Patent retrieval keyword extraction method and system based on technical requirements

Abstract

According to the patent retrieval keyword extraction method and system based on the technical requirements, multi-dimensional comprehensive evaluation is adopted for retrieval keyword analysis and extraction, synonyms and antonyms are generated, keyword combinations are optimized according to historical data of a patent library, the retrieval efficiency is improved, and the retrieval efficiency is improved. According to the keyword extraction method for multi-dimension comprehensive evaluation, an adaptive weight adjustment mechanism is introduced, and a machine learning or deep learning model is utilized to automatically adjust the weight according to historical feedback data of a user, so that the weight of each dimension is dynamically adjusted through a machine learning method; therefore, the keyword extraction effect is kept optimal under different technical requirement scenes; and in the synonym and antonym generation process, BERT or other context related models are utilized, the context of the technical requirements is analyzed, dynamic adjustment is performed based on the context, the actual context in the technical requirement document is better met, the problem that synonyms in a static dictionary are not suitable for the current context can be effectively avoided, and the retrieval effect is improved.

Description

Technical Field

[0001] The present invention relates to the field of patent retrieval, and in particular to a patent retrieval keyword extraction method and system based on technical requirements. Background Art

[0002] Patent retrieval is highly professional, and accurate keyword determination is crucial. Patent retrieval is the process of finding patents or non-patent literature related to a specific technical subject, which is particularly important for patent invalidation and infringement procedures. Inaccurate keyword selection will increase the workload of the search and make it difficult to find highly relevant documents. Therefore, it is particularly urgent to develop a keyword extraction method and system. The method and system need to be able to automatically analyze the real technical requirements based on the technical requirements documents written by technicians, and accurately extract patent search keywords to improve retrieval efficiency and accuracy. Summary of the invention

[0003] In view of the deficiencies in the prior art, the present invention provides a patent search keyword extraction method and system based on technical requirements, adopts multi-dimensional comprehensive evaluation for search keyword analysis and extraction, generates synonyms and antonyms, optimizes keyword combinations based on historical data of the patent library, and improves search efficiency.

[0004] In order to solve the above technical problems, the present invention is solved by the following technical solutions In order to achieve the above object, the present invention adopts the following technical solutions: A patent search keyword extraction method based on technical requirements includes the following steps: 1) Text preprocessing: Clean up the technical requirements documents written by technical personnel, including sentence segmentation, stop word filtering, stem extraction or word form restoration, and part-of-speech tagging; 2) Technical requirements analysis: Analyze the technical requirements in the document through natural language processing (NLP) technology, and use algorithms such as TF-IDF and Word2Vec to calculate the frequency and importance of vocabulary; 3) Keyword extraction: Extract keywords based on comprehensive evaluation of multiple dimensions including contextual semantic similarity, word co-occurrence frequency, domain specificity, and time sensitivity; 4) Synonym and antonym generation: Based on the dictionary or embedding model, the synonyms and antonyms related to the keywords are dynamically generated; 5) Keyword optimization: Optimize keyword combinations based on historical data from the patent database to improve retrieval efficiency.

[0005] Preferably, the text preprocessing includes: Stop word filtering: remove common meaningless words; Stemming or lemmatization: normalizing words to a standard form; Part-of-speech tagging: tag the part of speech of each word for subsequent analysis.

[0006] Preferably, the technical requirements analysis step calculates the frequency of vocabulary using the TF-IDF algorithm, and the formula is as follows:

[0007] Among them, t is the vocabulary, d is the document, N is the total number of documents, and {d'∈D:t∈d'} is the number of documents containing the keyword.

[0008] Preferably, the keyword extraction step is comprehensively scored based on the following formula: Final_Score(w)=λ1·Context_Score(w)+λ2·Co_occurrence_Score(w)+λ3·Domain_Specificity_Score(w)+λ4·Temporal_Score(w), where w is the candidate keyword, and λ1,λ2,λ3,λ4 are the weight coefficients of each dimension.

[0009] Preferably, the dimension weight coefficients λ1, λ2, λ3, λ4 can be automatically adjusted according to user feedback and retrieval effect through an adaptive adjustment module.

