Correlation analysis method of Chinese-Vietnamese news events integrating multi-granularity features

Abstract

The present invention relates to a method for analyzing the correlation between Chinese and Vietnamese news events by fusing multi-granularity features, and belongs to the field of natural language processing. The present invention performs correlation analysis between Chinese and Vietnamese news by fusing multi-granularity features, can measure similar texts from the event granularity, establish semantic connections between bilingual news texts at the event level, and use event features to perform news correlation analysis. The present invention first extracts event keywords, key sentences, and topic features from Chinese and Vietnamese news texts; secondly, it incorporates the above multi-granularity event information based on the semantics of the text; then, it uses a multi-head self-attention mechanism to extract the dependency relationship between event information at different granularities; finally, it calculates the Manhattan distance to obtain the similarity relationship between events in Chinese and Vietnamese news texts. The present invention has achieved good results in the task of analyzing the correlation between Chinese and Vietnamese news events, and provides support for tasks such as cross-language information retrieval, cross-language news event classification, and public opinion monitoring.

Description

Technical Field

[0001] The present invention relates to a Chinese-Vietnamese news event correlation analysis method integrating multi-granularity features, and belongs to the field of natural language processing. Background Art

[0002] The purpose of correlation analysis of Chinese and Vietnamese news events is to identify Vietnamese news texts related to Chinese news events. Directly calculating similarity between Chinese and Vietnamese news texts using text representations is ineffective. This is because Chinese and Vietnamese do not belong to the same language family and are quite different. There are issues with alignment of content and form between the different languages. Furthermore, the focus and reporting angles of the same event vary across countries. However, the keywords, key phrases, and themes of news events generally remain consistent. Using event information to conduct correlation analysis between Chinese and Vietnamese news events can yield more balanced and reasonable conclusions. This approach holds significant research value for future research in areas such as cross-language news event retrieval, cross-language news event classification, false news detection, and public opinion monitoring. Summary of the Invention

[0003] The present invention provides a method for analyzing the correlation between Chinese and Vietnamese news events by integrating multi-granularity features. By extracting event information of different granularities in news reports, the similarity of Chinese and Vietnamese news documents is comprehensively evaluated, thereby ensuring that the similarity is measured based on news events rather than the overall content of the news. The present invention improves the performance of the Chinese-Vietnamese news event text similarity calculation model.

[0004] The technical solution of the present invention is: a method for analyzing the correlation between Chinese and Vietnamese news events by integrating multi-granularity features. The specific steps of the method are as follows:

[0005] Step 1: Crawl Chinese and Vietnamese news documents and their keywords in various fields;

[0006] Step 2: Filter and screen the crawled Chinese and Vietnamese news documents and their keywords;

[0007] Step 3: Pre-train Chinese-Vietnamese cross-language word vectors;

[0008] Step 4: Extract features and perform embedding representation on Chinese-Vietnamese news documents at phrase, sentence, and topic levels;

[0009] Step 5: Learn various granular features through three layers of Stacked-LSTM to obtain the final feature vector;

[0010] Step 6: Calculate all attention values through a multi-head attention mechanism, then perform a linear transformation to obtain a news context representation with multi-granularity event features;

[0011] Step 7: Use Manhattan distance to calculate the actual document distance and obtain the semantic similarity of bilingual documents.

[0012] As a further solution of the present invention, in Step 3, first, the word embedding matrices X and Y are trained in the corpus of each language, the seed dictionary is represented as a binary matrix D, and the optimal mapping matrix W is found. * , so that the mapped Chinese word embedding matrix and Vietnamese word embedding matrix Minimize the sum of squared Euclidean distances between:

[0013]

[0014] in represents the i-th Chinese word embedding, W is the mapping matrix, Denotes the j-th Vietnamese word embedding. If the i-th Chinese word is aligned with the j-th Vietnamese word, then D ij =1;

[0015] Then, the word embedding matrices X and Y are normalized and centered preprocessing operations are performed: W is constructed as an orthogonal matrix to prevent the degradation of monolingual performance and provide better bilingual mapping. The optimized formula is:

[0016] W * =arg W maxTr(XWY T D T )

[0017] Where Tr(·) represents the sum of all elements on the main diagonal, W * is the optimal mapping matrix;

[0018] Finally, the dot product between the mapped source language embedding and the target language embedding is used as the similarity measure.

