Retrieval method and system based on semantic similarity

Abstract

The invention discloses a retrieval method and system based on semantic similarity, and relates to the field of patent retrieval, and the method comprises the following steps: S1, obtaining a retrieval keyword input by a user; s2, the retrieval keywords are preprocessed, wherein stop words are removed, and word forms are restored; s3, converting the preprocessed retrieval keywords into word vectors; s4, extracting a patent text from the patent database, performing preprocessing same as the retrieval keyword on the patent text, and then converting the patent text into a word vector; and S5, calculating the semantic similarity between the word vector of the retrieval keyword and the word vector of the patent text. The method has remarkable beneficial effects. The keyword and the patent text are converted into the word vectors, and retrieval is performed based on the semantic similarity, so that the problems that synonyms and synonyms cannot be accurately matched and complex semantics are difficult to mine in traditional keyword matching retrieval can be effectively solved, and the accuracy and comprehensiveness of patent retrieval are greatly improved.

Description

Technical Field

[0001] The present invention relates to the field of patent retrieval technology, and in particular to a retrieval method based on semantic similarity and a system thereof. Background Art

[0002] In the field of patent information retrieval, traditional patent search methods primarily rely on keyword matching. However, this approach has numerous limitations. For example, when users use different synonyms or near-synonyms to express the same technical concept, traditional methods may not accurately retrieve relevant patents. Furthermore, for some technical fields with complex semantics, simple keyword matching is difficult to uncover potentially relevant patent documents. As the number of patents continues to increase, users have placed higher demands on the accuracy and comprehensiveness of patent searches, and traditional search methods are no longer able to meet these demands. Summary of the Invention

[0003] The purpose of the present invention is to provide a retrieval method and system based on semantic similarity in order to solve the above problems.

[0004] The present invention achieves the above-mentioned purpose through the following technical solutions:

[0005] A retrieval method based on semantic similarity comprises the following steps:

[0006] S1: Get the search keywords entered by the user;

[0007] S2: Preprocessing the search keywords, including removing stop words and lemmatization;

[0008] S3: Convert the preprocessed search keywords into word vectors;

[0009] S4: Extract patent text from the patent database and perform search keywords on the patent text

[0010] After the same preprocessing, it is converted into word vectors;

[0011] S5: Calculate the semantic similarity between the search keyword word vector and the patent text word vector;

[0012] S6: Sort the patents according to semantic similarity and return the sorting results to the user.

[0013] Preferably, the pre-trained word vector model is a Word2Vec, GloVe or BERT model.

[0014] Preferably, the step of calculating semantic similarity adopts a cosine similarity algorithm, a Euclidean distance algorithm or a Manhattan distance algorithm for calculation.

[0015] Preferably, after sorting the patents, the method further includes: screening the sorting results to remove patents whose semantic similarity with the search keywords is lower than a preset threshold.

[0016] Preferably, after obtaining the search keywords input by the user, the method further includes: performing grammatical analysis on the search keywords to extract core concepts from the keywords.

[0017] The present invention also provides a retrieval system based on semantic similarity, comprising:

[0018] Input module, used to obtain search keywords input by users;

[0019] A preprocessing module, used for preprocessing the search keywords and the patent text extracted from the patent database;

[0020] Vector conversion module, used to convert pre-processed search keywords and patent text into word vectors;

[0021] Similarity calculation module, used to calculate the semantic similarity between the search keyword word vector and the patent text word vector;

[0022] Ranking module, used to sort patents based on semantic similarity;

[0023] The output module is used to return the sorting results to the user.

[0024] Preferably, the vector conversion module uses a pre-trained word vector model to perform word vector conversion.

[0025] In summary, the beneficial effects of the present invention are as follows: compared with traditional patent search methods, the present invention has significant beneficial effects. By converting keywords and patent texts into word vectors and searching based on semantic similarity, it is possible to effectively solve the problems of synonyms and near-synonyms that cannot be accurately matched and complex semantics that are difficult to mine in traditional keyword matching searches, thereby greatly improving the accuracy and comprehensiveness of patent searches. At the same time, the patent search system provided by the present invention, through a modular design, has clear division of labor and mutual cooperation among modules, and can efficiently realize patent searches based on semantic similarity, providing users with higher quality and more convenient patent search services, helping users to quickly and accurately obtain the required patent information, and promoting technological innovation and knowledge sharing. DETAILED DESCRIPTION

[0026] To make the objectives, technical solutions, and advantages of the present invention clearer, the technical solutions of the present invention will be described in detail below. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all of the embodiments. All other implementation manners obtained by those of ordinary skill in the art based on the embodiments in the present invention without creative efforts fall within the scope protected by the present invention.

