A method for calculating patent relevance based on a large model

Abstract

The application belongs to the technical field of AI models, and proposes a method for calculating the relevance of patents based on a large model. The patent invention point extraction algorithm based on the generative large model is used to cut each patent to be calculated to extract the overall invention point and the local invention point of each patent to be calculated. The patent text vectorization model is used to vectorize the overall invention point and the local invention point of each patent to be calculated to obtain the overall feature vector and the local feature vector of each patent to be calculated. The patent text vectorization model is used to improve the quality of patent document vectorization. The multi-vector matching algorithm is used to calculate the similarity of the local invention points, so as to calculate the local invention points of two patent files into the patent correlation degree, and improve the accuracy of the patent relevance calculation.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of AI model, and particularly relates to a method for calculating patent relevance based on a large model. BACKGROUND

[0002] At present, when searching for a patent, a semantic engine technology can be used to sort the search results by calculating the patent relevance, and to display the patents with higher relevance preferentially. This method indeed improves the efficiency of patent search to some extent, especially for more than 70% of the overall relevant patents. However, for the comparison of locally relevant patents, the overall similarity is low, and the sorting is inaccurate, which leads to missed detection and further leads to low credibility of the search results. Meanwhile, a piece of invention itself may have multiple invention points, and if only one vector is used to represent the patent, the relevance may not be high in the case of local relevance between two patents. Therefore, how to calculate the patent relevance based on a large model is a very important problem.

[0003] Based on this, the present application provides a method for calculating patent relevance based on a large model. SUMMARY

[0004] To solve the technical problems that the prior art cannot calculate the correlation degree according to the local similarity of two patent files, and the quality of patent document vectorization is low, the present application provides a method for calculating patent relevance based on a large model. The patent is cut according to the invention points by using a patent invention point extraction algorithm constructed based on a generative large model to generate overall invention points and local invention points, and a multi-vector matching algorithm is used to calculate the similarity of the local invention points, so as to realize the calculation of the local similarity of two patent files into the correlation degree of the patent. The quality of patent document vectorization is improved by using a large model of a patent text vectorization model, and finally the accuracy of patent relevance calculation is improved.

[0005] The present application provides a method for calculating patent relevance based on a large model, which comprises:

[0006] S1: Extracting overall and local invention points of the patent to be calculated: obtaining two patents to be calculated, constructing a patent invention point extraction algorithm based on a generative large model, cutting each patent to be calculated based on the patent invention point extraction algorithm to obtain the overall invention points and the local invention points corresponding to each patent to be calculated respectively;

[0007] S2: Invention point vectorization: using a patent text vectorization model constructed based on a large model to vectorize the overall invention points and the local invention points of each patent to be calculated respectively to obtain the overall feature vectors and the local feature vectors of each patent to be calculated, and to form a feature vector set of each patent to be calculated;

[0008] S3: Multi-vector matching and correlation calculation: using a multi-vector matching algorithm, extracting one feature vector from each of the two feature vector sets to form a matching pair, calculating the feature similarity of the matching pair of the two patents to be calculated, constructing a feature similarity matrix, and determining the patent correlation of the two patents to be calculated according to the feature similarity matrix.

[0009] Optionally, the S1 based on the patent invention point extraction algorithm, each patent to be calculated is cut to obtain the overall invention point and the local invention point corresponding to each patent to be calculated respectively.

[0010] S11: Extracting key terms: based on a pre-constructed patent knowledge graph, extracting key terms matching the patent knowledge graph from each patent to be calculated;

[0011] S12: Generating an overall invention point: by a first prompt word template, splicing the each patent to be calculated and the key terms extracted from the each patent to be calculated, and inputting into a first generative large model to obtain the overall invention point corresponding to the each patent to be calculated;

[0012] S13: Generating a local invention point: determining the number of times the key terms of the each patent to be calculated appear in each paragraph of the each patent to be calculated, and determining the local invention point corresponding to the each patent to be calculated according to the number of times of each paragraph determined and the weight corresponding to the key terms of the each patent to be calculated.

[0013] Optionally, in the S11, the nodes in the patent knowledge graph represent key terms, and the edges represent the relationship between the key terms.

[0014] Optionally, in the S13, the local invention point corresponding to the each patent to be calculated is determined according to the number of times of each paragraph determined and the weight corresponding to the key terms of the each patent to be calculated, and the local invention point corresponding to the each patent to be calculated is determined according to the number of times of each paragraph determined and the weight corresponding to the key terms of the each patent to be calculated.

[0015] S131: initializing the processing state of each paragraph of the each patent to be calculated to a first state;

[0016] S132: when the processing state of each paragraph of the each patent to be calculated is the first state, determining the importance of each paragraph of the each patent to be calculated according to the weight corresponding to the key terms of the each patent to be calculated and the number of times of the key terms of the each patent to be calculated appearing in each paragraph of the each patent to be calculated.

[0017] S133: combining adjacent paragraphs according to the importance of each paragraph of each patent to be calculated, determining the shortest paragraph group of each patent to be calculated as a patent segment, and the conditions for generating the patent segment are that the importance of the patent segment of each patent to be calculated is greater than half of the sum of the weights of all key terms of each patent to be calculated; the importance of the patent segment is generated by superimposing the importance corresponding to each paragraph in the patent segment;

[0018] S134: setting the processing state of the paragraph in the shortest paragraph group of each patent to be calculated to a second state, and adjusting the weight value of the key term hit in the shortest paragraph group of each patent to be calculated to half of the current weight;

[0019] S135: repeating steps S132-S134, recalculating the importance corresponding to each paragraph and determining the patent segment of the current period based on the adjusted weight value of the key term in S134, until the execution times reach the set threshold or no new patent segment can be generated, based on each patent segment of each patent to be calculated, using a second generative large model to determine each local invention point of each patent to be calculated.