[0019] As a further solution of the present invention, in Step 4, key phrases, key sentences and bilingual topics are selected as word-level, sentence-level and passage-level granular features, and TF-IDF, TextRank and bilingual LDA are used to extract each granular feature respectively.

[0020] (1) Key phrase feature extraction

[0021] Extract key phrases using TF-IDF. The calculation method is as follows:

[0022] tfidf i,j =tf i,j *idf i,j

[0023] Among them, tfidf i,jRepresents word frequency tf i,j and inverse document frequency (idf) i,j The larger the TF-IDF value, the more important the feature phrase is to the document. TF represents the frequency of a key phrase in the entire document. IDF stands for inverse document frequency. Document frequency refers to the number of times a key phrase appears in all documents in the entire corpus. Inverse document frequency, also known as inverse document frequency, is the reciprocal of document frequency and is primarily used to reduce the impact of common phrases across all documents that have little impact on the document.

[0024] The calculation formula for TF word frequency is as follows:

[0025]

[0026] Among them, n i,j is the characteristic phrase t i In the document d j The number of times it appears in k n k,j It is document d j The result of the calculation is the word frequency of a certain characteristic phrase.

[0027] The formula for calculating IDF is as follows:

[0028]

[0029] Where |D| represents the total number of documents in the corpus, Indicates the number of feature word phrases in the document. In case the phrase does not exist in the corpus, that is, the denominator is 0, use as the denominator.

[0030] (2) Key sentence feature extraction

[0031] Key sentences are extracted using TextRank. TextRank is a graph-based text processing and ranking model. Its main idea is to segment a document into several sentences, construct graph relationships with each sentence as a graph node, and iteratively calculate the ranking score of each graph node, ultimately obtaining a ranking of the important sentences in the document.

[0032] Specifically, each node represents a sentence in the document, and then an undirected graph is constructed with the nodes. The edges between the nodes represent the similarity weights of the two sentences. The calculation formula is as follows:

[0033]

[0034] Among them, WS(V i ) represents the weight of sentence i, d is the damping coefficient, the value is between 0 and 1, generally 0.85, In(Vi ) indicates pointing to node V i The node set, Out(V j ) represents node V j The set of nodes pointed to by the starting edge, W ij For node V i and node V j The weight value between WS(V j ) represents the weight of sentence j.

[0035] Each sentence in the text is treated as a node, and its similarity is calculated based on the number of co-occurrences of words. The formula is as follows:

[0036]

[0037] Among them, S i and S j Indicates a sentence, w k Represents the words in the sentence. The numerator represents the number of co-occurring words in the two sentences. The denominator represents the logarithm of the number of words in the sentence and then sums them up to reduce the influence of long sentences in the similarity calculation.

[0038] The above formula is used to calculate the similarity between any two nodes and construct a node connection graph. Then, the TextRank value of each node is iteratively calculated. Finally, the sentences corresponding to the nodes with the highest ranking values are selected as key sentences.

[0039] (3) Bilingual topic feature extraction

[0040] Bilingual LDA is used to detect the topic similarity between bilingual documents and extract document topics.

[0041] LDA is a probabilistic topic model that presents each document topic as a probability distribution, with each topic consisting of a probability distribution of terms. A document can contain multiple topics, with each word in the document being generated by one of these topics. Bilingual LDA is a bilingual extension of the LDA model, used to detect topic similarity between bilingual documents. Figure 2 It is the probability graph of bilingual LDA.

[0042] Where S and T represent the source language and target language respectively, K represents the number of topics, and Representative Theme Z K The word distribution under, M is the number of aligned documents, θ m is the topic distribution of document m, and Represents the number of words in document m, and is the implicit topic of the nth word in document m, and is the nth word in document m, α, β S and β T It is a super parameter.

