantic consistency of the outputs of the plurality of processing agents through joint training.

[0117] The embodiments of this application will be illustrated below with specific examples.

[0118] like Figure 6 As shown, this example proposes a patent document indexing system, which includes: a patent text input module 62, a text parsing and structuring module 64, a task splitting and scheduling module 66, a classification processing module 68, a keyword extraction module 70, an abstract generation module 72, a cross-task consistency evaluation module 74, a consistency arbitration and dynamic adjustment module 76, and an output module 78.

[0119] The text parsing and structuring module 64 is configured to: for different formats, call the corresponding parsing module to extract text from predetermined fields such as title, abstract, claims, and specification text; and simultaneously generate text content in scanned or image formats using Optical Character Recognition (OCR). The extracted text undergoes cleaning and format standardization processing, including removing invisible characters, redundant spaces or line breaks, and standardizing symbol formats and encoding standards. The processed text content is then constructed into structured data objects (such as JSON or database records). This structured text data serves as input for automatic indexing, supporting subtasks such as IPC prediction, keyword extraction, and abstract generation.

[0120] The task splitting and scheduling module 66 is configured to: after obtaining the structured text data of the original patent documents, split the indexing task into multiple sub-tasks based on the structured text data, wherein the multiple sub-tasks include at least: International Patent Classification (IPC) prediction, technical keyword extraction, and technical abstract generation. The IPC prediction task is assigned to the classification processing module 68, the technical keyword extraction task is assigned to the keyword extraction module 70, and the technical abstract generation task is assigned to the abstract generation module 72. The cross-task consistency evaluation module 74 is configured to: perform a cross-task semantic consistency evaluation on the initial indexing result set based on a pre-constructed semantic constraint relationship graph that associates the results of the multiple sub-tasks, and obtain a consistency score, wherein the semantic constraint relationship graph at least defines the semantic matching relationship between the IPC prediction result and the technical keyword set, and the semantic content coverage relationship between the technical keyword set and the technical abstract text. The consistency arbitration and dynamic adjustment module 76 is configured to: when the consistency score is lower than a preset threshold, analyze the contribution of each subtask result in the initial indexing result set to the overall semantic inconsistency, determine the target processing agent that caused the inconsistency based on the contribution, select an adjustment strategy and perform the adjustment operation according to the type of subtask performed by the target processing agent, and re-evaluate the consistency based on the new indexing result set after the adjustment operation is triggered.

[0121] The functions of each module in each system of the embodiments of this application can be found in the corresponding description in the above method, and they have corresponding beneficial effects, which will not be repeated here.

[0122] Figure 7 This is a block diagram of an electronic device used to implement embodiments of this application. For example... Figure 7 As shown, the electronic device includes a memory 701 and a processor 702. The memory 701 stores a computer program that can run on the processor 702. When the processor 702 executes the computer program, it implements the method described in the above embodiments. The number of memories 701 and processors 702 can be one or more.

[0123] The electronic device also includes:

[0124] The communication interface 703 is used to communicate with external devices and perform data exchange and transmission.

[0125] If the memory 701, processor 702, and communication interface 703 are implemented independently, they can be interconnected via a bus to communicate with each other. This bus can be an Industry Standard Architecture (ISA) bus, a Peripheral Component Interconnect (PCI) bus, or an Extended Industry Standard Architecture (EISA) bus, etc. This bus can be divided into address bus, data bus, control bus, etc. For ease of representation, Figure 7 The bus is represented by a single thick line, but this does not mean that there is only one bus or one type of bus.

[0126] Optionally, in a specific implementation, if the memory 701, processor 702, and communication interface 703 are integrated on a single chip, then the memory 701, processor 702, and communication interface 703 can communicate with each other through an internal interface.

[0127] This application provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the method provided in this application.

[0128] This application also provides a chip including a processor for calling and executing instructions stored in a memory, causing a communication device with the chip installed to perform the method provided in this application.

