Regulation retelling matching method and system based on deep learning and cosine similarity matching

Abstract

The invention belongs to the technical field of approximate semantic matching of texts, and particularly relates to a rule retelling matching method and system based on deep learning and cosine similarity matching. Comprising the following steps of: 1, constructing a graph structure of a text, expressing text information as a graph node mode, and expressing a relationship between vocabularies by using edges of a graph; 2, introducing a nuclear power professional vocabulary library; 3, modeling context vocabulary correlation by a language knowledge attention model, iteratively updating word representation and semantic representation, and fusing useful information of the word representation and the semantic representation; and 4, fusing word representation into a character level, matching two character sequences by using a bilateral multi-view matching mechanism, and predicting a relationship between two sentences by using a relationship classifier by using a sentence vector as an input. The method has the beneficial effects that the problem that accurate matching is difficult due to the fact that the repeated text and the original procedure text are the same in semantics and different in vocabularies during personnel retelling can be solved.

Description

Technical Field

[0001] The present invention belongs to the technical field of approximate semantic matching of texts, and in particular relates to a procedure paraphrase matching method and system based on deep learning and cosine similarity matching. Background Art

[0002] Before inspecting equipment, maintenance personnel at nuclear power plants will repeat the contents of the procedure documents to deepen their understanding of the contents of the procedure and avoid erroneous operations due to unfamiliarity. Usually, maintenance personnel reorganize the language according to their own understanding and express the meaning of the contents of the procedure. Therefore, there is a certain deviation between the repeated text and the actual text content. The traditional text matching method uses exact matching, that is, one-to-one correspondence between words, which cannot be applied to the repetition habits of nuclear power plant maintenance personnel. Secondly, nuclear power plant procedures are composed of short texts. The polysemy problem in short texts is more serious than that in long texts, because short texts usually have less contextual information, so it is difficult for the model to capture the correct semantics. Word-based models usually encounter some potential problems caused by word segmentation. Summary of the invention

[0003] The purpose of the present invention is to provide a procedure repetition matching method and system based on deep learning and cosine similarity matching, so as to solve the problem that the traditional exact matching method is inflexible and cannot achieve similar semantic matching.

[0004] The technical solution of the present invention is as follows: A procedure repetition matching method based on deep learning and cosine similarity matching comprises the following steps:

[0005] Step 1: Build a graph structure of the text, express the text information as graph nodes, and represent the relationship between words with graph edges;

[0006] Step 2: Introduction of nuclear power professional vocabulary database;

[0007] Step 3: The language knowledge attention model models the contextual vocabulary correlation, iteratively updates the word representation and semantic representation, and integrates each other’s useful information;

[0008] Step 4: Integrate word representation into character level and use bilateral multi-view matching mechanism to match two character sequences. The relation classifier takes sentence vector as input to predict the relationship between two sentences.

[0009] The step 1 includes the following:

[0010] Step 11: Contextual word embedding For each node x in the graph i , word w i The initial representation of is the attention pooling of the context character representation. Suppose the word w iIt consists of some consecutive character tokens {c t1 ,c t1+1 , …, c t2}, for each character c k (t1≤k≤t2), where k represents the character number, and the value range of the number is (t1, t2). t1 and t2 represent the first and last tokens after the characters are converted into tokens sequences. A two-layer feedforward network is used to calculate a feature-based score vector, and then the feature-based multi-dimensional softmax is used for normalization.

[0011] u k =MD-softmax k (FFN(c k )),

[0012] Among them, MD-softmax() is the function calculation formula;

[0013] Step 12: Transform the corresponding word vector c k With the normalized score u k Perform weighting to obtain the context word vector;

[0014]

[0015] Step 13: Use the Att-Pooling(·) pooling function to transform the context word vector

[0016] v i =Att-Pooling({c k |t1≤k≤t2}).

