A Chinese event trigger word extraction method based on multi-feature fusion

Abstract

The application discloses a Chinese event trigger word extraction method based on multi-feature fusion, and comprises the following steps: S1, acquiring a Chinese event extraction dataset; S2, building a Chinese event trigger word extraction network based on multi-feature fusion, wherein the network comprises a text coding layer, a feature fusion layer and a detection and classification layer; S3, inputting the data obtained in S1 into the Chinese event trigger word extraction network based on multi-feature fusion built in S2 after processing, and training the Chinese event trigger word extraction network; and S4, inputting a Chinese text to be extracted into the trained Chinese event trigger word extraction network for trigger word extraction, and outputting a detection result of the text trigger word. The application extracts the event trigger word of the input text, fully fuses the above features by using a convolutional neural network and an attention mechanism, acquires deep semantic information of the text, and thus improves the precision, recall rate and F1 value of the event trigger word extraction.

Description

Technical Field

[0001] This invention belongs to the field of natural language processing technology, specifically relating to a method for extracting Chinese event trigger words based on multi-feature fusion. Background Technology

[0002] With the rapid development of network media and technology, the internet has permeated all aspects of human life, becoming an indispensable part of it. Events are fundamental units of knowledge in human thought and memory, and also important storage media and dissemination methods in information processing and communication science. According to the ACE conference definition, an event refers to the objective fact of an individual or thing carrying out activities at a certain time and place. Among these, event trigger words are the most representative and characteristic words of the entire event, generally composed of verbs, nouns, or phrases. Event extraction technology helps to quickly grasp effective information from massive amounts of data, extracting event trigger words and arguments from unstructured text, thereby quickly grasping the central theme of the text. In recent years, the rapid development of deep learning technology in the field of natural language processing has effectively promoted research in the field of event extraction. Well-trained deep learning networks can identify the trigger words that best represent the current text, and the network as a whole also exhibits good performance. However, most research still focuses on event extraction from English. The linguistic characteristics of Chinese are different from those of English, and using character-level features does not conform to the characteristics of Chinese. Meanwhile, Chinese word segmentation may result in issues such as data sparsity and out-of-vocabulary (OOV) words, as well as neglecting the attention to difficult-to-segment positive samples during model training.

[0003] Yang Haiqin et al. disclosed an event extraction method and related apparatus in their patent application "An Event Extraction Method and Related Apparatus" (application number CN202210287573.1, publication number CN114385793A). This method first performs word encoding and position encoding on each word in the text to be extracted, obtaining corresponding word embedding vectors and position embedding vectors. Then, it combines these with the event type embedding vector corresponding to the event type to obtain a target embedding vector. Based on this target embedding vector, event parameters corresponding to the event type are extracted. Compared to word-level vectors, this method adds positional information, word segmentation information, and event type information, resulting in richer feature extraction. However, this method still has shortcomings: it does not extract specific information for the trigger word extraction model, feature extraction is still insufficient, and it does not address the difficulty in distinguishing positive samples during training. Summary of the Invention

[0004] To overcome the problems existing in the prior art, the present invention aims to provide a Chinese event trigger word extraction method based on multi-feature fusion. This method extracts event trigger words from the input text, primarily selecting word-level features, remotely supervised trigger word features, and part-of-speech features for analysis. It then fully integrates these features using convolutional neural networks and attention mechanisms. This process obtains deep semantic information from the text, thereby improving the precision, recall, and F1 score of event trigger word extraction.

[0005] To achieve the above objectives, the technical solution adopted by the present invention is as follows:

[0006] A method for extracting Chinese event trigger words based on multi-feature fusion includes the following steps;

[0007] S1: Obtain the Chinese event extraction dataset, which consists of a large amount of Chinese text, and mark the positions of the trigger words that best represent the semantics of the text;

[0008] S2: Construct a Chinese event trigger word extraction network based on multi-feature fusion. This network includes a text encoding layer, a feature fusion layer, and a detection and classification layer. The text encoding layer is used to obtain BERT character-level features, remote supervision trigger word features, word segmentation features, and part-of-speech features.