[0020] Optionally, the S135 based on each patent segment of each patent to be calculated, using a second generative large model to determine each local invention point of each patent to be calculated specifically includes:

[0021] When the patent segment of the patent to be calculated exceeds the limited length, the second prompt word template is spliced with the patent segment and input to the second generative large model to reduce the length of the patent segment through the second generative large model, to obtain the local invention point corresponding to the patent segment;

[0022] When the patent segment of the patent to be calculated does not exceed the limited length, the patent segment is taken as a local invention point.

[0023] Optionally, the first prompt word template in S12 includes first role information, first input requirements, first output requirements, a card slot corresponding to the first input content, and a card slot corresponding to the key term; the second prompt word template in S135 includes second role information, second output requirements, a card slot corresponding to the second input content, and a card slot corresponding to the key term.

[0024] Optionally, during the process of generating a local invention point in S13, the first local invention point generated is a main invention point, and the local invention points other than the main invention point are secondary invention points.

[0025] Optionally, the patent text vectorization model in S2 can be one of a general text vectorization large model and a fine-tuned general text vectorization large model.

[0026] S2 specifically includes:

[0027] inputting the overall invention points and the local invention points of each patent to be calculated into a patent text vectorization model constructed based on a large model, to obtain an overall feature vector and a local feature vector of each patent to be calculated;

[0028] constructing a feature vector set of each patent to be calculated according to the overall feature vector and the local feature vector of each patent to be calculated.

[0029] Optionally, in S3, a multi-vector matching algorithm is used, one feature vector combination is extracted from each of the two feature vector sets to form a matching pair, the feature similarity of all matching pairs of the two patents to be calculated is calculated, and a feature similarity matrix is constructed, which specifically includes:

[0030] S31: taking the two patents to be calculated as a first patent to be calculated and a second patent to be calculated respectively, and taking the overall feature vector and the local feature vector of the first patent to be calculated as each first feature vector, and taking the overall feature vector and the local feature vector of the second patent to be calculated as each second feature vector;

[0031] S32: combining each first feature vector in the feature vector set of the first patent to be calculated with each second feature vector in the feature vector set of the second patent to be calculated respectively to obtain each matching pair, and calculating the feature similarity between the first feature vector and the second feature vector in each matching pair respectively to construct a feature similarity matrix.

[0032] Optionally, S3 determines the patent relevance of the two patents to be calculated according to the feature similarity matrix, which specifically includes:

[0033] S33: determining the similarity weight corresponding to each matching pair according to the weight corresponding to the first feature vector and the second feature vector in each matching pair respectively;

[0034] S34: determining the maximum feature similarity in the feature similarity of each matching pair according to the feature similarity matrix, and determining the similarity weight corresponding to the maximum feature similarity as the maximum similarity weight;

[0035] S35: From the each matching pair, determine each matching pair satisfying a target condition as each target matching pair; each target matching pair includes a first feature vector and a second feature vector which are not the same, and the sum of the feature similarities of the each target matching pair is maximum;

[0036] S36: According to the maximum feature similarity, the maximum similarity weight, the feature similarity of each target matching pair and the similarity weight of each target matching pair, the patent correlation degree of the two patents to be calculated is calculated.

[0037] The above at least one technical solution adopted in the specification can achieve the following beneficial effects:

[0038] The present application proposes a method for calculating the correlation degree of a patent based on a large model, which realizes the consideration of local invention points when calculating the correlation degree of a patent, and the improvement of the quality of text vectorization. Specifically, the patent invention point extraction algorithm constructed based on the generative large model can be used to cut each obtained patent to be calculated to extract the overall invention point and the local invention point of each patent to be calculated, and the patent text vectorization model can be used to vectorize the overall invention point and the local invention point of each patent to be calculated to obtain the overall feature vector and the local feature vector of each patent to be calculated, so as to improve the quality of patent document vectorization through the patent text vectorization model. Then, the multi-vector matching algorithm is used to calculate the similarity of the local invention points, so as to realize the calculation of the local invention points of two patent files into the correlation degree of the patent. At the same time, the overall invention point and the local invention point extracted by the patent text vectorization model are vectorized to improve the quality of patent document vectorization, and finally the accuracy of the calculation of the correlation degree of the patent is improved.

[0039] When the overall invention point and the local invention point are obtained, the key terms matched with the patent knowledge graph can be extracted from each patent to be calculated through the patent knowledge graph, the nodes in the patent knowledge graph represent the key terms, and the edges represent the relationship between the key terms. Since the patent knowledge graph contains a large amount of search information input when searching for patents, i.e. contains a large amount of keywords (i.e. key terms) used when searching for patents, the key terms of each patent to be calculated can be proposed through the patent knowledge graph, which helps to improve the accuracy of the overall invention point and the local invention point determined subsequently.

[0040] When the overall invention point is obtained, each to-be-calculated patent and the key term of each to-be-calculated patent are spliced through the first prompt word template and input into the first generative large model to obtain the overall invention point corresponding to each to-be-calculated patent, so that the first generative large model extracts the overall invention point from each to-be-calculated patent through the first prompt word template and the key term, thereby improving the accuracy of determining the overall invention point. Moreover, the first prompt word template includes role information, input requirements, output requirements, a card slot corresponding to the input content, and a card slot corresponding to the key term. The structure of the first prompt word template can better prompt the first generative large model on how to extract the overall invention point, and limit the output of the first generative large model, so that the output overall invention point is more in line with the requirements.

