Business Identification System Based on Semantic Analysis
By designing a business recognition system based on semantic analysis, the problem of differences in the correlation determination and comparison file acquisition of semantic search systems in the prior art is solved, and more accurate semantic feature coding and search results recommendations are achieved, which improves the search accuracy and user experience.
Patent Information
- Application Number
- CN202411552837.7
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-03-14
- Publication Date
- 2025-04-01
- Estimated Expiration
- Not applicable · inactive patent
AI Technical Summary
The existing semantic retrieval system has differences in the number and location of correlation determination, comparison file acquisition, resulting in the selection of semantic words affecting the semantic interpretation effect, thereby reducing the search accuracy and user experience.
Design a business recognition system based on semantic analysis, including a search text preprocessing module, a patent text classification analysis module and a search result recommendation module. The system improves the accuracy of the correlation intensity calculation between patent texts and the recommendation of search results through search text preprocessing, text encoding generalization, patent text classification analysis and similarity matching value calculation.
Through this system, more accurate semantic feature coding can be obtained, similarity matching value calculations between patent texts can be improved, and search accuracy and semantic correlation expression effects can be improved, and user experience can be improved.
Smart Images
Figure CN119740581A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the field of artificial intelligence technology, and in particular to a business identification system based on semantic analysis. Background Art
[0002] Patent search is a complex task. Semantic search can assist professional searchers to a certain extent and improve work efficiency. At the same time, for the IPR of most enterprises, when there is a certain gap between search ability and database resources and professional searchers, semantic search tools can be used to narrow the gap, so as to better improve the innovation of enterprise patents. Although they can obtain comparative documents, different semantic search systems have great differences in the determination of relevance, the number and location of comparative documents. These differences ultimately lead to the selection of semantic words affecting the semantic interpretation effect. The interpretation deviation of semantic algorithms on semantic words will lead to biased search results, that is, in different fields, the meanings expressed by the same words may be completely different or may have different emphases. In the actual search process of users, it is found that sometimes the interpretation results of semantic words by semantic search systems will be different from our daily cognition. This interpretation deviation will cause users to fail to correctly find the patents they really want to search based on the input text. The retrieval accuracy of semantic analysis based on the search content is low, and the experience is not ideal. Therefore, it is necessary to design a business identification system based on semantic analysis to improve the expression effect of semantic association and the correctness of retrieval. Summary of the invention
[0004] The purpose of the present invention is to provide a business identification system based on semantic analysis to solve the problems raised in the above background technology.
[0005] In order to solve the above technical problems, the present invention provides the following technical solutions: a business identification system based on semantic analysis, including a retrieval text preprocessing module, a patent text classification analysis module, and a retrieval result recommendation module, characterized in that: the retrieval text preprocessing module is used to preprocess the patent text retrieved by the user; the patent text classification analysis module is used to perform classification analysis on the patent text for the generated retrieval word set and the expanded extended word set; the retrieval result recommendation module is used to calculate the correlation strength between patent texts using similarity matching values, and output recommendations for retrieval comparison results.
[0006] According to the above technical solution, the retrieval text preprocessing module includes: a retrieval formula generation module, a word vector acquisition module, and an undirected graph construction module; the retrieval formula generation module is used to automatically and intelligently generate a standard retrieval formula using the acquired retrieval keywords; the word vector acquisition module is used to obtain word vectors through BERT training; the undirected graph construction module is used to construct a keyword-based undirected graph according to the keyword vector table obtained after quantization.
[0007] According to the above technical solution, the patent text classification analysis module includes: a relationship feature generalization module, a domain ontology construction module, a convolutional neural network training module, and a similarity matching value calculation module. The relationship feature generalization module is used to generalize the relationship features at the character level, word level, and sentence level using text encoding; the domain ontology construction module is used to extract the subject concept using string matching, and use the matching strings as concept nodes to construct the domain ontology; the convolutional neural network training module is used to use the character-level convolutional neural network to input the preprocessed training set word vectors into the convolutional neural network for training; the similarity matching value calculation module is used to determine the similarity relationship between two documents and calculate the similarity matching value of the measured patent text.