[0017] The step 2 includes the following:

[0018] Step 21: For each word w i , the word sense set is represented as S(w i )={s i,1 ,s i,2 ,…,s i,K},s i,k Yes i The kth sememe of , which means its corresponding sememe is O(s i,k ), in order to obtain each meaning s i,k The embedding vector of , first obtains the multi-dimensional attention representation of each sememe:

[0019]

[0020] Among them, MD-GAT() is a graph structure calculation function based on features, where e is the embedding vector of the semantic primitive o; i is the ordinal number of the word; k is the ordinal number of the semantic primitive; O(s i,k ) indicates the original meaning; s i,k It means meaning item. Meaning item and semantic original are two meanings. Seme original means the smallest indivisible semantic unit in linguistics. Meaning item means that each word has rational meaning. If it has several meanings, it is expressed as several meaning items.

[0021] Step 22: For each sememe s i,k , whose embedding is obtained by attention pooling of all sememe representations

[0022]

[0023] The step 3 includes the following:

[0024] Step 31: Update semantic expression

[0025] Update the feature expression vector of the current word according to the context in the sentence, and each word meaning will first be updated from x i The useful information of the forward and backward aggregated words is calculated as follows:

[0026]

[0027] Among them, x i represents the i-th word; x j represents the jth word; fw represents the feedforward network; bw represents the reverse network; is the temporal feature vector of the l-1th word, is the hidden layer feature vector of the l-2th word, The feature vector representation of the preceding and following context of the lth word, where the two multidimensional attention functions MD-GAT(·) have different parameters;

[0028]

[0029] [·,·] represents the concatenation of vectors, using GRU to control the fusion of context information and current lexical semantic information;

[0030] Step 32: Update word expression

[0031] After multiple iterations, the final word representation contains not only the context word information but also semantic knowledge. For each sentence, h i Represent the final word representation

[0032]

[0033] The step 3 includes the following:

[0034] Step 41: After obtaining the word representation enhanced by the semantic knowledge of each sentence, the word information is fused into the characters;

[0035]

[0036] Among them, Att-Pooling(·) represents pooling calculation; represents the implicit vector of the i-th word in the attention layer; a represents the hidden layer of the attention layer; represents the weight value of the i-th word in the attention layer, Represents a set of weight values; Represents the word feature vector expression with enhanced attention weight; y t The feature vector enhanced by semantic knowledge; LayerNorm(·) represents layer normalization, c a t is the contextual character representation obtained using BERT. For each character c a t , using multi-dimensional attention from sentence C a and C b Aggregate information,

[0037]

[0038] The above multi-dimensional attention function MD-GAT(·) shares the same parameters, when the two sentences match exactly:

[0039]

[0040] Step 42: Compare using multi-angle cosine distance:

[0041]

[0042] Through P distances d1, d2, ..., d P , get the final character representation

[0043]

[0044] The final character representation contains three kinds of information: context information, word meaning knowledge, and character-level similarity. For each sentence C a or C b , use the attention pooling of all the final character representations of the sentence to get the sentence representation vector r a or b；

[0045] Step 43: Predict the similarity between the original text and the paraphrase sentence

[0046] p=FFN([c CLS ,r a ,r b ,r a ⊙r b ,|r a -r b |]),

[0047] Among them, FFN(·) is a feed-forward network with two hidden layers and a sigmoid activation after the output layer. The training goal is to minimize the binary cross entropy loss:

[0048]

[0049] A procedure repetition matching system based on deep learning and cosine similarity matching includes the following modules: a data preprocessing module, a feature extraction module, a context relevance calculation module, a cosine similarity calculation module, a matching decision module, a system integration and interaction module and a data interaction module.

