A patent matching method and device based on multi-agent cooperation

Abstract

The embodiment of the application discloses a patent matching method and device based on multi-agent cooperation, which is applied to the field of artificial intelligence, and the method comprises the following steps: obtaining a query text to be matched and a plurality of candidate texts to be compared; obtaining standardized structure information by performing standardization processing on the query text and the candidate texts based on a perception agent; obtaining a matching text set by matching the standardized structure information of the query text and the candidate texts based on an inference agent; the matching text set comprises at least one candidate text; obtaining directional information of the candidate text in the matching text set based on the inference agent, and determining the matching text corresponding to the query text from the matching text set according to the directional information of the candidate text. By using the application, the efficiency and accuracy of text matching can be improved.

Description

Technical Field

[0001] This application relates to the field of artificial intelligence technology, and in particular to a patent matching method and apparatus based on multi-agent collaboration. Background Technology

[0002] In the field of intellectual property management, patent matching is a crucial cognitive task aimed at identifying documents with overlapping technological innovations, thereby assisting in technology novelty searches, infringement risk assessments, and patent layout optimization. With the exponential growth in the scale of patent documents, patent matching technology based on Large Language Models (LLMs) has become the mainstream paradigm in current research and application. Currently, existing technologies mainly follow two mainstream architectural paths when processing the semantic representation and matching reasoning of patent documents: The first is a naive content matching method (based on end-to-end full text). This method typically concatenates the query patent and retrieved candidate patent documents (usually including full-text information such as abstracts) directly into a long text sequence, which is then directly input into the large language model. After receiving the full set of technical details, the model outputs the matching result based on the logical overlap in the text. However, because this method directly concatenates the full long text sequence, it contains a large amount of semantic noise, resulting in low matching efficiency. The second is a matching method based on structured information enhancement. This method aims to obtain more structured auxiliary information through a pre-extraction paradigm. The basic process involves using specific prompts to drive a large model to traverse its internal memory, extracting key entities or classification ontology (such as the technical field under the IPC classification) from the patent, and then directly using these extracted structured data for matching and comparison. However, this method relies solely on generalized and fragmented keywords for comparison, lacking support from micro-level underlying details. It can only achieve a rough screening of macro-level technical fields and is unable to resolve subtle differences in the underlying implementation path or micro-level algorithm logic of highly similar patents. The accuracy of patent matching needs to be improved. Summary of the Invention

[0003] This application provides a patent matching method, apparatus, computer device, and storage medium based on multi-agent collaboration, which can improve the efficiency and accuracy of text matching.

[0004] This application provides a patent matching method based on multi-agent cooperation, which may include: Retrieve the query text to be matched and multiple candidate texts to be compared; Based on the perceptual intelligent agent, the query text and the candidate text are standardized to obtain standardized structural information; A set of matched texts is obtained by matching the standardized structural information of the query text and the candidate texts based on the reasoning agent; the set of matched texts includes at least one candidate text. Based on the reasoning agent, the orientation information of candidate texts in the matching text set is obtained, and the matching text corresponding to the query text is determined from the matching text set according to the orientation information of the candidate texts.

[0005] In one feasible implementation, it further includes: Obtain the reasoning agent to be trained; The training database is obtained based on an expert intelligent agent. The training database includes multiple training data sets. Each training data set includes a query training data, a positive sample, and three negative samples. The training data set carries an inference logic chain. Based on the perceptual agent, the data in the training set is standardized to obtain the training standardized structure information; Based on the inference agent to be trained, the training matching result of the query training data is generated according to the standardized structural information, and the decision logic chain for generating the training matching result is also generated. A loss function is generated for the reasoning agent to be trained based on the training matching results, the reasoning logic chain, and the decision logic chain, and the parameters in the reasoning agent to be trained are adjusted based on the loss function. When the loss function converges, the adjusted inference agent to be trained is determined as the inference agent, and the parameters in the inference agent are the adjusted parameters.

[0006] In one feasible implementation, the step of generating a loss function for the inference agent to be trained based on the training matching result, the inference logic chain, and the decision logic chain, and adjusting the parameters in the inference agent to be trained based on the loss function, includes: The decision logic chain carries the total length of time steps and the token corresponding to each time step. The weight of each token is determined based on the reasoning logic chain and the decision logic chain. A first loss function is generated based on the total length of the time steps, the token corresponding to each time step, and the weight of each token; The reasoning logic chain includes the output probabilities of multiple output results. Based on the matching training results, the output probability of the correct result among the multiple output results is determined as the first output probability, and the output probability of the output result with the second highest output probability is determined as the second output probability. A second loss function is generated based on the first output probability, the second output probability, and the confidence boundary. The loss function of the inference agent to be trained is determined based on the first loss and the second loss, and the parameters in the inference agent to be trained are adjusted based on the loss function.

[0007] In one feasible implementation, the standardization process of the query text and the candidate text based on the perceptual agent to obtain standardized structural information includes: Keywords for the standardized processing of the perceptual agent are obtained, and the standardized processing is used to obtain standardized structural information; the structural information includes entity information, usage information, and classification information; the entity information, usage information, and classification information each have corresponding keywords; The perceptual agent is invoked to collect information from the query text using keywords, obtaining the entity information, usage information, and classification information corresponding to the query text. The perceptual agent is invoked to collect information about the candidate text using keywords, thereby obtaining the entity information, usage information, and classification information corresponding to the candidate text.

[0008] In one feasible implementation, it further includes: If the number of texts in the matching text set is 1, then the candidate texts in the matching text set are determined as the matching texts corresponding to the query text; If the number of texts in the matching text set is greater than 1, then the step of obtaining the orientation information of the candidate texts in the matching text set and determining the matching text corresponding to the query text from the matching text set based on the orientation information is executed.

[0009] In one feasible implementation, the step of matching the standardized structural information of the query text and candidate texts based on the reasoning agent to obtain a set of matched texts includes: The similarity is calculated based on the standardized structural information of the query text and the target candidate text using a reasoning agent; the target candidate text is any candidate text to be compared; the standardized structural information includes entity information, usage information, and classification information. When the similarity of the entity information, purpose information, and classification information of the query text and the target candidate text is all greater than the first similarity threshold, the target candidate text is added to the matching text set.

[0010] In one feasible implementation, the step of obtaining orientation information of candidate texts in the matching text set based on the reasoning agent, and determining the matching text corresponding to the query text from the matching text set based on the orientation information of the candidate texts, includes: The reasoning agent is invoked to obtain the orientation information of the first candidate text from the patent database, and the orientation information is added to the context of the first candidate text to obtain candidate context information; the first candidate text is any candidate text in the matching text set; The reasoning agent calculates the similarity between the first candidate context information and the context information of the query text. When the similarity between the first candidate context information and the context information of the query text is greater than the second similarity threshold, the first candidate text is determined as the matching text corresponding to the query text.

