Intellectual property multi-branch task processing method and device

Abstract

The application discloses a kind of intellectual property multi-branch task processing method and device, the method includes: user dialogue is carried out intention recognition, obtains the branch task of hit;If user intention hits retrieval element refining branch task, execute retrieval element refining branch task processing flow, including: user query key text set is matched with IPC classification number vector, by uncertainty word mark set, the matching result is handled, obtains recommended IPC classification number and retrieval element;If user intention hits expression writing branch task, execute expression writing branch task processing flow, including: analysis user preference portrait, determine retrieval strategy, according to retrieval element, query retrieval element knowledge graph, obtain associated retrieval element, and according to retrieval strategy, recommended IPC classification number, retrieval element and associated retrieval element, generate retrieval expression.The application carries out multi-branch task processing efficiency is high, response timely.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence and recommendation system technology, and in particular to a multi-branch task processing method and apparatus with intellectual property rights. Background Technology

[0002] In the field of intellectual property, users are increasingly demanding consultations and processing for various tasks, covering multiple areas and scenarios. Although some automated question-answering systems have emerged in the existing technology, most lack the ability to accurately identify user intent. Aside from answering some general knowledge questions, they are generally unable to automatically match corresponding processing tasks based on the question type. Furthermore, when user questions are vaguely worded or involve cross-functional business, it is even more difficult to accurately access relevant knowledge base resources or business interfaces, resulting in incomplete or even erroneous responses. In addition, for intellectual property-specific business scenarios, existing systems lack targeted branching mechanisms, failing to achieve intelligent and accurate responses.

[0003] Therefore, there is an urgent need for a system that relies on knowledge bases and standardized workflows to accurately identify user questions and match them to corresponding processing branches, thereby completing multi-branch intellectual property task processing. This would solve problems such as low efficiency and inaccurate response in traditional processing models and improve the intelligence level and service quality of intellectual property task processing. Summary of the Invention

[0004] In a first aspect, embodiments of the present invention provide a method for processing multi-branch tasks related to intellectual property, which uses precise intent recognition to match user questions to corresponding processing branches and complete multi-branch task processing for intellectual property-specific business scenarios, achieving high efficiency and timely response. The method includes:

[0005] Perform intent recognition on the input user dialogue to obtain the branch task of matching the user intent;

[0006] If the user intends to target the retrieval element extraction branch task, the retrieval element extraction branch task processing flow is executed. The retrieval element extraction branch task processing flow includes: generating a set of user query key texts and a set of uncertain word tags based on the user dialogue; matching the set of user query key texts with the IPC classification number vector to obtain the matching result; processing the matching result through the set of uncertain word tags to obtain the recommended IPC classification number; and obtaining the retrieval element based on the recommended IPC classification number.

[0007] If the user intends to hit the expression writing branch task, the expression writing branch task processing flow is executed. The expression writing branch task processing flow includes: executing the search element extraction branch task processing flow, obtaining recommended IPC classification numbers and search elements, analyzing user preference profiles, determining search strategies, querying the search element knowledge graph based on the search elements, obtaining related search elements, and generating search expressions based on the search strategies, recommended IPC classification numbers, search elements, and related search elements.

[0008] Secondly, embodiments of the present invention also provide an intellectual property multi-branch task processing device, used to match user questions to corresponding processing branches through accurate intent recognition, and to complete multi-branch task processing for intellectual property-specific business scenarios, with high efficiency and timely response. The device includes:

[0009] The intent recognition module is used to recognize the intent of the input user dialogue and obtain the branch tasks that match the user intent.

[0010] The retrieval element extraction branch task processing module is used to execute the retrieval element extraction branch task processing flow if the user's intent matches the retrieval element extraction branch task. The retrieval element extraction branch task processing flow includes: generating a set of user query key texts and a set of uncertain word tags based on the user dialogue; matching the set of user query key texts with the IPC classification number vector to obtain the matching result; processing the matching result through the set of uncertain word tags to obtain the recommended IPC classification number; and obtaining the retrieval element based on the recommended IPC classification number.

[0011] The expression writing branch task processing module is used to execute the expression writing branch task processing flow if the user's intent matches the expression writing branch task. The expression writing branch task processing flow includes: executing the search element extraction branch task processing flow, obtaining recommended IPC classification numbers and search elements, analyzing user preference profiles, determining search strategies, querying the search element knowledge graph based on the search elements, obtaining related search elements, and generating a search expression based on the search strategy, recommended IPC classification numbers, search elements, and related search elements.

[0012] Thirdly, embodiments of the present invention also provide a computer device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the above-mentioned multi-branch task processing method of intellectual property rights.

[0013] Fourthly, embodiments of the present invention also provide a computer-readable storage medium storing a computer program, which, when executed by a processor, implements the above-described multi-branch task processing method of the intellectual property rights.

[0014] Fifthly, embodiments of the present invention also provide a computer program product, the computer program product including a computer program, which, when executed by a processor, implements the above-mentioned intellectual property multi-branch task processing method.

[0015] The method proposed in this invention modularizes and streamlines patent retrieval and analysis tasks: decoupling complex patent information processing tasks into ordered branch tasks of retrieval element extraction and expression writing. This design enables the system to process users' vague natural language requirements step by step into precise, structured retrieval instructions, ensuring the standardization, controllability, and interpretability of the processing. A complete logical closed loop of "understanding-construction-execution" is formed: clearly defining the necessary path from "understanding the user's technical intent (element extraction)" to "constructing machine-readable instructions (expression writing)." The output of the retrieval element extraction branch (IPC classification number and retrieval elements) becomes the standardized input of the expression writing branch, avoiding the disconnect between intent understanding and retrieval execution in traditional processes, ensuring the logical consistency and predictability of the entire retrieval process. An "uncertainty term tag set" is introduced and utilized: not only extracting keywords but also identifying and tagging vague, broad, or uncertain expressions in user queries. By using this tag set to process IPC matching results, the strictness of the retrieval strategy can be dynamically adjusted. When faced with ambiguous requirements, the system automatically broadens the technical scope; when faced with precise requirements, it focuses on core classifications. This simulates the adaptive judgment ability of human experts when faced with incomplete information, greatly improving the system's intelligence level. It achieves collaborative intelligence between "IPC vector matching" and "uncertainty handling": combining semantic similarity-based IPC vector matching with logical rule-based uncertainty handling. This collaborative mechanism overcomes the limitations of single methods based solely on keyword matching or semantic similarity, enabling IPC classification number recommendations to capture the core idea while reasonably responding to the ambiguous boundaries of the query, resulting in more reasonable and comprehensive recommendations. It significantly shortens the "preparation time" from requirement submission to retrieval expression generation: automating the process replaces the tedious steps of manual IPC searching, synonym expansion, and concept hierarchy analysis required in traditional retrieval. Users only need to describe their technical requirements, and the system automatically outputs structured search elements and directly usable search expressions, compressing what might have been hours of professional preparation work into minutes or even seconds, greatly improving retrieval startup efficiency. It fundamentally improves the quality of search expression construction and the initial search hit rate: Because the search expression is generated based on system-recommended, uncertainty-optimized IPC classification numbers and multi-dimensional expanded search elements, it avoids common errors that easily occur during manual construction, such as missing IPC classification numbers, incomplete keywords, and improper logical relationships. This ensures that the initially constructed search expression has a high recall and precision rate, reducing the number of subsequent repeated debugging attempts and improving the overall accuracy and efficiency of the search process. It enables complex professional operations driven by natural dialogue: Users do not need to learn complex search syntax, IPC classification systems, or professional search strategies; they can drive the system to complete professional patent search preparatory work simply by describing technical problems, technical solutions, or technical fields using natural language.This breaks down the high barrier to entry for professional search tools, enabling R&D personnel and business managers who are not patent search experts to efficiently and accurately initiate patent information searches. It provides a clear and traceable technical processing path: the output includes not only the final result (search query) but also key intermediate products (such as recommended IPC classification numbers and extracted search elements). This gives users a sense of transparency and control; users can understand and verify the system's reasoning process, and make manual adjustments when necessary, enhancing user trust in the system's output and improving the collaborative experience. Attached Figure Description

[0016] To more clearly illustrate the technical solutions in the embodiments of the present invention or 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 of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort. In the drawings:

[0017] Figure 1 This is a flowchart of the intellectual property multi-branch task processing method in an embodiment of the present invention;

[0018] Figure 2 This is another flowchart of the intellectual property multi-branch task processing method in this embodiment of the invention;

[0019] Figure 3 This is another flowchart of the intellectual property multi-branch task processing method in the embodiments of the present invention;

[0020] Figure 4 This is a schematic diagram of the structure of the intellectual property multi-branch task processing device in an embodiment of the present invention;

[0021] Figure 5 This is a schematic diagram of a computer device in an embodiment of the present invention. Detailed Implementation

[0022] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the embodiments of the present invention will be further described in detail below with reference to the accompanying drawings. Here, the illustrative embodiments of the present invention and their descriptions are used to explain the present invention, but are not intended to limit the present invention.

