Construction method of patent-specific vector model and patent semantic retrieval method
By fine-tuning the open-source BGE-M3 vector model through multiple rounds and extracting 12-dimensional technical features, a patent-specific vector model was constructed, which solved the problem of low recall in patent retrieval, achieved efficient and accurate patent information retrieval, and improved recall rate and information utilization efficiency.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- TONGFANG KNOWLEDGE DIGITAL PUBLISHING TECH CO LTD
- Filing Date
- 2025-12-29
- Publication Date
- 2026-04-28
AI Technical Summary
Existing patent search technologies suffer from low recall rates. Traditional search methods are susceptible to ambiguity in technical terminology, leading to the omission of relevant patents and insufficient recall, which severely restricts the efficiency and depth of patent information utilization.
By fine-tuning the open-source BGE-M3 vector model in multiple rounds of similar patent retrieval, patent context, and data format adaptation categories, a patent-specific vector model is constructed. Furthermore, 12-dimensional technical features are extracted and vectors are transformed from patent texts. Combined with multi-dimensional vector similarity retrieval and weighted ranking, the patent semantic retrieval method is optimized.
It significantly improved the recall rate of patent semantic retrieval, with the TOP50 recall rate increasing by 41.35%, the TOP100 recall rate increasing by 41.91%, and the TOP400 recall rate increasing by 33.82%. Users can retrieve the information they need from massive amounts of data more quickly and accurately, improving work efficiency and data utilization.
Smart Images
Figure CN121935359A_ABST
Abstract
Description
Technical Field
[0001] This application belongs to the field of patent search technology, specifically relating to a method for constructing a patent-specific vector model and a patent semantic search method. Background Technology
[0002] Patent searching is a crucial method for obtaining patent information, playing an indispensable role in scientific research, innovation, and corporate competition. With the rapid development of technology and the dramatic increase in the number of patents, how to efficiently and accurately retrieve the required information from massive amounts of patent data has become a pressing issue.
[0003] Current patent search technologies generally suffer from a core pain point: low recall rates. Traditional search methods are easily affected by ambiguities in technical terminology, leading to the omission of a large number of relevant patents, and insufficient recall rates are commonplace. Figure 1 The recall rates of the top 50 products based on semantic retrieval are below 24.83%, the recall rates of the top 100 products are below 29.19%, and the recall rates of the top 400 products are below 37.44%, which seriously restricts the efficiency and depth of patent information utilization. Summary of the Invention
[0004] In view of this, this application provides a method for constructing a patent-specific vector model and a patent semantic retrieval method. By fine-tuning the open-source BGE-M3 vector model for similar patent retrieval, and then fine-tuning the fine-tuned vector model for patent context, highly similar patent retrieval, and data format adaptation, a patent-specific vector model is obtained. Patent semantic retrieval is then performed based on the patent-specific vector model to solve the technical problem of low efficiency and depth in the utilization of existing patent information.
[0005] The first aspect of this application provides a method for constructing a patent-specific vector model, which includes: Step S1: Using the first batch of batch cases and rejected EXY evidence pairs as the first training set, and using the loss function as the fine-tuning algorithm, the open-source BGE-M3 vector model is fine-tuned for similar patent retrieval to obtain the fine-tuned vector model.
[0006] Step S2: Randomly extract millions of uniform, unlabeled patent training corpora from multiple dimensions as the second training set. Use a loss function as a fine-tuning algorithm to fine-tune the fine-tuned vector model using patent context, obtaining a patent-specific vector base model. Multiple dimensions include the receiving office, IPC classification, patent type, case evidence ratio, and legal status.
[0007] Step S3: Using the second batch of batch cases and EX evidence pairs as the third training set, and employing a loss function as the fine-tuning algorithm, the patent-specific vector basic model is fine-tuned for high-similarity patent retrieval to obtain the patent-specific vector model. The number of EX evidence pairs in the batch is greater than the number of EXY evidence pairs in the batch of rejections.
[0008] Step S4: Use the multidimensional random sampling of millions of uniform unlabeled patents in the format of other suppliers in the second batch as the fourth training set, and perform data format adaptation fine-tuning on the patent-specific vector model to obtain the final patent-specific vector model.