[0009] S3: After processing the data obtained in S1, input it into the Chinese event trigger word extraction network based on multi-feature fusion built in S2 to train the Chinese event trigger word extraction network.

[0010] S4: Input the Chinese text to be extracted into the trained Chinese event trigger word extraction network to extract trigger words and output the detection results of text trigger words.

[0011] In S2, the connection order between the layers is as follows:

[0012] S201: The text encoding layer is responsible for extracting BERT character-level features, remote supervision trigger word features, word segmentation features, and part-of-speech features from the input text; it can fully extract the deep semantic representation of the input text and improve the performance of the trigger word extraction model.

[0013] S201a: After the input text S201 is segmented by the BERT word segmenter, it is input into the BERT pre-trained model to obtain BERT word vectors. These word vectors are composed of three parts: word embedding, segment embedding, and position embedding, which can obtain powerful semantic representation capabilities.

[0014] S201b: A remote supervision knowledge base is constructed using K-fold cross-validation to eliminate the influence of self-labels; the input text is compared with the constructed remote supervision knowledge base to find the remote supervision trigger words of the input text; one-hot encoding is used to obtain the remote supervision trigger word features of the input text through embedding;

[0015] S201c: First, web crawling technology is used to obtain texts from news websites in categories such as military, finance, politics, society, sports, science and technology, entertainment, law, and disaster / accident. Second, the external tool jieba is used to segment the obtained texts and sort them according to word frequency to create a Chinese dictionary. Finally, jieba is used to segment the input text of S201, and vector representations are obtained according to the word order in the dictionary. The segmentation features of the input text are obtained through Embedding.

[0016] S201d: Use the external tool jieba to analyze the part-of-speech of the input text of S201, find the verbs at the corresponding positions in the text and set the One-hot vectors at the corresponding positions to 1, encode them and input them into the Embedding layer to obtain the part-of-speech features of the input text;

[0017] S201a, S201b, S201c, and S201d all obtain BERT word-level information, remote supervision trigger word information, word segmentation information, and part-of-speech information respectively from the input text of S201.

[0018] S202: The feature fusion layer is responsible for extracting and fusing local features from various types of extracted information to obtain deep semantic information;

[0019] S202a: Uses a convolutional neural network to fuse word vectors from S201b, S201c, and S201d to extract local feature information from the input text;

[0020] S202b: The attention mechanism is used to fuse the word vectors of S202a and S201a in a certain proportion to obtain the fused features, which captures more critical information in the input text.

[0021] S203: Detection and classification layer, responsible for using the fused features obtained from S202b for the event-triggered word extraction task to obtain the final extraction result.

[0022] In step S202a, a convolutional neural network is used to fuse the word vectors of S201b, S201c, and S201d to extract local feature information of the input text. Specifically, the word vectors of S201b, S201c, and S201d are concatenated and input into convolutional neural networks composed of convolutional kernels of different scales (2×2, 3×3, and 4×4) to obtain three different word vectors. The three different word vectors are then concatenated and input into a fully connected layer and a Dropout layer.

[0023] S202b uses an attention mechanism to fuse the character vectors of S202a and S201a in a certain proportion to capture more critical information in the input text. The specific structural relationship is as follows: first branch → second branch → third branch, where:

[0024] The first branch generates an upper triangular occlusion matrix based on the dimension of the word vectors in S202a.