[0023] Figure 1 This is a task graph for the intellectual property multi-branch task processing method in an embodiment of the present invention. The method includes:

[0024] Step 101: Perform intent recognition on the input user dialogue to obtain the branch task where the user intent is hit;

[0025] Step 102: If the user intends to hit the retrieval element extraction branch task, execute the retrieval element extraction branch task processing flow. The retrieval element extraction branch task processing flow includes: generating a set of user query key texts and a set of uncertain word tags based on the user dialogue; matching the set of user query key texts with the IPC classification number vector to obtain the matching result; processing the matching result through the set of uncertain word tags to obtain the recommended IPC classification number; and obtaining the retrieval element based on the recommended IPC classification number.

[0026] Step 103: If the user intends to hit the expression writing branch task, execute the expression writing branch task processing flow. The expression writing branch task processing flow includes: executing the retrieval element extraction branch task processing flow, obtaining the recommended IPC classification number and retrieval elements, analyzing the user preference profile, determining the retrieval strategy, querying the retrieval element knowledge graph based on the retrieval elements, obtaining related retrieval elements, and generating a retrieval expression based on the retrieval strategy, the recommended IPC classification number, the retrieval elements, and the related retrieval elements.

[0027] The solution in this embodiment of the invention achieves modularization and process-orientation of patent retrieval and analysis tasks: decoupling the complex patent information processing task into ordered branch tasks of retrieval element extraction and expression writing. This design enables the system to process users' vague natural language requirements step by step into precise structured retrieval instructions, like an assembly line, ensuring the standardization, controllability, and interpretability of the processing. A complete logical closed loop of "understanding-construction-execution" is formed: clearly defining the necessary path from "understanding the user's technical intent (element extraction)" to "constructing machine-readable instructions (expression writing)". The output of the retrieval element extraction branch (IPC classification number and retrieval elements) becomes the standardized input of the expression writing branch, avoiding the problem of disconnect between intent understanding and retrieval execution in traditional processes, ensuring the logical consistency and predictability of the results throughout the retrieval process. An "uncertainty term tag set" is introduced and utilized: not only extracting keywords, but also identifying and tagging vague, broad, or uncertain expressions in the user's query. By using this tag set to process IPC matching results, the strictness of the retrieval strategy can be dynamically adjusted. When faced with ambiguous requirements, the system automatically broadens the technical scope; when faced with precise requirements, it focuses on core classifications. This simulates the adaptive judgment ability of human experts when faced with incomplete information, greatly improving the system's intelligence level. It achieves collaborative intelligence between "IPC vector matching" and "uncertainty handling": combining semantic similarity-based IPC vector matching with logical rule-based uncertainty handling. This collaborative mechanism overcomes the limitations of single methods based solely on keyword matching or semantic similarity, enabling IPC classification number recommendations to capture the core idea while reasonably responding to the ambiguous boundaries of the query, resulting in more reasonable and comprehensive recommendations. It significantly shortens the "preparation time" from requirement submission to retrieval expression generation: automating the process replaces the tedious steps of manual IPC searching, synonym expansion, and concept hierarchy analysis required in traditional retrieval. Users only need to describe their technical requirements, and the system automatically outputs structured search elements and directly usable search expressions, compressing what might have been hours of professional preparation work into minutes or even seconds, greatly improving retrieval startup efficiency. It fundamentally improves the quality of search expression construction and the initial search hit rate: Because the search expression is generated based on system-recommended, uncertainty-optimized IPC classification numbers and multi-dimensional expanded search elements, it avoids common errors that easily occur during manual construction, such as missing IPC classification numbers, incomplete keywords, and improper logical relationships. This ensures that the initially constructed search expression has a high recall and precision rate, reducing the number of subsequent repeated debugging attempts and improving the overall accuracy and efficiency of the search process. It enables complex professional operations driven by natural dialogue: Users do not need to learn complex search syntax, IPC classification systems, or professional search strategies; they can drive the system to complete professional patent search preparatory work simply by describing technical problems, technical solutions, or technical fields using natural language.This breaks down the high barrier to entry for professional search tools, enabling R&D personnel and business managers who are not patent search experts to efficiently and accurately initiate patent information searches. It provides a clear and traceable technical processing path: the output includes not only the final result (search query) but also key intermediate products (such as recommended IPC classification numbers and extracted search elements). This gives users a sense of transparency and control; users can understand and verify the system's reasoning process, and make manual adjustments when necessary, enhancing user trust in the system's output and improving the collaborative experience.

[0028] In step 101, intent recognition is performed on the input user dialogue to obtain the branch task where the user intent is matched;

[0029] In this embodiment of the invention, user dialogue can be analyzed using rules and regular expressions to quickly capture strong signals and identify direct instruction verbs such as "extract," "expand," "analyze," and "summarize," which indicate the extraction of search elements. Phrases like "write a search expression" and "construct a query" indicate expression writing. Phrases like "check" and "find patents" indicate dialogue-based retrieval. Phrases like "analyze trends" and "create a report" indicate in-depth analysis. Key interrogative words such as "how to search" and "what words to use" can also be identified, potentially indicating expression writing or the extraction of search elements.

[0030] Then, deep semantic feature extraction can be performed: using a finely tuned intent classification model (such as a BERT-based text classifier) ​​to semantically encode and classify the complete sentence input by the user.

[0031] The model outputs probability distributions for four preset branches (retrieval element extraction, expression writing, dialogue retrieval, and in-depth analysis).

[0032] Next, conversation context features are extracted: the current multi-turn dialogue history is analyzed. For example, if a user has just completed element extraction and then says "use this to write the search query," even if the instruction is ambiguous, it should strongly point to expression writing. The system checks whether intermediate products (such as element packages or expressions) already exist in the current conversation as supplementary evidence for intent determination. This results in an intent feature vector that integrates surface features, semantic probabilities, and contextual cues.

[0033] The intent feature vector is input into the decision rule engine, which adopts a hybrid decision-making strategy of "rule first, model as a fallback". The execution logic of the decision rule engine is as follows:

[0034] The first layer of determination involves strong instruction word rule matching. It first examines the set of surface keywords extracted from the intent feature vector. If the keywords contain words closely related to concept analysis, such as "extract," "synonym," "expansion," or "keyword," the user intent is immediately determined to be the "retrievable element extraction branch task," and the routing process ends here. Otherwise, if the keywords contain both "write" (or "generate," "construct") and "search expression" (or "expression," "query statement"), the user intent is immediately determined to be the "expression writing branch task," and the routing process ends here. Otherwise, if the keywords contain both "search" (or "find," "retrieve") and "patent" (or "document"), the user intent is immediately determined to be the "dialogue retrieval branch task," and the routing process ends here. Otherwise, if the keywords contain "analysis" and simultaneously contain words related to the analysis dimension, such as "trend," "report," "competitor," or "layout," the user intent is immediately determined to be the "deep analysis branch task," and the routing process ends here.

[0035] The second layer of judgment is based on the continuity of the conversation context rule. If the first layer of rules fails to make a clear judgment, the history of the current conversation will be checked. If it is found that a "reusable retrieval element package" has already been generated in the current conversation (i.e., the user has just completed it or the system has a usable element extraction result), and the intent feature vector includes execution instructions such as "write search expression" or "check", then it is determined to enter the "expression writing branch task" first. The purpose is to reuse the existing element package to directly generate the search expression. The process is the smoothest, and the routing process ends here. Otherwise, if it is found that a "candidate search expression" has already been generated in the current conversation, and the intent feature vector includes "execute", "run", or directly ask for "result", then it is determined to enter the "dialogue retrieval branch task" first. The purpose is to directly use the existing expression to perform the retrieval and return the result. The routing process ends here.