[0011] In one feasible implementation, it further includes: When the similarity between the first candidate context information and the context information of the query text is in an ambiguous state; The reasoning agent is invoked to obtain external retrieval enhancement information of the first candidate text from the patent database, and the external retrieval enhancement information is added to the context of the first candidate text to obtain enhanced candidate context information; When the similarity between the enhanced candidate context information and the context information of the query text is greater than a third similarity threshold, the first candidate text is determined as the matching text corresponding to the query text.

[0012] One embodiment of this application provides a patent matching device based on multi-agent cooperation, which may include: The text acquisition unit is used to acquire the query text to be matched and multiple candidate texts to be compared; The text processing unit is used to perform standardized processing on the query text and the candidate text based on the perceptual agent to obtain standardized structural information; The first text matching unit is used to match the standardized structural information of the query text and the candidate text based on the reasoning agent to obtain a set of matching texts; the set of matching texts includes at least one candidate text. The second text matching unit is used to obtain the orientation information of candidate texts in the matching text set based on the reasoning agent, and determine the matching text corresponding to the query text from the matching text set according to the orientation information of the candidate texts.

[0013] In one feasible implementation, a training unit is also included, for: Obtain the reasoning agent to be trained; The training database is obtained based on an expert intelligent agent. The training database includes multiple training data sets. Each training data set includes a query training data, a positive sample, and three negative samples. The training data set carries an inference logic chain. Based on the perceptual agent, the data in the training set is standardized to obtain the training standardized structure information; Based on the inference agent to be trained, the training matching result of the query training data is generated according to the standardized structural information, and the decision logic chain for generating the training matching result is also generated. A loss function is generated for the reasoning agent to be trained based on the training matching results, the reasoning logic chain, and the decision logic chain, and the parameters in the reasoning agent to be trained are adjusted based on the loss function. When the loss function converges, the adjusted inference agent to be trained is determined as the inference agent, and the parameters in the inference agent are the adjusted parameters.

[0014] In one feasible implementation, the training unit is used for: The decision logic chain carries the total length of time steps and the token corresponding to each time step. The weight of each token is determined based on the reasoning logic chain and the decision logic chain. A first loss function is generated based on the total length of the time steps, the token corresponding to each time step, and the weight of each token; The reasoning logic chain includes the output probabilities of multiple output results. Based on the matching training results, the output probability of the correct result among the multiple output results is determined as the first output probability, and the output probability of the output result with the second highest output probability is determined as the second output probability. A second loss function is generated based on the first output probability, the second output probability, and the confidence boundary. The loss function of the inference agent to be trained is determined based on the first loss and the second loss, and the parameters in the inference agent to be trained are adjusted based on the loss function.

[0015] In one feasible implementation, the text processing unit is configured to: Keywords for the standardized processing of the perceptual agent are obtained, and the standardized processing is used to obtain standardized structural information; the structural information includes entity information, usage information, and classification information; the entity information, usage information, and classification information each have corresponding keywords; The perceptual agent is invoked to collect information from the query text using keywords, obtaining the entity information, usage information, and classification information corresponding to the query text. The perceptual agent is invoked to collect information about the candidate text using keywords, thereby obtaining the entity information, usage information, and classification information corresponding to the candidate text.

[0016] In one feasible implementation, a detection unit is also included, for: If the number of texts in the matching text set is 1, then the candidate texts in the matching text set are determined as the matching texts corresponding to the query text; If the number of texts in the matching text set is greater than 1, then the step of obtaining the orientation information of the candidate texts in the matching text set and determining the matching text corresponding to the query text from the matching text set based on the orientation information is executed.

[0017] In one feasible implementation, the first text matching unit is configured to: The similarity is calculated based on the standardized structural information of the query text and the target candidate text using a reasoning agent; the target candidate text is any candidate text to be compared; the standardized structural information includes entity information, usage information, and classification information. When the similarity of the entity information, purpose information, and classification information of the query text and the target candidate text is all greater than the first similarity threshold, the target candidate text is added to the matching text set.

[0018] In one feasible implementation, the second text matching unit is used for: The reasoning agent is invoked to obtain the orientation information of the first candidate text from the patent database, and the orientation information is added to the context of the first candidate text to obtain candidate context information; the first candidate text is any candidate text in the matching text set; The reasoning agent performs similarity calculations on the context information of the first candidate context information and the context information of the query information. When the similarity between the first candidate context information and the context information of the query text is greater than the second similarity threshold, the first candidate text is determined as the matching text corresponding to the query text.

[0019] In one feasible implementation, the second text matching unit is further configured to: When the similarity between the first candidate context information and the query context information is in a fuzzy state; The reasoning agent is invoked to obtain external retrieval enhancement information of the first candidate text from the patent database, and the external retrieval enhancement information is added to the context of the first candidate text to obtain enhanced candidate context information; When the similarity between the enhanced candidate context information and the context information of the query text is greater than a third similarity threshold, the first candidate text is determined as the matching text corresponding to the query text.

[0020] One embodiment of this application provides a computer-readable storage medium storing a computer program adapted to be loaded by a processor and executed by the above-described method steps.

[0021] One embodiment of this application provides a computer device, including: a processor, a memory, and a network interface; the processor is connected to the memory and the network interface, wherein the network interface is used to provide network communication functions, the memory is used to store program code, and the processor is used to call the program code to execute the above-described method steps.

[0022] One embodiment of this application provides a computer program product or computer program, which includes computer instructions stored in a computer-readable storage medium. A processor of a computer device reads the computer instructions from the computer-readable storage medium and executes the computer instructions, causing the computer device to perform the method steps described above.