[0009] A second aspect of this application provides a patent semantic retrieval method, which includes: Step S10: Extract 12-dimensional technical feature text from the patent text database and clean it to construct 12 technical dimensions. The 12-dimensional technical feature text includes title, abstract, research field, claims, technical background, technical problem, invention purpose, invention content, embodiments, technical efficacy, description of drawings, and technical effects.
[0010] Step S20: Convert the 12-dimensional technical feature text into vectors using the final patent-specific vector model. The final patent-specific vector model is obtained using a patent-specific vector model construction method according to the first aspect of this application.
[0011] Step S30: Convert the input text to be searched or the independent claims of the patent corresponding to the patent number to be searched into vectors using the final patent-specific vector model, perform vector similarity search in the vector field of the 12-dimensional technical feature text, and output the results in a unique highest similarity sorted by all publication numbers.
[0012] Step S40: Based on the contribution rate of a single dimension to the total recall rate, calculate the recall rate weight of each dimension, and select the top 8 technical feature texts with the highest contribution rates from the 12-dimensional technical feature texts. The 8-dimensional technical feature texts include title, abstract, research field, claims, research question, invention content, embodiments, invention purpose and technical effect.
[0013] Step S50: Convert the input text to be searched or the independent claims of the patent corresponding to the patent number to be searched into vectors using the final patent-specific vector model, perform vector similarity search in 8 technical dimension vector fields, and output the results according to the weighted similarity of the unique highest corresponding dimension weight among all publication numbers.
[0014] In one specific implementation of this application, step S20 includes: Step S21: Use the final patent-specific vector model to convert the short text technical features in the 12-dimensional technical feature text into single-value vectors. The short text technical features include the title, abstract, and research field.
[0015] Step S22: Using the final patent-specific vector model, the long text technical features in the 12-dimensional technical feature text are sliced and converted into multi-value vectors. The long text technical features include claims, technical background, technical problem, invention purpose, invention content, embodiments, technical effects, description of drawings, and technical results.
[0016] In one specific implementation of this application, step S30 includes: Step S31: Convert the input text to be searched or the independent claims of the patent corresponding to the patent number to be searched into vectors using the final patent-specific vector model.
[0017] Step S32: In the vector fields of the 12-dimensional technical feature text, first retrieve the top N vectors with the highest similarity from each vector field, then merge the 12-dimensional search results according to the publication number. If multiple vectors are detected for the same publication number, take the highest similarity value, and then sort and output them according to the unique highest similarity of all publication numbers.
[0018] In one specific implementation of this application, step S50 includes: Step S51: Convert the input text to be searched or the independent claims of the patent corresponding to the patent number to be searched into vectors using the final patent-specific vector model.
[0019] Step S52: In the vector field of the 8-dimensional technical feature text, first retrieve the top N vectors with the highest similarity from each vector field, then merge the search results of the 8 dimensions according to the publication number. If multiple vectors are retrieved for the same publication number, take the highest similarity value after weighting the corresponding dimension weights, and then sort and output according to the unique highest similarity of all publication numbers.
[0020] A third aspect of this application provides a computer apparatus comprising a processor and a memory. The processor is used to execute a method for constructing a patent-specific vector model according to a first aspect of this application, and / or a method for patent semantic retrieval according to a second aspect of this application. The memory is used to store executable instructions of the processor.
[0021] The fourth aspect of this application provides a computer-readable storage medium storing executable instructions for a computer. When executed by a processor, the executable instructions implement a method for constructing a patent-specific vector model according to the first aspect of this application, and / or a patent semantic retrieval method according to the second aspect of this application.
[0022] The fifth aspect of this application provides a computer program product, including a computer program / instruction, which, when executed by a processor, implements a method for constructing a patent-specific vector model according to the first aspect of this application, and / or a patent semantic retrieval method according to the second aspect of this application.