[0050] The data preprocessing module implements the cleaning, formatting and word segmentation of the original data to prepare for the subsequent feature extraction, removes irrelevant characters through text cleaning, uses the word segmentation tool to perform Chinese word segmentation, receives the original text data, and outputs the preprocessed text data;

[0051] The feature extraction module extracts feature vectors from the preprocessed text, converts the text into a vector representation in a high-dimensional space using a word grid graph expression module, receives the preprocessed text data, and outputs a feature vector of the text;

[0052] The context relevance calculation module performs context relevance calculation on the feature vector to ensure the consistency and comparability of the data, uses the language knowledge attention model to express the word vector and the context, receives the feature vector output by the feature extraction module, and outputs the feature vector after the context information is associated;

[0053] The cosine similarity calculation module calculates the cosine similarity between two text feature vectors to measure the similarity of the texts, obtains the cosine similarity value by calculating the dot product of the two vectors and dividing it by the product of their respective moduli, receives the standardized feature vectors, and outputs a similarity score;

[0054] The matching decision module makes a decision on whether to match based on the similarity score. Based on the sentence matching layer and the relationship classifier of the method, a similarity threshold is set. Similarities above the threshold are considered as matches, and those below the threshold are considered as mismatches. The similarity score of the cosine similarity calculation module is received and the matching result is output;

[0055] The system integration and interaction module integrates the above modules into a complete system and handles the interaction with the user interface, using the system architecture to realize the interaction between the user and the system through the API or the graphical user interface;

[0056] The data interaction module manages the data flow between modules in the entire system process, including error handling and data feedback, and the data interaction relationship between modules.

[0057] The data preprocessing module receives the original text data and passes it to the feature extraction module after processing; the feature extraction module converts the text into a feature vector and passes it to the context relevance calculation module; the context relevance calculation module models the association relationship between the feature vector of the context and the current word, and then passes it to the cosine similarity calculation module; the cosine similarity calculation module calculates the similarity score and passes the result to the matching decision module; the matching decision module makes a matching decision based on the similarity score, and displays the result to the user through the system integration and interaction module.

[0058] It also includes a word grid graph representation module, a language knowledge attention model, a sentence matching layer, and a relation classifier.

[0059] The word grid graph expression module represents the nuclear power text as a graph structure, in which the relationship between nuclear power words can be represented by the edges of the graph, and outputs the initial context representation of each word and the initial semantic representation of each word meaning;

[0060] The language knowledge attention model iteratively updates word representation and semantic representation, and integrates useful information from each other;

[0061] The sentence matching layer and relation classifier first integrate word representation into character level, and then use bilateral multi-view matching mechanism to match two character sequences. The relation classifier takes sentence vector as input to predict the relationship between two sentences.

[0062] The beneficial effect of the present invention is that it combines a nuclear power professional vocabulary library, a procedure text graph structure representation and a pre-trained language model to solve the problem that when a person repeats, the repeated text has the same semantics as the original procedure text, but different vocabulary, making it difficult to accurately match. BRIEF DESCRIPTION OF THE DRAWINGS

[0063] Figure 1 This is a flow chart of a procedure repetition matching method based on deep learning and cosine similarity matching provided by the present invention. DETAILED DESCRIPTION

[0064] The present invention is further described in detail below with reference to the accompanying drawings and specific embodiments.

[0065] In order to integrate the semantic information of words, the present invention provides a procedure repetition matching method based on deep learning and cosine similarity matching, wherein the word grid graph feature expression expresses the text information as a graph structure, and the relationship between words can be represented by the edges of the graph. The language knowledge attention model can connect the context and handle the ambiguity of words, realize fuzzy matching of words in the text, and accurately match the semantics of the text.

[0066] A procedure repetition matching method based on deep learning and cosine similarity matching comprises the following steps:

[0067] Step 1: Construct a graph structure of the text, that is, express the text information as graph nodes, and the relationship between words is represented by the edges of the graph.

[0068] Step 11: Contextual word embedding For each node x in the graph i , word w i The initial representation of is the attention pooling of the context character representation. Suppose the word w i It consists of some consecutive character tokens {c t1 ,c t1+1 , …, c t2}, for each character c k (t1≤k≤t2), where k represents the character number, the value range of the number is (t1, t2), t1 and t2 represent the first and last tokens after the character is converted into a token sequence. A two-layer feedforward network (FFN) is used to calculate a feature-based score vector, and then normalized using a feature-based multidimensional softmax (MD-softmax).