[0023] In this embodiment, by acquiring the query text to be matched and multiple candidate texts to be compared, the query text and candidate texts are further standardized using a perceptual agent to obtain standardized structural information. This standardization process filters out redundant expressions, non-core technical details, and format noise in the original text, unifying the semantic expression paradigm of the patent text. Further, a matching text set is obtained by matching the standardized structural information of the query text and candidate texts using a reasoning agent; the matching text set includes at least one candidate text. Further, the orientation information of the candidate texts in the matching text set is acquired using the reasoning agent, and the matching text corresponding to the query text is determined from the matching text set based on the orientation information of the candidate texts. This method achieves text acquisition-matching-acquiring-re-matching, which not only reduces redundancy through structured information processing but also allows for re-matching using more detailed orientation information, improving the efficiency and accuracy of text matching. Attached Figure Description

[0024] To more clearly illustrate the technical solutions in the embodiments of this application 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 this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0025] Figure 1 This is a network architecture diagram of patent matching based on multi-agent collaboration provided in an embodiment of this application; Figure 2 This is a flowchart illustrating a patent matching method based on multi-agent collaboration provided in an embodiment of this application; Figure 3 This is a flowchart illustrating a patent matching method based on multi-agent collaboration provided in an embodiment of this application; Figure 4 This is an example schematic diagram of a patent matching method based on multi-agent collaboration provided in an embodiment of this application; Figure 5 This is a schematic diagram of the structure of a patent matching device based on multi-agent collaboration provided in an embodiment of this application; Figure 6 This is a schematic diagram of the structure of a computer device provided in an embodiment of this application. Detailed Implementation

[0026] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0027] Please see Figure 1 , Figure 1 This is a network architecture diagram of patent matching based on multi-agent collaboration provided in this application embodiment. The network architecture diagram may include a service server 100 and a user terminal cluster. The user terminal cluster may include user terminal 10a, user terminal 10b, ..., user terminal 10c. Communication connections may exist between user terminals in the cluster; for example, there is a communication connection between user terminal 10a and user terminal 10b, and a communication connection between user terminal 10b and user terminal 10c. Furthermore, any user terminal in the user terminal cluster may have a communication connection with the service server 100; for example, there is a communication connection between user terminal 10a and service server 100, and a communication connection between user terminal 10b and service server 100.

[0028] The aforementioned user terminal cluster (including user terminal 10a, user terminal 10b, and user terminal 10c) can all have the target application installed. Optionally, the target application may include an application with the function of displaying data information such as text, images, and videos.

[0029] Database 10d is used to store data such as text, video, and audio. For example, the database can store query text, candidate text, and the orientation information of query text and candidate text.

[0030] Optionally, the aforementioned user terminal can be one of the above-mentioned... Figure 1 Any user terminal selected in the user terminal cluster of the corresponding embodiment, for example, the user terminal can be the aforementioned user terminal 10b.

[0031] It is understood that the methods provided in this application embodiment can be executed by computer devices, including but not limited to terminals or servers. The business server 100 in this application embodiment can be a computer device, and the user terminals in the user terminal cluster can also be computer devices; this is not limited here. The aforementioned business server can be an independent physical server, a server cluster or distributed system composed of multiple physical servers, or a cloud server providing basic cloud computing services such as cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communication, middleware services, domain name services, security services, CDN, and big data and artificial intelligence platforms. The aforementioned terminals can include: smartphones, tablets, laptops, desktop computers, smart TVs, smart speakers, desktop computers, smartwatches, and other smart terminals with image recognition functions, but are not limited to these. The user terminals and the business server can be directly or indirectly connected via wired or wireless communication; this application does not impose any restrictions on this connection.

[0032] Furthermore, for ease of understanding, please refer to Figure 2 , Figure 2 This is a flowchart illustrating the patent matching method based on multi-agent collaboration provided in an embodiment of this application. This method can be implemented by a user terminal (e.g., the one described above). Figure 1 The operation can be performed by the user terminal shown, or by the user terminal and the service server (as described above). Figure 1 The business server 100 in the corresponding embodiment jointly executes the method. For ease of understanding, this embodiment uses the method executed by the aforementioned user terminal as an example for explanation. The patented matching method based on multi-agent collaboration may include at least the following steps S101-S104: S101, Obtain the query text to be matched and multiple candidate texts to be compared; Specifically, the user terminal obtains the query text to be matched and multiple candidate texts to be compared. It can be understood that the query text is the text that the user needs to match, and the candidate text is the text used to match the query text, that is, to select the text that meets the requirements from multiple candidate texts to match the query text. The aforementioned candidate texts can be texts that the user has pre-filtered according to conditions, or they can be randomly obtained from a large number of texts in the database. The query text and candidate texts can be pre-stored in the database by the user. For example, taking a patent document as an example, if the user needs to find a patent that matches the query patent from four candidate patents, and all four candidate patents and the query patent are stored in the database, then the user terminal can obtain the query patent to be matched and the four candidate patents to be compared from the database according to the acquisition instruction.

[0033] S102, Based on the perception agent, the query text and the candidate text are standardized to obtain standardized structural information; Specifically, the user terminal uses a perceptual agent to standardize the query text and the candidate text to obtain standardized structural information. It is understood that the perceptual agent is a lightweight, large-language-model-driven model used to convert text information without a fixed format into standardized structured information. This structured information can include multiple information blocks, such as entity information, classification information, functional information, etc. The following explanation uses a patent document as an example. For any given patent document, the perceptual agent can convert it into standardized structured information in the format "An entity for [purpose] (Classification number: [XX])". This structural information includes the patent's purpose, entity, and classification number information. Using a perceptual agent for standardization can filter redundant expressions, non-core technical details, and format noise in the original patent text, unifying the semantic expression paradigm of the patent text. This can effectively align the macro-technical fields of the candidate space and improve the signal-to-noise ratio of the information input.

[0034] S103, Based on the standardized structural information of the query text and the candidate text, a matching text set is obtained by matching the query text and the candidate text using the reasoning agent; the matching text set includes at least one candidate text. Specifically, the user terminal uses an inference agent to match the standardized structural information of the query text and candidate texts to obtain a set of matching texts. It is understood that the set of matching texts includes at least one candidate text. The inference agent is a pre-trained lightweight large language model that completes the matching of query text and candidate texts through an inference mechanism. Specifically, the inference agent can calculate the similarity between the standardized information of the query patent and candidate patents, determine the matching result based on the similarity, and generate the corresponding set of matching texts. The similarity calculation can use cosine similarity. The following example uses a query patent and four candidate patents A, B, C, and D. The query patent and candidate patents obtain standardized structural information through a perception agent, specifically including information in two dimensions: entity and function. The inference agent performs deduction based on the entity and function, completing a preliminary matching assessment between the query patent and each candidate patent. It is confirmed that the entities and uses of candidates C and D are significantly different from the query patent and can be directly excluded. The entities of candidates A and B are consistent with the query patent, so A and B are potential matches and cannot be definitively distinguished using existing structured information. Therefore, candidate patents A and B are determined as the set of matching texts.

[0035] S104, based on the reasoning agent, obtain the orientation information of the candidate texts in the matching text set, and determine the matching text corresponding to the query text from the matching text set according to the orientation information of the candidate texts.