[0041] When the local invention point is obtained, the shortest paragraph group of each to-be-calculated patent is calculated as a patent segment based on the weight of the key term of each to-be-calculated patent and the number of times the key term appears in each paragraph of each to-be-calculated patent, so as to cover the key term more and improve the accuracy of the subsequently generated local invention point. Moreover, the local invention point of each to-be-calculated patent is determined based on the patent segment of each to-be-calculated patent. In addition, when the local invention point is determined based on the patent segment, the second prompt word template is spliced with the patent segment when the patent segment exceeds the limited length and is input into the second generative large model, so that the length of the patent segment is reduced by the second generative large model to obtain the local invention point corresponding to the patent segment, so that the length of the generated local invention point meets the input requirements of subsequent vectorization. Moreover, the second prompt word template includes role information, output requirements, a card slot corresponding to the input content, and a card slot corresponding to the key term. The second prompt word template prompts the second generative large model on how to reduce the input patent segment to ensure that the output local invention point meets the output requirements and does not lose any key information.

[0042] When the patent correlation degree of two to-be-calculated patents is determined according to the calculated feature similarity, not only the feature similarity between each feature vector is considered, but also the importance of each feature vector is considered. That is, the patent correlation degree of two to-be-calculated patents can be determined according to the similarity weight of each matching pair and the feature similarity of each matching pair, so as to improve the accuracy of the determined patent correlation degree, solve the problem of local correlation between patents, improve the efficiency and recall rate of comparative design search, greatly improve the retrieval efficiency, and improve the quality of the retrieval result. BRIEF DESCRIPTION OF DRAWINGS

[0043] Figure 1 A flowchart of a method for calculating a patent correlation degree based on a large model is provided in the specification;

[0044] Figure 2A schematic diagram of a method for calculating patent relevance based on a large model is provided in the present specification. DETAILED DESCRIPTION

[0045] The present application will be further described in detail below in combination with the drawings and specific embodiments.

[0046] The present specification provides a method for calculating patent relevance based on a large model, specifically as shown in Figure 1 Figure 1 A flowchart of a method for calculating patent relevance based on a large model is provided in the present specification, specifically including the following steps:

[0047] S1: Extracting the overall and local invention points of the patent to be calculated: obtaining two patents to be calculated, constructing a patent invention point extraction algorithm based on a generative large model, and cutting each patent to be calculated based on the patent invention point extraction algorithm to obtain the overall and local invention points corresponding to each patent to be calculated.

[0048] In the present specification, the device for calculating patent relevance can first extract the overall and local invention points of the patent to be calculated, i.e., obtain two patents to be calculated, construct a patent invention point extraction algorithm based on a generative large model, and cut each patent to be calculated based on the patent invention point extraction algorithm to obtain the overall and local invention points corresponding to each patent to be calculated. The device for calculating patent relevance can be a server, a system, or an electronic device such as a desktop computer or a notebook computer. For ease of description, the method for calculating patent relevance based on a large model provided in the present specification will be described below with the server as the execution subject. The type of the patent to be calculated can be an invention or a utility model, and the present specification does not make specific limitations.

[0049] ​Currently, based on the attention mechanism model, the self-attention mechanism needs to pay attention to all other positions at each position, so long text sequences will bring large calculation and storage costs. Therefore, there is a certain limit to the length of the text, which is generally 512, although it can be solved by modifying the configuration and source code, but the cost performance is very low, such as greatly reducing the training efficiency. In processing long text sequences, the commonly used strategies are blocking, truncation, sliding window, etc., but they will all lose the context information of the sequence. Assuming that the sequence length supported by the model completely meets the requirements of the length of the patent text, such as 100,000 words, but since the correlation between patent texts is mostly local rather than global, the whole patent is not considered to be vectorized, but needs to be split according to the characteristics of the patent text, so as to fundamentally solve the problem of local correlation. Based on this, the server can first obtain two patents to be calculated, and construct a patent invention point extraction algorithm based on the generative large model, based on the patent invention point extraction algorithm, cut each patent to be calculated respectively to obtain the corresponding overall invention point and local invention point of each patent to be calculated. Specifically, when cutting each patent to be calculated based on the patent invention point extraction algorithm to obtain the corresponding overall invention point and local invention point of each patent to be calculated, the server can perform the following steps:

[0050] S11: Extract key terms: based on the pre-constructed patent knowledge graph, extract key terms matching the patent knowledge graph from each patent to be calculated.

[0051] S12: Generate overall invention points: splice each patent to be calculated and the extracted key terms of each patent to be calculated through the first prompt word template, and input into the first generative large model to obtain the corresponding overall invention points of each patent to be calculated.

[0052] S13: Generate local invention points: determine the number of times each key term of each patent to be calculated appears in each paragraph of each patent to be calculated, and determine the corresponding local invention points of each patent to be calculated according to the determined number of times of each paragraph and the corresponding weight of each key term of each patent to be calculated.

[0053] The patent knowledge graph in S11 above is constructed based on the search log information of the patent search system. The nodes in the patent knowledge graph represent key terms, and the edges represent the relationships between the key terms. The patent search system accumulates a large amount of search log information. In addition, to avoid loss of search log information, the search log information in the patent search system can be regularly transferred every month. The search log information contains a large amount of valuable information, i.e., the search log information includes a large amount of search information used when searching for patents, i.e., patent search formulas, such as (IC='F21V21 / 30' OR IC='F21V21 / 14') AND ('support' AND / SEN 'rod' AND / SEN 'base'), where IC represents the international patent classification number, 'F21V21 / 30' and 'F21V21 / 14' represent the classification numbers, and ((CPC='B28B11 / 245')) AND (('temperature control' OR 'temperature sensor') AND ('air cooler' OR 'water cooler') AND 'cooling water pipe') AND (USE='reduce concrete structure cracks'), where CPC represents the cooperative patent classification, 'B28B11 / 245' represents the classification number, and USE represents the use or purpose of the technical solution described in the patent. The server can construct a patent knowledge graph based on the search log information, and can extract key terms from the patent to be calculated based on the patent knowledge graph. Specifically, in S11 above, the server can extract key terms that match the nodes in the patent knowledge graph from each patent to be calculated based on the pre-constructed patent knowledge graph. The number of key terms extracted from each patent to be calculated that best represent each patent to be calculated is not more than a preset number, which can be pre-set, such as 30. The patent invention point extraction algorithm is constructed based on a generative large model, which at least includes the first generative large model, and can also include the second generative large model described below.