[0129] This application also provides a chip, including: an input interface, an output interface, a processor, and a memory. The input interface, output interface, processor, and memory are connected through an internal connection path. The processor is used to execute code in the memory. When the code is executed, the processor is used to execute the method provided in the application embodiment.

[0130] It should be understood that the aforementioned processor can be a Central Processing Unit (CPU), or other general-purpose processors, Digital Signal Processors (DSPs), Application Specific Integrated Circuits (ASICs), Field-Programmable Gate Arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. General-purpose processors can be microprocessors or any conventional processor. It is worth noting that the processor can be a processor supporting Advanced Reduced Instruction Set Machines (ARM) architecture.

[0131] Further, optionally, the aforementioned memory may include read-only memory and random access memory. The memory may be volatile memory or non-volatile memory, or may include both. Non-volatile memory may include read-only memory (ROM), programmable read-only memory (PROM), erasable programmable read-only memory (EPROM), electrically erasable programmable read-only memory (EEPROM), or flash memory. Volatile memory may include random access memory (RAM), which serves as an external cache. By way of example, but not limitation, many forms of RAM are available. Examples include Static Random Access Memory (SRAM), Dynamic Random Access Memory (DRAM), Synchronous DRAM (SDRAM), Double Data Rate Synchronous DRAM (DDR SDRAM), Enhanced Synchronous DRAM (ESDRAM), Synchronous Link DRAM (SLDRAM), and Direct Rambus RAM (DR RAM).

[0132] In the above embodiments, implementation can be achieved, in whole or in part, through software, hardware, firmware, or any combination thereof. When implemented in software, it can be implemented, in whole or in part, as a computer program product. A computer program product includes one or more computer instructions. When the computer program instructions are loaded and executed on a computer, all or part of the processes or functions according to this application are generated. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions can be stored in a computer-readable storage medium or transferred from one computer-readable storage medium to another.

[0133] In the description of this specification, the references to terms such as "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of this application. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples. Moreover, without contradiction, those skilled in the art can combine and integrate the different embodiments or examples described in this specification, as well as the features of those different embodiments or examples.

[0134] Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Thus, a feature defined as "first" or "second" may explicitly or implicitly include at least one of that feature. In the description of this application, "a plurality of" means two or more, unless otherwise explicitly specified.

[0135] Any process or method described in the flowchart or otherwise herein can be understood as representing a module, segment, or portion of code comprising one or more executable instructions for implementing a particular logical function or process. Furthermore, the scope of the preferred embodiments of this application includes additional implementations in which functions may be performed not in the order shown or discussed, including substantially simultaneously or in reverse order depending on the functionality involved.

[0136] The logic and / or steps described in the flowchart or otherwise herein, for example, can be considered as a sequenced list of executable instructions for implementing logical functions, and can be embodied in any computer-readable medium for use by, or in conjunction with, an instruction execution system, apparatus or device (such as a computer-based system, a processor-included system or other system that can fetch and execute instructions from, an instruction execution system, apparatus or device).

[0137] It should be understood that various parts of this application can be implemented using hardware, software, firmware, or a combination thereof. In the above embodiments, multiple steps or methods can be implemented using software or firmware stored in memory and executed by a suitable instruction execution system. All or part of the steps of the methods in the above embodiments can be implemented by a program instructing related hardware, the program being stored in a computer-readable storage medium, which, when executed, includes one or a combination of the steps of the method embodiments.

[0138] Furthermore, the functional units in the various embodiments of this application can be integrated into a processing module, or each unit can exist physically separately, or two or more units can be integrated into a module. The integrated module can be implemented in hardware or as a software functional module. If the integrated module is implemented as a software functional module and sold or used as an independent product, it can also be stored in a computer-readable storage medium. This storage medium can be a read-only memory, a disk, or an optical disk, etc.