[0036] Specifically, the user terminal obtains the orientation information of candidate texts in the matching text set based on the inference agent. Based on the orientation information of the candidate texts, it determines the matching text corresponding to the query text from the matching text set. It can be understood that the orientation information is a file containing more details corresponding to the candidate text and the query text. For example, the orientation information corresponding to a patent document could be the original patent abstract or the patent claims. The orientation file can be stored at the same address as the candidate file and the query file, for example, it can be stored in a database. After obtaining the orientation information, the inference agent can calculate the similarity between the orientation information of the query patent and the candidate patent, and further determine the matching result corresponding to the query text from the matching text set based on the aforementioned similarity. When the reasoning agent acquires targeted information, it only acquires the minimum necessary information and never indiscriminately requests the full text. For example, in the instance of step S103, the reasoning agent only needs to distinguish the technical differences between candidate patents A and B. Therefore, when acquiring targeted information, it only acquires the original summaries of candidate A and candidate B, without acquiring any information about the query patent, candidate C, or D, thus minimizing semantic noise caused by irrelevant text. The reasoning agent retrieves the original summaries of candidate patents A and B from the patent database, completing a single environmental information interaction. The reasoning agent seamlessly integrates the newly acquired original summaries into the current reasoning context, re-matches, determines the matching text corresponding to the query text from the matching text set, and finally confirms that the core technical solution of candidate A completely matches the entity and function of the query patent, while the core technical use of candidate B does not match. The final decision result is "A", and patent A is the matching result of the query patent.

[0037] In this embodiment, by acquiring the query text to be matched and multiple candidate texts to be compared, the query text and candidate texts are further standardized using a perceptual agent to obtain standardized structural information. This standardization process filters out redundant expressions, non-core technical details, and format noise in the original text, unifying the semantic expression paradigm of the patent text. Further, a matching text set is obtained by matching the standardized structural information of the query text and candidate texts using a reasoning agent; the matching text set includes at least one candidate text. Further, the orientation information of the candidate texts in the matching text set is acquired using the reasoning agent, and the matching text corresponding to the query text is determined from the matching text set based on the orientation information of the candidate texts. This method achieves text acquisition-matching-acquiring-re-matching, which not only reduces redundancy through structured information processing but also allows for re-matching using more detailed orientation information, improving the efficiency and accuracy of text matching.

[0038] Please see Figure 3 , Figure 3This is a flowchart illustrating the patent matching method based on multi-agent collaboration provided in an embodiment of this application. This method can be implemented by a user terminal (e.g., the one described above). Figure 1 The operation can be performed by the user terminal shown, or by the user terminal and the service server (as described above). Figure 1 The business server 100 in the corresponding embodiment jointly executes the method. For ease of understanding, this embodiment uses the method executed by the aforementioned user terminal as an example for explanation. The patented matching method based on multi-agent collaboration may include at least the following steps S201-S205: S201, Train and generate an inference agent based on the training data set; Specifically, the user terminal acquires the inference agent to be trained, and obtains a training database based on the expert agent. The training database includes multiple training data sets, each containing one query training data set, one positive sample, and three negative samples. Each training data set carries an inference logic chain, which is the inference process by which the inference agent matches the query training data and samples. The generation of the training data sets is explained below; the training data consists of patent texts. Specifically, 1500 query patents are randomly selected from the database. A dense retrieval model (BGE-base) is used to calculate the similarity between the query patent and all samples in the sample database. The sample with the highest similarity is taken as the positive sample, and its similarity is denoted as . Given a positive sample, based on the calculated similarity results, the similarity between the sample and the queried patent is obtained for all remaining samples in the sample database. And based on the relative semantic gap (relative semantic gap) Define three discrete negative sample pools: the high-difficulty negative sample pool ( The negative sample pool (with minimal semantic difference between positive and negative samples, similar techniques, and the highest ambiguity) is of medium difficulty. The semantic difference between the samples and positive samples is moderate, and there is a relatively clear technical boundary distinction; a low-difficulty negative sample pool ( (The semantic difference between the positive and negative samples is relatively large, with significant technical differentiation and the lowest ambiguity). By controlling the composition ratio of negative samples, we divide the data into three difficulty levels: A, B, and C. Each level contains one positive sample and three negative samples. The specific rules are as follows: Difficulty A (high ambiguity, requiring additional information) requires sampling one sample from each of the high-difficulty, medium-difficulty, and low-difficulty negative sample pools, while satisfying a double constraint—① the relative semantic difference between high-difficulty and medium-difficulty negative samples. ② The relative semantic gap between medium-difficulty negative samples and low-difficulty negative samples If multiple sampling attempts fail to meet the constraints, the difficulty level is automatically downgraded to B. Difficulty B (boundary game, requiring additional information as needed) requires sampling one sample from the high-difficulty negative sample pool and two samples from the low-difficulty negative sample pool. If the high-difficulty negative sample pool has no valid samples or sampling fails, the difficulty level is downgraded to C. Difficulty C (low ambiguity, no additional information required) requires sampling one sample from the medium-difficulty negative sample pool and two samples from the low-difficulty negative sample pool. If sampling fails, the sample for that query patent is marked as invalid and removed. Finally, the candidate lists of one positive sample and three negative samples generated for each query patent are randomly shuffled to eliminate positional bias caused by sorting, ultimately forming the data training set.

[0039] Furthermore, the user terminal standardizes the data in the training set based on the perceptual agent to obtain training standardization structure information. Based on the inference agent to be trained, it generates the training matching result of the query training data and the decision logic chain for generating the training matching result according to the standardization structure information. The decision logic chain is the inference process of matching training data during training. Based on the training matching result, the inference logic chain and the decision logic chain, a loss function of the inference agent to be trained is generated. The parameters in the inference agent to be trained are adjusted according to the loss function. The loss function mentioned above consists of two parts: fine-grained dynamic weighted cross-entropy and decision boundary alignment loss. The specific calculation process is as follows.

[0040] The decision logic chain carries the total length of time steps and the token corresponding to each time step. The weight of each token is determined based on the reasoning logic chain and the decision logic chain. A first loss function is generated based on the total length of the time steps, the token corresponding to each time step, and the weight of each token; The first loss function is the fine-grained dynamic weighted cross-entropy, calculated as follows:

[0041] in, For fine-grained dynamic weighted cross-entropy, Represents the total length of the time step. Representing the Tokens generated at each time step Represents all context prefixes preceding this token. Representative parameters are The predicted probability of the model. This represents the dynamic weight assigned based on the specific responsibilities of the current token. It includes four levels and sets differentiated values: basic structure label weight ( (corresponding to format tags such as analysis, acquisition, and results), the default value is set to 0.5; internal thought chain weight ( (corresponding to the deduction process text), the default value is set to 1.2, preserving the focus on the reasoning logic; routing action weight ( (Select either get or result), the default setting is 2.0, which enhances the learning of decision-making actions; and the target execution weight ( (corresponding to specific acquisition targets such as matching results or targeted data), the default value is set to 3.0, which serves as the highest priority weight to ensure accurate fitting of core execution actions.

