Synonymous sentence recognition method based on semantic feature analysis
By combining machine translation and BERT-Whitening technology with feature analysis of the original sentence and the translated text, the problem of poor performance of deep learning models in synonym recognition is solved, achieving more efficient synonym recognition and resource conservation.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-12-04
- Publication Date
- 2026-03-10
AI Technical Summary
Existing deep learning models do not perform well in synonym recognition, making it difficult to support large-scale practical applications, and there is limited room for improvement.
By employing machine translation and BERT-Whitening technology, combined with feature analysis of the original sentence and its translation, and through preprocessing, embedded representation, and similarity calculation, the performance of synonym recognition is improved.
It improves the accuracy and efficiency of synonym recognition, reduces resource consumption, optimizes data representation capabilities, and adapts to practical application needs.
Smart Images
Figure CN121638249A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of document processing technology, and in particular relates to a method for identifying synonyms based on semantic feature analysis. Background Technology
[0002] Synonym recognition refers to identifying word and phrase pairs with synonymous relationships. It is a fundamental technology in information resource management and services. In information retrieval, this technology helps improve the relevance of search results for phrase or word-based queries; in automated question-answering services, it helps improve the semantic matching effect between user questions and question-answer pairs in the resource database; and in academic misconduct detection services, it helps identify academic misconduct such as paraphrasing or plagiarism. Current synonym recognition methods are mainly based on the application of deep learning technology. Most mainstream deep learning models have been adjusted and improved for synonym recognition, improving the performance compared to earlier solutions. However, the results are still not ideal and cannot effectively support large-scale practical applications. Until further breakthroughs in natural language processing and machine learning, the potential for improvement in synonym recognition performance under existing technological approaches is relatively limited. Summary of the Invention
[0003] The purpose of this invention is to address the shortcomings of existing technologies and meet practical application needs. This invention proposes a semantic sentence recognition method that fully utilizes machine translation, BERT-Whitening, and other technologies, combining the features exhibited by synonyms in the original sentence and its translation text to perform synonym analysis, thereby obtaining semantic representations of words and sentences in various forms. Based on these representations, the method determines whether words and sentences are synonymous, thus improving the effectiveness of synonym sentence recognition.
[0004] To achieve the above objectives, the present invention adopts the following technical solution.
[0005] A synonym recognition method based on semantic feature analysis includes the following steps:
[0006] Step 1, sentence preprocessing, refers to segmenting and concatenating the input text to create a sentence set; standardizing the sentences, removing invalid data, and performing word segmentation.
[0007] Step 2, candidate synonym identification, refers to extracting clue words for candidate synonyms based on the word segmentation results; screening the preliminary candidate set containing clue words for preliminary filtering; and establishing a sentence set index.
[0008] Step 3, Embedded Text Representation, refers to using machine translation technology to generate translations and then embedding the original sentences and translations into embedded representations to obtain multi-form text representation results.
[0009] In a further improvement or preferred implementation of the aforementioned synonym recognition method based on semantic feature analysis, step A1, specifically, involves segmenting and concatenating the input text: preliminary segmentation based on standard text symbols, using text symbols in the text content as the basic dividing structure for text rows and columns; establishing a sentence set specifically refers to obtaining a preliminary set of long sentences to be recognized based on the segmentation results.
[0010] A further improvement or preferred implementation of the aforementioned synonym recognition method based on semantic feature analysis also includes a step of expanding the sentence set, specifically including:
[0011] a. Segmentation based on length threshold: For natural sentences containing multiple clauses, the sentence is segmented, and the clauses with a length exceeding a certain threshold are treated as independent sentences to be identified. The clauses within the natural sentence are then concatenated to expand the set of sentences to be identified.
[0012] b. Expand the set of sentences to be identified by splicing consecutive natural sentences into groups, taking paragraphs as the unit.
[0013] A further improvement or preferred implementation of the aforementioned synonym recognition method based on semantic feature analysis includes step A2, which involves standardizing the sentence by removing parentheses, parenthetical content, footnotes, endnotes, and citation superscripts from the sentence.
[0014] In a further improvement or preferred embodiment of the aforementioned synonym recognition method based on semantic feature analysis, the word segmentation process in step A2 further includes: segmenting each sentence to be recognized into words, and combining adjacent word segmentation results into compound words if they can be combined into compound words.