[0025] The second branch consists of a multi-head attention mechanism, concatenating the outputs of each head to obtain the output vector. The input Q matrix is ​​the BERT word vector obtained from S201a, the K and V matrices are vectors obtained from S202a, and the occlusion matrix is ​​the matrix obtained from the first branch; the formula used is:

[0026] MultiHead(Q,K,V)=Concat(head1,head2,...,head h W o

[0027] head i =Attention(QW i Q ,KW i K VW i V )

[0028] Among them, head i W represents the i-th attention head. i Q W i K and W i V These are the transformation matrices of matrices Q, K, and V, respectively, and W... o It is the concatenated transformation matrix, head i The calculation is also known as the scaling dot product attention mechanism, and the specific formula is:

[0029]

[0030] Where T is the transpose symbol, K Td represents the transpose of matrix K. k Let d represent the dimension of the last dimension of matrix K; from the above formula, we know that d q =d k That is, the dimension of the last dimension of matrix Q is equal to the dimension of the last dimension of matrix K;

[0031] The third branch consists of two fully connected layers and one Dropout layer in sequence, with the input being the vector obtained from the second branch.

[0032] The S203 detection and classification layer is responsible for using the fused features for the event trigger word extraction task to obtain the final extraction result. Specifically, the fused feature vector in S202 is input into two binary classifiers, the trigger word start position classifier and the trigger word end position classifier, to find the start and end positions of the trigger words in the input text.

[0033] Specifically, S3 is:

[0034] S301: Set up the environment required for the target detection network;

[0035] S302: Download the BERT-wwm pre-trained model;

[0036] S303: Load the pre-trained model from S302, process the data obtained from S1, and input it into the Chinese event trigger word extraction network based on multi-feature fusion built in S2, and train it using a new weighted focus loss function.

[0037] S303 loads the pre-trained model from S302, processes the data obtained from S1, and inputs it into the Chinese event trigger word extraction network based on multi-feature fusion built by S2, and trains it using a new weighted focus loss function. The specific loss function expression is as follows:

[0038] L = α i ×L FL (p i ,y i )

[0039] Where, α i This is the weight factor for the current text, and its value is the proportion of the number of event types corresponding to the current text to the total number of events in the dataset. i For the output predicted by the network, y i It is the actual label value, L FL (*) represents the focus loss function, and the specific formula is as follows:

[0040]

[0041] Where p is the true output of the model, (1-p t) γ γ is the adjustment factor, and it is an adjustable parameter. The larger the value of γ, the smaller the proportion of easily distinguishable samples in the loss, and the more the model focuses on the proportion of difficult-to-distinguish samples.

[0042] The beneficial effects of this invention are:

[0043] First, based on the unique characteristics of the Chinese language, this invention acquires remote supervision trigger word information, word segmentation information, and part-of-speech information of the text on the basis of the BERT pre-trained model, and constructs remote supervision word vectors, word segmentation word vectors, and part-of-speech word vectors, which enriches the feature selection for Chinese event trigger word extraction to a certain extent. At the same time, this method incorporates the idea of ​​remote supervision into the trigger word extraction task, which can alleviate the problems of data sparsity and out-of-vocabulary words after Chinese word segmentation.

[0044] Second, this invention uses a combination of convolutional neural networks and attention mechanisms to extract deep semantic features from text. Multi-scale convolutional neural networks can extract various local textual features, while the attention mechanism can retain more important contextual information based on the extracted features, effectively improving the performance of the trigger word extraction network.

[0045] Third, this invention uses a novel weighted focus loss function. Compared to the cross-entropy loss function and the focus loss function, the loss function of this invention can assign a weight to each text according to the different distributions of the data, thereby distinguishing the importance between samples and allowing the model to pay more attention to difficult-to-distinguish positive samples during training. Attached Figure Description

[0046] Figure 1 This is a flowchart illustrating the implementation of the present invention.

[0047] Figure 2 This is a flowchart illustrating the construction of the model for this invention.

[0048] Figure 3 This is a flowchart of the training network of the present invention.

[0049] Figure 4 This is a comparison diagram showing the effects of the present invention. Detailed Implementation

[0050] The present invention will now be described in further detail with reference to the accompanying drawings and embodiments.

[0051] First, the terminology used in one or more embodiments of the present invention will be explained.

[0052] An event is a change in a situation or state that occurs at a specific point in time or within a specific geographical area and involves one or more actions by one or more characters.