[0042] Furthermore, the inference logic chain includes the output probabilities of multiple output results. Based on the matching training results, the output probability of the correct result among the multiple output results is determined as the first output probability, and the output probability of the output result with the second highest output probability is determined as the second output probability. A second loss function is generated based on the first output probability, the second output probability, and the confidence boundary. The second loss function is the decision boundary alignment loss, calculated as follows:

[0043] in, To align the loss to the decision boundary, This indicates the confidence level boundary, which defaults to 1.0. The probability of the correct result must be at least ahead of the second-highest output result. Only then can they be exempted from punishment.

[0044] Furthermore, a loss function for the inference agent to be trained is determined based on the first loss and the second loss, and the parameters in the inference agent to be trained are adjusted according to the loss function.

[0045]

[0046] in, For loss function, This represents the weighting coefficient used to balance the fine-grained dynamic weighted cross-entropy loss with the decision boundary alignment loss; the default value is 1.0.

[0047] When the loss function converges, the adjusted inference agent to be trained is determined as the inference agent, and the parameters in the inference agent are the adjusted parameters.

[0048] S202, Obtain the query text to be matched and multiple candidate texts to be compared; S203, Based on the perception agent, the query text and the candidate text are standardized to obtain standardized structural information; Specifically, the user terminal performs standardized processing on the query text and the candidate text based on the perceptual agent to obtain standardized structural information. This means the user terminal obtains keywords from the standardized processing of the perceptual agent, and this standardization is used to obtain standardized structural information. The structured information includes entity information, usage information, and classification information; each of these has corresponding keywords. Further, the perceptual agent is invoked to collect information from the query text using these keywords, obtaining the entity information, usage information, and classification information corresponding to the query text. Similarly, the perceptual agent is invoked to collect information from the candidate text using these keywords, obtaining the entity information, usage information, and classification information corresponding to the candidate text. For example, for a patent document, the perceptual agent can convert it into standardized structured information in the format "An entity for [purpose] (Classification number: [XX])", where the structured information includes entity information, usage information, and classification information.

[0049] S204, Based on the reasoning agent, the standardized structural information of the query text and the candidate text is matched to obtain a set of matching texts; Specifically, the user terminal uses an inference agent to match the standardized structural information of the query text and candidate texts to obtain a set of matched texts. This means that the inference agent calculates the similarity between the standardized structural information of the query text and the target candidate text; the target candidate text is any candidate text to be compared; the standardized structural information includes entity information, usage information, and classification information; the similarity between the query text and each candidate text is calculated, and when the similarity of the entity information, usage information, and classification information of the query text and the target candidate text is all greater than a first similarity threshold, the target candidate text is added to the set of matched texts. The first similarity threshold can be preset.

[0050] It should be noted that if the number of texts in the matched text set is 1, then the candidate texts in the matched text set are determined as the matching texts corresponding to the query text; that is, the subsequent process of obtaining targeted information for matching is not performed, and the matching file is directly determined. If the number of texts in the matched text set is greater than 1, then the step of obtaining the targeted information of the candidate texts in the matched text set and determining the matching texts corresponding to the query text from the matched text set based on the targeted information is performed.

[0051] S205, based on the reasoning agent, obtain the orientation information of the candidate texts in the matching text set, and determine the matching text corresponding to the query text from the matching text set according to the orientation information of the candidate texts.

[0052] Specifically, the user terminal invokes the reasoning agent to obtain the orientation information of the first candidate text from the patent database, and adds the orientation information to the context of the first candidate text to obtain candidate context information; the first candidate text is any candidate text in the matching text set; the reasoning agent calculates the similarity between the first candidate context information and the context information of the query text; that is, it calculates the similarity between the context information of each candidate text and the context information of the query text. When the similarity between the first candidate context information and the context information of the query text is greater than a second similarity threshold, the first candidate text is determined as the matching text corresponding to the query text. The second similarity threshold can be preset. When the similarity between the first candidate context information and the context information of the query text is in an ambiguous state, the aforementioned ambiguous state is the state in which the matching text cannot be determined based on the orientation information, which usually means that the orientation information is not accurate enough, or the candidate text similarity is high. Furthermore, the reasoning agent is invoked to retrieve external retrieval enhancement information of the first candidate text from the patent database, and this external retrieval enhancement information is added to the context of the first candidate text to obtain enhanced candidate context information. The aforementioned external enhancement information is textual information with a high degree of correlation to the directional information; for example, the directional information of the candidate text is the corresponding original summary, and the external enhancement information can be background technology associated with the summary information. When the similarity between the enhanced candidate context information and the context information of the query text is greater than a third similarity threshold, the first candidate text is determined as the matching text corresponding to the query text.

[0053] Wherein, step S202 of the embodiment of the present invention is referred to Figure 2 The specific description of step S101 in the illustrated embodiment will not be repeated here.

[0054] In this embodiment, by acquiring the query text to be matched and multiple candidate texts to be compared, the query text and candidate texts are further standardized using a perceptual agent to obtain standardized structural information. This standardization process filters out redundant expressions, non-core technical details, and format noise in the original text, unifying the semantic expression paradigm of the patent text. Further, a matching text set is obtained by matching the standardized structural information of the query text and candidate texts using a reasoning agent; the matching text set includes at least one candidate text. Further, the orientation information of the candidate texts in the matching text set is acquired using the reasoning agent, and the matching text corresponding to the query text is determined from the matching text set based on the orientation information of the candidate texts. This method achieves text acquisition-matching-acquiring-re-matching, which not only reduces redundancy through structured information processing but also allows for re-matching using more detailed orientation information, improving the efficiency and accuracy of text matching.

[0055] The following will combine Figure 4 The following describes specific implementation scenarios provided in the embodiments of this application, such as... Figure 4 As shown.