[0023] The beneficial effects of the technical solution in this application are as follows: Experimental verification shows that the patent semantic retrieval method based on the final patent-specific vector model constructed in this application significantly improves the recall rate for data retrieval, achieving a TOP50 recall rate of 76.33%, a TOP100 recall rate of 86.08%, and a TOP400 recall rate of 100%. Compared to using the open-source BGE-M3 vector model, using the final dedicated vector model further improves the TOP50 recall rate by 41.35%, the TOP100 recall rate by 41.91%, and the TOP400 recall rate by 33.82%. Therefore, this method not only significantly improves the recall rate of the final patent-specific vector model at different TOPK values but also further verifies the effectiveness of the multi-round fine-tuning strategy for optimizing vector model performance. In practical applications, this means that users can retrieve the information they need from massive amounts of data more quickly and accurately, greatly improving work efficiency and data utilization. Furthermore, this multi-round fine-tuning method and multi-dimensional vector weighted retrieval strategy have good versatility and scalability, and can be applied to vector model optimization needs in different fields and scenarios, as well as diverse retrieval scenarios such as patent examination, patent novelty search, patent invalidation, and infringement prevention. Attached Figure Description
[0024] Figure 1 The diagram shows a comparison of the test results for the recall rates of various existing products.
[0025] Figure 2 The diagram shown is a flowchart illustrating a method for constructing a patent-specific vector model according to an embodiment of this application.
[0026] Figure 3 The diagram shown is a schematic flowchart of a patent semantic retrieval method provided in an embodiment of this application.
[0027] Figure 4 The diagram shows a comparison of the test results of the patent-specific vector model of this application and the recall rates of existing products.
[0028] Figure 5 The diagram shows the stability test results of the fine-tuned vector model. Detailed Implementation
[0029] 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.
[0030] At least one embodiment of this application provides a method for constructing a patent-specific vector model, referencing Figure 2 The method for constructing the patent-specific vector model includes the following steps.
[0031] Step S1: Using the first batch of batch cases and rejected EXY evidence pairs as the first training set, and using the loss function as the fine-tuning algorithm, the open-source BGE-M3 vector model is fine-tuned for similar patent retrieval to obtain the fine-tuned vector model.
[0032] For example, a batch of rejected EXY evidence could be 30,000 pieces of rejected EXY evidence.
[0033] The loss function uses the BGM-M3 open-source loss element: Where p∗ and P′ represent the positive and negative samples corresponding to query q, respectively; s(∙) is {S dense (∙), S lex (∙),S mul Any function in (∙)}.
[0034] Step S2: Randomly extract millions of uniform, unlabeled patent training corpora from multiple dimensions as the second training set. Use a loss function as a fine-tuning algorithm to fine-tune the fine-tuned vector model using patent context, obtaining a patent-specific vector base model. Multiple dimensions include the receiving office, IPC classification, patent type, case evidence ratio, and legal status.
[0035] For example, a uniform unlabeled patent training corpus could consist of 2 million patent data points. This uniform unlabeled patent training corpus can be automatically generated.
[0036] Step S3: Using the second batch of batch cases and EX evidence pairs as the third training set, and employing a loss function as the fine-tuning algorithm, the patent-specific vector basic model is fine-tuned for high-similarity patent retrieval to obtain the patent-specific vector model. The number of EX evidence pairs in the batch is greater than the number of EXY evidence pairs in the batch of rejections.
[0037] For example, a batch of EX evidence could be 300,000 EX patent citation evidence pairs.
[0038] Step S4: Use the multidimensional random sampling of millions of uniform unlabeled patents in the format of other suppliers in the second batch as the fourth training set, and perform data format adaptation fine-tuning on the patent-specific vector model to obtain the final patent-specific vector model.
[0039] For example, the second batch of batch cases used 1.2 million new training samples (Law & Business data; regularity rule optimization, large model extraction rule optimization).
[0040] In the above embodiments, step S1 enables the fine-tuned vector model to initially possess the ability to identify similar patents in a manner similar to that of a human; step S2 enables the patent-specific vector base model to adapt to the characteristics of patent description language text; step S3 enables the patent-specific vector model to possess the ability to identify highly similar patents in a manner similar to that of a human; and step S4 enables the final patent-specific vector model to adapt to patent data in multiple formats.