[0043] Use Cibbs Sampling to distribute the topic θ m and word distribution and The approximate solution is as follows:

[0044]

[0045] Where l represents language, Theme Z k The number of times word v appears in , is the Dirichlet prior of the term v, is the probability distribution of term v in topic k. For a new corpus, keep Unchanged, the topic distribution is obtained through Gibbs Sampling:

[0046]

[0047] in, Represents topic Z k In the document m l Probability of occurrence.

[0048] As a further solution of the present invention, in Step 5, a twin recurrent neural network structure is used to calculate the similarity of Chinese and Vietnamese news. The twin recurrent neural network is composed of two neural networks with the same structure, and the network can learn to distinguish whether two inputs are similar; the main part of the twin recurrent neural network is composed of two Stacked-LSTMs, which share parameters. The inputs are Chinese and Vietnamese word vectors and multi-granularity word vector sequences, respectively. The three-layer Stacked-LSTM learns various granularity features at the word level, sentence level, and topic level to obtain the final feature vector.

[0049] As a further solution of the present invention, in Step 6, all attention values are firstly vectorized by vertical connection, and then a news context representation with multi-granularity event features is obtained by linear transformation; the calculation formula is as follows.

[0050] H′ n =Multi_self_attention(H n )

[0051] Among them, Multi_self_attention is the mapping function; H′ nis the context feature representation of the n-th word with multi-granularity event information; the principle of the multi-head self-attention mechanism is to scale the dot product attention, and the attention value is obtained by dot product calculation. The calculation formula is as follows:

[0052]

[0053] Among them, Q, K, and V are the Query, Ke, and Value matrices for calculating self-attention, respectively. k is the dimension of the key, Used to prevent the inner product of Q and K from being too large.

[0054] As a further solution of the present invention, in Step 7, Manhattan distance is used to calculate the actual distance due to the different structures and semantic differences between Chinese and Vietnamese news texts.

[0055] Specifically, in Step 6, the Chinese and Vietnamese texts are respectively obtained through the twin network structure H' zh and H' vi On this basis, the sigmoid function is used for normalization to calculate the similarity of the bilingual text P = Similarity(H' zh , H' vi ), the formula is as follows:

[0056] P=Similarity(H' zh , H' vi )=1-sigmoid(manhattan(H' zh , H' vi ))

[0057] Among them, when P>0.6, it is predicted that Chinese and Vietnamese news are related, H' zh It is a Chinese document context feature representation with multi-granularity event information; H' vi It is a Vietnamese document context feature representation with multi-granularity event information.

[0058] The beneficial effects of the present invention are:

[0059] 1. The method proposed in this paper integrates multi-granularity event features to extract word-level, sentence-level, and topic-level features from news events. It uses Stacked-LSTM and a multi-head attention mechanism to learn various multi-granularity features and strengthen semantic representation, effectively alleviating the problem of news texts describing the same event but with different semantics.

[0060] 2. In the corresponding model of the method of the present invention, the similarity of the input texts of the two languages is measured by sharing weights through the twin neural network, solving the problem of cross-language understanding; effectively improving the accuracy of the semantic representation of news event texts, thereby improving the performance of the Chinese-Vietnamese news event text similarity calculation model. BRIEF DESCRIPTION OF THE DRAWINGS

[0061] Figure 1 Schematic diagram of the model structure of the present invention;

[0062] Figure 2 This is a bilingual LDA structure diagram;

[0063] Figure 3 This is a twin recurrent neural network diagram based on Stacked-LSTM. DETAILED DESCRIPTION

[0064] Example 1: Figure 1-Figure 3 As shown in FIG, a method for analyzing the correlation between Chinese and Vietnamese news events by integrating multi-granularity features is described. The specific steps of the method are as follows:

[0065] Step 1: Crawl Chinese and Vietnamese news documents and their keywords in various fields;

[0066] Step 2: Filter and screen the crawled Chinese and Vietnamese news documents and their keywords;

[0067] Step 3: Pre-train Chinese-Vietnamese cross-language word vectors;

[0068] Step 4: Extract features and perform embedding representation on Chinese-Vietnamese news documents at phrase, sentence, and topic levels;

[0069] Step 5: Learn various granular features through three layers of Stacked-LSTM to obtain the final feature vector;

[0070] Step 6: Calculate all attention values through a multi-head attention mechanism, then perform a linear transformation to obtain a news context representation with multi-granularity event features;

[0071] Step 7: Use Manhattan distance to calculate the actual document distance and obtain the semantic similarity of bilingual documents.