[0056] Regarding text matching technology, this application employs a multi-agent collaborative method (DART) for action-aware reasoning. Specifically, it uses one original query patent text to be matched and four original candidate patent texts to be compared. The user terminal responds to the command, retrieving one original query patent text and four original candidate patent texts (A, B, C, and D) from the database. It then invokes a perceptual agent driven by a lightweight large language model to uniformly transform the unformatted original patent text into standardized structural information in the format "An entity for [purpose] (Classification XX)". Simultaneously, it outputs corresponding structured information for the four candidate patents. Through standardization, the perceptual agent filters redundant expressions, non-core technical details, and format noise from the original patent text, unifying the semantic expression paradigm of the patent text. This effectively aligns the macro-technical fields of the candidate space and improves the signal-to-noise ratio of the input information.

[0057] Subsequently, the reasoning agent queries the standardized structured information of the patent and four candidate patents, without additional full text details. In each round of interaction, the reasoning agent first enters the matching phase, generating a complete internal thought chain. Based on the two core dimensions of entity and function, it completes a preliminary matching degree assessment between the query patent and each candidate patent, identifying irrelevant candidates that can be directly eliminated and potential candidates with matching possibilities. For example, if the reasoning agent completes deduction in the matching phase and confirms that the core entities and uses of candidates C and D are significantly different from the query patent and can be directly eliminated, while the core entities of candidates A and B are consistent with the query patent, then A and B are potential matches and cannot be definitively distinguished using the existing structured information, the reasoning agent then enters the precise targeted information retrieval phase. The reasoning agent initiates targeted information requests, obtaining only the minimum necessary information to resolve the current reasoning ambiguity, and never indiscriminately retrieving the full text. Its action space pre-sets three types of targeted information targets: the original abstract of the query patent, the original abstract of the specified candidate patent (candidate A / B / C / D), and the original abstract of the specified candidate patent (candidate A / B / C / D). The inference agent retrieves abstracts and external retrieval augmentation information (RAG) only for specific obfuscated objects, obtaining their original technical details. In this example, the inference agent only needs to distinguish the technical differences between candidate A and candidate B. Therefore, during the targeted information retrieval phase, it only requests the original abstracts of candidate A and candidate B, without requesting any information about the queried patent, candidates C and D, or triggering external RAGs. The retrieval process minimizes semantic noise from irrelevant text. Upon receiving a retrieval command, the system retrieves the original technical details of the corresponding target from the patent database and returns them to the reasoning agent, completing a single environmental information interaction. After the target technical details are returned, the reasoning agent seamlessly integrates the newly acquired information into the current reasoning context and re-enters the analysis and matching phase. It then performs a secondary evaluation and deep technical logic comparison of the overall evidence chain. This multi-agent dynamic closed loop of "perception-analysis-acquisition-reanalysis" will continue to operate in multiple rounds. Each round only supplements the minimum necessary information required to resolve the current ambiguity until the reasoning agent matches a decision that meets the requirements. Finally, the final matching result is output. In this example, after obtaining the original summaries of candidates A and B, the reasoning agent completes a deep comparison in the secondary analysis phase, confirming that the core technical solution of candidate A completely matches the core entity and function of the queried patent, while the core technical use of candidate B does not match. Finally, the final decision result is confirmed as "A", completing this patent matching reasoning.Please refer to Table 1, which compares the performance of our proposed method (DART) with three other methods (Chain-of-Thought, Retrieval-Augmented Generation, and MemGraph) on three datasets (GLM-4-9B-Chat, Qwen2.5-14B-Instruct, and Llama-3.1-8B-Instruct). The table shows that our proposed method, DART, comprehensively outperforms the current state-of-the-art method MemGraph and traditional CoT and RAG methods on four mainstream large-scale model bases, achieving a maximum accuracy improvement of 18.2 percentage points, effectively addressing the core technical shortcomings of existing methods.

[0058] Table 1. Performance Comparison of Various Methods for Patent Matching Datasets

[0059] Please see Figure 5 , Figure 5 This is a schematic diagram of a patent matching device based on multi-agent collaboration provided in an embodiment of this application. The patent matching device based on multi-agent collaboration can be a computer program (including program code) running on a computer device; for example, the patent matching device based on multi-agent collaboration is an application software. This device can be used to execute the corresponding steps in the method provided in the embodiments of this application. Figure 5 As shown, the patent matching device 1 based on multi-agent collaboration in this application embodiment may include: a text acquisition unit 11, a text processing unit 12, a first text matching unit 13, and a second text matching unit 14.

[0060] The text acquisition unit 11 is used to acquire the query text to be matched and multiple candidate texts to be compared; Text processing unit 12 is used to perform standardized processing on the query text and the candidate text based on the perceptual agent to obtain standardized structural information; The first text matching unit 13 is used to match the standardized structural information of the query text and the candidate text based on the reasoning agent to obtain a matching text set; the matching text set includes at least one candidate text. The second text matching unit 14 is used to obtain the orientation information of candidate texts in the matching text set based on the reasoning agent, and determine the matching text corresponding to the query text from the matching text set according to the orientation information of the candidate texts.

[0061] In one feasible implementation, a training unit is also included, for: Obtain the reasoning agent to be trained; The training database is obtained based on an expert intelligent agent. The training database includes multiple training data sets. Each training data set includes a query training data, a positive sample, and three negative samples. The training data set carries an inference logic chain. Based on the perceptual agent, the data in the training set is standardized to obtain the training standardized structure information; Based on the inference agent to be trained, the training matching result of the query training data is generated according to the standardized structural information, and the decision logic chain for generating the training matching result is also generated. A loss function is generated for the reasoning agent to be trained based on the training matching results, the reasoning logic chain, and the decision logic chain, and the parameters in the reasoning agent to be trained are adjusted based on the loss function. When the loss function converges, the adjusted inference agent to be trained is determined as the inference agent, and the parameters in the inference agent are the adjusted parameters.

[0062] In one feasible implementation, the training unit is used for: The decision logic chain carries the total length of time steps and the token corresponding to each time step. The weight of each token is determined based on the reasoning logic chain and the decision logic chain. A first loss function is generated based on the total length of the time steps, the token corresponding to each time step, and the weight of each token; The reasoning logic chain includes the output probabilities of multiple output results. Based on the matching training results, the output probability of the correct result among the multiple output results is determined as the first output probability, and the output probability of the output result with the second highest output probability is determined as the second output probability. A second loss function is generated based on the first output probability, the second output probability, and the confidence boundary. The loss function of the inference agent to be trained is determined based on the first loss and the second loss, and the parameters in the inference agent to be trained are adjusted based on the loss function.

[0063] In one feasible implementation, the text processing unit is configured to: Keywords for the standardized processing of the perceptual agent are obtained, and the standardized processing is used to obtain standardized structural information; the structural information includes entity information, usage information, and classification information; the entity information, usage information, and classification information each have corresponding keywords; The perceptual agent is invoked to collect information from the query text using keywords, obtaining the entity information, usage information, and classification information corresponding to the query text. The perceptual agent is invoked to collect information about the candidate text using keywords, thereby obtaining the entity information, usage information, and classification information corresponding to the candidate text.