[0041] At least one embodiment of this application also provides a patent semantic retrieval method, see reference. Figure 3 The semantic retrieval method of this patent includes the following steps.
[0042] Step S10: Extract 12-dimensional technical feature text from the patent text database and clean it to construct 12 technical dimensions. The 12-dimensional technical feature text includes title, abstract, research field, claims, technical background, technical problem, invention purpose, invention content, embodiments, technical efficacy, description of drawings, and technical effects.
[0043] For example, 12-dimensional technical feature text can be extracted from patent text database data using regular expressions, large models, and other methods.
[0044] Step S20: Convert the 12-dimensional technical feature text into vectors using the final patent-specific vector model. The final patent-specific vector model is obtained using a patent-specific vector model construction method according to an embodiment of this application.
[0045] For example, for 53 million Chinese characters in the patent text database, multi-dimensional technical feature text is extracted using methods such as regularization and fine-tuning of the technical feature extraction model. The final patent-specific vector model is then used to transform the multi-dimensional technical feature text into vectors.
[0046] Step S30: Convert the input text to be searched or the independent claims of the patent corresponding to the patent number to be searched into vectors using the final patent-specific vector model, perform vector similarity search in the vector field of the 12-dimensional technical feature text, and output the results in a unique highest similarity sorted by all publication numbers.
[0047] For example, vector retrieval uses the QVECTOR method supported by KBASE, and vector distance uses L2 Euclidean distance.
[0048] Step S40: Based on the contribution rate of each dimension to the total recall rate, calculate the recall rate weight of each dimension, and select the top 8 technical feature texts with the highest contribution rates from the 12-dimensional technical feature texts. The 8-dimensional technical feature texts include title, abstract, research field, claims, research question, invention content, embodiments, invention purpose and technical effects.
[0049] Step S50: Convert the input text to be searched or the independent claims of the patent corresponding to the patent number to be searched into vectors using the final patent-specific vector model, perform vector similarity search in 8 technical dimension vector fields, and output the results according to the weighted similarity of the unique highest corresponding dimension weight among all publication numbers.
[0050] In at least one embodiment of this application, step S20 includes: Step S21: Use the final patent-specific vector model to convert the short text technical features in the 12-dimensional technical feature text into single-value vectors. The short text technical features include the title, abstract, and research field.
[0051] Step S22: Using the final patent-specific vector model, the long text technical features in the 12-dimensional technical feature text are sliced and converted into multi-value vectors. The long text technical features include claims, technical background, technical problem, invention purpose, invention content, embodiments, technical effects, description of drawings, and technical results.
[0052] In at least one embodiment of this application, step S30 includes: Step S31: Convert the input text to be searched or the independent claims of the patent corresponding to the patent number to be searched into vectors using the final patent-specific vector model.
[0053] Step S32: In the vector fields of the 12-dimensional technical feature text, first retrieve the top N vectors with the highest similarity from each vector field, then merge the 12-dimensional search results according to the publication number. If multiple vectors are detected for the same publication number, take the highest similarity value, and then sort and output them according to the unique highest similarity of all publication numbers.
[0054] For example, the top N vectors could be the top 400 vectors.
[0055] In at least one embodiment of this application, step S50 includes: Step S51: Convert the input text to be searched or the independent claims of the patent corresponding to the patent number to be searched into vectors using the final patent-specific vector model.
[0056] Step S52: In the vector field of the 8-dimensional technical feature text, first retrieve the top N vectors with the highest similarity from each vector field, then merge the search results of the 8 dimensions according to the publication number. If multiple vectors are retrieved for the same publication number, take the highest similarity value after weighting the corresponding dimension weights, and then sort and output according to the unique highest similarity of all publication numbers.
[0057] For example, the top N vectors could be the top 400 vectors.
[0058] It should be noted that the recall rate = number of recalled evidence / total number of applications. The recall rate can also be called the detection rate.
[0059] Below, we will provide detailed examples of the recall and stability test results of the retrieval results obtained by the patent semantic retrieval method, using specific embodiments as examples.