[0008] According to the above technical solution, the retrieval result recommendation module includes: an association strength calculation module, a relevance score calculation module, and a retrieval result ranking output module. The association strength calculation module is used to calculate the association strength between patent texts using similarity matching values; the relevance score calculation module is used to comprehensively calculate the relevance score in each recommended patent; the retrieval result ranking output module is used to combine the two relevance scores to obtain the final ranking result and perform ranking recommendation.
[0009] According to the above technical solution, the method steps for operating the business identification system include: Step 1: Preprocess the patent text retrieved by the user, extract keywords to generate search word sets and extended word sets; Step 2: Based on the keyword vector table obtained after quantization, use text encoding to generalize the relationship features at the character level, word level, and sentence level; Step 3: Classify and analyze the patent text based on the generated search word set and the expanded extended word set to obtain significant features; Step 4: Use similarity matching values to calculate the correlation strength between patent texts and output recommendations for search comparison results.
[0010] According to the above technical solution, the implementation method of preprocessing the patent text retrieved by the user and extracting keywords to generate a search word set and an extended word set includes: Step 11: Obtain the search keywords and search methods input by the user, that is, automatically and intelligently generate a standard search formula using the obtained search keywords; Step 12: Preprocess the retrieved patent text data, extract keywords to generate a search word set, use the search word set to perform search expansion through the keyword network to obtain an extended word set, and obtain word vectors through BERT training; Step 13: Construct a keyword-based undirected graph based on the keyword vector table obtained after quantization to represent the relationship between keywords and words, and use the preprocessed word vectors to construct a patent text keyword network.
[0011] According to the above technical solution, the generalization process steps include: Step 21: Patent text data contains multiple information types, including patent content, citation relationships and time sequence. The relationship characteristics between patent texts are constructed based on the information types contained. Step 22: Determine the domain theme of the patent text based on the keyword vector table obtained after quantification; Step 23: Use string matching to extract the subject concept, and use the matching strings as concept nodes to construct the domain ontology; Step 24: Use natural language processing methods to extract concepts from the unstructured patent text to be searched and perform concept search; Step 25: Screen the concepts by calculating the relevance method to obtain the concepts and the relationships between them; Step 26: Finally, BERT is used to obtain patent text features. When extracting matching information, RNN is used to encode the context of the previous patent text, and the convolutional network is used to extract the dependency relationship between terms to form quantized character-level, word-level, and sentence-level relationship features.
[0012] According to the above technical solution, the step of classifying and analyzing the patent text includes: Step 31: Use a character-level convolutional neural network to input the preprocessed training set word vectors into the convolutional neural network for training, and extract the patent text features of the output training set; Step 32: Classify using softmax classifier; Step 33: Use the convolution filter in the convolutional neural network to convolve the input text features to obtain text features and extract local features of the patent text; Step 34: Use the convolutional neural network to set convolution kernels of different lengths, multiply the word vector and the convolution formula, perform maximum pooling, and take the maximum value output in the convolution layer result as the most significant feature; Step 35: The queried patent text and the related patent documents are used as the input of the model through the text convolutional neural network, and two intermediate vectors of the same dimension are obtained to determine the similarity relationship between the two documents and calculate the similarity matching value of the measured patent text.
[0013] According to the above technical solution, the process steps of recommending the search and comparison results include: Step 41: Using semantic similarity and citation association to define technical association and technical overlap between patents; Step 42: Calculate the association strength between patent texts using the similarity matching value. The higher the similarity between patent texts, the stronger the association between them. Step 43: Obtain the technical keywords involved in the candidate recommended patents through the similarity matching value, and then comprehensively calculate the relevance score of each recommended patent according to the search keywords entered by the user during the search, and use it as a score, and extract candidate results from multiple shards in parallel based on this score; Step 44: Based on the obtained sentences containing technical keywords, the sentences are put into the search system for retrieval, and the patents with the top n relevance scores are taken as candidate recommended patents. Each main claim in the claims of the candidate patents is taken out for similarity comparison with the comparative documents containing technical keywords; Step 45: Based on each shard, select the top n patents ranked once as the best candidates for this search, then add weights to the top candidate results according to the candidate results, calculate the secondary relevance score, combine the two relevance scores to get the final ranking result, and make ranking recommendations.