[0072] As a further solution of the present invention, in Step 3, first, the word embedding matrices X and Y are trained in the corpus of each language, the seed dictionary is represented as a binary matrix D, and the optimal mapping matrix W is found. * , so that the mapped Chinese word embedding matrix and Vietnamese word embedding matrix Minimize the sum of squared Euclidean distances between:

[0073]

[0074] in represents the i-th Chinese word embedding, W is the mapping matrix, Denotes the j-th Vietnamese word embedding. If the i-th Chinese word is aligned with the j-th Vietnamese word, then D ij =1;

[0075] Then, the word embedding matrices X and Y are normalized and centered preprocessing operations are performed: W is constructed as an orthogonal matrix to prevent the degradation of monolingual performance and provide better bilingual mapping. The optimized formula is:

[0076] W * =arg W maxTr(XWY T D T )

[0077] Where Tr(·) represents the sum of all elements on the main diagonal, W * is the optimal mapping matrix;

[0078] Finally, the dot product between the mapped source language embedding and the target language embedding is used as the similarity measure.

[0079] As a further solution of the present invention, in Step 4, key phrases, key sentences and bilingual topics are selected as word-level, sentence-level and passage-level granular features, and TF-IDF, TextRank and bilingual LDA are used to extract each granular feature respectively.

[0080] (1) Key phrase feature extraction

[0081] Extract key phrases using TF-IDF. The calculation method is as follows:

[0082] tfidf i,j =tf i,j *idf i,j

[0083] Among them, tfidf i,j Represents word frequency tf i,j and inverse document frequency (idf) i,j The larger the TF-IDF value, the more important the feature phrase is to the document. TF represents the frequency of a key phrase in the entire document. IDF stands for inverse document frequency. Document frequency refers to the number of times a key phrase appears in all documents in the entire corpus. Inverse document frequency, also known as inverse document frequency, is the reciprocal of document frequency and is primarily used to reduce the impact of common phrases across all documents that have little impact on the document.

[0084] The calculation formula for TF word frequency is as follows:

[0085]

[0086] Among them, n i,j is the characteristic phrase t i In the document d j The number of times it appears in k n k,j It is document d j The result of the calculation is the word frequency of a certain characteristic phrase.

[0087] The formula for calculating IDF is as follows:

[0088]

[0089] Where |D| represents the total number of documents in the corpus, Indicates the number of feature word phrases in the document. In case the phrase does not exist in the corpus, that is, the denominator is 0, use as the denominator.

[0090] (2) Key sentence feature extraction

[0091] Key sentences are extracted using TextRank. TextRank is a graph-based text processing and ranking model. Its main idea is to segment a document into several sentences, construct graph relationships with each sentence as a graph node, and iteratively calculate the ranking score of each graph node, ultimately obtaining a ranking of the important sentences in the document.

[0092] Specifically, each node represents a sentence in the document, and then an undirected graph is constructed with the nodes. The edges between the nodes represent the similarity weights of the two sentences. The calculation formula is as follows:

[0093]

[0094] Among them, WS(V i ) represents the weight of sentence i, d is the damping coefficient, the value is between 0 and 1, generally 0.85, In(V i ) indicates pointing to node V i The node set, Out(V j ) represents node V j The set of nodes pointed to by the starting edge, W ij For node V i and node V j The weight value between WS(V j ) represents the weight of sentence j.

[0095] Each sentence in the text is treated as a node, and its similarity is calculated based on the number of co-occurrences of words. The formula is as follows:

[0096]

[0097] Among them, S i and S j Indicates a sentence, w k Represents the words in the sentence. The numerator represents the number of co-occurring words in the two sentences. The denominator represents the logarithm of the number of words in the sentence and then sums them up to reduce the influence of long sentences in the similarity calculation.

[0098] The above formula is used to calculate the similarity between any two nodes and construct a node connection graph. Then, the TextRank value of each node is iteratively calculated. Finally, the sentences corresponding to the nodes with the highest ranking values are selected as key sentences.