[0060] The open-source BGE-M3 vector model was used to perform vector transformation on a database of 53 million Chinese patents. A test was conducted on 16,739 cases (16,269 cases with only foreign language evidence and 14,735 cases with only Chinese evidence after excluding those with only foreign language evidence), using step S50 for retrieval. The results are as follows. Figure 4 As shown. By Figure 4 It can be seen that the recall rate of the TOP50 is 34.98%, the recall rate of the TOP100 is 44.17%, and the recall rate of the TOP400 is 66.18%.
[0061] To test the performance of the fine-tuned model at each stage, a test environment with a patent database of 1.82 million was built. To test the stability of the vector model at each stage, a four-level gradient test database was extracted from the 1.82 million patent database, progressively increasing in size.
[0062] [Method] Based on the TOP400 / 600 / 800 / 1000 of each single dimension of the open-source BGE-M3 vector, a subset was collected from the 1.82 million test database respectively.
[0063] [Standard test cases] 862.
[0064] The open-source BGE-M3 vector model was used to perform vector transformation on the test patent database. The test was conducted with 862 cases, and the recall rates for the TOP50, TOP100, and TOP400 patents were 63.15-63.15%, 76.36-76.61%, and 102.79-105.82%, respectively.
[0065] The fine-tuned vector model was used to perform vector transformation on a test database of 1.82 million patents, and tested with 862 cases. The recall rate of the TOP50 patents increased by 13-14%, the recall rate of the TOP100 patents increased by 12-13%, and the recall rate of the TOP400 patents increased by (-0.1)-1.2%.
[0066] Stability tests were performed on the fine-tuned vector model: Combination Figure 5 It can be seen that the retrieval and recall capabilities of the fine-tuned vector model are basically stable.
[0067] Performance testing of the recall rate of the patent-specific vector base model: In a small sample environment of 862 cases, compared with the open-source BGE-M3 vector model: the recall rate of TOP50 increased by 15-17%, the recall rate of TOP100 increased by 12-14%, and the recall rate of TOP400 increased by 0.5-0.7%. Compared with the fine-tuned vector model: the recall rate of TOP50 increased by 2.2-2.9%, the recall rate of TOP100 increased by 0.5-1.1%, and the recall rate of TOP400 decreased by (-3.6) to (-1.2)%.
[0068] Stability tests on the patent-specific vector base model revealed that its performance is generally stable.
[0069] Performance testing of the patent-specific vector model: In a small sample environment of 862 cases, compared to the open-source BGE-M3 vector model: TOP50 recall improved by 19.40-20.85%, TOP100 recall improved by 16.73-18.31%, and TOP400 recall improved by 0.24-3.27%. Compared to the fine-tuned vector model: TOP50 recall improved by 3.64-5.45%, TOP100 recall improved by 3.63-5.94%, and TOP400 recall improved by 0.24-0.73%.
[0070] Stability tests on the patent-specific vector model revealed that its performance is generally stable.
[0071] Performance testing of the final patent-specific vector model was conducted: In a small sample environment of 862 cases, compared with the open-source BGE-M3 vector model: the recall rate of TOP50 improved by 23.40-20.85%, the recall rate of TOP100 improved by 20.97-18.31%, and the recall rate of TOP400 improved by 0.48-1.21%. Compared with the patent-specific vector model: the recall rate of TOP50 improved by 2.55-0%, the recall rate of TOP100 improved by 2.66-0.12%, and the recall rate of TOP400 improved by 0.37-(-2.06)%.
[0072] Stability testing of the final patent-specific vector model revealed a slight decrease in recall, with a decrease of 1%, from the minimum test set to the maximum test set. Considering the overall improvement in recall achieved by the final patent-specific vector model, this decrease is within an acceptable range.
[0073] For a database of 57 million Chinese patents, vector transformation was performed using the final patent-specific vector model. Given the massive data volume, parallel tables were used for storage, and cloud-based parallel retrieval was executed. The recall rate was tested using the 862 cases mentioned above.