[0036] The third layer of judgment is a fallback based on semantic model probability. If the judgments based on rules and context in the above two layers are not triggered, the probability distribution calculated by the deep semantic model in step one will be relied upon to extract the branch intent with the highest probability judged by the semantic model and check its confidence score. If the highest confidence score exceeds a preset threshold (for example, greater than 0.6), the user intent is determined to be the branch task recommended by the semantic model, and the routing process ends here. If the highest confidence score does not exceed the threshold, it means that the model cannot determine it either.

[0037] The fourth layer of processing is a default clarification query. A branch will not be forcibly selected if and only if all the above three layers of decision logic fail to reach a clear, high-confidence routing conclusion. In this case, the output is not a branch identifier, but a user-oriented, structured clarification question designed to guide the user to clarify their needs. For example, a response could be: "Would you like me to help you extract keywords and classification numbers for your search, or would you prefer I write a search expression for you directly?"

[0038] After going through the above four layers of logic, the final output is a clear "hit branch task identifier" (one of the four: retrieval element extraction, expression writing, dialogue retrieval, and in-depth analysis), along with an explanation of the basis for this decision (such as "matching based on strong instruction word rules").

[0039] If the process proceeds to the fourth level, a clarification guide instruction will be output instead of a branch indicator, and the process will restart from step one based on the user's subsequent response.

[0040] In one embodiment, the retrieval elements include at least one of the following: element words of the subject in the user dialogue, synonyms, hypernyms, hyponyms, translated words, IPC classifications, technical fields, and background technologies.

[0041] In one embodiment, the processing flow for the retrieval element extraction branch task includes:

[0042] The user dialogue is converted into a structured description, generating a set of key texts for user queries. Each key text in the set of key texts for user queries includes the keyword type and the keyword content.

[0043] Analyze ambiguous expressions in user dialogues to generate a set of uncertain word tags;

[0044] Using keyword types, all keyword content in the user query key text set is encoded into user query key text vectors. The similarity of user query key text vectors with the semantic vectors of each IPC classification number is calculated in turn to obtain the first IPC classification candidate list in the matching results. The semantic vector of each IPC classification number is generated based on the background text encoding of the IPC classification number.

[0045] Extract the core keywords from the keyword content of all user query key texts;

[0046] Match all core keywords with the keywords corresponding to each IPC classification number to obtain the second IPC classification candidate list in the matching results;

[0047] By fusing the first IPC classification candidate list and the second IPC classification candidate list with the uncertain word tag set, a sorted set of IPC classification numbers is obtained;

[0048] Select a first preset number of IPC classification numbers from the sorted set of IPC classification numbers as recommended IPC classification numbers.

[0049] In this embodiment of the invention, for example, if a user's dialogue is "I want to find some patents about high-end battery management, similar to those of industry leaders, that can effectively balance the voltage of individual battery cells, possibly using active balancing technology, preferably from recent years," then the user's key search text can be summarized as follows:

[0050] Technical topic: High-end battery management system (BMS);

[0051] Technical issue: Balancing the voltage of individual battery cells;

[0052] Technical means: Active balancing technology may be used;

[0053] Technical benchmark: Level of similar industry leaders;

[0054] Time requirement: Recent years;

[0055] Among them, "Technology Theme" is the keyword type, and "High-end Battery Management System (BMS)" is the keyword content.

[0056] In this embodiment of the invention, when generating the user query key text vector V_query, each keyword content is first encoded into a sub-vector. Then, according to the weights corresponding to different keyword types, all keyword content is weighted and encoded to obtain the user query key text vector. For example, the weight of technical topic is 0.3, the weight of technical issue is 0.2, the weight of technical means is 0.2, etc., and the sum of the weights of all keyword types is 1.

[0057] Taking IPC group H02J 7 / 00 as an example, the semantic vector of each IPC classification number is constructed by concatenating the following parts:

[0058] IPC category title: "Circuit devices for charging or discharging battery packs";

[0059] IPC's official definition and notes: "This group contains circuitry specifically designed for battery packs, including control of charging or discharging the battery pack, and circuitry for balancing the states of individual battery cells within the battery pack..."

[0060] Typical patent abstract (example): "This invention discloses a battery balancing circuit and method, which transfers energy from a high-voltage battery cell to a low-voltage battery cell by controlling a switching network and a capacitor, thereby achieving lossless balancing..."

[0061] The aforementioned background text is directly input into the same pre-trained model and encoded into a fixed semantic vector V_ipc_H02J_7_00.

[0062] When calculating semantic similarity, the cosine similarity formula can be used to obtain the semantic similarity between the user query key text vector and the semantic vector of each IPC classification number. The first IPC classification candidate list is formed by sorting the semantic similarity from high to low. The first IPC classification candidate list can have a preset length, depending on the actual situation.

[0063] The extracted core keywords include battery management (core theme), battery (core object), balancing (core function / action), voltage (core parameter / object attribute), active balancing (specific technical means), and individual cell (limited object, battery cell), etc. These can be encoded into core keyword vectors through a pre-trained large model. The keywords corresponding to each IPC classification number are also encoded into vectors. The extraction method for the keywords corresponding to each IPC classification number is the same as the method for extracting core keywords, and will not be repeated here. The steps for matching the core keyword vector with the keyword vector corresponding to each IPC classification number include:

[0064] For each core keyword, query the keywords corresponding to each IPC classification number to return a series of IPC classification numbers associated with each core keyword, as well as a keyword matching strength score (based on the statistical probability of the core keyword appearing in the IPC classification definition and typical patents), which can be represented as a triple of (core keyword, IPC_j, association strength). Sort the obtained IPC classification numbers according to the keyword matching strength score from high to low to obtain the second IPC classification candidate list.

[0065] For example, the keyword "battery balancing" might return {H02J 7 / 00: 0.95, H01M 10 / 42: 0.80}.

[0066] Each uncertainty tag in the uncertainty tag set includes multiple uncertainty word types and uncertainty word content.

[0067] The set of uncertain word tags can be organized as follows: [

[0069] {

[0070] "Fuzzy Type": "Technical Level Fuzzy"

[0071] "Anchor point": "Similar to the high-end of an industry leader".

[0072] "Processing Guidelines": "Technical specifications and architecture need to be expanded. Expansion directions: 1) High-performance indicators (such as high precision, high efficiency, and high reliability); 2) Advanced architecture (such as distributed, modular, and redundant design); 3) Advanced functions (such as state estimation, thermal management collaboration, and cloud monitoring)."

[0073] },

[0074] {

[0075] "Fuzzy Type": "Performance Degree Fuzzy"

[0076] "Anchor point": "relatively good land"

[0077] "Processing Guidelines": "Performance dimensions need to be quantitatively expanded. Expansion directions: equalization accuracy (e.g., voltage difference threshold <10mV), equalization efficiency (>95%), equalization speed (dynamic, real-time), equalization strategy (adaptive, predictive)."

[0078] },

[0079] {

[0080] "Fuzzy type": "Uncertainty about technical means"

[0081] "Anchor point": "Active balancing technology may have been used."

[0082] "Processing Guidelines": "Technical solutions need to be explored and compared. Expansion directions: 1) Clarify active balancing technologies (such as those based on DC-DC converters, inductors, and capacitor arrays); 2) Supplement their alternative or complementary technologies (passive balancing, hybrid balancing); 3) Cover specific circuit topologies (flying capacitors, switching inductors, etc.)."

[0083] },

[0084] {

[0085] "Fuzzy Type": "Fuzzy Time Range"

[0086] "Anchor point": "In recent years"

[0087] "Processing Guidelines": "The time range must be clearly defined and adapted to the patent publication cycle. The default extension is: application date or publication date to [current year - 5] to the current year. This is also presented to the user as an adjustable parameter."

[0088] } ]

[0090] Among them, "anchor point" is an uncertain word type, and "in recent years" is an uncertain word content.

[0091] In this embodiment of the invention, the first preset quantity can be 5 to 10.