[0069] u k =MD-softmax k (FFN(c k )),

[0070] Among them, MD-softmax() is the function calculation formula.

[0071] Step 12: Transform the corresponding word vector c k With the normalized score u k Weighted, get the context word vector.

[0072]

[0073] Step 13: Use the Att-Pooling(·) pooling calculation function to transform the context word vector.

[0074] v i =Att-Pooling({c k|t1≤k≤t2}).

[0075] Step 2: Introduction of nuclear power professional vocabulary

[0076] Step 21: For each word w i , the word sense set is represented as S(w i )={s i,1 ,s i,2 ,…,s i,K}. i,k Yes i The kth sememe of , which means its corresponding sememe is O(s i,k ). In order to obtain each meaning s i,k The embedding vector of , first obtains the multi-dimensional attention representation of each sememe:

[0077]

[0078] Among them, MD-GAT() is a graph structure calculation function based on features, where e is the embedding vector of the semantic primitive o; i is the ordinal number of the word; k is the ordinal number of the semantic primitive; O(s i,k ) indicates the original meaning; s i,k Indicates meaning item. Meaning item and semantic original are two meanings. Semematic original indicates the smallest indivisible semantic unit in linguistics; meaning item indicates that each word has rational meaning. If it has several meanings, it is expressed as several meaning items.

[0079] Step 22: For each sememe s i,k , whose embedding is obtained by attention pooling of all sememe representations:

[0080]

[0081] Step 3: The language knowledge attention model models the contextual vocabulary correlation, iteratively updates the word representation and semantic representation, and fuses each other’s useful information.

[0082] Step 31: Update semantic expression

[0083] Update the feature expression vector of the current word according to the context in the sentence. Each word meaning will first be updated from x i The useful information of the forward and backward aggregated words is calculated as follows:

[0084]

[0085] Among them, x i represents the i-th word; x j represents the jth word; fw represents the feedforward network; bw represents the reverse network; is the temporal feature vector of the l-1th word, is the hidden layer feature vector of the l-2th word, The feature vector expression of the preceding and following context of the lth word. The two multidimensional attention functions MD-GAT(·) have different parameters.

[0086]

[0087] [·,·] represents the concatenation of vectors. It is worth noting that m is not used directly. i，k , the reason is m i,k It only contains context information, and a gate, namely GRU, is needed to control the fusion of context information and current vocabulary semantic information.

[0088] Step 32: Update word expression

[0089] After multiple iterations, the final word representation contains not only contextual word information but also semantic knowledge.

[0090] For each sentence, use h i Represent the final word representations respectively.

[0091]

[0092] Step 4: Integrate word representation into character level and use bilateral multi-view matching mechanism to match two character sequences. The relation classifier takes sentence vector as input and predicts the relationship between two sentences.

[0093] Step 41: After obtaining the word representations enhanced with semantic knowledge of each sentence, these word information are fused into characters.

[0094]

[0095] Among them, Att-Pooling(·) represents pooling calculation; represents the implicit vector of the i-th word in the attention layer; a represents the hidden layer of the attention layer; represents the weight value of the i-th word in the attention layer, Represents a set of weight values; Represents the word feature vector expression with enhanced attention weight; y t The feature vector enhanced by semantic knowledge; LayerNorm(·) represents layer normalization, c a t is the contextual character representation obtained using BERT. For each character c a t , using multi-dimensional attention from sentence C a and C b Aggregate information.

[0096]

[0097] The above multi-dimensional attention function MD-GAT(·) shares the same parameters. Through this sharing mechanism, the model has a good property that when two sentences match completely:

[0098]

[0099] Step 42: Compare using multi-angle cosine distance:

[0100]

[0101] Through P distances d1, d2, ..., d P , we can get the final character representation

[0102]

[0103] The final character representation contains three kinds of information: context information, word meaning knowledge, and character-level similarity. For each sentence C a or C b , use the attention pooling of all the final character representations of the sentence to get the sentence representation vector r a or b .