[0054] In S12, the server can splice the to-be-calculated patent and the extracted key terms of the to-be-calculated patent by a first prompt word template, and input into a first generative large model to obtain the overall invention point corresponding to the to-be-calculated patent. The first prompt word template can include first role information, first input requirements, first output requirements, a card slot corresponding to the first input content, and a card slot corresponding to the key terms. The first role information can be pre-set, and can be “Now you are a professional patent examiner, and you need to use your professional knowledge to read and understand the following long text patent, and extract 3-6 invention points and implementation steps that best represent the novelty and creativity of the patent”. The first input requirements and the first output requirements can also be pre-set. The first input requirements can be “the input content is the patent text”, and the first output requirements can be “filtering known theories and technologies, the output content must come from the input content, and the output content should as much as possible revolve around the given key terms”. Therefore, the first prompt word template can be “Now you are a professional patent examiner, and you need to use your professional knowledge to read and understand the following long text patent, and extract 3-6 invention points and implementation steps that best represent the novelty and creativity of the patent, but you must pay attention to the following four points: (1) filtering known theories and technologies; (2) the output content must come from the input content; (3) the output content should as much as possible revolve around the given key terms, and the given key terms are: [key terms]; (4) the output word number is about 500 words. The patent text you need to process is: [patent text], and the extracted overall invention point is ”. The “[key terms]” is a card slot corresponding to the key terms, which is used to add the key terms corresponding to the to-be-calculated patent, and the number of key terms is pre-set. The “[patent text]” is a card slot corresponding to the first input content, which is used to add the patent text corresponding to the to-be-calculated patent. It should be noted that the first role information, the first input requirements, the first output requirements, and the first prompt word template can also be other contents. The first role information, the first input requirements, the first output requirements, and the first prompt word template shown above are only one example, and the specific content of the first role information, the first input requirements, the first output requirements, and the first prompt word template is not limited in the present application. The first generative large model can be an existing large model, such as the Kimi large model, or a large model fine-tuned from a general large model, which is not limited in the present application.

[0055] In S13, the server can determine the number of times the key terms of each to-be-calculated patent appear in each paragraph of the to-be-calculated patent. The paragraphs corresponding to each to-be-calculated patent can be the paragraphs corresponding to the specification of each to-be-calculated patent.

[0056] In the above S13, when determining the local invention point corresponding to each patent to be calculated based on the determined number of times corresponding to each paragraph and the weight corresponding to each key term of the patent to be calculated, the server may perform the following steps:

[0057] S131: Initialize the processing state corresponding to each paragraph of each patent to be calculated to the first state.

[0058] S132: When the processing state of each paragraph of each patent to be calculated is the first state, the importance of each paragraph of each patent to be calculated is determined according to the weight corresponding to the key term of each patent to be calculated and the number of times the key term of each patent to be calculated appears in each paragraph of each patent to be calculated.

[0059] S133: Combine adjacent paragraphs according to the importance corresponding to each paragraph of each patent to be calculated, determine the shortest paragraph group of each patent to be calculated, and use it as the patent fragment. The conditions for generating patent fragments are: the importance of each patent fragment of each patent to be calculated is greater than half of the sum of the weights of all key terms of each patent to be calculated; the importance of the patent fragment is generated by superimposing the importance corresponding to each paragraph in the patent fragment.

[0060] S134: Setting the processing status of the paragraphs in the shortest paragraph group of each patent to be calculated to the second status, and adjusting the weight value of the key term hit in the shortest paragraph group of each patent to be calculated to half of the current weight.

[0061] S135: Repeat steps S132-S134, and based on the weight values ​​of the key terms adjusted in S134, recalculate the importance of each paragraph and determine the patent judgment of the current cycle until the number of executions reaches the set threshold or new patent fragments cannot be generated. Based on the patent fragments of each patent to be calculated, the second generative large model is used to determine the local invention points of each patent to be calculated.

[0062] The first state in the above S131 can be preset, and the first state can be represented by 0, which can be specifically represented as , p represents the number of paragraphs, Specifically, the server may set the processing status of each paragraph of each patent to be calculated to the first state for each patent to be calculated, so as to initialize the processing status of each paragraph of the patent to be calculated.

[0063] In S132, the server determines the importance of each paragraph of the patent to be calculated according to the weight of the key term corresponding to the key term of the patent to be calculated and the number of times the key term appears in the paragraph when the processing state of the paragraph is the first state. Then in S133, the server combines adjacent paragraphs in the patent to be calculated according to the importance of each paragraph of the patent to be calculated to determine the shortest paragraph group of the patent to be calculated as a patent fragment. In the above, the weight of each key term of the patent to be calculated is pre-set (i.e. initial weight), that is , q represents the number of key terms, represents the weight of key term j. In the above, when determining the importance of each paragraph of the patent to be calculated according to the weight of the key term corresponding to the key term of the patent to be calculated and the number of times the key term appears in the paragraph, the following formula (1) can be used for calculation:

[0064] (1)

[0065] wherein, M i represents the importance of the paragraph, cnt ij represents the number of times the key term appears in the paragraph.