[0092] In one embodiment, the first IPC classification candidate list includes multiple IPC classification numbers and corresponding semantic similarity scores; the second IPC classification candidate list includes multiple IPC classification numbers and corresponding keyword matching strength scores; each uncertainty tag in the uncertainty word tag set includes multiple uncertainty word types; the uncertainty word types include fuzzy types;

[0093] By fusing the first and second IPC classification candidate lists using the uncertain word tag set, a sorted set of IPC classification numbers is obtained, including:

[0094] After normalizing the semantic similarity score and keyword matching strength score, the union of the first IPC classification candidate list and the second IPC classification candidate list is used to obtain the basic candidate set.

[0095] The fuzziness weights of each fuzzy type in the uncertain word tag set are statistically analyzed, and the sum of the fuzziness weights of all fuzzy types is calculated as the tag weight;

[0096] The label weight is used as the semantic channel weight, and the value obtained by subtracting the label weight from 1 is used as the keyword channel weight.

[0097] For each IPC classification number in the basic candidate set, the weighted fusion score of the IPC classification number is calculated based on the semantic similarity score and the corresponding semantic channel weight, the keyword matching strength score and the corresponding keyword channel weight.

[0098] The basic candidate set is reordered from high to low according to the weighted fusion score, and the sorted IPC classification number set is output.

[0099] In this embodiment of the invention, the scores (SA and KB) in the first IPC classification candidate list A and the second IPC classification candidate list B are respectively Min-Max normalized so that they fall in the [0,1] interval, and the normalized scores SA_norm and KB_norm are obtained.

[0100] Take the union of all IPC classification numbers from lists A and B to form the basic candidate set U. Calculate the weight of the "fuzziness level" for each fuzzy type in the uncertain word tag set. For example:

[0101] Vague technical references and uncertain technical means → high weight (e.g., 0.4 each);

[0102] Performance level is ambiguous → weight is moderate (e.g., 0.2);

[0103] Ambiguous time range → low weight (e.g., 0.1);

[0104] The overall uncertainty score U_score is calculated as Σ(weights of each label). The higher the U_score, the more ambiguous the user query is.

[0105] Define the semantic channel weight α = U_score (i.e., the higher the uncertainty, the larger α is);

[0106] Define the keyword channel weight β = 1 - U_score (i.e., the higher the certainty, the larger β is);

[0107] Ensure α + β = 1;

[0108] For each IPC classification number ipc_i in the base set U:

[0109] Get individual score:

[0110] Get SA_norm(ipc_i) from the first IPC classification candidate list A, or record it as 0 if it does not exist.

[0111] Obtain KB_norm(ipc_i) from the second IPC category candidate list B, or record it as 0 if it does not exist.

[0112] Calculate the weighted fusion score F_base(ipc_i):

[0113] F_base(ipc_i)=α×SA_norm(ipc_i)+β×KB_norm(ipc_i).

[0114] In one embodiment, before reordering the basic candidate set according to the weighted fusion score from high to low and outputting the sorted IPC classification number set, the method further includes:

[0115] Based on the co-occurrence relationship of all IPC classifiers in the basic candidate set, the co-occurrence enhancement score of each IPC classifier is calculated;

[0116] The collaborative filtering score is calculated based on the weighted fusion score and co-occurrence enhancement score for each IPC classification number.

[0117] The basic candidate set is reordered according to the weighted fusion score from high to low, and the sorted IPC classification set is output, including:

[0118] The basic candidate set is reordered according to the scores after collaborative filtering from high to low, and the sorted IPC classification number set is output.

[0119] In this embodiment of the invention, the basic score is corrected by utilizing the co-occurrence relationship of IPC classification numbers in historical patent data (two classification numbers often appear in the same patent) in order to discover implicit associations.

[0120] Constructing co-occurrence enhancement scores:

[0121] For each ipc_i, find its Top-N most frequent co-occurring IPC classification numbers Co_occur(ipc_i) (from the pre-computed co-occurrence matrix).

[0122] Calculate the co-occurrence enhancement factor C_boost(ipc_i):

[0123] C_boost(ipc_i)=γ×Σ[F_base(ipc_j)×sim(ipc_i,ipc_j)]foripc_jinCo_occur(ipc_i)

[0124] Where: γ is the co-occurrence adjustment coefficient (e.g., 0.1), which controls the enhancement magnitude; sim(ipc_i,ipc_j) is the co-occurrence strength or semantic relevance (pre-calculated value); and the summation part is essentially a weighted average of the scores of the "neighbors" of ipc_i.

[0125] Calculate the score F_cf(ipc_i) after collaborative filtering:

[0126] F_cf(ipc_i)=F_base(ipc_i)+C_boost(ipc_i)

[0127] In one embodiment, the uncertain word type further includes processing guidelines;

[0128] Before reordering the basic candidate set according to the collaborative filtering scores from high to low and outputting the sorted IPC classification set, the following steps are also included:

[0129] For each uncertainty label, if the IPC classification number in the base candidate set belongs to the IPC classification number mapped to the processing guidelines for that uncertainty label, generate a targeted fine-tuning score for the IPC classification number.

[0130] The score after collaborative filtering for each IPC classification number is added to the targeted fine-tuning score to obtain the final score;

[0131] The basic candidate set is reordered from high to low according to the scores after collaborative filtering, and the sorted IPC classification number set is output.

[0132] In this embodiment of the invention, a mapping rule (predefined or determined in real time) can be established from the "processing guide" to the IPC classification number:

[0133] For example, for the label "Ambiguous technical level: technical indicators and architecture need to be extended", the known related IPC classification numbers include: G05B (control), G06F (data processing), and H04L (digital transmission) subclasses involving "intelligent" and "adaptive".

[0134] For example, when the label "Technical means uncertain: technical solutions need to be explored" is applied, the system will also pay attention to related classification numbers such as H02J7 / 00 (active balancing) and its alternative technology H01M10 / 44 (charging method).

[0135] Calculate the targeted fine-tuning bonus D_boost(ipc_i):

[0136] For each uncertainty tag, if ipc_i belongs to the IPC classification number mapped to by the tag's "processing guidelines", then a fixed directional fine-tuning score δ (e.g., 0.05) is added, or a directional fine-tuning score is obtained proportionally to the relevance.

[0137] D_boost(ipc_i) is the sum of the scores for all relevant tags.

[0138] Calculate the final score F_final(ipc_i):

[0139] F_final(ipc_i)=F_cf(ipc_i)+D_boost(ipc_i)

[0140] In this embodiment of the invention, after sorting by F_final(ipc_i) in descending order, a score threshold (e.g., 0.1) is set to remove classification numbers with too low F_final values. The maximum output preset number N (e.g., Top-10) can be set.

[0141] It can also generate explanatory metadata (optional but important): labeling the main contributing channel (semantic A / keyword B / co-occurrence enhancement / targeted fine-tuning) for each output IPC classification number, and recording the main uncertainty label types affecting its ranking. It outputs a preset number of recommended IPC classification numbers, with each entry containing (IPC classification number, final score).

[0142] In one embodiment, based on the recommended IPC classification number, the search elements are obtained, including:

[0143] Based on each recommended IPC classification number, basic elements are extracted from the IPC classification number interpretation library and packaged into an IPC knowledge background package;

[0144] Add the IPC knowledge background package, user query key text, uncertainty word tag set, and retrieval element requirements to the instruction template to obtain the large model instruction;

[0145] The large model instructions are input into the large language model to obtain the search elements.

[0146] In this embodiment of the invention, the basic elements include:

[0147] Official title and explanation: The precise technical scope definition of this IPC classification.

[0148] Typical technical means: Keywords of common technical solutions that fall into this category.

[0149] Common technical issues and effects: This category of technologies typically aims to solve problems and achieve results.

[0150] Direct hierarchical relationship: its parent class, child class classification number and brief description.

[0151] The instruction template can be represented as:

[0152] Role: You are a senior patent search analyst.

[0153] Task: Based on users' technical needs and combined with professional patent classification knowledge, extract multi-dimensional technical elements for patent retrieval.

[0154] User technical requirements description: [Key text for user queries]

[0155] When analyzing the above requirements, please pay special attention to addressing the following ambiguous or undefined aspects, and expand upon them accordingly:

[0156] [Set of Uncertainty-Based Word Tags]

[0157] To accurately understand the technical scope, please refer to the following relevant International Patent Classification (IPC) technical background information:

[0158] Enhanced contextual hints (i.e., textual integration of IPC knowledge background packages)

[0159] Please strictly adhere to the following eight dimensions when outputting structured results to ensure logical connections between elements:

[0160] 1. Key words: Core technical feature words derived from user needs and IPC background.