[0015] In a further improvement or preferred implementation of the aforementioned synonym recognition method based on semantic feature analysis, the clue words in step two are obtained by removing function words, quantifiers, and degree words from the word segmentation results;
[0016] The preliminary filtering refers to selecting statements containing at least one clue word from the statement set to form a preliminary candidate set. For each statement, candidate statements with a length difference of more than 50% are filtered out, and the remaining statements are used as candidate synonyms.
[0017] The establishment of the statement set index refers to extracting all terms from the statement set to form an index dictionary; and establishing an inverted record table for each term in the index dictionary, wherein the inverted record table stores the statement ID of the term and its position code in the statement.
[0018] In a further improvement or preferred implementation of the aforementioned synonym recognition method based on semantic feature analysis, step three specifically includes:
[0019] C1. Machine translation-based translation generation refers to the translation of multilingual sentences using a translation model based on deep neural networks.
[0020] C2. BERT-Whitening-based vectorization refers to text vectorization based on the BERT pre-trained model.
[0021] C3, Sentence Vector Whitening, refers to extracting the original sentence vector set obtained from the BERT pre-trained model in step C2, and transforming the sentence vectors output by the model on the test set data based on the weight matrix and the mean vector set to obtain the whitened sentence vectors.
[0022] C4. Semantic storage of multiple representation results refers to the semantic storage of text using tuples based on the sentence vectors to generate results.
[0023] In a further improvement or preferred implementation of the aforementioned synonym recognition method based on semantic feature analysis, in step three,
[0024] Step C2 also includes: based on the BERT pre-trained model, modifying the input text using different pre-trained mask templates, designing a segmentation instruction template Prompt to provide input to the model to guide it in generating synonymous sentences, using the designed Prompt template to randomly replace words in the sentence that do not have connection meaning to enhance the explicitness of the sentence; and weighting the word vectors in the hidden layer sentence based on the average vector of the last layer or special words to output the sentence vector.
[0025] Step C3 specifically includes: extracting the original sentence vector set obtained from the BERT pre-trained model in step C2, and randomly selecting sentence vectors to establish a training set. ,
[0026] Calculate the mean vector set corresponding to the training set. ;
[0027] Furthermore, the covariance matrix of the training set is obtained. ;
[0028] Define a set of sentence vectors that conforms to a standard normal distribution. ,make ;in, This is the weight matrix;
[0029] The orthogonal matrix obtained by SVD decomposition of the covariance matrix and diagonal array Then the weight matrix ;
[0030] The sentence vectors output by the test set data model are transformed based on the weight matrix and the mean vector set to obtain the whitened sentence vectors.
[0031] In step C4, the tuple is represented as: (sen, (ori, em-vec), (trans, em-vec), (trans, em-vec), ...); where the parameter sen stores the content of the sentence to be recognized, em-vec stores the sentence vector, and in (ori, em-vec) ori indicates that the sentence vector of the original sentence is stored here; in (trans, em-vec) the value of the parameter trans is the language of the translation, and em-vec is its corresponding sentence vector.
[0032] A further improvement or preferred implementation of the aforementioned synonym recognition method based on semantic feature analysis, specifically includes the following steps in step four:
[0033] D1. Similarity Calculation: Calculate the similarity between the original sentences of the statement to be processed and the candidate synonyms, as well as between pairs of translations in the same language using the same translation tool. For the statement to be processed (a) and the candidate synonym (b), calculate the sentence vectors of the two corresponding statements. and And further calculate the cosine similarity. ;in Represents the magnitude of the sentence vector;
[0034] D2. Similarity fusion refers to the fusion of the similarity between the original words / sentences and the similarity between the translated texts; the similarity results of the original words / sentences are directly retained, and the maximum similarity of the translated texts is taken as the translation similarity.
[0035] Represented as ;
[0036] Where sim(a,b)m represents the similarity between the original statements a and b to be analyzed. Let be the similarity of the wording and phrases in the translations of the original sentences a and b under the i-th language category;
[0037] D3. Synonym judgment refers to setting similarity thresholds for the original sentence and the translations in different languages based on the obtained similarity results, and judging whether the similarity between the sentences exceeds the threshold; then, the judgment results of the two sides are merged and judged, and if either side judges it as synonym, it is regarded as a synonym relationship.
[0038] In a further improvement or preferred implementation of the aforementioned synonym recognition method based on semantic feature analysis, the original word-sentence similarity threshold is set to a lower value compared to the translations in each language.