[0099] (3) Bilingual topic feature extraction

[0100] Bilingual LDA is used to detect the topic similarity between bilingual documents and extract document topics.

[0101] LDA is a probabilistic topic model that presents each document topic as a probability distribution, with each topic consisting of a probability distribution of terms. A document can contain multiple topics, with each word in the document being generated by one of these topics. Bilingual LDA is a bilingual extension of the LDA model, used to detect topic similarity between bilingual documents. Figure 2 It is the probability graph of bilingual LDA.

[0102] Where S and T represent the source language and target language respectively, K represents the number of topics, and Representative Theme Z K The word distribution under, M is the number of aligned documents, γ m is the topic distribution of document m, and Represents the number of words in document m, and is the implicit topic of the nth word in document m, and is the nth word in document m, α, β s and β T It is a super parameter.

[0103] Gibbs Sampling is used to distribute the topic θ m and word distribution and The approximate solution is as follows:

[0104]

[0105] Where l represents language, Theme Z k The number of times word v appears in , is the Dirichlet prior of the term v, is the probability distribution of term v in topic k. For a new corpus, keep Unchanged, the topic distribution is obtained through GibbsSampling:

[0106]

[0107] in, Represents topic Z k In the document m l Probability of occurrence.

[0108] As a further solution of the present invention, in Step 5, a twin recurrent neural network structure is used to calculate the similarity of Chinese and Vietnamese news. The twin neural network is composed of two neural networks with the same structure. The network can learn to distinguish whether two inputs are similar. The main part of the twin recurrent neural network is composed of two Stacked-LSTMs, which share parameters. The inputs are Chinese and Vietnamese word vectors and multi-granularity word vector sequences. The three-layer Stacked-LSTM learns various granularity features at the word level, sentence level, and topic level to obtain the final feature vector, such as Figure 3 shown.

[0109] The formula for calculating the hidden state of the three-layer Stacked-LSTM is as follows:

[0110] h 1,t =LSTM1(x t ,h 1,t-1 )

[0111] h 2,t =LSTM2(h 1,t ,h 2,t-1 )

[0112] h 3,t =LSTM3(h 2,t ,h 3,t-1 )

[0113] Among them, h n,t is the hidden state of the word vector sequence at the nth layer at time t.

[0114] As a further solution of the present invention, in Step 6, all attention values are firstly vectorized by vertical connection, and then a news context representation with multi-granularity event features is obtained by linear transformation; the calculation formula is as follows.

[0115] H′ n =Multi_self_attention(H n )

[0116] Among them, Multi_self_attention is the mapping function; H′ n is the context feature representation of the n-th word with multi-granularity event information; the principle of the multi-head self-attention mechanism is to scale the dot product attention, and the attention value is obtained by dot product calculation. The calculation formula is as follows:

[0117]

[0118] Where Q, K, and V are the query, key, and value matrices for calculating self-attention, respectively. k is the dimension of the key, Used to prevent the inner product of Q and K from being too large.

[0119] The multi-head self-attention mechanism can learn more important information in the sequence by repeatedly performing multiple attention calculations, then concatenating all attention values and performing a linear transformation to obtain a news context representation with multi-granularity event features. The vector corresponding to the output of the nth word is calculated as follows:

[0120] head i =Attention(QWi i Q ,KW i K ,VW i V )

[0121] H=MultiHead(Q,K,V)=Concat(head1,head2,...,head h )W 0

[0122] The weight matrix W 0 ∈d*d,Wi i Q ∈d / h*d,W i K ∈d / h*d,W i V ∈d / h*d, h represents the number of heads and d represents the feature dimension.

[0123] As a further solution of the present invention, in Step 7, Manhattan distance is used to calculate the actual distance due to the different structures and semantic differences between Chinese and Vietnamese news texts.