[0010] Preferably, the synonyms and antonyms are generated based on dynamic expansion of the context, and the similarity between the keyword and the synonym candidate is calculated by the BERT model, and the formula is as follows:

[0011] Among them, w is the keyword, si is the synonym candidate, and D is the technical requirement document.

[0012] Preferably, the most relevant synonyms and antonyms are dynamically screened out according to a similarity threshold Threshold.

[0013] Preferably, the keyword optimization step includes: 1) Patent library frequency matching: count the frequency of occurrence of each keyword in patent documents; 2) Retrieval performance prediction: Calculate the retrieval performance of keywords and give priority to keyword combinations with higher performance.

[0014] A patent search keyword extraction system based on technical requirements, comprising: 1) Text preprocessing module: used to clean and format technical requirement documents; 2) Technical requirements analysis module: used to analyze the technical requirements in the document and identify key technical points; 3) Keyword extraction module: extract technical keywords based on multiple dimensions; 4) Synonym and antonym generation module: dynamically generate synonyms and antonyms related to keywords; 5) Keyword optimization module: optimize keywords and generate the final search keyword list.

[0015] Preferably, the keyword extraction module dynamically adjusts the weight coefficient of each dimension through an adaptive weight adjustment mechanism to improve the accuracy of keyword extraction.

[0016] Compared with the prior art, the present invention has the following beneficial effects: The present invention discloses a patent search keyword extraction method and system based on technical requirements, which adopts multi-dimensional comprehensive evaluation for the search keyword analysis and extraction, generates synonyms and antonyms, optimizes keyword combinations according to the historical data of the patent library, and improves the search efficiency.

[0017] The present invention provides a keyword extraction method for comprehensive evaluation of multiple dimensions, and introduces an adaptive weight adjustment mechanism. It uses machine learning or deep learning models to automatically adjust weights based on users' historical feedback data. The weights of each dimension are dynamically adjusted through machine learning methods to ensure that the effect of keyword extraction remains optimal under different technical demand scenarios.

[0018] The present invention utilizes BERT or other context-related models in the generation of synonyms and antonyms, analyzes the context of technical requirements, and dynamically adjusts based on the context to be more in line with the actual context in the technical requirement document. This can effectively avoid the problem that static dictionary synonyms are not suitable for the current context and improve the retrieval effect. DETAILED DESCRIPTION

[0019] The following description is used to disclose the present invention so that those skilled in the art can implement the present invention. The preferred embodiments described below are only examples, and those skilled in the art can think of other obvious variations. The basic principles of the present invention defined in the following description can be used for other embodiments, variations, improvements, equivalents, and other technical solutions that do not deviate from the spirit and scope of the present invention.

[0020] Example: This application provides a patent search keyword extraction method based on technical requirements, comprising the following steps: 1) Text preprocessing: Clean up the technical requirements documents written by technical personnel, including sentence segmentation, stop word filtering, stem extraction or word form restoration, and part-of-speech tagging; 2) Technical requirements analysis: Analyze the technical requirements in the document through natural language processing (NLP) technology, and use algorithms such as TF-IDF and Word2Vec to calculate the frequency and importance of vocabulary; 3) Keyword extraction: Extract keywords based on comprehensive evaluation of multiple dimensions including contextual semantic similarity, word co-occurrence frequency, domain specificity, and time sensitivity; 4) Synonym and antonym generation: Based on the dictionary or embedding model, the synonyms and antonyms related to the keywords are dynamically generated; 5) Keyword optimization: Optimize keyword combinations based on historical data from the patent library to improve retrieval efficiency.

[0021] In this application, text preprocessing includes: Stop word filtering: remove common meaningless words; Stemming or lemmatization: normalizing words to a standard form; Part-of-speech tagging: tag the part of speech of each word for subsequent analysis.

[0022] The steps include: Input: Technical requirements document written by technical personnel; deal with: Sentence segmentation, Stop word filtering, Stemming or Lemmatization, Part-of-speech tagging; Output: preprocessed text.

[0023] In this application, the technical requirements analysis step calculates the frequency of vocabulary using the TF-IDF algorithm, and the formula is as follows:

[0024] Among them, t is the vocabulary, d is the document, N is the total number of documents, and {d'∈D:t∈d'} is the number of documents containing the word.