[0066] The shortest paragraph group is composed of consecutive paragraphs, and the server can determine the shortest paragraph group of the patent to be calculated according to the importance of each paragraph of the patent to be calculated using any pre-set algorithm. In addition, the condition for generating the patent fragment is that the importance of the patent fragment of each patent to be calculated is greater than half of the sum of the weights of all key terms of each patent to be calculated, and the importance of the patent fragment is generated by superimposing the importance of each paragraph in the patent fragment.

[0067] In S134, the server can set the processing state of each paragraph in the shortest paragraph group of the patent to be calculated to a second state, and adjust the weight value of the key term hit in the shortest paragraph group of the patent to be calculated to half of the current weight, after determining the shortest paragraph group of the patent to be calculated for each paragraph of the patent to be calculated and each paragraph of the patent to be calculated.

[0068] In S135, the server can determine the local invention points of each patent to be calculated based on the patent segments of each patent to be calculated by using the second generative large model.

[0069] In addition, in order to avoid the length of the generated local invention points exceeding the input length of the subsequent patent text vectorization model, when the patent segment of the patent to be calculated exceeds the limited length, the server can splice the second prompt word template and the patent segment (i.e., the patent segment exceeding the limited length) and input it into the second generative large model to reduce the length of the patent segment through the second generative large model to obtain the local invention points corresponding to the patent segment. When the patent segment of the patent to be calculated does not exceed the limited length, the patent segment is taken as the local invention point.

[0070] Specifically, the server can splice the second prompt word template with each patent segment of each patent to be calculated when the patent segment exceeds the limited length, and input to the second generative large model to reduce the length of the patent segment by the second generative large model to obtain the local invention point corresponding to the patent segment. When the patent segment does not exceed the limited length, the patent segment is taken as the local invention point. The limited length can be pre-set, and the limited length can be 512 characters. The second generative large model can be the first generative large model, or other general large models, or large models fine-tuned from general large models, which are not limited in the specification. The second prompt word template is pre-set, and the second prompt word template includes second role information, second output requirements, card slots corresponding to second input content, and card slots corresponding to key terms. The second role information can be pre-set, and the second role information can be "Now you are a professional patent examiner, you need to use your professional knowledge to read and understand the following long text patent, and reduce the text to within 500 words". The second output requirement can also be pre-set, and the second output requirement can be "The output content must come from the input content, and the output content should as much as possible contain the given key terms". In addition, the second output requirement can also include the number of words of the output content, such as "The output content must be within 500 words". Therefore, the second prompt word template can be "Now you are a professional patent examiner, you need to use your professional knowledge to read and understand the following long text patent, and reduce the text to within 500 words, but please pay attention to the following two points: (1) The output content must come from the input content; (2) The output content should as much as possible contain the given key terms, the given key terms are:

key terms

patent segment

key terms

patent segment

[0071] S2: Invention point vectorization: using a patent text vectorization model constructed based on a large model, respectively vectorizing the overall invention point and the local invention point of each patent to be calculated to obtain the overall feature vector and the local feature vector of each patent to be calculated, and forming a feature vector set of each patent to be calculated.

[0072] In the specification, the server can invent point vectorization, that is, using a patent text vectorization model constructed based on a large model, respectively vectorizing the overall invention points and the local invention points of each patent to be calculated to obtain the overall feature vector and the local feature vector of each patent to be calculated, and constructing a feature vector set of each patent to be calculated. Specifically, the overall invention points and the local invention points of each patent to be calculated are respectively input into the patent text vectorization model constructed based on the large model to obtain the overall feature vector and the local feature vector of each patent to be calculated. According to the overall feature vector and the local feature vector of each patent to be calculated, a feature vector set of each patent to be calculated is constructed. The patent text vectorization model is constructed based on a large model, and the patent text vectorization model can be one of a general text vectorization large model and a fine-tuned general text vectorization large model, and the general text vectorization large model is a Transformer model based on an attention mechanism.

[0073] In addition, the server can input the overall invention points of each patent to be calculated into the patent text vectorization model constructed based on the large model to obtain the overall feature vector, and input the local invention points of the patent to be calculated into the patent text vectorization model constructed based on the large model to obtain the local feature vector. Then, the feature vector set of the patent to be calculated is constructed based on the overall feature vector and the local feature vector of the patent to be calculated.

[0074] S3: Multi-vector matching and correlation calculation: using a multi-vector matching algorithm, one feature vector is extracted from each of the two feature vector sets to form a matching pair, the feature similarity of all matching pairs of the two patents to be calculated is calculated, a feature similarity matrix is constructed, and the patent correlation of the two patents to be calculated is determined according to the feature similarity matrix.

[0075] In the specification, the server can multi-vector matching and correlation calculation, that is, using a multi-vector matching algorithm, one feature vector is extracted from each of the two feature vector sets to form a matching pair, the feature similarity of all matching pairs of the two patents to be calculated is calculated, a feature similarity matrix is constructed, and the patent correlation of the two patents to be calculated is determined according to the feature similarity matrix. The above two feature vector sets are the feature vector sets corresponding to the above two patents to be calculated, and the above all matching pairs are the overall feature vector and the local feature vector of any one of the two patents to be calculated and the overall feature vector and the local feature vector of the other patent to be calculated except the any one of the patents to be calculated.