[0053] Trigger words: These are the core words that indicate the occurrence of an event, often verbs or nouns. A string matching method is used to check if the trigger words in the remote supervision knowledge base exist in the input text. If they do, they are recorded as remote supervision trigger words in the input text.

[0054] Argument: The participants in an event, mainly composed of entities, values, and time.

[0055] Event extraction: This involves extracting event information from unstructured text and presenting it in a structured form. The core problem that event extraction tasks need to solve is extracting information from unstructured, semi-structured, and structured text.

[0056] Reference Figure 1 This example can be divided into four steps, and the specific implementation steps are as follows:

[0057] Step 1: Obtain the Chinese event extraction dataset.

[0058] Download the Baidu Qianyan dataset DuEE1.0. This dataset contains 9 major event categories and 65 event types, including 4150 event arguments and 121 predefined argument roles, as well as 14854 sentences containing event information, of which there are 11958 training data, 1498 validation data and 1489 test data.

[0059] Step 2: Build a novel deep learning-based Chinese trigger word extraction network. (Refer to...) Figure 2 The flowchart for building the model in this invention, in order of network structure, consists of: a text encoding layer, a feature fusion layer, and a detection and classification layer. The key steps are as follows:

[0060] 2.1) The text encoding layer is responsible for extracting BERT character-level features, remote supervision trigger word features, word segmentation features, and part-of-speech features from the input text. The text is input into the BERT pre-trained model to obtain BERT character vectors; a remote supervision knowledge base is constructed using K-fold cross-validation to build the remote supervision trigger words for the input text; (all samples are divided into k equal-sized subsets; these k subsets are iterated sequentially, with the current subset used as the validation set and all other samples used as the training set, and the trigger words of all other samples used as the remote supervision trigger word knowledge base for the current subset;)

[0061] One-hot encoding is input into the Embedding layer to obtain remotely supervised word vectors. Web crawling is used to obtain approximately 15,000 texts categorized as military, finance, politics, society, sports, technology, entertainment, law, and disaster / accident. After word segmentation, these texts are sorted by word frequency to construct a word segmentation dictionary. Word vectors are then constructed based on the word order in the dictionary after word segmentation of the input text. Finally, these vectors are input into the Embedding layer to obtain the word vectors of the input text. Observing that event trigger words often exist in the form of verbs, jieba is used to obtain the verbs in the input text. After one-hot encoding, these verbs are input into the Embedding layer to obtain the part-of-speech features of the input text.

[0062] 2.2) Feature Fusion Layer: This layer is responsible for extracting and fusing local features from various types of extracted information to obtain deep semantic information. It uses a convolutional neural network to fuse remotely supervised word vectors, word segmentation vectors, and part-of-speech vectors to extract local feature information from the input text. An attention mechanism is used to fuse BERT word vectors with the word vectors extracted by the convolutional neural network in a certain proportion to capture more important information from the input text.

[0063] 2.2.1) The convolutional neural network consists of convolutional kernels of different scales, namely 2×2, 3×3 and 4×4, as well as a fully connected layer and a Dropout layer. The input is remote supervision word vectors, word segmentation word vectors and part-of-speech word vectors. After convolution at three different scales, the word vectors are concatenated and then input into the fully connected layer and Dropout layer to extract different local information of the text and enhance the performance of network extraction.

[0064] 2.2.2) The attention mechanism fuses the BERT word vectors with the word vectors extracted by the convolutional neural network in a certain proportion. The specific structural relationship is as follows: first branch → second branch → third branch, where:

[0065] The first branch generates an upper triangular occlusion matrix based on the dimension of the input word vectors;

[0066] The second branch consists of a multi-head attention mechanism, concatenating the outputs of each head to obtain the output vector. The input Q matrix is ​​the BERT word vector, the K and V matrices are the word vectors extracted by the convolutional neural network, and the occlusion matrix is ​​the matrix obtained from the first branch. The formula used is:

[0067] MultiHead(Q,K,V)=Concat(head1,head2,...,head h W o

[0068] head i =Attention(QW i Q,KW i K VW i V )

[0069] Among them, head i W represents the i-th attention head. i Q W i K and W i V These are the transformation matrices of matrices Q, K, and V, respectively, and W... o It is the transformation matrix after concatenation. i The calculation is also known as the scaling dot product attention mechanism, and the specific formula is:

[0070]

[0071] Where T is the transpose symbol, K T d represents the transpose of matrix K. k Let d represent the dimension of the last dimension of matrix K. From the above equation, we know that d... q =d k That is, the dimension of the last dimension of matrix Q is equal to the dimension of the last dimension of matrix K.

[0072] The third branch consists of two fully connected layers and one Dropout layer in sequence, with the input being the vector obtained from the second branch.

[0073] 2.3) The detection and classification layer is responsible for using the fused features for the event trigger word extraction task to obtain the final extraction result. This layer consists of two binary classifiers: a start position classifier and a trigger word end position classifier, which detect the start and end positions corresponding to the trigger words in the input text, respectively.

[0074] Step 3: Train the Chinese event trigger word extraction network. (Refer to...) Figure 3 The flowchart for training the network in this invention, following the network structure order, is as follows: setting up the environment required for the object detection network, downloading the BERT-wwm pre-trained model, loading the pre-trained model, and training the network using a novel weighted focus loss function. The specific loss function expression is:

[0075] L = α i ×L FL (p i ,y i )

[0076] Where, α i This is the weight factor for the current text, and its value is the proportion of the number of event types corresponding to the current text to the total number of events in the dataset. iFor the output predicted by the network, y i It is the actual label value, L FL (*) represents the focus loss function, and the specific formula is as follows:

[0077]

[0078] Where p is the true output of the model, (1-p t ) γ γ is the adjustment factor, and it is an adjustable parameter. The larger the value of γ, the smaller the proportion of easily distinguishable samples in the loss, and the more the model focuses on the proportion of difficult-to-distinguish samples.

[0079] Step 4: Input the Chinese text to be extracted and obtain the text trigger word detection results.

[0080] Step 2 generates two word vectors of the same length as the input text, representing the probability that the corresponding positions are the start and end positions of the trigger word. Simultaneously, positions in the word vectors exceeding a threshold (set to 0.5) are set to 1, and others are set to 0, thus obtaining the start and end positions of the trigger word in the input text.

[0081] The effects of the present invention will be further described below with reference to simulation experiments.

[0082] 1. Experimental conditions

[0083] The hardware test platform for the simulation experiment of this invention is: CPU is AMD Ryzen 9 3900XT, GPU is GeForce RTX 3090 (24GB), and memory is 64GB.

[0084] The evaluation metrics for the experiment are: precision (P), recall (R), and F1 score, calculated using the following formulas:

[0085]

[0086]

[0087]

[0088] Wherein, TP represents the number of trigger words predicted by the model that are also trigger words in the label, FP represents the number of trigger words predicted by the model that are not trigger words in the label, and FN represents the number of trigger words that exist in the label but are not among the trigger words predicted by the model.

[0089] 2. Experimental content and results analysis:

[0090] The simulation experiments of this invention were conducted using the method described herein on the Baidu Qianyan DuEE1.0 dataset. The BERT-wwm model, developed by the Harbin Institute of Technology-iFlytek Joint Laboratory, was used as the baseline model to train the trigger word extraction network. The trained network was then used to annotate the trigger words in the input text, yielding the trigger word extraction results. Simultaneously, the performance of the event extraction algorithms ERNIE-CRF and PLMEE on the event trigger word extraction task was compared.