[0104] Step 43: Predict the similarity between the original text and the paraphrase sentence.

[0105] p=FFN([c CLS ,r a ,r b ,r a ⊙r b ,|r a -r b |]),

[0106] Where FFN(·) is a feed-forward network with two hidden layers and a sigmoid activation after the output layer. The training goal is to minimize the binary cross entropy loss:

[0107]

[0108] A procedure repetition matching system based on deep learning and cosine similarity matching, including the following modules:

[0109] 1. Data preprocessing module

[0110] Function: Responsible for the cleaning, formatting and word segmentation of raw data to prepare for subsequent feature extraction.

[0111] Implementation method: Remove irrelevant characters through text cleaning, and use word segmentation tools (such as Jieba word segmentation) to perform Chinese word segmentation.

[0112] Data interaction: Receive raw text data and output preprocessed text data.

[0113] 2. Feature Extraction Module

[0114] Function: Extract feature vectors from preprocessed text.

[0115] Implementation method: Use the word grid graph expression module proposed in this method to convert the text into a vector representation in a high-dimensional space.

[0116] Data interaction: Receive preprocessed text data and output the feature vector of the text.

[0117] 3. Context relevance calculation module

[0118] Function: Calculate contextual relevance of feature vectors to ensure data consistency and comparability.

[0119] Implementation: Use the language knowledge attention model in this method to express word vectors and contexts.

[0120] Data interaction: Receive the feature vector output by the feature extraction module, and output the feature vector after associating it with context information.

[0121] 4. Cosine similarity calculation module

[0122] Function: Calculate the cosine similarity between two text feature vectors to measure the similarity of the texts.

[0123] Implementation: The cosine similarity value is obtained by calculating the dot product of two vectors and dividing it by the product of their respective moduli.

[0124] Data interaction: Receive the standardized feature vector and output the similarity score.

[0125] 5. Matching decision module

[0126] Function: Make a match decision based on the similarity score.

[0127] Implementation method: Based on the sentence matching layer and relationship classifier proposed in this method, a similarity threshold is set. Similarities above the threshold are considered matches, and similarities below the threshold are considered mismatches.

[0128] Data interaction: Receive the similarity score from the cosine similarity calculation module and output the matching result.

[0129] 6. System integration and interaction module

[0130] Function: Integrate the above modules into a complete system and handle the interaction with the user interface.

[0131] Implementation method: Use system architecture design to implement user-system interaction through API or graphical user interface (GUI).

[0132] Data interaction: manage the data flow between modules in the entire system process, including error handling and data feedback.

[0133] Data interaction between modules:

[0134] The data preprocessing module receives the original text data and passes it to the feature extraction module after processing.

[0135] The feature extraction module converts the text into a feature vector and passes it to the context relevance calculation module.

[0136] The context relevance calculation module models the association relationship between the feature vector of the context and the current word, and then passes it to the cosine similarity calculation module.

[0137] The cosine similarity calculation module calculates the similarity score and passes the result to the matching decision module.

[0138] The matching decision module makes matching decisions based on the similarity scores and displays the results to the user through the system integration and interaction modules.

[0139] The present invention comprises three parts: a word grid graph expression module, a language knowledge attention model, a sentence matching layer and a relation classifier.

[0140] Word grid graph expression module: It represents nuclear power text as a graph structure, where the relationship between nuclear power words can be represented by the edges of the graph. It outputs the initial context representation of each word and the initial semantic representation of each word meaning.

[0141] Language Knowledge Attention Model: Iteratively updates word representation and semantic representation, and fuses each other’s useful information.

[0142] Sentence matching layer and relation classifier: First, the word representation is integrated into the character level, and then the two character sequences are matched using the bilateral multi-view matching mechanism. The relation classifier takes the sentence vector as input and predicts the relationship between the two sentences.