[0014] Compared with the prior art, the beneficial effects achieved by the present invention are as follows: the present invention pre-processes the patent text retrieved by the user, extracts keywords to generate a search word set and an extended word set, and then generalizes the relationship features at the character level, word level, and sentence level based on the keyword vector table obtained after quantization using text encoding; and performs classification analysis on the patent text to obtain significant features; it is able to obtain more accurate semantic feature encoding, and obtain better results when calculating the similarity scores of words and sentences in the patent text; finally, the similarity matching value is used to calculate the correlation strength between patent texts, and a recommendation of the search comparison results is output; and better search results can be obtained by sorting the retrieved patent documents, thereby improving the search accuracy of semantic analysis based on the search content and the expression effect of semantic association. BRIEF DESCRIPTION OF THE DRAWINGS
[0015] The accompanying drawings are used to provide a further understanding of the present invention and constitute a part of the specification. Together with the embodiments of the present invention, they are used to explain the present invention and do not constitute a limitation of the present invention. In the accompanying drawings: Figure 1A schematic diagram of the module composition of a business identification system based on semantic analysis provided in Embodiment 2 of the present invention. DETAILED DESCRIPTION
[0016] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.
[0017] Embodiment 1: This embodiment can be applied to the scenario of patent search semantic analysis, and the method specifically includes the following steps: Step 1: Preprocess the patent text retrieved by the user, extract keywords to generate search word sets and extended word sets; In the embodiments of the present invention, specific implementation methods include: Step 11: Obtain the search keywords and search methods input by the user, that is, automatically and intelligently generate a standard search formula using the obtained search keywords; Step 12: Preprocess the retrieved patent text data, extract keywords to generate a search word set, use the search word set to perform search expansion through the keyword network to obtain an extended word set, and obtain word vectors through BERT training; Step 13: Construct a keyword-based undirected graph based on the keyword vector table obtained after quantization to represent the relationship between keywords and words, and use the preprocessed word vectors to construct a patent text keyword network; Specifically, the quantization process includes: initializing an undirected graph, adding keywords to the graph, obtaining the similarity between two word nodes according to a given vector matrix, using the similarity as the correlation between the two nodes, setting a similarity threshold, and defining the two words as having correlation when the similarity is greater than the threshold, establishing an undirected edge between the two nodes, and using the similarity between the word nodes as the weight of the edge to represent the correlation between the two nodes, updating the edge set and weights of the undirected graph, and using the last updated undirected graph as the keyword network.
[0018] Step 2: Based on the keyword vector table obtained after quantization, use text encoding to generalize the relationship features at the character level, word level, and sentence level; In the embodiment of the present invention, the process of generalization includes: Step 21: Patent text data contains multiple information types, including patent content, citation relationships and time sequence. The relationship characteristics between patent texts are constructed based on the information types contained. Step 22: Determine the domain theme of the patent text based on the keyword vector table obtained after quantification; Step 23: Use string matching to extract the subject concept, and use the matching strings as concept nodes to construct the domain ontology; Step 24: Use natural language processing methods to extract concepts from the unstructured patent text to be searched and perform concept search; Step 25: Screen the concepts by calculating the relevance method to obtain the concepts and the relationships between them; Step 26: Finally, BERT is used to obtain patent text features. When extracting matching information, RNN is used to encode the context of the previous patent text, and the convolutional network is used to extract the dependency relationship between terms to form quantized character-level, word-level, and sentence-level relationship features.