[0091] like Figure 4 As shown, Ours-Trigger represents the trigger word extraction method of this invention, and BERT-wwm represents a trigger word extraction method based on BERT-wwm without employing multi-feature, convolutional neural networks, and attention mechanisms. It can be seen that the method of this invention performs best in the event trigger word extraction task, with higher precision, recall, and F1 score than other models. Comparing Experiments 1 and 2, it can be seen that the ERNIE model outperforms BERT-wwm in both recall and F1 score. In addition to using BERT's character-level features, ERNIE also uses entity and word-level features for knowledge integration, effectively capturing the relationships between words and thus performing better in the event extraction task. Comparing Experiments 1 and 3, it can be seen that BERT outperforms PLMEE in both recall and F1 score in event trigger word extraction, thanks to BERT's powerful semantic representation capabilities. In the comparative experiments, PLMEE, like the method proposed in this chapter, uses the idea of ​​remote supervision to build a knowledge base, while the method in this chapter outperforms PLMEE in all evaluation metrics. Therefore, the method of this invention can better incorporate prior knowledge from remote supervision and has stronger semantic feature extraction capabilities. Comparison of Experiments 2 and 4 shows that, in addition to using word-level features, the method of this invention also employs part-of-speech feature information, using a wider variety of features and possessing richer feature representation capabilities.

[0092] Experimental results show that the method of this invention performs best on the DuEE1.0 dataset, effectively extracting deep semantic information from text and demonstrating good performance in the trigger word extraction task.

[0093] The above is merely a specific example of the present invention to enable those skilled in the art to understand the present invention. However, the present invention is not limited to the scope of this specific example. For those skilled in the art, various changes are obvious as long as they are within the spirit and scope of the present invention as defined and determined by the appended claims. All inventions utilizing the concept of the present invention are within the protection scope of the present invention.

Claims

1. A Chinese event trigger extraction method based on multi-feature fusion, characterized in that, The method comprises the following steps: S1: obtaining a Chinese event extraction dataset composed of a large amount of Chinese texts, and marking the positions of trigger words that best represent the semantics of the texts; S2: building a Chinese event trigger word extraction network based on multi-feature fusion, the network comprising a text encoding layer, a feature fusion layer and a detection and classification layer, wherein the text encoding layer is used to obtain BERT word-level features, remote supervision trigger word features, word segmentation features and part-of-speech features; S3: inputting the data obtained in S1 after processing into the Chinese event trigger word extraction network based on multi-feature fusion built in S2 to train the Chinese event trigger word extraction network; S4: inputting a Chinese text to be extracted into the trained Chinese event trigger word extraction network to extract trigger words, and outputting the detection result of the text trigger words; In S2, the connection order between the layers is as follows: S201: the text encoding layer is used to extract BERT word-level features, remote supervision trigger word features, word segmentation features and part-of-speech features of the input text; S201a: after the input text in S201 is segmented by the BERT segmenter, the BERT pre-training model is used to obtain BERT word vectors, which are composed of word embedding, segment embedding and position embedding; S201b: a K-fold cross-validation method is used to build a remote supervision knowledge base to exclude the influence of its own labels; the input text is compared with the built remote supervision knowledge base to find the remote supervision trigger words of the input text; One-hot encoding is used to obtain the remote supervision trigger word features of the input text via Embedding; S201c: first, the crawler technology is used to obtain texts of military, financial, political, social, sports, technological, entertainment, judicial and disaster / accident categories on news websites; Secondly, the texts obtained are segmented by using an external tool jieba, and a Chinese dictionary is made according to the word frequency order; Finally, the input text is segmented by using jieba, and a vector representation is obtained according to the word order in the Chinese dictionary, and the word segmentation features of the input text are obtained via Embedding; S201d: the part-of-speech of the input text in S201 is analyzed by using the external tool jieba, the verbs corresponding to the positions of the text are found, and the One-hot vectors corresponding to the positions are set to 1, and then the part-of-speech features of the input text are obtained after encoding and inputting into the Embedding layer; S201a, S201b, S201c and S201d are used to obtain BERT word-level information, remote supervision trigger word information, word segmentation information and part-of-speech information of the input text in S201, respectively; S202: the feature fusion layer is used to extract and fuse local features of the extracted information to obtain deep semantic information; S202a: the word vectors of S201b, S201c and S201d are fused by using a convolutional neural network to extract local feature information of the input text; S202b: using an attention mechanism to fuse the word vectors of S202a and S201a in a certain proportion, obtaining the fused features, and capturing more key information in the input text; S203: a detection classification layer, responsible for using the fused features obtained by S202b for event trigger word extraction tasks, to obtain the final extraction results. 2.The Chinese event trigger extraction method based on multi-feature fusion according to claim 1, wherein, In S202a, the convolutional neural network is used to fuse the word vectors of S201b, S201c and S201d to extract the local feature information of the input text. Specifically, the word vectors of S201b, S201c and S201d are spliced and input into a convolutional neural network composed of 2x2, 3x3 and 4x4 different scale convolution kernels to obtain three different word vectors. The three different word vectors are spliced and input into a fully connected layer and a Dropout layer. 3.The Chinese event trigger extraction method based on multi-feature fusion according to claim 1, characterized in that, In S202b, the attention mechanism is used to fuse the word vectors of S202a and S201a in a certain proportion to capture more key information in the input text. The specific structure relationship is as follows: first branch→second branch→third branch, wherein: The first branch generates an upper triangular occlusion matrix according to the dimension of the word vector in S202a. The second branch is composed of a multi-head attention mechanism, and the output vector is obtained by splicing the output of each head. The input Q matrix is the BERT word vector obtained by S201a, the K matrix and the V matrix are the vectors obtained by S202a, and the occlusion matrix is the matrix obtained by the first branch. The formula used is: MultiHead(Q, K, V) = Concat(head1, head2,..., head h )W o head i = Attention(QW i Q ,KW i K ,VW i V ) wherein head i represents the i-th attention head, W i Q , W i K and W i V are transformation matrices of matrices Q, K, V respectively, W o is the concatenated transformation matrix, and the calculation of head i is also called scaled dot-product attention mechanism, and the specific formula is as follows: where K T denotes the transpose of matrix K, d k denotes the dimension of the last dimension of matrix K; from the above equation, d q = d k , i.e. the dimension of the last dimension of matrix Q is equal to the dimension of the last dimension of matrix K; The third branch is sequentially composed of two fully connected layers and a Dropout layer, and the input is the vector obtained by the second branch.