Claims

1. A procedure repetition matching method based on deep learning and cosine similarity matching, characterized in that: The steps include: Step 1: Build a graph structure of the text, express the text information as graph nodes, and represent the relationship between words with graph edges; Step 2: Introduction of nuclear power professional vocabulary database; Step 3: The language knowledge attention model models the contextual vocabulary correlation, iteratively updates the word representation and semantic representation, and integrates each other’s useful information; Step 4: Integrate word representation into character level and use bilateral multi-view matching mechanism to match two character sequences. The relation classifier takes sentence vector as input to predict the relationship between two sentences.

2. A procedure repetition matching method based on deep learning and cosine similarity matching as claimed in claim 1, characterized in that: The step 1 includes the following: Step 11: Contextual word embedding For each node x in the graph i , word w i The initial representation of is the attention pooling of the context character representation. Suppose the word w i It consists of some consecutive character tokens {c t1 , c t1+1 , …, c t2 }, for each character c k (t1≤k≤t2), where k represents the character number, the value range of the number is (t1, t2), t1 and t2 represent the first and last tokens after the characters are converted into tokens sequence, and a two-layer feedforward network is used to calculate a feature-based score vector, and then normalized using a feature-based multidimensional softmax. u k =MD-softmax k (FFN(c k )), Among them, MD-softmax ( ) is the function calculation formula; Step 12: Transform the corresponding word vector c k With the normalized score u k Perform weighting to obtain the context word vector; Step 13: Use the Att-Pooling(·) pooling function to transform the context word vector v i =Att-Pooling({c k |t1≤k≤t2}).。 3. The method for repetition matching based on deep learning and cosine similarity matching according to claim 1, characterized in that: The step 2 includes the following: Step 21: For each word w i , the word sense set is represented as S(w i )={s i,1 ,s i,2 ,…,s i,K },s i,k Yes i The kth sememe of , which means its corresponding sememe is O(s i,k ), in order to obtain each meaning s i,k The embedding vector of , first obtains the multi-dimensional attention representation of each sememe: Among them, MD-GAT() is a graph structure calculation function based on features, where e is the embedding vector of the semantic primitive o; i is the ordinal number of the word; k is the ordinal number of the semantic primitive; O(s i,k ) indicates the original meaning; s i,k It means meaning item. Meaning item and semantic original are two meanings. Seme original means the smallest indivisible semantic unit in linguistics. Meaning item means that each word has rational meaning. If it has several meanings, it is expressed as several meaning items. Step 22: For each sememe s i,k , whose embedding is obtained by attention pooling of all sememe representations 4. The method for repetition matching based on deep learning and cosine similarity matching according to claim 1, characterized in that: The step 3 includes the following: Step 31: Update semantic expression Update the feature expression vector of the current word according to the context in the sentence, and each word meaning will first be updated from x i The useful information of the forward and backward aggregated words is calculated as follows: Among them, represents the i-th word; represents the j-th word; fw represents the feedforward network; bw represents the reverse network; is the temporal feature vector of the th word, is the hidden layer feature vector of the word, The feature vector representation of the preceding and following contexts of the th word, where the two multidimensional attention functions MD-GAT(·) have different parameters; [·,·] represents the concatenation of vectors, using GRU to control the fusion of context information and current lexical semantic information; Step 32: Update word expression After multiple iterations, the final word representation contains not only the context word information but also semantic knowledge. For each sentence, h i Represent the final word representation 5. The method for repetition matching based on deep learning and cosine similarity matching according to claim 1, characterized in that: The step 3 includes the following: Step 41: After obtaining the word representation enhanced by the semantic knowledge of each sentence, the word information is fused into the characters; Among them, Att-Pooling(·) represents pooling calculation; represents the implicit vector of the i-th word in the attention layer; a represents the hidden layer of the attention layer; represents the weight value of the i-th word in the attention layer, Represents a set of weight values; Represents the word feature vector expression with enhanced attention weight; y t The feature vector enhanced by semantic knowledge; LayerNorm(·) represents layer normalization, c a t is the contextual character representation obtained using BERT. For each character c a t , using multi-dimensional attention from sentence C a and C b Aggregate information, The above multi-dimensional attention function MD-GAT(·) shares the same parameters, when the two sentences match exactly: Step 42: Compare using multi-angle cosine distance: Through P distances d1, d2, ..., d P , get the final character representation The final character representation contains three kinds of information: context information, word meaning knowledge, and character-level similarity. For each sentence C a or C b , use the attention pooling of all the final character representations of the sentence to get the sentence representation vector r a or b ; Step 43: Predict the similarity between the original text and the paraphrase sentence p=FFN([c CLS r a ,r b ,r a ☉r b ,|r a -r b |])), Among them, FFN(·) is a feed-forward network with two hidden layers and a sigmoid activation after the output layer. The training goal is to minimize the binary cross entropy loss:

6. A procedure repetition matching system based on deep learning and cosine similarity matching, characterized in that: It includes the following modules: data preprocessing module, feature extraction module, context relevance calculation module, cosine similarity calculation module, matching decision module, system integration and interaction module and data interaction module.

7. A procedure repetition matching system based on deep learning and cosine similarity matching as claimed in claim 6, characterized in that: The data preprocessing module implements the cleaning, formatting and word segmentation of the original data to prepare for the subsequent feature extraction, removes irrelevant characters through text cleaning, uses the word segmentation tool to perform Chinese word segmentation, receives the original text data, and outputs the preprocessed text data; The feature extraction module extracts feature vectors from the preprocessed text, converts the text into a vector representation in a high-dimensional space using a word grid graph expression module, receives the preprocessed text data, and outputs a feature vector of the text; The context relevance calculation module performs context relevance calculation on the feature vector to ensure the consistency and comparability of the data, uses the language knowledge attention model to express the word vector and the context, receives the feature vector output by the feature extraction module, and outputs the feature vector after the context information is associated; The cosine similarity calculation module calculates the cosine similarity between two text feature vectors to measure the similarity of the texts, obtains the cosine similarity value by calculating the dot product of the two vectors and dividing it by the product of their respective moduli, receives the standardized feature vectors, and outputs a similarity score; The matching decision module makes a decision on whether to match based on the similarity score. Based on the sentence matching layer and the relationship classifier of the method, a similarity threshold is set. Similarities above the threshold are considered as matches, and those below the threshold are considered as mismatches. The similarity score of the cosine similarity calculation module is received and the matching result is output; The system integration and interaction module integrates the above modules into a complete system and handles the interaction with the user interface, using the system architecture to realize the interaction between the user and the system through the API or the graphical user interface; The data interaction module manages the data flow between modules in the entire system process, including error handling and data feedback, and the data interaction relationship between modules.

8. The procedure repetition matching system based on deep learning and cosine similarity matching according to claim 6, characterized in that: The data preprocessing module receives the original text data and passes it to the feature extraction module after processing; The feature extraction module converts the text into a feature vector and passes it to the context relevance calculation module; the context relevance calculation module models the association between the context feature vector and the current word, and then passes it to the cosine similarity calculation module; the cosine similarity calculation module calculates the similarity score and passes the result to the matching decision module; The matching decision module makes matching decisions based on the similarity scores and displays the results to the user through the system integration and interaction module.

9. The procedure repetition matching system based on deep learning and cosine similarity matching according to claim 6, characterized in that: It also includes a word grid graph representation module, a language knowledge attention model, a sentence matching layer, and a relation classifier.

10. The procedure repetition matching system based on deep learning and cosine similarity matching according to claim 9, characterized in that: The word grid graph expression module represents the nuclear power text as a graph structure, in which the relationship between nuclear power words can be represented by the edges of the graph, and outputs the initial context representation of each word and the initial semantic representation of each word meaning; The language knowledge attention model iteratively updates word representation and semantic representation, and integrates useful information from each other; The sentence matching layer and relation classifier first integrate word representation into character level, and then use bilateral multi-view matching mechanism to match two character sequences. The relation classifier takes sentence vector as input to predict the relationship between two sentences.

Images