[0039] Its beneficial effects are as follows:
[0040] (1) Based on preprocessing and machine translation technology, efficient data processing of words, sentences and their translations was realized, providing a technical foundation for the practical application of synonym recognition methods;
[0041] (2) Based on the semantic representation of generated words and sentences and their translations, it is possible to obtain semantic representation results of words and sentences in various forms. At the same time, it makes full use of the synonym features in the original words and sentences and the translated words to further enhance the accuracy of synonym recognition and recognition path, enhance the judgment of whether words and sentences are synonymous, and improve the effect of synonym recognition.
[0042] (2) By preprocessing the input text by segmentation, standardization and word segmentation, the complexity of subsequent word and sentence recognition and synonym judgment is effectively reduced, the resource consumption and technical difficulty of the solution are reduced, and the necessary foundation is provided for the practical application of the technical solution.
[0043] (3) Whitening is introduced into the embedded representation. Spatial transformation operation is performed on the sentence vectors obtained based on BERT to map them to the standard normal distribution, which enhances the semantic representation ability of words and sentences, optimizes the appearance features of data, and makes them more recognizable. Attached Figure Description
[0044] Figure 1 This is a schematic diagram of the scheme principle of the synonym recognition method of this application. Detailed Implementation
[0045] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention. Furthermore, the technical features involved in the various embodiments of this invention described below can be combined with each other as long as they do not conflict with each other.
[0046] This invention relates to a synonym recognition method based on semantic feature analysis. It is mainly used to provide a synonym recognition and analysis method that fully considers the current state of computer technology, and achieves a practical application through various optimization approaches and improvement schemes. This method consumes less system hardware and software resources and can be effectively implemented in various online and offline environments.
[0047] The basic principle of this application is as follows: Figure 1 As shown, the following section provides a detailed explanation of its basic steps.
[0048] With the rapid development of the expression and use of online language, the text expression forms in various industries and scenarios are becoming increasingly complex due to the need for text expression effects and various personalized needs. However, in the process of text synonym recognition, a large amount of content in words and sentences does not directly affect the synonym judgment and analysis, but the existence of this content will greatly consume the system's software and hardware resources, and at the same time limit the synonym recognition effect of words and sentences. In order to reduce the application cost and resource consumption of the solution, and reduce the dependence on the actual application technology and hardware equipment, it is necessary to perform corresponding preprocessing.
[0049] Step 1: Word and Sentence Preprocessing
[0050] The input text is segmented and concatenated. First, it is segmented by natural sentences, that is, by periods, exclamation marks, and question marks, to obtain a preliminary set of words and sentences to be identified. Then, natural sentences containing multiple clauses are segmented, and clauses exceeding a certain threshold in length are treated as independent words and sentences to be identified. The clauses within the natural sentences are then concatenated to expand the set of words and sentences to be identified. Finally, the set of words and sentences to be identified is expanded by concatenating consecutive natural sentences into groups, on a paragraph basis.
[0051] Standardizing words and sentences requires tailored methods based on the characteristics of the text being processed. For example, academic texts need to remove sentence headings, parentheses and their contents, as well as footnotes, endnotes, and citation superscripts to avoid interference. Based on this, each sentence to be identified is segmented into words to support the identification of candidate synonyms.
[0052] Based on this, each sentence to be identified is segmented into words, thus providing support for the identification of candidate synonyms. During this process, if adjacent segmentation results can be combined into compound words, they should be combined to further reduce the size of candidate synonyms and improve the efficiency of synonym identification.
[0053] A1. Segment and concatenate the input text.
[0054] Preliminary segmentation based on standard text symbols: The indispensable text symbols in the text content constitute the basic division structure of the text rows and columns. Preliminary segmentation based on standard text symbols means that the input text fragments are initially segmented according to standard text symbols such as periods, exclamation marks, and question marks to obtain a preliminary set of long sentences to be identified.
[0055] A2. Standardize wording and sentence structure.
[0056] It is necessary to design targeted methods based on the characteristics of the text to be processed. For example, for academic texts, it is necessary to remove the ordinal number at the beginning of the sentence, the parentheses and the content inside the parentheses, as well as indirect or auxiliary text information such as footnotes, endnotes, and citation superscripts, in order to avoid interference from the above information.