[0124] Specifically, in Step 6, the Chinese and Vietnamese texts are respectively obtained through the twin network structure H' zh and H' vi On this basis, the sigmoid function is used for normalization to calculate the similarity of the bilingual text P = Similarity(H' zh , H' vi ), the formula is as follows:

[0125] P=Similarity(H' zh , H' vi )=1-sigmoid(manhattan(H' zh , H' vi ))

[0126] Among them, when P>0.6, it is predicted that Chinese and Vietnamese news are related, H' zh It is a Chinese document context feature representation with multi-granularity event information; H' vi It is a Vietnamese document context feature representation with multi-granularity event information.

[0127] To illustrate the effectiveness of the present invention, three comparative experiments were conducted. Three currently mainstream cross-language document similarity models, namely the TF-IDF+LDA model, the Doc2vec+BiLSTM model, and the Siamese Bert model, were selected as baseline models for comparison with the present invention. The results in Table 2 demonstrate the effectiveness of the present invention.

[0128] The event-driven Chinese-Vietnamese news relevance analysis model proposed in this paper achieves the best performance compared to the other three models. It can be seen that incorporating phrase-level, sentence-level, and document-level topic features significantly improves the calculation of cross-language text similarity for news events.

[0129] The model of the present invention has an F1 value 6.6% higher than that of the TF-IDF+LDA model. This is because TF-IDF represents text at the word level, and sentences and phrases have richer and more accurate text semantic features than words, so the text representation effect is good.

[0130] Compared with the Doc2vec+BiLSTM model, the F1 value is 4.4% higher. This is because Doc2vec can only capture the semantics of documents well when trained on a large amount of data. However, this method is not ideal for Vietnamese, where the corpus is scarce.

[0131] The Siamese BERT model uses the pre-trained model BERT to obtain good language features and obtains more interactive matching information through the interactive attention mechanism, achieving good results. The F1 value is 2.2% lower than that of the model of the present invention.

[0132] Table 1 is the experimental data statistics

[0133] training set Test set Validation set 4500 1000 1000

[0134] Table 2 shows the experimental results of different models

[0135]

[0136]

[0137] To verify the impact of key phrases, key sentences, and topic features, as well as the use of Stacked-LSTM and multi-head self-attention mechanisms on model performance, we conducted ablation experiments. The base model is a twin network structure without any key phrases, key sentences, or bilingual topic features, and uses a single-layer LSTM without a multi-head self-attention mechanism. The experimental results are shown in Table 3.

[0138] Table 3 shows the comparison results of ablation experiments

[0139] Model P(％) R(％) <![CDATA[F1(%)]]> Base Model 55.9 53.6 54.8 Base Model+Phrase 60.1 57.3 58.7 Base Model+Phrase+Sentence 62.5 59.4 60.9 Base Model+Phrase+Topic 63.8 61.2 62.5 Base Model+Phrase+Sentence+Topic 64.3 62.1 63.2 Base Model + phrase + sentence + topic + Stacked LSTM 66.2 64.1 65.1 Base Model + phrase + sentence + topic + Stacked LSTM + multi-head self-attention mechanism 67.1 65.3 66.2

[0140] Experimental results indicate that the basic model, which lacks multi-granularity features and directly uses a single-layer LSTM twin network for similarity calculation, performs poorly. Incorporating phrase features into the basic model increases the F1 score by 3.9%. This is because keyphrases contain more complete semantic information than keywords and can more accurately represent textual information, demonstrating the effectiveness of keyphrases in calculating similarity between Chinese and Vietnamese news events. Furthermore, incorporating sentence or topic features separately improves model performance. Incorporating phrase, sentence, and topic features simultaneously into the basic model significantly improves performance, achieving an F1 score 8.4% higher than the basic model.

[0141] After incorporating phrase, sentence, and topic features into the basic model, the F1 score increased by another 1.9% by replacing the single-layer LSTM network with a Stacked LSTM. This demonstrates that the multi-layer LSTM is more capable of extracting features of varying granularity and can learn more abstract features, such as the underlying topic information within the text. Furthermore, the addition of a multi-head self-attention mechanism increased the F1 score by another 1.1%. This is because this mechanism helps the network capture richer information and enhances information at varying granularity, resulting in a significant improvement in performance.