[0139] The above description is merely an exemplary embodiment of this application, but the scope of protection of this application is not limited thereto. Any person skilled in the art can easily conceive of various variations or substitutions within the technical scope described in this application, and these should all be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A method for indexing patent documents, characterized in that, include: S1: After obtaining the structured text data of the original patent documents, the indexing task is divided into multiple sub-tasks based on the structured text data. The multiple sub-tasks include at least: International Patent Classification (IPC) prediction, technical keyword extraction, and technical abstract generation. S2: Invoke multiple processing agents to execute the multiple subtasks in parallel, and generate an initial indexing result set containing the results of each subtask, wherein the initial indexing result set includes at least: IPC prediction results, a set of technical keywords, and technical summary text; S3: Based on the semantic constraint relationship graph that associates the results of the multiple sub-tasks, perform cross-task semantic consistency evaluation on the initial indexing result set to obtain a consistency score. The semantic constraint relationship graph at least defines the semantic matching relationship between the IPC prediction results and the set of technical keywords, as well as the semantic content coverage relationship between the set of technical keywords and the technical summary text. S4: When the consistency score is lower than a preset threshold, analyze the contribution of each subtask result in the initial indexing result set to the overall semantic inconsistency, and determine the target processing agent that causes the inconsistency based on the contribution. S5: Select an adjustment strategy and perform an adjustment operation based on the type of subtask executed by the target processing agent; after the adjustment operation is triggered, return to S3 and re-evaluate the consistency based on the new indexing result set; S6: When the consistency score is not lower than the preset threshold, output the current indexing result set as the final indexing result.

2. The indexing method according to claim 1, characterized in that, S3 includes: S31: The IPC prediction result is mapped to a first semantic vector through a shared semantic coding model, the set of technical keywords is mapped to a second semantic vector, and the similarity between the first semantic vector and the second semantic vector is calculated as the score of the semantic matching relationship; S32: Identify text segments in the technical summary text that have semantic equivalence in the set of technical keywords, and calculate the score of the semantic content coverage relationship based on the proportion of the text segments to the total number of the set of technical keywords; S33: The consistency score is obtained based on the score of the semantic matching relationship and the score of the semantic content coverage relationship.

3. The indexing method according to claim 1, characterized in that, S4 includes: S41: Apply a preset perturbation to the results of each subtask in the initial indexing result set; S42: Calculate the change in consistency score before and after the disturbance; S43: Using the change as a contribution value, sort the agents performing each sub-task according to the contribution value; S44: Select the agent with the highest contribution value as the target processing agent.

4. The indexing method according to claim 1, characterized in that, S5 includes: S51: Select an adjustment strategy from a predefined set of priority adjustment strategies and execute the adjustment operation corresponding to the adjustment strategy.

5. The indexing method according to claim 4, characterized in that, When the adjustment strategy is model switching, S51 includes: S511: Select an alternative model for the target processing agent from the model pool based on a model selection function, wherein the model selection function calculates a selection score based on the historical performance data of each candidate model, and the historical performance data includes the accuracy, response time and computational cost when processing similar or related patent documents.

6. The indexing method according to claim 1, characterized in that, S5 includes: S52: Obtain historical adjustment strategies and the corresponding adjustment effects of the historical adjustment strategies; S53: Sort the historical adjustment strategies according to their adjustment effects, and select the historical adjustment strategy with the highest adjustment effect as the adjustment strategy.

7. The indexing method according to claim 1, characterized in that, Also includes: S7: Map each subtask result in the initial indexing result set to the corresponding node; S8: Establish edges between the nodes to represent semantic associations between tasks, and set weights for the edges for consistency scoring, wherein the weights are updated through the consistency scoring; S9: Combine the nodes and edges to construct the semantic constraint relationship graph.

8. The indexing method according to claim 1, characterized in that, Before S2, it also includes: S21: Perform complexity analysis on the structured text data to obtain the complexity analysis results; S22: Based on the complexity analysis results, select a suitable processing agent.