[0076] Specifically, when using the multi-vector matching algorithm, one feature vector in each of the two feature vector sets is extracted to form a matching pair, the feature similarity of all matching pairs of the two patents to be calculated is calculated, and the feature similarity matrix is constructed, the following steps can be performed:

[0077] S31: The two patents to be calculated are taken as a first patent to be calculated and a second patent to be calculated respectively, and the global feature vector and the local feature vector of the first patent to be calculated are taken as each first feature vector, and the global feature vector and the local feature vector of the second patent to be calculated are taken as each second feature vector.

[0078] S32: Each first feature vector in the feature vector set of the first patent to be calculated is combined with each second feature vector in the feature vector set of the second patent to be calculated to obtain each matching pair, and the feature similarity between the first feature vector and the second feature vector in each matching pair is calculated to construct the feature similarity matrix.

[0079] Each first feature vector includes the global feature vector and the local feature vector of the first patent to be calculated, and the feature vector set of the first patent to be calculated includes each first feature vector. Each second feature vector includes the global feature vector and the local feature vector of the second patent to be calculated, and the feature vector set of the second patent to be calculated includes each second feature vector. Each matching pair includes one first feature vector and one second feature vector. When obtaining each matching pair, the server can combine each first feature vector with each second feature vector to obtain each matching pair. For example, assuming that there are 5 first feature vectors, i.e., V w , V m , V o1 , V o2 , and 4 second feature vectors, i.e., U w , U m , U o1 , U o2 , there are 20 matching pairs, i.e., V w and U w , V w and U m , V w and U o1 , V w and U o2 , V m and U w , V m and U m , V m and U o1 , V m and U o2 , V o1 and U w . w , V m , V o1 , V o2 , and 4 second feature vectors, i.e., U w , U m , U o1 , U o2 , there are 20 matching pairs, i.e., V w and U w , V w and U m , V w and U o1 , V w and U o2 , V m and U w , V m and U m , V m and U o1 , V m and U o2 , V o1 and U w . w , V m , V o1 , V o2 , and 4 second feature vectors, i.e., U w , U m , U o1 , U o2 , there are 20 matching pairs, i.e., V w and U w , V w and U m , V w and U o1 , V w and U o2 , V m and U w , V m and U m , V m and U o1 , V m and U o2 , V o1 and U w .and U w , V o1 and U m , V o1 and U o1 , V o1 and U o2 , V o2 and U w , V o2 and U m , V o2 and U o1 , V o2 and U o2 , V o3 and U w , V o3 and U m , V o3 and U o1 , V o3 and U o2 The feature similarity can be a cosine similarity between the first feature vector and the second feature vector, and the server can calculate the feature similarity of each matching pair through a cosine similarity algorithm. The local invention point is also included in the similarity calculation, that is, the local similarity affects the degree of association of the patents. The feature similarity matrix is composed of the feature similarities of each matching pair.

[0080] When determining the patent correlation degree of the two patents to be calculated according to the feature similarity matrix, the server can perform the following steps:

[0081] S33: According to the weight corresponding to the first feature vector and the second feature vector in each matching pair, determine the similarity weight corresponding to each matching pair.

[0082] S34: According to the feature similarity matrix, determine the maximum feature similarity in the feature similarity of each matching pair, and determine the similarity weight corresponding to the maximum feature similarity as the maximum similarity weight.

[0083] S35: From each matching pair, determine each matching pair that meets the target condition as the target matching pair; the first feature vector and the second feature vector included in each target matching pair are not the same, and the sum of the feature similarities of each target matching pair is the largest.

[0084] S36: According to the maximum feature similarity, the maximum similarity weight, the feature similarity of each target matching pair, and the similarity weight of each target matching pair, calculate the patent correlation degree of the two patents to be calculated.

[0085] The weight corresponding to the first feature vector and the second feature vector in each matching pair is pre-set, such as the weight of each first feature vector can be W w , Wm , W o1 , W o2 , W o3 The weight of each second feature vector can be W w , W m , W o1 , W o2 , wherein W w =1.0, W m =0.9, W o1 = W o2 = W o3 =0.8. The feature similarity of each matching pair can be S1, S2, S3, …, S 20 , and the corresponding similarity weight can be W1, W2, W3, …, W 20 . When calculating the similarity weight, taking the similarity weight of the matching pair including V w and U m as an example, .

[0086] The similarity weight of each matching pair is the product of the weight of the first feature vector and the weight of the second feature vector in each matching pair. Specifically, in S33, the server can multiply the weight of the first feature vector and the weight of the second feature vector in each matching pair as the similarity weight of the matching pair. The maximum feature similarity is the maximum of the feature similarities of each matching pair, and the maximum similarity weight is the similarity weight of the matching pair corresponding to the maximum feature similarity. The maximum feature similarity can be represented as Smax, and the maximum similarity weight can be represented as Wmax. . .

[0087] The target condition in S35 can be pre-set, and the target condition can be that the first feature vector and the second feature vector included in each target matching pair selected are different, and the feature similarity of each target matching pair selected is the largest. That is, the server can first determine the first number of the first feature vector of the first patent to be calculated, and determine the second number of the second feature vector of the second patent to be calculated, compare the size of the first number and the second number, and take the smaller number as the target number. Based on each matching pair, a combination group of matching pairs including the target number is constructed, and the first feature vector and the second feature vector of each matching pair in the constructed combination group are different. Calculate the sum of the feature similarities of each matching pair in each combination group, and take the sum of the feature similarities as the total feature similarity, and take the combination group corresponding to the maximum total feature similarity as the target combination group, and take each matching pair in the target combination group as each target matching pair satisfying the target condition.