[0161] 2. Synonyms: The above-mentioned elements are commonly expressed as synonyms or near-synonyms in patent documents.

[0162] 3. Superordinate terms: These summarize the broader technical category to which the element term belongs.

[0163] 4. Sub-terms: List the specific implementation methods, components, or more detailed technologies of the element terms.

[0164] 5. Translation terms: Provide corresponding standard English terms.

[0165] 6. IPC Classification: List the directly related IPC classification numbers (refer to the background above, or add other related classifications).

[0166] 7. Technical Field: Define the application or research field to which this technology belongs.

[0167] 8. Background Technology: Describe the traditional technical problems, defects, or technical bottlenecks that this technology aims to solve.

[0168] In one embodiment, a search expression is generated based on the search strategy, recommended IPC classification number, search elements, and related search elements, including:

[0169] A three-layer structure of objectives, constraints, and examples is used to construct the cue word skeleton;

[0170] Based on user dialogue, generate a target description and add it to the target's keyword skeleton;

[0171] Add the search strategy, recommended IPC classification number, search elements, and related search elements as constraints to the prompt keyword skeleton;

[0172] Generate a sample search expression and add a suggestion keyword skeleton to obtain the suggestion keywords;

[0173] The prompt words are input into a pre-trained large model to generate at least one retrieval expression.

[0174] In this embodiment of the invention, user preference profiles are analyzed. If the user preference is highly accurate, the strategy is "precise retrieval", and the retrieval elements are mainly connected using "AND".

[0175] If a user's history shows that they frequently perform full searches, the strategy is "full coverage," with extensive use of "OR" connections between synonyms / hypothetical terms and "AND" connections between search elements.

[0176] Constructing the prompt keyword skeleton: adopting a three-layer structure of "target-constraint-example" and pre-embedding strategies;

[0177] Objective: The task is defined as "generating executable search expressions for the patent database".

[0178] Constraints (Dynamic Part):

[0179] Format: Specify the fields that must be included (e.g., TI / AB / CLMS, IPC, release date), logical operators (AND / OR / NOT), and the rules for using parentheses.

[0180] Strategy: Embed a "retrieval strategy" (e.g., "Please adopt a 'comprehensive coverage' strategy to ensure that synonyms and subordinate concepts of core technologies are fully expanded").

[0181] Search elements: embedded search elements and related search elements.

[0182] IPC: Embedded recommended IPC classification number.

[0183] Scope: Extract filtering criteria such as time range (e.g., "last three years") and patent type (e.g., "invention patent") from user preference profiles.

[0184] Example: Here is an example of a search expression.

[0185] In this embodiment of the invention, a retrieval element knowledge graph is pre-constructed. The retrieval element knowledge graph uses IPC classification numbers and inspection elements as nodes to establish four types of key association edges:

[0186] Conceptual hierarchy edges: Establish relationships such as is_a (is a kind of) and synonym_of (synonym to) between nodes of "element word", "synonym", "superordinate word" and "hypothesis word" to form a concept tree. For example: "active balancing" -- is_a --> "battery balancing"; "active balancing" -- synonym_of --> "active balancing".

[0187] Technology attribution edge: Establish the belongs_to relationship between the "Element Word" node and the "IPC Classification" and "Technology Field" nodes. For example: "Active Balancing" -- belongs_to --> "H02J7 / 00"; "SOC Estimation" -- belongs_to --> "Battery Management System (BMS) Field".

[0188] Problem-solving edge: Establishes a solves relationship between the "Element Word" node and the "Problem" node described in the "Background Technology" section. For example: "Active Balancing" -- solves --> "Battery pack capacity degradation inconsistency".

[0189] For core keywords entered by users (such as "new energy"), the search element knowledge graph can automatically return its subordinate sub-domains (solar energy, wind energy, hydrogen energy), related technologies (energy storage technology, smart grid) and core components (photovoltaic panels, wind turbine blades), ensuring that the suggested words cover the entire field.

[0190] The weight of the associated edge can be derived from the contribution of the retrieved element to the IPC classification number.

[0191] When obtaining related search elements, the weights between search elements are used to find related search elements whose weights on the edges corresponding to the search elements are greater than a weight threshold. Alternatively, all related search elements can be found directly.

[0192] In this embodiment of the invention, the obtained prompt word can be represented as:

[0193] [Objective] Please generate a professional search expression that can be directly used in the patent database. Do not use parentheses in the keywords in the search expression. Generate the search expression directly without any comments or prompts. The current date should be the automatically obtained system time.

[0194]

constraint

[0195] 1. Format: It must contain three parts connected by AND: the keyword part in (TI,AB,CLMS), the IPC category part, and the time / type filter part.

[0196] 2. Strategy: Please adopt the "Precise Search" strategy. The core function "Active Balancing" must be present in TI or AB.

[0197] 3. Search Elements: Please be sure to use the following search elements. {Insert search elements and related search elements}

[0198] 4. IPC: Please select the 3-5 most relevant classification numbers from the following set and connect them using OR: {Insert IPC classification number set}.

[0199] 5. Scope: Limited to "invention patents with an application date after 2020".

[0200]

Example

[0201] In addition to the above objectives, other requirements may be specified, such as:

[0202] 1) If the user's input involves a field, such as "new energy field", then the patent IPC number related to "new energy" should be output first, followed by the search query IPC = (number recommended by the model); if the user's input does not involve a patent IPC number, then no output is needed.

[0203] 2) If there is an IPC classification number, the generated IPC classification number will be compared with the patent classification knowledge base to optimize the result.

[0204] IPC Classification Mapping: Based on the technical field and core solution, match the patent classification thesaurus provided by the customer (the patent classification specified by the user should be used first, and if not, it will be automatically derived); mark the correspondence between patent classification and technical content (e.g., "A01B1 / 20 corresponds to the combination of different types of manual tools [2006.01]").

[0205] 3) Only check and optimize the IPC classification number; do not change or remove other logical expressions [e.g., tf=(computer)]. Furthermore, when there are multiple IPC classification numbers, only retain the one with the broadest coverage. For example: [G06F, G06F15 / 10], only retain G06F.

[0206] 4) Never use "\n" in the search query.

[0207] 5) Determine if the user input includes: [I. Text Content Class] Title, Abstract, Claims, Description. If so, add a tacd field containing the title, abstract, claims, and description, connected using OR. If the search results are less than 10, add them up to 10.

[0208] 6) If the tacd field is needed, it should be replaced in the original position of [I. Text Content Class] in the output.

[0209] 7) If the user input contains a date, carefully determine which date field should be selected for the search. Specific rules are as follows:

[0210] When the user mentions "application" in the conversation (such as application date, application year, application number, etc.), select the application-related date field: ad (application date), ady (application year), adym (application month and year).

[0211] When a user mentions "public" in the conversation (such as the date of publication, publication number, etc.), select the relevant date field for publication: pd (date of publication), pdy (year of publication), pdym (month and year of publication).

[0212] When the user mentions "authorization" (such as authorization date, authorization number, etc.), select the authorization-related date field: gd (authorization date).

[0213] When the user mentions "priority" in the conversation, select the priority-related field: prd (priority date).

[0214] If multiple date types are mentioned, they are processed according to the following priority: application date > publication date > authorization date > priority date, and only the most matching date field is selected.

[0215] 8) If you can clearly identify the type of the search content, use the search field under the corresponding type in [Searchable Fields].

[0216] For example, if the search content is related to an address, select the search field with "[Address]" as the type. If the search content is related to people, select the search field with "[Company & Related People]" as the type, and so on.

[0217] 9) It is strictly forbidden to mix search content with completely unrelated search fields. For example, "pa" refers to the applicant (patentee) and is unrelated to the address, so the "pa" field cannot be used.

[0218] In this embodiment of the invention, after generating at least one retrieval expression, the expression syntax can be automatically parsed to verify whether it meets the constraints (such as field usage and IPC number validity) and to mark the source of each part of the expression (such as: a keyword comes from a synonym of the element package, and an IPC comes from the recommendation set).

[0219] In one embodiment, the method further includes:

[0220] Analyze user preference profiles to obtain core user preference data, which includes the technical fields of interest.

[0221] Add the IPC classification number of the technical field of interest and the corresponding search elements to the prompt word skeleton.