[0057] Based on this, each sentence to be identified is segmented into words to support the identification of candidate synonyms. During the segmentation process, if adjacent segmentation results can be combined into compound words, they should be combined to further reduce the size of candidate synonyms and improve the efficiency of synonym identification.
[0058] Traditional technical solutions typically rely directly on the word segmentation results of the original text and employ various semantic recognition and analysis techniques for processing. However, due to limitations in the capacity and coverage of the original text database, it is difficult to guarantee effective and comprehensive synonym analysis processing. Using a complete corpus involves a massive amount of basic data, which is not suitable for general applications. Therefore, it is necessary to optimize this process accordingly, as described in the solution adopted in step two of this application.
[0059] Step 2: Candidate Synonym Identification
[0060] B1. Based on the word segmentation results in step one, filter out function words, quantifiers, and degree words, and use the remaining content words as clue words to find candidate synonyms.
[0061] B2. Obtain synonyms for all content words based on the thesaurus, thereby expanding the set of clue words and improving the comprehensiveness of candidate synonym sentences;
[0062] B3. Based on this, select sentences containing at least one clue word from the sentence set to form a preliminary candidate set. Utilize the characteristic that synonym sentences are similar in length, filter out candidate sentences with a length difference of more than 50% for each sentence, and use the remaining sentences as candidate synonym sentences.
[0063] To improve the efficiency of candidate synonym discovery based on clue words, it also includes methods for building a sentence set index to facilitate synonym retrieval based on clue words, specifically including:
[0064] Extract all terms from the statement set to form an index dictionary; create an inverted record table for each term in the index dictionary, which stores the statement ID of the term and its position code in the statement; by creating the index dictionary, data processing efficiency can be effectively improved and system computing power consumption can be reduced during subsequent model training and optimization, while improving model efficiency, reducing training and optimization time, further enhancing the compatibility and application capability of this solution, and reducing application costs.
[0065] Step 3: Embedded text representation.
[0066] Due to the differences in text representation and the characteristics of character shapes and meanings, words and sentences in different linguistic environments have significantly different semantic features. It is difficult to achieve a comprehensive synonym recognition scheme that meets the needs of multiple languages and multiple contexts by relying on only one or a few semantic analysis or image recognition technologies. In order to improve the comprehensiveness of synonym recognition, it is proposed to generate translations and perform embedded representations using machine translation technology while directly embedding the original sentence, thereby obtaining text representation results in multiple forms, making full use of the expressive power of synonym features in different translated texts, and expanding the synonym recognition scheme.
[0067] The specific implementation process includes:
[0068] C1. Translation generation based on machine translation
[0069] Translation models based on deep neural networks are used to complete multilingual sentence translation. In current practical applications, the most effective translation model is the one based on deep neural networks. The basic characteristic of this model is that the richer the training data, the more ideal the translation effect. Therefore, in the selection of the target language, priority should be given to foreign languages with relatively rich bilingual data. Using a single language may not achieve the best results. Therefore, the implementation strategy of multilingual translation can better realize the synonym analysis of sentences.
[0070] C2. Vectorized representation based on BERT-Whitening
[0071] Text vectorization representation based on BERT pre-trained model;
[0072] Based on the BERT pre-trained model, the input text is modified using different pre-trained mask templates. A word segmentation indicator template, Prompt, is designed to provide input to the model to guide it in generating synonyms and sentences. Connectors in the designed Prompt template are used to randomly replace words in the sentences that do not have connecting meaning to enhance the explicitness of words and sentences. The word vectors in the hidden layer are weighted based on the average vector of the last layer or special words to output the sentence vector.
[0073] C3. Sentence vector whitening processing
[0074] Extract the original sentence vector set obtained from the BERT pre-trained model in step C2, and randomly select sentence vectors to build a training set. ,
[0075] Calculate the mean vector set corresponding to the training set. ;
[0076] Furthermore, the covariance matrix of the training set is obtained. ;
[0077] To solve for the sentence vector set that conforms to the standard normal distribution ,make ;in, This is the weight matrix;
[0078] The orthogonal matrix obtained by SVD decomposition of the covariance matrix and diagonal array Then the weight matrix ;
[0079] The sentence vectors output by the test set data model are transformed based on the weight matrix and the mean vector set to obtain the whitened sentence vectors.