[0027] The patent retrieval method based on semantic similarity provided by the present invention comprises the following specific steps:

[0028] First, obtain the retrieval keywords input by the user. The user inputs the keywords related to the technology to be retrieved through the retrieval interface. These keywords can be one or more words, used to express the technical topics that the user is interested in.

[0029] Next, perform preprocessing on the retrieval keywords. The preprocessing process includes removing stop words and word form reduction. Stop words refer to words that appear frequently in the text but have little effect on expressing semantics, such as "de", "le", "zai", etc. Removing stop words can reduce redundant information and improve the retrieval efficiency. Word form reduction is to restore different forms of words (such as the past tense of verbs, plural forms of nouns, etc.) to their basic forms for subsequent unified semantic analysis.

[0030] Then, convert the preprocessed retrieval keywords into word vectors. A word vector is a representation method that maps words to a vector space. In this way, the semantic information of words can be quantified into vector form for convenient mathematical calculation and comparison. Here, a pre-trained word vector model is used for conversion. Common pre-trained word vector models include Word2Vec, GloVe, and BERT, etc. These models can learn the semantic relationships between words through training on large-scale text data, and thus accurately convert words into vectors with semantic information.

[0031] After that, extract patent texts from the patent database. The patent database stores a large number of patent documents, containing rich technical information. Perform the same preprocessing steps on the extracted patent texts as the retrieval keywords, that is, remove stop words and word form reduction, and then convert them into word vectors.

[0032] Next, the semantic similarity between the search keyword word vector and the patent text word vector is calculated. This can be done using a variety of algorithms, such as the cosine similarity algorithm, the Euclidean distance algorithm, or the Manhattan distance algorithm. The cosine similarity algorithm measures the similarity between two vectors by calculating the cosine value of the angle between them. The smaller the angle, the closer the cosine value is to 1, indicating a higher semantic similarity between the two vectors. The Euclidean distance algorithm calculates the straight-line distance between two vectors in vector space; the closer the distance, the higher the semantic similarity. The Manhattan distance algorithm calculates the sum of the differences between the two vectors in each dimension. Similarly, the smaller the distance, the higher the semantic similarity.

[0033] Finally, the patents are sorted based on semantic similarity and the ranking results are returned to the user. Patents with higher similarity are ranked higher in the ranking, allowing users to prioritize viewing patent documents most relevant to their search intent.

[0034] The above method can be further optimized. For example, after ranking patents, the sorted results can be screened to remove patents whose semantic similarity to the search keywords falls below a preset threshold. This can further improve the relevance and quality of the search results and reduce the time users spend browsing irrelevant patents. After obtaining the search keywords entered by the user, grammatical analysis of the search keywords can be performed to extract the core concepts within the keywords, helping to more accurately understand the user's search intent and thus improve search accuracy.

[0035] Based on the above patent search method, the present invention also provides a patent search system based on semantic similarity, which includes the following modules:

[0036] The input module is used to obtain the search keywords input by the user and provide the data source for the entire search process. Through this module, users can input the technical topics they are interested in into the system in the form of keywords.

[0037] The preprocessing module is responsible for preprocessing search keywords and patent text extracted from the patent database. By removing stop words and performing lemmatization, it improves the quality and consistency of the data, laying the foundation for subsequent semantic analysis.

[0038] The vector conversion module uses a pre-trained word vector model to convert the pre-processed search keywords and patent text into word vectors, realizing the conversion from text to semantic vectors, so that the semantic information of the text can be processed and compared in mathematical form.

[0039] The similarity calculation module uses the cosine similarity algorithm, Euclidean distance algorithm, or Manhattan distance algorithm to calculate the semantic similarity between the search keyword word vector and the patent text word vector, and quantify the semantic relevance between different patents and the search keyword.

[0040] The sorting module sorts patents according to semantic similarity, placing patents with high similarity at the front so that users can view the most relevant patent documents first.