[0074] Vector retrieval uses the QVECTOR method supported by KBASE, and vector distance is calculated using L2 Euclidean distance.
[0075] The search interface input parameters are: n=50, k=3200, nprobe=120. Here, n represents the number of results returned for a single form field, k represents the maximum number of results returned for the comprehensive search, and nprobe represents the number of cluster centers that are the nearest neighbors to the vector being queried.
[0076] Table 1 Recall Test Results Table 2 Comparison of recall rates with a test database of 1.82 million At least one embodiment of this application also provides a computer device including a processor and a memory. The processor is used to execute a method for constructing a patent-specific vector model and / or a patent semantic retrieval method provided in any of the above embodiments of this application. The memory is used to store executable instructions of the processor, such as application programs. The number of processors can be one or more. The application programs stored in the memory can include one or more modules, each corresponding to a set of instructions. Furthermore, the processor is configured to execute instructions to perform the above-described method for constructing a patent-specific vector model and / or the patent semantic retrieval method.
[0077] The computer device may also include a power supply component configured for power management, a wired or wireless network interface configured to connect the computer device to a network, and an input / output (I / O) interface. The computer device can operate on an operating system stored in memory, such as Windows Server. TM Mac OSX TM Unix TM Linux TM FreeBSD TM Or similar.
[0078] At least one embodiment of this application also provides a computer-readable storage medium storing executable instructions for a computer. When executed by a processor, the executable instructions implement a method for constructing a patent-specific vector model, and / or a patent semantic retrieval method, as provided in any of the above embodiments of this application.
[0079] A non-transitory computer-readable storage medium, wherein instructions in the storage medium, when executed by a processor of the computer device, enable the computer device to perform the method for constructing the patent-specific vector model and / or the patent semantic retrieval method. The method for constructing the patent-specific vector model and / or the patent semantic retrieval method is executed by a proxy program.
[0080] Those skilled in the art will recognize that the algorithmic steps of the various examples described in conjunction with the embodiments disclosed in this application can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.
[0081] At least one embodiment of this application also provides a computer program product, including a computer program / instruction, which, when executed by a processor, implements a method for constructing a patent-specific vector model provided in any of the above embodiments of this application, and / or a patent semantic retrieval method.
[0082] If the aforementioned functions are implemented as software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or a portion of the technical solution, can be embodied in the form of a computer program product. This computer program product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the vector model fine-tuning method of the various embodiments of this application. The aforementioned storage medium includes various media capable of storing program verification codes, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.
[0083] It should be noted that the combination of the technical features in the embodiments of this application is not limited to the combination methods described in the embodiments of this application or the combination methods described in specific embodiments. All technical features described in this application can be freely combined or combined in any way, unless they contradict each other.
[0084] As indicated in this application and claims, unless the context clearly indicates otherwise, the words “a,” “an,” and / or “the” do not specifically refer to the singular and may also include the plural. Generally speaking, the term “comprising” only indicates that it includes the explicitly identified steps and elements, which do not constitute an exclusive list, and the method or apparatus may also include other steps or elements.
[0085] The above description is merely a preferred embodiment of this application and is not intended to limit this application. Any modifications or equivalent substitutions made within the spirit and principles of this application should be included within the protection scope of this application.
Claims
1. A method for constructing a patent-specific vector model, characterized in that, include: Step S1: Using the first batch of batch cases and rejected EXY evidence pairs as the first training set, and using the loss function as the fine-tuning algorithm, the open-source BGE-M3 vector model is fine-tuned for similar patent retrieval to obtain the fine-tuned vector model. Step S2: Randomly extract millions of uniform unlabeled patent training corpora from multiple dimensions as the second training set. Use the loss function as a fine-tuning algorithm to fine-tune the fine-tuned vector model in the patent context category to obtain the patent-specific vector basic model. Multiple dimensions include the receiving office, IPC classification, patent type, case evidence ratio and legal status. Step S3: Using the second batch of batch cases and EX evidence pairs as the third training set, and using the loss function as the fine-tuning algorithm, the patent-specific vector basic model is fine-tuned for high-similarity patent retrieval to obtain the patent-specific vector model. The number of batch EX evidence pairs is greater than the number of batch rejection EXY evidence pairs. Step S4: Use the multidimensional random sampling of millions of uniform unlabeled patents in the format of other suppliers in the second batch as the fourth training set, and perform data format adaptation fine-tuning on the patent-specific vector model to obtain the final patent-specific vector model.