[0064] In one feasible implementation, a detection unit is also included, for: If the number of texts in the matching text set is 1, then the candidate texts in the matching text set are determined as the matching texts corresponding to the query text; If the number of texts in the matching text set is greater than 1, then the step of obtaining the orientation information of the candidate texts in the matching text set and determining the matching text corresponding to the query text from the matching text set based on the orientation information is executed.

[0065] In one feasible implementation, the first text matching unit is configured to: The similarity is calculated based on the standardized structural information of the query text and the target candidate text using a reasoning agent; the target candidate text is any candidate text to be compared; the standardized structural information includes entity information, usage information, and classification information. When the similarity of the entity information, purpose information, and classification information of the query text and the target candidate text is all greater than the first similarity threshold, the target candidate text is added to the matching text set.

[0066] In one feasible implementation, the second text matching unit is used for: The reasoning agent is invoked to obtain the orientation information of the first candidate text from the patent database, and the orientation information is added to the context of the first candidate text to obtain candidate context information; the first candidate text is any candidate text in the matching text set; The reasoning agent performs similarity calculations on the context information of the first candidate context information and the context information of the query information. When the similarity between the first candidate context information and the context information of the query text is greater than the second similarity threshold, the first candidate text is determined as the matching text corresponding to the query text.

[0067] In one feasible implementation, the second text matching unit is further configured to: When the similarity between the first candidate context information and the query context information is in a fuzzy state; The reasoning agent is invoked to obtain external retrieval enhancement information of the first candidate text from the patent database, and the external retrieval enhancement information is added to the context of the first candidate text to obtain enhanced candidate context information; When the similarity between the enhanced candidate context information and the context information of the query text is greater than a third similarity threshold, the first candidate text is determined as the matching text corresponding to the query text.

[0068] In this embodiment, by acquiring the query text to be matched and multiple candidate texts to be compared, the query text and candidate texts are further standardized using a perceptual agent to obtain standardized structural information. This standardization process filters out redundant expressions, non-core technical details, and format noise in the original text, unifying the semantic expression paradigm of the patent text. Further, a matching text set is obtained by matching the standardized structural information of the query text and candidate texts using a reasoning agent; the matching text set includes at least one candidate text. Further, the orientation information of the candidate texts in the matching text set is acquired using the reasoning agent, and the matching text corresponding to the query text is determined from the matching text set based on the orientation information of the candidate texts. This method achieves text acquisition-matching-acquiring-re-matching, which not only reduces redundancy through structured information processing but also allows for re-matching using more detailed orientation information, improving the efficiency and accuracy of text matching.

[0069] Please see Figure 6 , Figure 6 This is a schematic diagram of the structure of a computer device provided in an embodiment of this application. Figure 6 As shown, the computer device 1000 may include: at least one processor 1001, such as a CPU; at least one network interface 1004; a user interface 1003; a memory 1005; and at least one communication bus 1002. The communication bus 1002 is used to enable communication between these components. The user interface 1003 may include a display screen, and optionally, it may also include a standard wired interface or a wireless interface. The network interface 1004 may optionally include a standard wired interface or a wireless interface (such as a Wi-Fi interface). The memory 1005 may be random access memory (RAM) or non-volatile memory (NVM), such as at least one disk storage device. Optionally, the memory 1005 may also be at least one storage device located remotely from the aforementioned processor 1001. Figure 6 As shown, the memory 1005, which serves as a computer storage medium, may include an operating system, a network communication module, a user interface module, and application programs.

[0070] exist Figure 6In the computer device 1000 shown, the network interface 1004 provides network communication functions, the user interface 1003 is mainly used to provide an input interface for the user, and the processor 1001 can be used to call the application programs stored in the memory 1005 to achieve the above. Figures 2-4 The description of the patent matching method based on multi-agent collaboration in any of the corresponding embodiments will not be repeated here.

[0071] It should be understood that the computer device 1000 described in the embodiments of this application can execute the foregoing text. Figures 2-4 The description of the patent matching method based on multi-agent cooperation in any corresponding embodiment can also be performed as described above. Figure 5 The description of the patent matching device based on multi-agent cooperation in the corresponding embodiments will not be repeated here. Furthermore, the beneficial effects of using the same method will also not be repeated.

[0072] Furthermore, it should be noted that this application embodiment also provides a computer-readable storage medium, which stores a computer program executed by the aforementioned multi-agent cooperative patent matching device, and the computer program includes program instructions. When the processor executes the program instructions, it can execute the aforementioned... Figures 2-4 The description of the patent matching method based on multi-agent collaboration in any corresponding embodiment is already provided and will not be repeated here. Furthermore, the beneficial effects of using the same method will also not be repeated. For technical details not disclosed in the computer-readable storage medium embodiments related to this application, please refer to the description of the method embodiments of this application. As an example, program instructions may be deployed to execute on a single computing device, or on multiple computing devices located in one location, or on multiple computing devices distributed across multiple locations and interconnected via a communication network. These multiple computing devices distributed across multiple locations and interconnected via a communication network can constitute a blockchain system.

[0073] Those skilled in the art will understand that all or part of the processes in the above embodiments can be implemented by a computer program instructing related hardware. The program can be stored in a computer-readable storage medium, and when executed, it can include the processes of the embodiments of the above methods. The computer-readable storage medium can be a patent matching device based on multi-agent collaboration provided in any of the foregoing embodiments, or an internal storage unit of the device, such as a hard drive or memory of an electronic device. The computer-readable storage medium can also be an external storage device of the electronic device, such as a plug-in hard drive, smart media card (SMC), secure digital (SD) card, flash card, etc., equipped on the electronic device. The computer-readable storage medium can also include magnetic disks, optical disks, read-only memory (ROM), or random access memory, etc. Furthermore, the computer-readable storage medium can include both internal storage units and external storage devices of the electronic device. The computer-readable storage medium is used to store the computer program and other programs and quantities required by the electronic device. The computer-readable storage medium can also be used to temporarily store data that has been output or will be output.