[0142] Table 4 shows the comparison of the correlation analysis results of Chinese and Vietnamese news events

[0143]

[0144]

[0145] As shown in Table 4, the corresponding model of the method of the present invention adds features of different granularities to obtain more semantic information based on the text representation and the characteristics of news events, effectively improving the accuracy of the semantic representation of news event texts, thereby improving the performance of the Chinese-Vietnamese news event text similarity calculation model.

[0146] The specific embodiments of the present invention are described in detail above with reference to the accompanying drawings. However, the present invention is not limited to the above embodiments. Various changes can be made within the knowledge of ordinary technicians in this field without departing from the scope of the present invention.

Claims

1. A method for analyzing the correlation between Chinese and Vietnamese news events by integrating multi-granularity features, characterized by: The specific steps of the method are as follows: Step 1: Crawl Chinese and Vietnamese news documents and their keywords in various fields; Step 2: Filter and screen the crawled Chinese and Vietnamese news documents and their keywords; Step 3: Pre-train Chinese-Vietnamese cross-language word vectors; Step 4: Extract features at phrase level, sentence level and topic level from Chinese-Vietnamese news documents. express; Step 5, pass through the three layers Learn various granular features and obtain the final feature vector; Step 6: Calculate all attention values through a multi-head attention mechanism, then perform a linear transformation to obtain a news context representation with multi-granularity event features; Step 7: Use Manhattan distance to calculate the actual document distance and obtain the semantic similarity of bilingual documents; In Step 3, we first train the word embedding matrix in the corpus of each language. and , representing the seed dictionary as a binary matrix , find the optimal mapping matrix , so that the mapped Chinese word embedding matrix and Vietnamese word embedding matrix Minimize the sum of squared Euclidean distances between: ； in Indicates the Chinese word embeddings, is the mapping matrix, Indicates the Vietnamese word embedding, if the Chinese words and Vietnamese words are aligned, then ; Then the word embedding matrix and Perform normalization and centering preprocessing operations: It is constructed as an orthogonal matrix to prevent the degradation of monolingual performance and provide better bilingual mapping. The optimized formula is: ； in represents the sum of all elements on the main diagonal, is the optimal mapping matrix; Finally, the dot product between the mapped source language embedding and the target language embedding is used as the similarity measure; In Step 5, a twin recurrent neural network structure is used to calculate the similarity between Chinese and Vietnamese news. The twin neural network consists of two neural networks with the same structure. The network can learn to distinguish whether two inputs are similar. The main part of the twin recurrent neural network consists of two The inputs are Chinese and Vietnamese word vectors and multi-granularity word vector sequences, three layers Learn various granular features at word level, sentence level, and topic level to obtain the final feature vector.

2. The method for analyzing the correlation between Chinese and Vietnamese news events by integrating multi-granularity features according to claim 1 is characterized by: In Step 4, key phrases, key sentences and bilingual topics are selected as word-level, sentence-level and paragraph-level granular features, respectively. 、 , bilingual Extract each particle size feature.

3. The method for analyzing the correlation between Chinese and Vietnamese news events by integrating multi-granularity features according to claim 1 is characterized by: In Step 6, all attention values are first concatenated using a vertical concatenation method, and then linearly transformed to obtain a news context representation with multi-granularity event features. The calculation formula is as follows: ； in, is the mapping function; For the Contextual feature representation of words with multi-granularity event information; The principle of the multi-head self-attention mechanism is to scale the dot product attention and obtain the attention value through dot product calculation. The calculation formula is as follows: ； in Calculate the self-attention 、 、 matrix, is the dimension of the key, Used to avoid The inner product is too large.

4. The method for analyzing the correlation between Chinese and Vietnamese news events by integrating multi-granularity features according to claim 1 is characterized by: In Step 7, due to the different structures and semantic differences between Chinese and Vietnamese news texts, Manhattan distance is used to calculate the actual distance; Specifically, in Step 6, the Chinese and Vietnamese texts are obtained through the twin network structure. and , based on this The function is normalized to calculate the similarity of bilingual texts , the formula is as follows: ； Among them, when When predicting the correlation between Chinese and Vietnamese news, It is a context feature representation of Chinese documents with multi-granularity event information; It is a Vietnamese document context feature representation with multi-granularity event information.

Images