[0019] Step 3: Classify and analyze the patent text based on the generated search word set and the expanded extended word set to obtain significant features; In an embodiment of the present invention, performing classification analysis of patent texts includes: Step 31: Use a character-level convolutional neural network to input the preprocessed training set word vectors into the convolutional neural network for training, and extract the patent text features of the output training set; Step 32: Classify using softmax classifier; Step 33: Use the convolution filter in the convolutional neural network to convolve the input text features to obtain text features and extract local features of the patent text; Step 34: Use the convolutional neural network to set convolution kernels of different lengths, multiply the word vector and the convolution formula, perform maximum pooling, and take the maximum value output in the convolution layer result as the most significant feature; Step 35: The queried patent text and the related patent documents are used as the input of the model through the text convolutional neural network, and two intermediate vectors of the same dimension are obtained to determine the similarity relationship between the two documents and calculate the similarity matching value of the measured patent text.
[0020] Step 4: Use the similarity matching value to calculate the correlation strength between patent texts and output the recommendation of search comparison results; In the embodiment of the present invention, the process of recommending the search comparison results includes: Step 41: Using semantic similarity and citation association to define technical association and technical overlap between patents; Step 42: Calculate the association strength between patent texts using the similarity matching value. The higher the similarity between patent texts, the stronger the association between them. Step 43: Obtain the technical keywords involved in the candidate recommended patents through the similarity matching value, and then comprehensively calculate the relevance score of each recommended patent according to the search keywords entered by the user during the search, and use it as a score, and extract candidate results from multiple shards in parallel based on this score; Step 44: Based on the obtained sentences containing technical keywords, the sentences are put into the search system for retrieval, and the patents with the top n relevance scores are taken as candidate recommended patents. Each main claim in the claims of the candidate patents is taken out for similarity comparison with the comparative documents containing technical keywords; Step 45: Based on each shard, select the top n patents ranked once as the best candidates for this search, then add weights to the top candidate results according to the candidate results, calculate the secondary relevance score, combine the two relevance scores to get the final ranking result, and make ranking recommendations.
[0021] Embodiment 2: Embodiment 2 of the present invention provides a service identification system based on semantic analysis. Figure 1 A schematic diagram of the module composition of a business identification system based on semantic analysis provided in the second embodiment of the present invention, such as Figure 1 As shown, the system includes: The search text preprocessing module is used to preprocess the patent text searched by the user; Patent text classification analysis module, used to perform classification analysis of patent texts based on the generated search word set and the expanded extended word set; The retrieval result recommendation module is used to calculate the correlation strength between patent texts using similarity matching values and output recommendations for retrieval comparison results.
[0022] In some embodiments of the present invention, the retrieval text preprocessing module includes: A search formula generation module is used to automatically and intelligently generate a standard search formula using the acquired search keywords; The word vector acquisition module is used to obtain word vectors through BERT training; The undirected graph construction module is used to construct an undirected graph based on keywords according to the keyword vector table obtained after quantization.
[0023] In some embodiments of the present invention, the patent text classification analysis module includes: The relation feature generalization module is used to generalize the relation features at the character level, word level, and sentence level using text encoding; The domain ontology construction module is used to extract the subject concepts by string matching and use the matching strings as concept nodes to construct the domain ontology; A convolutional neural network training module is used to input the preprocessed training set word vectors into the convolutional neural network for training using a character-level convolutional neural network; The similarity matching value calculation module is used to determine the similarity relationship between two documents and calculate the similarity matching value of the measured patent text.
[0024] In some embodiments of the present invention, the search result recommendation module includes: An association strength calculation module is used to calculate the association strength between patent texts using similarity matching values; A relevance score calculation module is used to comprehensively calculate the relevance score of each recommended patent; The search result ranking output module is used to combine the two relevance scores to obtain the final ranking result and make ranking recommendations.
[0025] It should be noted that, in this article, relational terms such as first and second, etc. are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Moreover, the terms "include", "comprise" or any other variants thereof are intended to cover non-exclusive inclusion, so that a process, method, article or device including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, article or device.
[0026] Finally, it should be noted that the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art can still modify the technical solutions described in the aforementioned embodiments or replace some of the technical features therein by equivalents. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention shall be included in the protection scope of the present invention.