[0222] User preference profiles include, but are not limited to, search keywords, filtering conditions (such as time range, technical field), and interactive feedback (such as clicking, saving, or disliking search results).

[0223] In this embodiment of the invention, the search expression can also be used to perform a rapid simulated search in a historical patent dataset or abstract library to generate estimated indicators:

[0224] Completeness bias: The breadth of use of "OR" and the number of hyponyms in the search expression.

[0225] Precision criterion: Statistical analysis of the use of "AND" in search expressions and the constraints of core terms in the TI field.

[0226] Result magnitude prediction: Predict the magnitude of the number of patent documents returned.

[0227] Generate personalized recommendation descriptions: Generate a description for each search expression and associate it with user preferences.

[0228] Example: "Expression A uses a 'precise search' strategy, limiting core keywords to the title. The estimated results are accurate but fewer in number, suitable for you to quickly locate core patents. Expression B expands the application scenario (electric vehicles), and the estimated results are more comprehensive, suitable for you to conduct a comprehensive patent analysis."

[0229] Final output: the primary search expression (1, comprehensive recommendation), alternative search expressions (1-2), and a personalized evaluation and recommendation report including estimated indicators, applicable scenario descriptions, and ingredient traceability.

[0230] Figure 2This is another flowchart of the dialogue retrieval branch task processing method in an embodiment of the present invention. In one embodiment, the branch task further includes a dialogue retrieval branch task.

[0231] The method further includes:

[0232] Step 104: If the user's intent is to hit the dialogue retrieval branch task, execute the dialogue retrieval branch task processing flow. The dialogue retrieval branch task processing flow includes: executing the expression writing branch task processing flow, obtaining the retrieval expression, calling the retrieval system API interface through the retrieval expression, and obtaining the retrieval result.

[0233] In this embodiment of the invention, the processing flow of the expression writing branch task (branch two) is automatically invoked to obtain the search expression. If the user has already generated a search expression through branch two, the historical search expression can be read directly; if not, branch two is triggered to generate the search expression.

[0234] The search results include the total number of search entries, the names, publication (announcement) numbers, and publication dates of the top 10 patents. These search results can be fed into a pre-trained large model for natural language polishing to ensure the results are clear and easy to understand. Polishing commands can be:

[0235] Role: You are a patent analysis expert. Please convert the following search results into an oral report for technical personnel.

[0236] Search background: {Source explanation of the insertion expression}.

[0237] Key technical points (optional): {Insert core concept clusters from the retrieval element package}.

[0238] Raw data:

[0239] A total of {totalHits} related patents were found.

[0240] The top N most relevant patents are listed below: {patentList}.

[0241] Report requirements:

[0242] 1. Introduction Summary: Use one sentence to summarize the search scope and overall situation.

[0243] 2. Key Patent Interpretation: For the first 3-5 patents, explain in layman's terms "what problem it solves" and "what its core solution is".

[0244] 3. Technological Trend Insights: Observe the application dates and applicants of these patents, and point out any obvious time trends or major R&D institutions.

[0245] 4. Follow-up suggestions: Based on the quantity and quality of the results, provide 1-2 follow-up action suggestions (such as "The results are accurate, the full text can be downloaded for detailed reading" or "There are too many results, it is recommended to add the qualifier 'electric vehicle'").

[0246] 5. Remain objective; all statements must be based on the given data.

[0247] In one embodiment, the branch task further includes summarizing and analyzing the branch task;

[0248] Figure 3 This is another flowchart of the dialogue retrieval branch task processing method in an embodiment of the present invention. The method further includes:

[0249] Step 105: If the user intent matches the summary and analysis branch task, execute the summary and analysis branch task processing flow. The summary and analysis branch task processing flow includes: obtaining the associated search scope based on the user intent; searching the analysis item knowledge base to obtain the analysis item that matches the user intent; executing the dialogue search branch task processing flow to obtain the search results; and calling the search system API interface to obtain the structured analysis results through the search results, the analysis item, and the associated search scope.

[0250] In this embodiment of the invention, the structured analysis results returned by the interface can also be fed into a pre-trained large model to build a report framework and refine the content, generating an analysis report that includes analysis conclusions, data support, and visual descriptions, and finally fed back to the user in the form of a document or dialogue. The data dimensions of the analysis items include at least patent trend analysis, applicant ranking, and regional distribution dimensions.

[0251] Report polishing instructions can be:

[0252] Role: You are a senior intellectual property analyst.

[0253] Task: Based on the following multi-dimensional data analysis, write a professional, concise, and insightful patent analysis report.

[0254] Analysis Background and Needs: The user's need is "{user dialogue}", to analyze the patent data of {number of patents}.

[0255] The original analysis data is as follows:

[0256] 1. [Technology Trend Analysis Results]: {Insert trend data from the search results, such as a list of annual application volumes}.

[0257] 2. [Analysis Results of Major Applicants]: {Insert applicant ranking and application volume data}.

[0258] 3. [Technical Composition Analysis Results]: {Insert IPC Classification Distribution Data}.

[0259] Report writing requirements:

[0260] 1. Framework: The report should include three parts: "Summary of Core Conclusions", "Detailed Analysis and Findings", and "Recommendations".

[0261] 2. Data-driven: All conclusions must be based on the above data, and key conclusions must cite specific data.

[0262] 3. Insight and Extraction: Not only describe the data, but also interpret the meaning behind the data (e.g., "The decline in the number of applications in year A may be related to the maturity of the technology in this field").

[0263] 4. Visual description: Clearly describe key chart trends in text (e.g., "Patent applications grew rapidly from year B, peaked in year C, and then declined slightly").

[0264] 5. Language: Professional but easy to understand.

[0265] In one embodiment, the method further includes:

[0266] If the user's intent does not match the branch task, the user's intent and dialogue are input into the pre-trained large model to generate the task processing result.

[0267] In this year's invention embodiments, the user's intent may also be patent translation, patent interpretation, etc. In this case, a pre-trained large model can be directly called to obtain the task processing results.

[0268] In this embodiment of the invention, based on the core design concept of "standardization covering general scenarios and personalization adapting to special needs," a dual-track workflow system of "built-in workflow + extension mechanism" is constructed. This ensures efficient processing of routine intellectual property tasks while flexibly responding to new business needs. The aforementioned four types of sub-tasks constitute the core functions of the standardized built-in workflow. Furthermore, administrators are allowed to customize new task branches based on new business scenarios (such as trademark search and copyright registration consultation), and configure the reuse relationship between branches and existing modules (e.g., when adding a "trademark search" branch, the element extraction logic of branch one is reused).

[0269] This invention also proposes an intellectual property multi-branch task processing device, the principle of which is similar to the intellectual property multi-branch task processing method, and will not be described in detail here.

[0270] Figure 4 This is a schematic diagram of the structure of the intellectual property multi-branch task processing device in an embodiment of the present invention, including:

[0271] The intent recognition module 401 is used to perform intent recognition on the input user dialogue and obtain the branch tasks where the user intent is hit.

[0272] The retrieval element extraction branch task processing module 402 is used to execute the retrieval element extraction branch task processing flow if the user's intent matches the retrieval element extraction branch task. The retrieval element extraction branch task processing flow includes: generating a set of user query key texts and a set of uncertain word tags based on the user dialogue; matching the set of user query key texts with the IPC classification number vector to obtain the matching result; processing the matching result through the set of uncertain word tags to obtain the recommended IPC classification number; and obtaining the retrieval element based on the recommended IPC classification number.

[0273] The expression writing branch task processing module 403 is used to execute the expression writing branch task processing flow if the user's intent hits the expression writing branch task. The expression writing branch task processing flow includes: executing the search element extraction branch task processing flow, obtaining the recommended IPC classification number and search elements, analyzing the user preference profile, determining the search strategy, querying the search element knowledge graph according to the search elements, obtaining related search elements, and generating a search expression according to the search strategy, the recommended IPC classification number, the search elements and related search elements.

[0274] In one embodiment, the retrieval element extraction branch task processing module is used for:

[0275] The user dialogue is converted into a structured description, generating a set of key texts for user queries. Each key text in the set of key texts for user queries includes the keyword type and the keyword content.

[0276] Analyze ambiguous expressions in user dialogues to generate a set of uncertain word tags;

[0277] Using keyword types, all keyword content in the user query key text set is encoded into user query key text vectors. The similarity of user query key text vectors with the semantic vectors of each IPC classification number is calculated in turn to obtain the first IPC classification candidate list in the matching results. The semantic vector of each IPC classification number is generated based on the background text encoding of the IPC classification number.