[0074] The terms "first," "second," etc., used in the claims, description, and drawings of this invention are used to distinguish different objects, not to describe a particular order. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover non-exclusive inclusion. For example, a process, method, system, product, or apparatus that comprises a series of steps or units is not limited to the listed steps or units, but may optionally include steps or units not listed, or may optionally include other steps or units inherent to these processes, methods, products, or apparatuses. References to "embodiment" herein mean that a particular feature, structure, or characteristic described in connection with an embodiment may be included in at least one embodiment of the invention. The presentation of this phrase in various places in the specification does not necessarily refer to the same embodiment, nor is it a separate or alternative embodiment mutually exclusive with other embodiments. It will be explicitly and implicitly understood by those skilled in the art that the embodiments described herein can be combined with other embodiments. The term "and / or" as used in this specification and the appended claims refers to any combination of one or more of the associated listed items and all possible combinations, and includes such combinations.

[0075] Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, computer software, or a combination of both. To clearly illustrate the interchangeability of hardware and software, the components and steps of the various examples have been generally described in terms of functionality in the foregoing description. Those skilled in the art can implement the described functions using different methods for each specific application, but such implementations should not be considered beyond the scope of this invention.

[0076] In the various embodiments of this application, the functional units can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit.

[0077] The above-disclosed embodiments are merely preferred embodiments of this application and should not be construed as limiting the scope of this application. Therefore, any equivalent variations made in accordance with the claims of this application shall still fall within the scope of this application.

Claims

1. A patent matching method based on multi-agent cooperation, characterized in that, include: Retrieve the query text to be matched and multiple candidate texts to be compared; Based on the perceptual intelligent agent, the query text and the candidate text are standardized to obtain standardized structural information; A set of matched texts is obtained by matching the standardized structural information of the query text and the candidate texts based on the reasoning agent; the set of matched texts includes at least one candidate text. Based on the reasoning agent, the orientation information of candidate texts in the matching text set is obtained, and the matching text corresponding to the query text is determined from the matching text set according to the orientation information of the candidate texts.

2. The method according to claim 1, characterized in that, Also includes: Obtain the reasoning agent to be trained; The training database is obtained based on an expert intelligent agent. The training database includes multiple training data sets, each of which includes a query training data, a positive sample, and three negative samples. The training data set carries an inference logic chain. Based on the perceptual agent, the data in the training dataset is standardized to obtain the training standardized structure information; Based on the inference agent to be trained, the training matching result of the query training data is generated according to the standardized structural information, and the decision logic chain for generating the training matching result is also generated. A loss function is generated for the reasoning agent to be trained based on the training matching results, the reasoning logic chain, and the decision logic chain, and the parameters in the reasoning agent to be trained are adjusted based on the loss function. When the loss function converges, the adjusted inference agent to be trained is determined as the inference agent, and the parameters in the inference agent are the adjusted parameters.

3. The method according to claim 2, characterized in that, The step of generating a loss function for the reasoning agent to be trained based on the training matching results, the inference logic chain, and the decision logic chain, and adjusting the parameters in the reasoning agent to be trained based on the loss function, includes: The decision logic chain carries the total length of time steps and the token corresponding to each time step. The weight of each token is determined based on the reasoning logic chain and the decision logic chain. A first loss function is generated based on the total length of the time steps, the token corresponding to each time step, and the weight of each token; The reasoning logic chain includes the output probabilities of multiple output results. Based on the matching training results, the output probability of the correct result among the multiple output results is determined as the first output probability, and the output probability of the output result with the second highest output probability is determined as the second output probability. A second loss function is generated based on the first output probability, the second output probability, and the confidence boundary. The loss function of the inference agent to be trained is determined based on the first loss and the second loss, and the parameters in the inference agent to be trained are adjusted based on the loss function.

4. The method according to claim 1, characterized in that, The standardization process based on the perceptual agent to obtain standardized structural information from the query text and the candidate text includes: Keywords for the standardized processing of a perceptual agent are obtained, and the standardized processing is used to obtain standardized structural information; the standardized structural information includes entity information, usage information, and classification information; the entity information, usage information, and classification information each have corresponding keywords; The perceptual agent is invoked to collect information from the query text using keywords, obtaining the entity information, usage information, and classification information corresponding to the query text. The perceptual agent is invoked to collect information about the candidate text using keywords, thereby obtaining the entity information, usage information, and classification information corresponding to the candidate text.

5. The method according to claim 1, characterized in that, The matching text set obtained by matching the standardized structural information of the query text and candidate texts based on the reasoning agent includes: The similarity is calculated based on the standardized structural information of the query text and the target candidate text using a reasoning agent; the target candidate text is any candidate text to be compared; the standardized structural information includes entity information, usage information, and classification information. When the similarity of the entity information, purpose information, and classification information of the query text and the target candidate text is all greater than the first similarity threshold, the target candidate text is added to the matching text set.

6. The method according to claim 1, characterized in that, The step of obtaining the orientation information of candidate texts in the matching text set based on the reasoning agent, and determining the matching text corresponding to the query text from the matching text set based on the orientation information of the candidate texts, includes: The reasoning agent is invoked to obtain the orientation information of the first candidate text from the patent database, and the orientation information is added to the context of the first candidate text to obtain candidate context information; the first candidate text is any candidate text in the matching text set; The reasoning agent calculates the similarity between the first candidate context information and the context information of the query text. When the similarity between the first candidate context information and the context information of the query text is greater than the second similarity threshold, the first candidate text is determined as the matching text corresponding to the query text.

7. The method according to claim 6, characterized in that, Also includes: When the similarity between the first candidate context information and the context information of the query text is in an ambiguous state; The reasoning agent is invoked to obtain external retrieval enhancement information of the first candidate text from the patent database, and the external retrieval enhancement information is added to the context of the first candidate text to obtain enhanced candidate context information; When the similarity between the enhanced candidate context information and the context information of the query text is greater than a third similarity threshold, the first candidate text is determined as the matching text corresponding to the query text.

8. A patent matching device based on multi-agent collaboration, characterized in that, include: The text acquisition unit is used to acquire the query text to be matched and multiple candidate texts to be compared; The text processing unit is used to perform standardized processing on the query text and the candidate text based on the perceptual agent to obtain standardized structural information; The first text matching unit is used to match the standardized structural information of the query text and the candidate text based on the reasoning agent to obtain a set of matching texts; the set of matching texts includes at least one candidate text. The second text matching unit is used to obtain the orientation information of candidate texts in the matching text set based on the reasoning agent, and determine the matching text corresponding to the query text from the matching text set according to the orientation information of the candidate texts.

9. A computer device, characterized in that, include: Processor, memory, and network interface; The processor is connected to the memory and the network interface, wherein the network interface is used to provide network communication functions, the memory is used to store program code, and the processor is used to call the program code to execute the method according to any one of claims 1-7.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program adapted to be loaded by a processor and to execute the method of any one of claims 1-7.

Images