Claims
1. A business identification system based on semantic analysis, comprising a search text preprocessing module, a patent text classification analysis module, and a search result recommendation module, characterized in that: The search text preprocessing module is used to preprocess the patent text retrieved by the user; the patent text classification analysis module is used to perform classification analysis on the patent text based on the generated search word set and the expanded extended word set; the search result recommendation module is used to calculate the correlation strength between patent texts using the similarity matching value and output the recommendation of the search comparison result; The method steps for operating the business identification system include: Step 1: Preprocess the patent text retrieved by the user, extract keywords to generate search word sets and extended word sets; Step 2: Based on the keyword vector table obtained after quantization, use text encoding to generalize the relationship features at the character level, word level, and sentence level; Step 3: Classify and analyze the patent text based on the generated search word set and the expanded extended word set to obtain the most significant features; Step 4: Use the similarity matching value to calculate the correlation strength between patent texts and output the recommendation of search comparison results; The implementation method of preprocessing the patent text retrieved by the user and extracting keywords to generate a search word set and an extended word set includes: Step 11: Obtain the search keywords and search methods input by the user, that is, automatically and intelligently generate a standard search formula using the obtained search keywords; Step 12: Perform data preprocessing on the retrieved patent text, extract keywords to generate a search word set, use the search word set to perform search expansion through the keyword network to obtain an extended word set, and obtain the training set word vector through BERT training; Step 13: Construct a keyword-based undirected graph based on the keyword vector table obtained after quantization to represent the relationship between keywords and words, and use the preprocessed word vectors to construct a patent text keyword network; The steps of the generalization process include: Step 21: The patent text contains multiple information types, including patent content, citation relationship and time sequence, and the relationship characteristics between the patent texts are constructed according to the information types contained; Step 22: Determine the domain theme of the patent text based on the keyword vector table obtained after quantification; Step 23: Use string matching to extract the subject concept, and use the matching strings as concept nodes to construct the domain ontology; Step 24: Use natural language processing methods to extract concepts from the unstructured patent text to be searched and perform concept search; Step 25: Screen the concepts by calculating the relevance method to obtain the concepts and the relationships between the concepts; Step 26: Finally, BERT is used to obtain patent text features. RNN is used to encode the patent text context when extracting matching information. The convolutional network is used to extract the dependency relationship between terms to form quantized character-level, word-level, and sentence-level relationship features. The step of classifying and analyzing the patent text includes: Step 31: Use a character-level convolutional neural network to input the preprocessed training set word vectors into the convolutional neural network for training, and extract the patent text features of the output training set; Step 32: Classify using softmax classifier; Step 33: Use the convolution filter in the convolutional neural network to convolve the input text features to obtain text features and extract local features of the patent text; Step 34: Use the convolutional neural network to set convolution kernels of different lengths, multiply the word vector and the convolution formula, perform maximum pooling, and take the maximum value output in the convolution layer result as the most significant feature; Step 35: Using the text convolutional neural network to take the queried patent text and the related patent text as the input of the model, obtain two intermediate vectors of the same dimension, determine the similarity relationship between the two texts and calculate the similarity matching value of the measured patent text; The process steps of recommending the search and comparison results include: Step 41: Using semantic similarity and citation association to define technical association and technical overlap between patents; Step 42: Calculate the association strength between patent texts using the similarity matching value. The higher the similarity between patent texts, the stronger the association between them. Step 43: Obtain the technical keywords involved in the candidate recommended patents through the similarity matching value, and then comprehensively calculate the relevance score of each candidate recommended patent according to the search keywords entered by the user during the search, and use it as a score, and extract candidate results from multiple shards in parallel based on this score; Step 44: Based on the obtained sentences containing technical keywords, the sentences are put into the search system for retrieval, and the patents with the top n relevance scores are taken as candidate recommended patents. Each main claim in the claims of the candidate patents is taken out for similarity comparison with the comparative documents containing technical keywords; Step 45: Based on each shard, select the top n patents ranked once as the best candidates for this search, then add weights to the top candidate results according to the candidate results, calculate the secondary relevance score, combine the two relevance scores to get the final ranking result, and make ranking recommendations.