9. The indexing method according to any one of claims 1 to 8, characterized in that, Also includes: S11: Perform self-supervised consistency enhancement pre-training on the multiple processing agents, wherein the pre-training process includes: constructing positive and negative sample pairs using manually indexed historical patent documents, and optimizing the semantic consistency of the outputs of the multiple processing agents through joint training.

10. An indexing system for patent documents, characterized in that, include: The module includes a task scheduling module, a multi-agent parallel processing module, a semantic consistency evaluation module, a decision-making module, and an output module. The output of the task scheduling module is connected to the input of the multi-agent parallel processing module. It is configured to, after obtaining the structured text data of the original patent document, split the indexing task into multiple sub-tasks based on the structured text data. The multiple sub-tasks include at least: International Patent Classification (IPC) prediction, technical keyword extraction, and technical abstract generation. The output of the multi-agent parallel processing module is connected to the input of the semantic consistency evaluation module and is configured to execute the multiple sub-tasks in parallel to generate an initial indexing result set containing the results of each sub-task. The initial indexing result set includes at least: IPC prediction results, a set of technical keywords, and technical summary text. The output of the semantic consistency evaluation module is connected to the input of the decision module and is configured to perform cross-task semantic consistency evaluation on the initial indexing result set based on the semantic constraint relationship graph that associates the results of the multiple sub-tasks, and obtain a consistency score. The semantic constraint relationship graph at least defines the semantic matching relationship between the IPC prediction results and the set of technical keywords, as well as the semantic content coverage relationship between the set of technical keywords and the technical summary text. The output of the decision module is connected to the input of the semantic consistency evaluation module and the input of the output module. It is configured to, when the consistency score is lower than a preset threshold, analyze the contribution of each subtask result in the initial indexing result set to the overall semantic inconsistency, determine the target processing agent causing the inconsistency based on the contribution, select an adjustment strategy and execute an adjustment operation according to the type of subtask performed by the target processing agent, re-evaluate the consistency based on the new indexing result set after the adjustment operation is triggered, and output the current indexing result set as the final indexing result when the consistency score is not lower than the preset threshold.

11. The indexing system according to claim 10, characterized in that, The semantic consistency evaluation module is further configured to: map the IPC prediction result to a first semantic vector and the set of technical keywords to a second semantic vector using a shared semantic coding model; calculate the similarity between the first and second semantic vectors as a score for the semantic matching relationship; identify text segments in the technical abstract text that have semantic equivalence to the set of technical keywords; calculate a score for the semantic content coverage relationship based on the proportion of the text segments to the total number of the set of technical keywords; and obtain the consistency score based on the score for the semantic matching relationship and the score for the semantic content coverage relationship. The decision module is also configured to: apply a preset perturbation to the results of each subtask in the initial indexing result set; and calculate the change in consistency score before and after the perturbation. The change is used as a contribution value, and the agents executing each sub-task are ranked according to the contribution value; the agent with the highest contribution value is selected as the target processing agent. as well as, From a predefined set of priority adjustment strategies, an adjustment strategy is selected and the corresponding adjustment operation is executed. When the adjustment strategy is model switching, the execution of the corresponding adjustment operation includes: selecting an alternative model for the target processing agent from a model pool based on a model selection function, wherein the model selection function calculates a selection score based on historical performance data of each candidate model, and the historical performance data includes accuracy, response time, and computational cost when processing similar or related patent documents; or, obtaining historical adjustment operations and their corresponding adjustment effects; sorting the historical adjustment operations according to their adjustment effects, and selecting the historical adjustment operation with the highest adjustment effect ranking as the adjustment operation.

12. An electronic device, characterized in that, It includes a memory, a processor, and a computer program stored in the memory, wherein the processor, when executing the computer program, implements a patent document indexing method as described in any one of claims 1 to 9.

Images