[0088] In the above S36, the server can first calculate the first weighted similarity of the two patents to be calculated according to the feature similarity of each target matching pair and the similarity weight of each target matching pair. Specifically, the following formula (2) can be used for calculation:

[0089] (2)

[0090] wherein, S1 represents the first weighted similarity, S1 represents the first weighted similarity, S1 represents the first weighted similarity, k S1 represents the first weighted similarity, S1 represents the first weighted similarity, S1 represents the first weighted similarity,

[0091] Then, the product of the maximum feature similarity and the maximum similarity weight is calculated, and the product is taken as the second weighted similarity. Then, the patent relevance of the two patents to be calculated is calculated according to the first weighted similarity and the second weighted similarity. Specifically, the following formula (3) can be used for calculation:

[0092] (3)

[0093] wherein, S represents the patent relevance, S1 represents the first weighted similarity, S1 represents the first weighted similarity, S1 represents the first weighted similarity, * S1 represents the first weighted similarity, W1 represents the weight corresponding to the first weighted similarity, W2 represents the weight corresponding to the second weighted similarity, and W1 and W2 are pre-set. W1 can be 0.3 and W2 can be 0.7.

[0094] In some embodiments of the present specification, the first local invention point generated in the process of generating the local invention point in the above S13 is the main invention point, and the local invention points other than the main invention point are the secondary invention points. For an invention patent, the number of secondary invention points is not more than three, and for a utility model patent, the number of secondary invention points is zero. The overall feature vector corresponding to the overall invention point can be represented by V w or U w , the feature vector corresponding to the main invention point can be represented by V m or U m , and the feature vector corresponding to the secondary invention point can be represented by V o1 , V o2 , V o3 or U o1 , U o2 .

[0095] In some embodiments of the present specification, as shown in Figure 2 Figure 2 is a schematic diagram of a method for calculating patent relevance based on a large model provided in the present specification. The two patents to be calculated in the above are the patent to be calculated 1 (i.e., the first patent to be calculated in the above) and the patent to be calculated 2 (i.e., the second patent to be calculated in the above) in Figure 2 . The server can determine the overall invention points and the local invention points of the patent to be calculated 1 and the patent to be calculated 2 respectively based on the patent invention point extraction algorithm, and determine the first feature vectors (i.e., the first feature vectors 1-n) of the patent to be calculated 1 and the second feature vectors (i.e., the second feature vectors 1-m) of the patent to be calculated 2 respectively based on the patent text vectorization model, and then determine each matching pair according to the first feature vectors and the second feature vectors, and calculate the feature similarity of each matching pair. Figure 2 D.1-1~D.1-m, D.2-1~D.2-m, D.3-1~D.3-m …… D.n-1~D.n-m in Figure 2 “Calculate the best match” in Figure 2 The feature similarity of each target matching pair shown in Figure 2 is marked in gray. Then, the server can calculate the patent relevance between the patent to be calculated 1 and the patent to be calculated 2, i.e., “Calculate the patent relevance” in Figure 2 , which can be specifically as steps S33-S36 described above.

[0096] ​In some embodiments of the present disclosure, a test patent pair can be obtained, the test patent pair comprising a first patent and a second patent related to the first patent. The first 400 third patents related to the first patent in the test patent pair are searched based on a patent database by using a conventional patent search method, and the first 400 fourth patents related to the first patent in the test patent pair are searched based on the patent database by using the method shown in S1-S3. It should be noted that the above two patents to be calculated can be the first patent and any one patent in the patent database. When the patent relevance of the two patents to be calculated exceeds the threshold, it is indicated that the two patents to be calculated are relevant, and the other patent in the two patents to be calculated except the first patent is related to the first patent. On the contrary, when the patent relevance of the two patents to be calculated does not exceed the threshold, it is indicated that the two patents to be calculated are not relevant, and the other patent in the two patents to be calculated except the first patent is not related to the first patent. The proportion of the third patents hitting the second patent is determined and taken as the first proportion, and the first proportion is about 50%, and the proportion of the fourth patents hitting the second patent is determined and taken as the second proportion, and the second proportion is about 65%. Obviously, the test result (i.e. the second proportion) by using the method shown in S1-S3 is better than the test result (i.e. the first proportion) by using the conventional patent search method, so it is indicated that the patent invention point extraction algorithm, the patent text vectorization model and the multi-vector matching algorithm improve the accuracy of patent relevance calculation.

[0097] It should be noted that the above specific embodiments can enable those skilled in the art to more fully understand the present invention, but do not limit the present invention in any way. Therefore, although the present invention has been described in detail with reference to the drawings and examples, those skilled in the art should understand that modifications or equivalent replacements can still be made to the present invention, in short, all technical solutions and improvements that do not deviate from the spirit and scope of the present invention should be covered in the protection scope of the patent of the present invention.