[0278] Extract the core keywords from the keyword content of all user query key texts;

[0279] Match all core keywords with the keywords corresponding to each IPC classification number to obtain the second IPC classification candidate list in the matching results;

[0280] By fusing the first IPC classification candidate list and the second IPC classification candidate list with the uncertain word tag set, a sorted set of IPC classification numbers is obtained;

[0281] Select a first preset number of IPC classification numbers from the sorted set of IPC classification numbers as recommended IPC classification numbers.

[0282] In one embodiment, the first IPC classification candidate list includes multiple IPC classification numbers and corresponding semantic similarity scores; the second IPC classification candidate list includes multiple IPC classification numbers and corresponding keyword matching strength scores; each uncertainty tag in the uncertainty word tag set includes multiple uncertainty word types; the uncertainty word types include fuzzy types;

[0283] The retrieval element extraction branch task processing module is used for:

[0284] After normalizing the semantic similarity score and keyword matching strength score, the union of the first IPC classification candidate list and the second IPC classification candidate list is used to obtain the basic candidate set.

[0285] The fuzziness weights of each fuzzy type in the uncertain word tag set are statistically analyzed, and the sum of the fuzziness weights of all fuzzy types is calculated as the tag weight;

[0286] The label weight is used as the semantic channel weight, and the value obtained by subtracting the label weight from 1 is used as the keyword channel weight.

[0287] For each IPC classification number in the basic candidate set, the weighted fusion score of the IPC classification number is calculated based on the semantic similarity score and the corresponding semantic channel weight, the keyword matching strength score and the corresponding keyword channel weight.

[0288] The basic candidate set is reordered from high to low according to the weighted fusion score, and the sorted IPC classification number set is output.

[0289] In one embodiment, the retrieval element extraction branch task processing module is used for:

[0290] Before reordering the basic candidate set according to the weighted fusion score from high to low and outputting the sorted IPC classification number set, the co-occurrence enhancement score of each IPC classification number is calculated based on the co-occurrence relationship of all IPC classification numbers in the basic candidate set.

[0291] The collaborative filtering score is calculated based on the weighted fusion score and co-occurrence enhancement score for each IPC classification number.

[0292] The basic candidate set is reordered according to the scores after collaborative filtering from high to low, and the sorted IPC classification number set is output.

[0293] In one embodiment, the uncertain word type further includes processing guidelines;

[0294] The retrieval element extraction branch task processing module is used for:

[0295] Before reordering the basic candidate set according to the scores after collaborative filtering from high to low and outputting the sorted set of IPC classification numbers, for each uncertainty label, if the IPC classification number in the basic candidate set belongs to the IPC classification number mapped to by the processing guidelines of that uncertainty label, a targeted fine-tuning score for the IPC classification number is generated.

[0296] The score after collaborative filtering for each IPC classification number is added to the targeted fine-tuning score to obtain the final score;

[0297] The basic candidate set is reordered according to the final score from high to low, and the sorted IPC classification number set is output.

[0298] In one embodiment, the retrieval element extraction branch task processing module is used for:

[0299] Based on each recommended IPC classification number, basic elements are extracted from the IPC classification number interpretation library and packaged into an IPC knowledge background package;

[0300] Add the IPC knowledge background package, user query key text, uncertainty word tag set, and retrieval element requirements to the instruction template to obtain the large model instruction;

[0301] The large model instructions are input into the large language model to obtain the search elements.

[0302] In one embodiment, the expression writing branch task processing module is used for:

[0303] A three-layer structure of objectives, constraints, and examples is used to construct the cue word skeleton;

[0304] Based on user dialogue, generate a target description and add it to the target's keyword skeleton;

[0305] Add the search strategy, recommended IPC classification number, search elements, and related search elements as constraints to the prompt keyword skeleton;

[0306] Generate a sample search expression and add a suggestion keyword skeleton to obtain the suggestion keywords;

[0307] The prompt words are input into a pre-trained large model to generate at least one retrieval expression.

[0308] In one embodiment, the branch task further includes a dialogue retrieval branch task;

[0309] In one embodiment, the device further includes a dialogue retrieval branch task processing module 404, for:

[0310] If the user's intent is to match the dialogue retrieval branch task, the dialogue retrieval branch task processing flow is executed. The dialogue retrieval branch task processing flow includes: executing the expression writing branch task processing flow, obtaining the retrieval expression, and calling the retrieval system API interface through the retrieval expression to obtain the retrieval results.

[0311] In one embodiment, the branch task further includes summarizing and analyzing the branch task;

[0312] In one embodiment, the apparatus further includes a summary and analysis branch task processing module 405, used for:

[0313] If the user's intent matches the summary and analysis branch task, the summary and analysis branch task processing flow is executed. The summary and analysis branch task processing flow includes: obtaining the associated search scope based on the user's intent; searching the analysis item knowledge base to obtain the analysis item that matches the user's intent; executing the dialogue search branch task processing flow to obtain the search results; and calling the search system API interface to obtain structured analysis results through the search results, the analysis item, and the associated search scope.

[0314] The proposed solution in this invention modularizes and streamlines patent retrieval and analysis tasks: decoupling complex patent information processing tasks into ordered branch tasks of retrieval element extraction and expression writing. This design enables the system to process users' vague natural language requirements step by step into precise, structured retrieval instructions, ensuring the standardization, controllability, and interpretability of the processing. A complete logical closed loop of "understanding-construction-execution" is formed: clearly defining the necessary path from "understanding the user's technical intent (element extraction)" to "constructing machine-readable instructions (expression writing)." The output of the retrieval element extraction branch (IPC classification number and retrieval elements) becomes the standardized input of the expression writing branch, avoiding the disconnect between intent understanding and retrieval execution in traditional processes, ensuring the logical consistency and predictability of the entire retrieval process. An "uncertainty term tag set" is introduced and utilized: not only extracting keywords but also identifying and tagging vague, broad, or uncertain expressions in user queries. By using this tag set to process IPC matching results, the strictness of the retrieval strategy can be dynamically adjusted. When faced with ambiguous requirements, the system automatically broadens the technical scope; when faced with precise requirements, it focuses on core classifications. This simulates the adaptive judgment ability of human experts when faced with incomplete information, greatly improving the system's intelligence level. It achieves collaborative intelligence between "IPC vector matching" and "uncertainty handling": combining semantic similarity-based IPC vector matching with logical rule-based uncertainty handling. This collaborative mechanism overcomes the limitations of single methods based solely on keyword matching or semantic similarity, enabling IPC classification number recommendations to capture the core idea while reasonably responding to the ambiguous boundaries of the query, resulting in more reasonable and comprehensive recommendations. It significantly shortens the "preparation time" from requirement submission to retrieval expression generation: automating the process replaces the tedious steps of manual IPC searching, synonym expansion, and concept hierarchy analysis required in traditional retrieval. Users only need to describe their technical requirements, and the system automatically outputs structured search elements and directly usable search expressions, compressing what might have been hours of professional preparation work into minutes or even seconds, greatly improving retrieval startup efficiency. It fundamentally improves the quality of search expression construction and the initial search hit rate: Because the search expression is generated based on system-recommended, uncertainty-optimized IPC classification numbers and multi-dimensional expanded search elements, it avoids common errors that easily occur during manual construction, such as missing IPC classification numbers, incomplete keywords, and improper logical relationships. This ensures that the initially constructed search expression has a high recall and precision rate, reducing the number of subsequent repeated debugging attempts and improving the overall accuracy and efficiency of the search process. It enables complex professional operations driven by natural dialogue: Users do not need to learn complex search syntax, IPC classification systems, or professional search strategies; they can drive the system to complete professional patent search preparatory work simply by describing technical problems, technical solutions, or technical fields using natural language.This breaks down the high barrier to entry for professional search tools, enabling R&D personnel and business managers who are not patent search experts to efficiently and accurately initiate patent information searches. It provides a clear and traceable technical processing path: the output includes not only the final result (search query) but also key intermediate products (such as recommended IPC classification numbers and extracted search elements). This gives users a sense of transparency and control; users can understand and verify the system's reasoning process, and make manual adjustments when necessary, enhancing user trust in the system's output and improving the collaborative experience.