2. A patent semantic retrieval method, characterized in that, include: Step S10: Extract 12-dimensional technical feature text from the patent text library data, clean it, and construct 12 technical dimensions. The 12-dimensional technical feature text includes title, abstract, research field, claims, technical background, technical problem, invention purpose, invention content, embodiments, technical effects, description of drawings and technical effects. Step S20: Use the final patent-specific vector model to convert the 12-dimensional technical feature text into vectors. The final patent-specific vector model is obtained by the method for constructing a patent-specific vector model as described in claim 1. Step S30: Convert the input text to be searched or the independent claims of the patent corresponding to the patent number to be searched into vectors using the final patent-specific vector model, perform vector similarity search in the vector field of the 12-dimensional technical feature text, and output the results in a unique highest similarity sorted by all publication numbers. Step S40: Based on the contribution rate of a single dimension to the total recall rate, calculate the recall rate weight of each dimension, and select the top 8 technical feature texts with the highest contribution rates from the 12-dimensional technical feature texts. The 8-dimensional technical feature texts include title, abstract, research field, claims, research question, invention content, embodiments, invention purpose and technical effect. Step S50: Convert the input text to be searched or the independent claims of the patent corresponding to the patent number to be searched into vectors using the final patent-specific vector model, perform vector similarity search in 8 technical dimension vector fields, and output the results according to the weighted similarity of the unique highest corresponding dimension weight among all publication numbers.
3. The patent semantic retrieval method according to claim 2, characterized in that, Step S20 includes: Step S21: Use the final patent-specific vector model to convert the short text technical features in the 12-dimensional technical feature text into single-value vectors. The short text technical features include the title, abstract, and research field. Step S22: Using the final patent-specific vector model, the long text technical features in the 12-dimensional technical feature text are sliced and converted into multi-value vectors. The long text technical features include claims, technical background, technical problem, invention purpose, invention content, embodiments, technical effects, description of drawings and technical effects.
4. The patent semantic retrieval method according to claim 2, characterized in that, Step S30 includes: Step S31: Convert the input text to be searched or the independent claims of the patent corresponding to the patent number to be searched into vectors using the final patent-specific vector model; Step S32: In the vector fields of the 12-dimensional technical feature text, first retrieve the top N vectors with the highest similarity from each vector field, then merge the 12-dimensional search results according to the publication number. If multiple vectors are detected for the same publication number, take the highest similarity value, and then sort and output them according to the unique highest similarity of all publication numbers.
5. The patent semantic retrieval method according to claim 2, characterized in that, Step S50 includes: Step S51: Convert the input text to be searched or the independent claims of the patent corresponding to the patent number to be searched into vectors using the final patent-specific vector model; Step S52: In the vector field of the 8-dimensional technical feature text, first retrieve the top N vectors with the highest similarity from each vector field, then merge the search results of the 8 dimensions according to the publication number. If multiple vectors are retrieved for the same publication number, take the highest similarity value after weighting the corresponding dimension weights, and then sort and output according to the unique highest similarity of all publication numbers.
6. A computer device, characterized in that, include: A processor for executing the method for constructing a patent-specific vector model as described in claim 1, and / or, the patent semantic retrieval method as described in any one of claims 2 to 5; and Memory for storing the executable instructions of the processor.
7. A computer-readable storage medium having executable instructions stored thereon, characterized in that, When the executable instructions are executed by the processor, they implement the method for constructing a patent-specific vector model as described in claim 1, and / or the patent semantic retrieval method as described in any one of claims 2 to 5.
8. A computer program product comprising a computer program / instructions, characterized in that, When the computer program / instruction is executed by the processor, it implements the method for constructing a patent-specific vector model as described in claim 1, and / or, the patent semantic retrieval method as described in any one of claims 2 to 5.