Claims

1. A method for calculating patent relevance based on a large model, characterized in that: include: S1: Extracting the overall and local invention points of the patent to be calculated: Obtain two patents to be calculated, build a patent invention point extraction algorithm based on the generative large model, and then segment each patent to be calculated based on the patent invention point extraction algorithm to obtain the overall invention point and local invention point corresponding to each patent to be calculated; S2: Invention point vectorization: Using the patent text vectorization model built based on the large model, the overall invention point and the local invention point of each patent to be calculated are vectorized respectively to obtain the overall feature vector and the local feature vector of each patent to be calculated, forming the feature vector set of each patent to be calculated; S3: Multi-vector matching and correlation calculation: Using a multi-vector matching algorithm, extract one feature vector from each of the two feature vector sets to form a matching pair, calculate the feature similarity of all the matching pairs of the two patents to be calculated, construct a feature similarity matrix, and determine the patent correlation of the two patents to be calculated based on the feature similarity matrix; In S1, based on the patent invention point extraction algorithm, each patent to be calculated is segmented to obtain the overall invention points and local invention points corresponding to each patent to be calculated, specifically including: S11: Extract key terms: Based on the pre-built patent knowledge graph, extract key terms that match the patent knowledge graph from each patent to be calculated; S12: Generate overall invention points: Use the first prompt word template to combine each patent to be calculated and the extracted key terms of each patent to be calculated, and input them into the first generative model to obtain the overall invention points corresponding to each patent to be calculated; S13: Generate local invention points: Determine the number of times the key terms of each patent to be calculated appear in each paragraph of each patent to be calculated, and determine the local invention points corresponding to each patent to be calculated based on the determined number of times corresponding to each paragraph and the weight corresponding to the key terms of each patent to be calculated; The step S13 of determining the local invention point corresponding to each patent to be calculated according to the determined number of times corresponding to each paragraph and the weight corresponding to the key term of each patent to be calculated specifically includes: S131: Initializing the processing state corresponding to each paragraph of each patent to be calculated to a first state; S132: When the processing state of each paragraph of each patent to be calculated is the first state, determining the importance of each paragraph of each patent to be calculated according to the weight corresponding to the key term of each patent to be calculated and the number of times the key term of each patent to be calculated appears in each paragraph of each patent to be calculated; S133: Adjacent paragraphs are combined according to the importance corresponding to each paragraph of each patent to be calculated, and the shortest paragraph group of each patent to be calculated is determined as a patent fragment. The patent fragment is generated under the following conditions: the importance of the patent fragment of each patent to be calculated is greater than half of the sum of the weights of all key terms of each patent to be calculated; and the importance of the patent fragment is generated by adding the importance corresponding to each paragraph in the patent fragment; S134: setting the processing state of the paragraphs in the shortest paragraph group of each patent to be calculated to the second state, and adjusting the weight values ​​of the key terms hit in the shortest paragraph group of each patent to be calculated to half of the current weight; S135: Repeat steps S132-S134, and based on the weight values ​​of the key terms adjusted in S134, recalculate the importance of each paragraph and determine the patent fragments of the current cycle, until the number of executions reaches the set threshold or new patent fragments cannot be generated, and then use the second generative large model based on the patent fragments of each patent to be calculated to determine the local invention points of each patent to be calculated.

2. The method for calculating patent relevance based on a large model according to claim 1, characterized in that: The nodes in the patent knowledge graph in S11 represent key terms, and the edges represent the relationships between the key terms.

3. The method for calculating patent relevance based on a large model according to claim 1, characterized in that: In S135, based on each patent fragment of each patent to be calculated, the second generative model is used to determine each local invention point of each patent to be calculated, which specifically includes: When the patent fragment of the patent to be calculated exceeds the specified length, the second prompt word template is spliced ​​with the patent fragment and input into the second generative model to shorten the length of the patent fragment through the second generative model to obtain the local invention point corresponding to the patent fragment; When the patent segment of the patent to be calculated does not exceed the limited length, the patent segment is regarded as a local invention point.

4. The method for calculating patent relevance based on a large model according to claim 3, characterized in that: The first prompt word template in S12 includes first role information, first input requirement, first output requirement, a slot corresponding to the first input content, and a slot corresponding to the key term; the second prompt word template in S135 includes second role information, second output requirement, a slot corresponding to the second input content, and a slot corresponding to the key term.

5. The method for calculating patent relevance based on a large model according to claim 1, characterized in that: In the process of generating local invention points in S13, the first local invention point generated is the main invention point, and the local invention points other than the main invention point are secondary invention points.

6. The method for calculating patent relevance based on a large model according to claim 1, characterized in that: The patent text vectorization model in S2 may be one of a general text vectorization large model and a fine-tuned general text vectorization large model; The S2 specifically includes: Input the overall invention point and the local invention point of each patent to be calculated into the patent text vectorization model constructed based on the large model, respectively, to obtain the overall feature vector and the local feature vector of each patent to be calculated; A feature vector set of each patent to be calculated is constructed based on the overall feature vector and the local feature vector of each patent to be calculated.

7. The method for calculating patent relevance based on a large model according to claim 1, characterized in that: In S3, a multi-vector matching algorithm is used to extract a feature vector from each of the two feature vector sets to form a matching pair, and the feature similarity of the matching pairs of the two patents to be calculated is calculated. The feature similarity matrix is ​​constructed specifically including: S31: The two patents to be calculated are respectively used as a first patent to be calculated and a second patent to be calculated, and the overall feature vector and the local feature vector of the first patent to be calculated are used as first feature vectors, and the overall feature vector and the local feature vector of the second patent to be calculated are used as second feature vectors; S32: Combine each first eigenvector in the first eigenvector set of the patent to be calculated with each second eigenvector in the second eigenvector set of the patent to be calculated to obtain matching pairs, and calculate the feature similarity between the first eigenvector and the second eigenvector in each matching pair to construct a feature similarity matrix.

8. The method for calculating patent relevance based on a large model according to claim 7, characterized in that: Determining the patent relevance of the two patents to be calculated according to the feature similarity matrix in S3 specifically includes: S33: Determine a similarity weight corresponding to each matching pair according to the weights corresponding to the first feature vector and the second feature vector in each matching pair; S34: determining the maximum feature similarity among the feature similarities of the matching pairs according to the feature similarity matrix, and determining a similarity weight corresponding to the maximum feature similarity, and using the similarity weight as the maximum similarity weight; S35: From the matching pairs, determining matching pairs that meet the target condition and using them as target matching pairs; each target matching pair includes a different first feature vector and a different second feature vector, and the sum of the feature similarities of the target matching pairs is the largest; S36: Calculate the patent relevance of the two patents to be calculated based on the maximum feature similarity, the maximum similarity weight, the feature similarity of each target matching pair, and the similarity weight of each target matching pair.

Images