[0025] The steps include: Input: preprocessed text; Processing: Analyze the frequency and importance of words in the text through the above TF-IDF (TermFrequency-InverseDocumentFrequency) or Word2Vec model. Among them, Word2Vec calculates the similarity between words by mapping words to vector space.

[0026] In this application, the keyword extraction module uses a comprehensive algorithm to extract keywords in combination with multiple dimensions. Each candidate keyword will be scored based on the following dimensions:

[0027] Dimension 1: Contextual Semantic Similarity Through contextual embedding models such as BERT, the semantic relevance of vocabulary and technical requirements text as a whole is measured. The higher the relevance, the greater the keyword weight.

[0028] Formula: Context_Score(w,D)=cosine_similarity(BERT(w),BERT(D)) Among them, w is the candidate keyword and D is the technical requirement document.

[0029] Dimension 2: Word co-occurrence frequency The word co-occurrence matrix is ​​used to calculate the frequency with which each word co-occurs with other words in the document. Words with high co-occurrence frequency are more likely to be core technical points.

[0030] formula:

[0031] Among them, w1 and w2 are word pairs, and W is the word set in the document.

[0032] Dimension 3: Domain Specificity Identify high-frequency words related to the technical field through data in the domain dictionary or patent library to ensure that the extracted keywords are highly relevant to the field.

[0033] formula:

[0034] Dimension 4: Time Sensitivity By analyzing the time trends in the technology field, emerging technologies or current hot words are extracted first.

[0035] formula:

[0036] Finally, the keyword extraction step extracts the most valuable keywords by weightedly combining the scores of the above dimensions.

[0037] Final_Score(w)=λ·Context_Score(w)+λ2·Co_occurence_Score(w)+λ3·Domain_Specificity_Score(u)+λ4·Temporal_Score(w) Among them, λ, λ2, λ3, λ4, λ5 are the weighting coefficients of each dimension.

[0038] It should be noted that the weight coefficients of each of the above dimensions (context semantics, co-occurrence frequency, domain specificity, etc.) are fixed or manually adjusted, which may not always maintain the best performance in different technical fields and different needs. To this end, an adaptive weight adjustment mechanism is introduced to automatically adjust the weights based on the user's historical feedback data using machine learning or deep learning models. The specific contents are as follows:

[0039] In the adaptive adjustment module of dimension weights, the weights of each dimension are dynamically adjusted through machine learning methods (such as gradient descent or reinforcement learning) to ensure that the effect of keyword extraction remains optimal under different technical requirement scenarios.

[0040] Algorithm: Gradient descent based weight adjustment enter D: technical requirements document; wi: candidate keywords.

[0041] λ1,λ2,...,λn: initial weights for each dimension (initialized randomly or set equal).

[0042] S1: User feedback score (used to adjust weight).

[0043] step (1) Initialization weights: The initial weights λ1,λ2,...,λn are usually set equal, or preliminarily estimated through historical data. Formula:

[0044]

[0045] (2) Calculate the scores of each dimension: For each candidate keyword wi, calculate its scores in different dimensions (such as context similarity, co-occurrence frequency, domain specificity, etc.), recorded as , where k represents the score of the kth dimension.

[0046] (3) User feedback acquisition: After keyword extraction is completed, a score S is assigned based on the user’s feedback on the extraction results, with the score range being [0, 1]. The higher the score, the better the quality of the extracted keywords.

[0047] (4) Loss function definition: Define the loss function L, which is used to evaluate the gap between the current keyword extraction quality and user feedback. Formula:

[0048]

[0049] Where m is the total number of candidate keywords.

[0050] (5) Gradient calculation: Calculate the gradient of each dimension weight using the gradient descent algorithm. Formula:

[0051]

[0052] (6) Weight update: Update the weight of each dimension according to the gradient. α is the learning rate. Formula:

[0053]

[0054] (7) Iterative adjustment: Repeat steps 3-6 until the loss function converges, that is, the extraction result is highly consistent with the user feedback.

[0055] Output The dynamically adjusted weights of each dimension are λ1,λ2,...,λn.

[0056] Optimal keyword extraction results.