[0041] The output module returns the sorting results to the user, allowing the user to intuitively obtain the retrieved patent information.

[0042] Example:

[0043] Suppose a user wants to search for patents related to "smart watches".

[0044] First, the user enters "smart watch" as a search keyword through the input module in the search interface.

[0045] The input module passes the keyword to the preprocessing module, which preprocesses the keyword "smartwatch." Since this keyword does not contain stop words or word forms that need to be restored, it directly enters the vector conversion module. The vector conversion module uses the pretrained Word2Vec model to convert "smartwatch" into its corresponding word vector.

[0046] At the same time, the preprocessing module extracts all patent texts from the patent database, performs preprocessing operations such as removing stop words and lemmatization on these patent texts, and then converts them into word vectors by the vector conversion module.

[0047] Then, the similarity calculation module uses the cosine similarity algorithm to calculate the cosine similarity between the word vector of "smart watch" and the word vector of each patent text, and obtains the semantic similarity value between each patent and the search keyword.

[0048] The sorting module sorts patents according to their semantic similarity values, placing patents with higher similarity values at the front.

[0049] Finally, the output module returns the sorted patent results to the user, who can browse these patent documents and view technical information related to "smart watches".

[0050] In real-world applications, users may enter more complex search keywords, such as "smartwatch with health monitoring function." In this case, the preprocessing module removes stop words, performs lemmatization on the keywords, and extracts core concepts such as "smartwatch" and "health monitoring function" through grammatical analysis. Then, the preprocessing module performs word embedding conversion, semantic similarity calculation, ranking, and output according to the aforementioned process.

[0051] Different software and hardware implementations can be used for each module of the patent search system. For example, the input module can be implemented based on a web form or a mobile application interface; the preprocessing module can use the NLTK (Natural Language Toolkit) or spaCy library in the Python language to complete operations such as stop word removal and lemmatization; the vector conversion module can call the TensorFlow or PyTorch framework to load pre-trained word vector models; the similarity calculation module can use the NumPy library to calculate algorithms such as cosine similarity; the sorting module can use common sorting algorithms such as quick sort or merge sort; and the output module can present the search results to the user through a web page display or a mobile application interface. The various modules of the system can be deployed on a server and interact with the user terminal through the network to achieve efficient and stable patent search services.

[0052] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any modifications or substitutions that can be easily conceived by a person skilled in the art within the technical scope disclosed in the present invention should be included in the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be based on the scope of protection of the claims.

Claims

1. A retrieval method based on semantic similarity, characterized in that: The following steps are involved: S1: Get the search keywords entered by the user; S2: Preprocessing the search keywords, including removing stop words and lemmatization; S3: Convert the preprocessed search keywords into word vectors; S4: Extract patent text from the patent database, perform the same preprocessing as the search keywords on the patent text, and then convert it into a word vector; S5: Calculate the semantic similarity between the search keyword word vector and the patent text word vector; S6: Sort the patents according to semantic similarity and return the sorting results to the user.

2. The retrieval method based on semantic similarity according to claim 1, characterized in that: The pre-trained word vector model is Word2Vec, GloVe or BERT model.

3. The retrieval method based on semantic similarity according to claim 2, characterized in that: The step of calculating semantic similarity adopts cosine similarity algorithm, Euclidean distance algorithm or Manhattan distance algorithm for calculation.

4. The retrieval method based on semantic similarity according to claim 1, characterized in that: After sorting the patents, the process also includes: screening the sorting results to remove patents whose semantic similarity with the search keywords is lower than a preset threshold.

5. The retrieval method based on semantic similarity according to claim 4, characterized in that: After obtaining the search keywords input by the user, the process also includes: performing grammatical analysis on the search keywords and extracting the core concepts in the keywords.

6. A retrieval system based on semantic similarity, characterized in that: include: Input module, used to obtain search keywords input by users; A preprocessing module, used for preprocessing the search keywords and the patent text extracted from the patent database; Vector conversion module, used to convert pre-processed search keywords and patent text into word vectors; Similarity calculation module, used to calculate the semantic similarity between the search keyword word vector and the patent text word vector; Ranking module, used to sort patents based on semantic similarity; The output module is used to return the sorting results to the user.

7. The retrieval system based on semantic similarity according to claim 6, characterized in that: The vector conversion module uses a pre-trained word vector model to perform word vector conversion.