[0315] This invention also provides a computer device. Figure 5 This is a schematic diagram of a computer device in an embodiment of the present invention. The computer device 500 includes a memory 510, a processor 520, and a computer program 530 stored in the memory 510 and executable on the processor 520. When the processor 520 executes the computer program 530, it implements the above-described method.

[0316] This invention also provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the above-described method.

[0317] This invention also provides a computer program product, which includes a computer program that, when executed by a processor, implements the above-described method.

[0318] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0319] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0320] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0321] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0322] The specific embodiments described above further illustrate the purpose, technical solution, and beneficial effects of the present invention. It should be understood that the above descriptions are merely specific embodiments of the present invention and are not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for processing multi-branch tasks related to intellectual property, characterized in that, include: Perform intent recognition on the input user dialogue to obtain the branch task of matching the user intent; If the user intends to target the retrieval element extraction branch task, the retrieval element extraction branch task processing flow is executed. The retrieval element extraction branch task processing flow includes: generating a set of user query key texts and a set of uncertain word tags based on the user dialogue; matching the set of user query key texts with the IPC classification number vector to obtain the matching result; processing the matching result through the set of uncertain word tags to obtain the recommended IPC classification number; and obtaining the retrieval element based on the recommended IPC classification number. If the user intends to hit the expression writing branch task, the expression writing branch task processing flow is executed. The expression writing branch task processing flow includes: executing the search element extraction branch task processing flow, obtaining recommended IPC classification numbers and search elements, analyzing user preference profiles, determining search strategies, querying the search element knowledge graph based on the search elements, obtaining related search elements, and generating search expressions based on the search strategies, recommended IPC classification numbers, search elements, and related search elements.

2. The method according to claim 1, characterized in that, The processing flow for the retrieval element extraction branch task includes: The user dialogue is converted into a structured description, generating a set of key texts for user queries. Each key text in the set of key texts for user queries includes the keyword type and the keyword content. Analyze ambiguous expressions in user dialogues to generate a set of uncertain word tags; Using keyword types, all keyword content in the user query key text set is encoded into user query key text vectors. The similarity of user query key text vectors with the semantic vectors of each IPC classification number is calculated in turn to obtain the first IPC classification candidate list in the matching results. The semantic vector of each IPC classification number is generated based on the background text encoding of the IPC classification number. Extract the core keywords from the keyword content of all user query key texts; Match all core keywords with the keywords corresponding to each IPC classification number to obtain the second IPC classification candidate list in the matching results; By fusing the first IPC classification candidate list and the second IPC classification candidate list with the uncertain word tag set, a sorted set of IPC classification numbers is obtained; Select a first preset number of IPC classification numbers from the sorted set of IPC classification numbers as recommended IPC classification numbers.

3. The method according to claim 2, characterized in that, The first IPC classification candidate list includes multiple IPC classification numbers and their corresponding semantic similarity scores; the second IPC classification candidate list includes multiple IPC classification numbers and their corresponding keyword matching strength scores; each uncertainty tag in the uncertainty word tag set includes multiple uncertainty word types; the uncertainty word types include fuzzy types; By fusing the first and second IPC classification candidate lists using the uncertain word tag set, a sorted set of IPC classification numbers is obtained, including: After normalizing the semantic similarity score and keyword matching strength score, the union of the first IPC classification candidate list and the second IPC classification candidate list is used to obtain the basic candidate set. The fuzziness weights of each fuzzy type in the uncertain word tag set are statistically analyzed, and the sum of the fuzziness weights of all fuzzy types is calculated as the tag weight; The label weight is used as the semantic channel weight, and the value obtained by subtracting the label weight from 1 is used as the keyword channel weight. For each IPC classification number in the basic candidate set, the weighted fusion score of the IPC classification number is calculated based on the semantic similarity score and the corresponding semantic channel weight, the keyword matching strength score and the corresponding keyword channel weight. The basic candidate set is reordered from high to low according to the weighted fusion score, and the sorted IPC classification number set is output.

4. The method according to claim 3, characterized in that, Before reordering the basic candidate set according to the weighted fusion score from high to low and outputting the sorted IPC classification set, the following steps are also included: Based on the co-occurrence relationship of all IPC classifiers in the basic candidate set, the co-occurrence enhancement score of each IPC classifier is calculated; The collaborative filtering score is calculated based on the weighted fusion score and co-occurrence enhancement score for each IPC classification number. The basic candidate set is reordered from high to low according to the weighted fusion score, and the sorted IPC classification number set is output.

5. The method according to claim 4, characterized in that, The uncertain word type also includes processing guidelines; Before reordering the basic candidate set according to the collaborative filtering scores from high to low and outputting the sorted IPC classification set, the following steps are also included: For each uncertainty label, if the IPC classification number in the base candidate set belongs to the IPC classification number mapped to the processing guidelines for that uncertainty label, generate a targeted fine-tuning score for the IPC classification number. The score after collaborative filtering for each IPC classification number is added to the targeted fine-tuning score to obtain the final score; The basic candidate set is reordered from high to low according to the scores after collaborative filtering, and the sorted IPC classification number set is output.

6. The method according to claim 1, characterized in that, Based on the recommended IPC classification number, the search elements are obtained, including: Based on each recommended IPC classification number, basic elements are extracted from the IPC classification number interpretation library and packaged into an IPC knowledge background package; Add the IPC knowledge background package, user query key text, uncertainty word tag set, and retrieval element requirements to the instruction template to obtain the large model instruction; The large model instructions are input into the large language model to obtain the search elements.

7. The method according to claim 1, characterized in that, Based on the search strategy, recommended IPC classification numbers, search elements, and related search elements, a search expression is generated, including: A three-layer structure of objectives, constraints, and examples is used to construct the cue word skeleton; Based on user dialogue, generate a target description and add it to the target's keyword skeleton; Add the search strategy, recommended IPC classification number, search elements, and related search elements as constraints to the prompt keyword skeleton; Generate a sample search expression and add a suggestion keyword skeleton to obtain the suggestion keywords; The prompt words are input into a pre-trained large model to generate at least one retrieval expression.

8. The method according to claim 1, characterized in that, The branch tasks also include the dialogue retrieval branch task; The method further includes: If the user's intent is to match the dialogue retrieval branch task, the dialogue retrieval branch task processing flow is executed. The dialogue retrieval branch task processing flow includes: executing the expression writing branch task processing flow, obtaining the retrieval expression, and calling the retrieval system API interface through the retrieval expression to obtain the retrieval results.

9. The method according to claim 1, characterized in that, The branch tasks also include summarizing and analyzing the branch tasks; The method further includes: If the user's intent matches the summary and analysis branch task, the summary and analysis branch task processing flow is executed. The summary and analysis branch task processing flow includes: obtaining the associated search scope based on the user's intent; searching the analysis item knowledge base to obtain the analysis item that matches the user's intent; executing the dialogue search branch task processing flow to obtain the search results; and calling the search system API interface to obtain structured analysis results through the search results, the analysis item, and the associated search scope.

10. A multi-branch task processing device for intellectual property rights, characterized in that, include: The intent recognition module is used to recognize the intent of the input user dialogue and obtain the branch tasks that match the user intent. The retrieval element extraction branch task processing module is used to execute the retrieval element extraction branch task processing flow if the user's intent matches the retrieval element extraction branch task. The retrieval element extraction branch task processing flow includes: generating a set of user query key texts and a set of uncertain word tags based on the user dialogue; matching the set of user query key texts with the IPC classification number vector to obtain the matching result; processing the matching result through the set of uncertain word tags to obtain the recommended IPC classification number; and obtaining the retrieval element based on the recommended IPC classification number. The expression writing branch task processing module is used to execute the expression writing branch task processing flow if the user's intent matches the expression writing branch task. The expression writing branch task processing flow includes: executing the search element extraction branch task processing flow, obtaining recommended IPC classification numbers and search elements, analyzing user preference profiles, determining search strategies, querying the search element knowledge graph based on the search elements, obtaining related search elements, and generating a search expression based on the search strategy, recommended IPC classification numbers, search elements, and related search elements.

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

12. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program that, when executed by a processor, implements the method of any one of claims 1 to 9.

13. A computer program product, characterized in that, The computer program product includes a computer program that, when executed by a processor, implements the method of any one of claims 1 to 9.

Images