[0057] In the present application, the synonym and antonym generation step is performed by inputting the extracted keywords; using WordNet or a pre-built domain synonym dictionary to generate synonyms and antonyms related to the keywords, and semantic verification is performed on the synonyms and antonyms to ensure that they conform to the context of the technical field.

[0058] At present, synonym generation relies on static dictionaries or pre-trained models, which may ignore the specific meanings of words in different technical contexts. For example, the synonyms of "anti-corrosion" in the fields of chemical materials and construction may be different. To this end, BERT or other context-related models are used to analyze the context of technical requirements and dynamically generate more accurate synonyms in different fields and scenarios. The specific contents are as follows:

[0059] Overview of the context-based dynamic synonym expansion module In the context of technical requirements documents, synonyms and antonyms of keywords may have different meanings. The context-based dynamic synonym expansion module uses context-related language models (such as BERT) to dynamically generate synonyms and antonyms according to the current technical background.

[0060] Algorithm: Dynamic Synonym Expansion Based on BERT enter w: extracted candidate keywords; D: Technical requirements document.

[0061] step 1. Keyword embedding generation: Use the pre-trained BERT model to generate the context embedding vector of the candidate keyword w. Embedding(w,D)=BERT(w,D)

[0062] Here BERT(w,D) represents the semantic embedding of keyword w in the context of technical document D.

[0063] 2. Synonym candidate generation: Get a preliminary synonym list {s1, s2, ..., sn} from a predefined synonym dictionary (such as WordNet) or a domain thesaurus.

[0064] 3. Synonym embedding calculation: Calculate the embedding vector of each synonym candidate si in the document context.

[0065] Embedding(si,D)=BERT(si,D) 4. Semantic similarity calculation: Based on cosine similarity, calculate the similarity between each synonym candidate and the keyword in the current context. Calculation formula:

[0066]

[0067] 5. Dynamic screening: Screen synonyms based on similarity scores and keep the most relevant synonyms, {si|Similarity(w,si)>Threshold} The threshold is an adjustable parameter and can be set by the user as required.

[0068] 6. Antonym expansion: Similarly, antonyms are generated through domain antonym dictionaries or using embedding models.

[0069] Output A list of synonyms and antonyms generated dynamically by context.

[0070] The above synonyms and antonyms are dynamically adjusted based on the context, which is more in line with the actual context in the technical requirements document. This can effectively avoid the problem that static dictionary synonyms are not suitable for the current context and improve the retrieval effect.

[0071] In this application, the keyword optimization step includes: 1) Patent library frequency matching: count the frequency of occurrence of each keyword in patent documents; 2) Retrieval performance prediction: Calculate the retrieval performance of keywords and give priority to keyword combinations with higher performance.

[0072] Specifically, the patent library frequency matching step is: input the initially extracted keywords, synonyms and antonyms, match them with the documents in similar fields in the patent library, and count the frequency of each keyword in the historical patent documents; The formula is:

[0073] count(w) is the number of times the keyword appears in the patent database, and total_docs is the total number of patent documents.

[0074] Specifically, the search performance prediction step is as follows: input the keywords after the frequency matching described above, predict the search performance of each keyword based on the historical search results and literature feedback, and give priority to keyword combinations that can cover more relevant patents; The formula is: Retrieval_Performance(w)=Precision(w)×Recall(w) Precision and recall are evaluated based on the performance of keywords in the search results.

[0075] Specifically, it also includes the keyword combination optimization step: combining the coverage of patent documents, keyword matching accuracy and recall rate, and finding the optimal keyword combination through heuristic algorithms (such as genetic algorithms and simulated annealing algorithms); The formula is:

[0076] W is the set of all possible keywords, and K is the optimal keyword subset.

[0077] Based on the above method, this application also discloses a patent search keyword extraction system based on technical requirements, including: 1) Text preprocessing module: used to clean and format technical requirement documents; 2) Technical requirements analysis module: used to analyze the technical requirements in the document and identify key technical points; 3) Keyword extraction module: extract technical keywords based on multiple dimensions; 4) Synonym and antonym generation module: dynamically generate synonyms and antonyms related to keywords; 5) Keyword optimization module: optimize keywords and generate the final search keyword list.