[0080] C3. Semantic storage of multiple representations of results;
[0081] Based on the sentence vector generation results, a tuple (sen, (ori, em-vec), (trans, em-vec), (trans, em-vec), ...) is established to represent the text and store it. Here, the parameter sen stores the content of the word / sentence to be identified, em-vec stores the sentence vector, and in (ori, em-vec), ori indicates that the sentence vector of the original word / sentence is stored here; in (trans, em-vec), the value of the trans parameter is the language of the translation, and em-vec is its corresponding sentence vector.
[0082] Step 4: Identifying Synonyms
[0083] The purpose of this step is to calculate the semantic similarity between the statement to be processed and the candidate statements, and on this basis, to determine whether there is a synonym relationship between the two. Its implementation steps can be divided into three steps: cosine similarity calculation, similarity fusion and synonym judgment.
[0084] D1. Similarity Calculation
[0085] Calculate the similarity between the original words and sentences to be processed and the candidate synonyms, as well as between pairs of translations in the same language using the same translation tool;
[0086] Specifically: Based on the sentence vectors determined in step three, for the sentence to be analyzed (a) and the candidate synonym (b), calculate the sentence vectors of the two corresponding sentences respectively. and And further calculate the cosine similarity. ;in Represents the magnitude of the sentence vector;
[0087] D2, Similarity Fusion
[0088] Based on the initial similarity calculation, the calculation results need to be merged to generate a similarity result;
[0089] Given the significant differences between the original text and the translation, the similarity fusion is divided into two categories: original text similarity and translation similarity. The former directly retains the similarity results, while the latter takes the maximum similarity as the translation similarity.
[0090] Let sim(a,b) be the similarity between words and phrases a and b, sim(a,b)m be the similarity between the original words and phrases, and sim(a,b)1, sim(a,b)2, ..., sim(a,b)n-1, sim(a,b)n be the similarity results of n translations.
[0091] but ;
[0092] D3. Synonym Judgment
[0093] Based on the similarity results, similarity thresholds are set for the original words and sentences and the translations in different languages to determine whether the similarity between words and sentences exceeds the threshold.
[0094] Then, the judgment results of the two parties are merged and judged. If either party judges them as synonyms, they are considered to be synonymous.
[0095] In particular, the threshold setting is a key factor affecting the synonym recognition effect; in order to improve the recognition effect, the original word and sentence similarity threshold is set to a lower value compared to the translation of each language, so as to maximize the recall rate of synonym recognition while better ensuring accuracy.
[0096] Setting a relatively higher similarity threshold for the translated text can achieve a certain level of synonym recall without causing false positives, especially when the similarity between the original words and sentences is low. This improves the comprehensiveness of synonym recognition without significantly affecting accuracy.
[0097] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit the scope of protection of the present invention. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the essence and scope of the technical solutions of the present invention.
Claims
1. A method for identifying synonymous sentences based on semantic feature analysis, characterized in that, Comprise the following steps: Step one, sentence preprocessing, refers to the input text segmentation and splicing, the establishment of sentence set; the sentence is standardized, invalid data and word segmentation processing are removed; Step two, candidate synonymous sentence recognition, refers to the extraction of clue words from the word segmentation results; the preliminary candidate set containing the clue words is screened and preliminarily filtered to establish the sentence set index; Step three, embedded text representation, refers to generating translation by using machine translation technology, and embedding the original sentence and translation respectively to obtain multi-modal text representation results. 2.The semantic feature analysis based synonymous sentence identification method according to claim 1, wherein, In step A1, The segmentation and splicing of the input text specifically refers to: based on the preliminary segmentation of standard text symbols, the text symbols in the text content are used as the basic division structure of text rows and columns for segmentation; the establishment of sentence set specifically refers to: according to the segmentation results, a preliminary long sentence set to be recognized is obtained. 3.The semantic feature analysis based synonymous sentence identification method according to claim 2, characterized in that, It also includes the step of expanding the sentence set, specifically including: a. Based on the length threshold, the paragraph is segmented, the natural sentence containing multiple clauses is segmented, the clause whose length exceeds a certain threshold is taken as an independent sentence to be recognized, and the clauses in the natural sentence are spliced to expand the sentence set to be recognized; b. In units of paragraphs, the expansion of the sentence set to be recognized is realized by splicing the continuous natural sentences. 4.The semantic feature analysis based synonymous sentence identification method according to claim 1, wherein, The step A2, the standardization processing of the sentence includes: removing the brackets, the content in the brackets, the footnotes, the endnotes and the quotation marks in the sentence. 