4. The Chinese event trigger extraction method based on multi-feature fusion according to claim 1, characterized in that, In S203, the detection classification layer is responsible for using the fused features for event trigger word extraction tasks to obtain the final extraction results. Specifically, the fused feature vector in S202 is input into two binary classifiers, trigger word start position classifier and trigger word end position classifier, to find the start position and end position of the trigger word in the input text.

5. The Chinese event trigger extraction method based on multi-feature fusion according to claim 1, characterized in that, S3 specifically includes: S301: setting the environment required by the target detection network; S302: downloading the BERT-wwm pre-training model; S303: loading the pre-training model of S302, and inputting the data obtained by S1 after processing into the Chinese event trigger word extraction network based on multi-feature fusion built by S2, and training using a new weighted focal loss function.

6. The Chinese event trigger extraction method based on multi-feature fusion according to claim 5, characterized in that, In S303, the pre-training model of S302 is loaded, and the data obtained by S1 after processing is input into the Chinese event trigger word extraction network based on multi-feature fusion built by S2, and trained using the weighted focal loss function. The specific loss function expression is: L = a i x L FL (p i , y i ) where, α i is the weight factor of the current text, which is the proportion of the number of event types corresponding to the current text in the total number of events in the data set; p i is the output of the network prediction, y i is the true label value, and L FL (*) is the focal loss function, and the specific formula is: where p is the real output of the model, (1-p t ) γ is the adjustment factor, and γ is an adjustable parameter; the greater the value of γ, the smaller the proportion of easy-to-distinguish samples in the loss, and the more the model focuses on the proportion of difficult-to-distinguish samples.

Images