[0078] The keyword extraction module dynamically adjusts the weight coefficients of each dimension through an adaptive weight adjustment mechanism to improve the accuracy of keyword extraction.

[0079] Those skilled in the art should understand that the embodiments of the present invention shown in the above description are only examples and do not limit the present invention. The purpose of the present invention has been fully and effectively achieved. The functions and structural principles of the present invention have been demonstrated and explained in the embodiments, and the embodiments of the present invention may be deformed or modified in any way without departing from the principles.

Claims

1. A patent search keyword extraction method based on technical requirements, characterized by: The following steps are involved: 1) Text preprocessing: Clean up the technical requirements documents written by technical personnel, including sentence segmentation, stop word filtering, stem extraction or word form restoration, and part-of-speech tagging; 2) Technical requirements analysis: Analyze the technical requirements in the document through natural language processing (NLP) technology, and use algorithms such as TF-IDF and Word2Vec to calculate the frequency and importance of vocabulary; 3) Keyword extraction: Extract keywords based on comprehensive evaluation of multiple dimensions including contextual semantic similarity, word co-occurrence frequency, domain specificity, and time sensitivity; 4) Synonym and antonym generation: Based on the dictionary or embedding model, the synonyms and antonyms related to the keywords are dynamically generated; 5) Keyword optimization: Optimize keyword combinations based on historical data from the patent library to improve retrieval efficiency.

2. A patent search keyword extraction method based on technical requirements according to claim 1, characterized in that: The text preprocessing includes: Stop word filtering: remove common meaningless words; Stemming or lemmatization: normalizing words to a standard form; Part-of-speech tagging: tag the part of speech of each word for subsequent analysis.

3. The patent search keyword extraction method based on technical requirements according to claim 1 is characterized by: The technical requirements analysis step calculates the frequency of words through the TF-IDF algorithm, and the formula is as follows: ； Among them, t is the vocabulary, d is the document, N is the total number of documents, and {d'∈D:t∈d'} is the number of documents containing the keyword.

4. The patent search keyword extraction method based on technical requirements according to claim 1 is characterized by: The keyword extraction step is comprehensively scored based on the following formula: Final_Score(w)=λ1·Context_Score(w)+λ2·Co_occurrence_Score(w)+λ3·Domain_Specificity_Score(w)+λ4·Temporal_Score(w), where w is the candidate keyword, and λ1,λ2,λ3,λ4 are the weight coefficients of each dimension.

5. The patent search keyword extraction method based on technical requirements according to claim 4 is characterized by: The dimension weight coefficients λ1, λ2, λ3, and λ4 can be automatically adjusted according to user feedback and retrieval effects through an adaptive adjustment module.

6. The patent search keyword extraction method based on technical requirements according to claim 1 is characterized by: The generation of synonyms and antonyms is based on dynamic expansion of the context, and the similarity between keywords and synonym candidates is calculated through the BERT model. The formula is as follows: ； Among them, w is the keyword, si is the synonym candidate, and D is the technical requirement document.

7. The patent search keyword extraction method based on technical requirements according to claim 6 is characterized by: The most relevant synonyms and antonyms are dynamically filtered out according to the similarity threshold Threshold.

8. The patent search keyword extraction method based on technical requirements according to claim 1 is characterized by: The keyword optimization steps include: 1) Patent library frequency matching: count the frequency of occurrence of each keyword in patent documents; 2) Retrieval performance prediction: Calculate the retrieval performance of keywords and give priority to keyword combinations with higher performance.

9. A patent search keyword extraction system based on technical requirements using the method described in any one of claims 1 to 8, characterized in that: include: 1) Text preprocessing module: used to clean and format technical requirement documents; 2) Technical requirements analysis module: used to analyze the technical requirements in the document and identify key technical points; 3) Keyword extraction module: extract technical keywords based on multiple dimensions; 4) Synonym and antonym generation module: dynamically generate synonyms and antonyms related to keywords; 5) Keyword optimization module: optimize keywords and generate the final search keyword list.

10. A patent search keyword extraction system based on technical requirements according to claim 9, characterized in that: The keyword extraction module dynamically adjusts the weight coefficients of each dimension through an adaptive weight adjustment mechanism to improve the accuracy of keyword extraction.

Images