5.The semantic feature analysis based synonymous sentence identification method according to claim 1, wherein, The word segmentation processing in step A2 also includes: for each sentence to be recognized, if the adjacent word segmentation results can be combined into a compound word, it will be combined. 6.The semantic feature analysis based synonymous sentence identification method according to claim 1, wherein, The clue word in step two is obtained by removing the function word, measure word and degree word in the word segmentation results; The preliminary filtering refers to screening at least one sentence containing a clue word from the sentence set to form a preliminary candidate set, and filtering out the candidate sentences with a length difference of more than 50% for each sentence, and taking the remaining sentences as candidate synonymous sentences; The establishment of sentence set index refers to extracting all word items in the sentence set to form an index dictionary; An inverted record table is established for each word item in the index dictionary, which saves the word item ID and the position code in the sentence. 7.The semantic feature analysis based synonymous sentence identification method according to claim 1, wherein, The step three specifically includes: C1, translation generation based on machine translation, refers to the completion of multilingual sentence translation based on deep neural network translation model; C2, vectorization representation based on BERT-Whitening, refers to text vectorization representation based on BERT pre-training model; C3, sentence vector whitening processing, refers to extracting the original sentence vector set obtained based on the BERT pre-training model in step C2, transforming the sentence vector output by the test set data model based on the weight matrix and mean vector set to obtain the sentence vector after whitening processing; C4, semantic storage of multiple representation results, refers to establishing semantic storage using multiple groups to represent text according to the sentence vector generation results. 8.The semantic feature analysis based synonymous sentence identification method according to claim 7, characterized in that, In step three, The step C2 further comprises: on the basis of the BERT pre-training model, input text is modified by using different pre-training mask templates, a word segmentation instruction template Prompt is designed to provide input to the model to guide it to generate synonymous sentences, and a designed Prompt template connecting word is used to replace words without connecting meaning in the sentence to enhance the visualization of the sentence; and a weighted processing is performed on the word vector in the hidden layer based on the last layer average vector or special word to output a sentence vector; The step C3 specifically comprises: extracting the original sentence vector set obtained based on the BERT pre-training model in the step C2, and randomly extracting the sentence vector to establish a training set , computing a mean vector set corresponding to the training set ; Further obtain the covariance matrix of the training set ; defining a set of sentence vectors conforming to a standard normal distribution , let ; wherein, is a weight matrix; An orthogonal matrix obtained by performing SVD decomposition on the covariance matrix and a diagonal matrix , then the weight matrix ; The sentence vector output by the model is transformed based on the weight matrix and the mean vector set to obtain a white processed sentence vector; In step C4, the multi-element group is in the form of (sen, (ori, em-vec), (trans, em-vec), (trans, em-vec), ……); wherein, the parameter sen stores the content of the sentence to be recognized, and the em-vec stores the sentence vector; the ori in (ori, em-vec) indicates that the sentence vector of the original word sentence is stored here; the trans parameter in (trans, em-vec) is the language of the translation, and the em-vec is the corresponding sentence vector. 9.The semantic feature analysis based synonymous sentence identification method according to claim 1, wherein, The step four specifically comprises the following steps: D1、Similarity calculation, calculate the similarity between the original word sentence of the to-be-processed sentence and the candidate synonymous sentence and the same language translation of the same translation tool one by one; for the to-be-analyzed to-be-processed sentence a and the candidate synonymous sentence b, the sentence vectors corresponding to the two sentences are calculated respectively and , and further calculate the cosine similarity ; wherein denotes the modulus of the sentence vector; D2, similarity fusion, refers to fusing the original word sentence similarity and the translation similarity; the original word sentence similarity directly retains the similarity result, and the maximum value of the translation similarity is taken as the translation similarity; represents ; Wherein sim(a,b)m is the similarity of the original sentences a and b to be analyzed, sim(a,b)i is the similarity of the translated sentences a and b of the i-th language. D3, synonym judgment, refers to setting the original word sentence similarity threshold and the translation similarity threshold of different languages based on the obtained similarity result, judging whether the similarity between the sentences exceeds the threshold, and then fusing the judgment results of the two parties, regarding it as a synonymous relationship as long as one party judges it as synonymous. 10.The method of claim 9, wherein, The original word sentence similarity threshold is set to a lower value than the translation similarity threshold of each language.