A Graph Neural Network Dialogue Emotion Recognition Method with Noise Reduction and Error Correction
By designing the context screening module and feature error correction module, the problem of insufficient noise interference and error correction in dialogue emotion recognition is solved, and higher recognition accuracy and robustness are achieved.
Patent Information
- Application Number
- CN202310161583.5
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2023-02-24
- Publication Date
- 2025-08-05
- Estimated Expiration
- 2043-02-24
AI Technical Summary
The existing graph neural networks are susceptible to noise contexts in dialogue emotion recognition and lack effective error correction mechanisms, resulting in insufficient recognition accuracy of the model.
A context screening module was designed to evaluate semantic correlation and information volume, and used the relationship graph neural network to learn contextual information, and integrated semantic and emotional features through the feature correction module, controlled error information, and classified it with a fully connected network.
It effectively reduces the impact of noise context, improves the accuracy and robustness of the model in dialogue emotion recognition, and improves the emotional recognition performance.
Smart Images
Figure CN116127968B_ABST
Abstract
Description
Technical Field
[0001] The present invention belongs to the field of natural language processing and relates to a graph neural network conversation emotion recognition method with noise reduction and error correction. Background Art
[0002] Conversational emotion recognition differs from traditional text-based emotion recognition in that it is influenced by multiple factors. The same utterance can express different emotions in different contexts or from different speakers. In psychological research, Li et al. attributed the emotional variations of speakers during conversations to two primary factors: self-dependence and speaker-dependence. Self-dependence refers to the influence of a speaker's own emotions, while speaker-dependence refers to the mutual influence of the emotions of both parties in the conversation. Given these factors, accurately identifying the emotion of utterances is one of the challenges of conversational emotion recognition.
[0003] To account for the impact of context and speaker relationships on the sentiment of utterances, researchers Hazarika, Majumder, and others used recurrent neural networks (RNNs) to capture the temporal order of conversations and memorize the historical contextual information of different speakers. Due to the long-term dependency problem of RNNs, even when considering both the temporal order of conversations and the influence of context, these RNN-based models perform poorly when capturing distant context. Furthermore, Li et al. combined external knowledge bases to learn commonsense knowledge from utterances, improving their language understanding capabilities. However, these approaches complicate conversation modeling and their ability to capture distant context remains limited. The emergence of graph neural networks (GNNs) and their variants has, to some extent, alleviated the long-term dependency problem of RNNs. Due to their powerful ability to process relational data, GNNs are increasingly being used by researchers for conversational sentiment recognition. Recently, Song, Gao, and others combined GNNs with pre-trained language models to contextually understand the semantic and syntactic information in conversations, achieving state-of-the-art performance.
[0004] GNN-based emotion recognition methods view a conversation as a graph, with nodes representing utterances and edges representing utterance relationships. Conversational relationships are established based on self-dependence and inter-speaker dependence. Semantic similarity is then used to initialize edge weights in the graph, representing the influence between utterances. However, using all context to establish utterance dependencies can easily lead to weakly related or even irrelevant contextual utterances being linked. Considering too little context can lead to poor model performance in everyday conversations. These methods overlook another potentially influential factor in conversations: information content. Intuitively, if an utterance is sufficiently informative to significantly alter others' perceptions and induce emotional changes, then it should be considered in emotion recognition. However, some utterances often lack clear emotional tendencies and contribute little to emotion perception. Linking these contextual references wastes computational resources. Utterances with low semantic relevance and information content are referred to as noise, which can interfere with the model's emotion recognition performance.
[0005] GNN-based models not only easily introduce noise when establishing discourse dependencies, but are also susceptible to noise during the learning process, leading to errors in learned feature information. Lian et al. used graph-relational neural networks to capture interpersonal and discourse relationships to correct discourse features. This, to a certain extent, mitigates prediction errors caused by the model's inadequate understanding of conversational relationships. To our knowledge, there is currently a lack of research on error-correction systems for conversational emotion recognition.
[0006] Based on the above discussion, a graph neural network with noise reduction and error correction capabilities is proposed. To reduce the noise introduced by context during model training, a context filtering module is designed to establish dependencies between utterances. This module simultaneously considers the semantic relevance and information content of the context, ensuring that utterances are highly relevant and informative. This not only avoids the loss of long-range contextual information but also reduces noise interference with the model. Furthermore, a feature correction method is designed to further improve model performance. Feature correction combines the semantic features obtained through preprocessing with the sentiment features obtained through graph processing. A gating mechanism is used to control erroneous information and correct some errors in the features. Finally, experiments are conducted on a commonly used conversational dataset, including comparisons with state-of-the-art methods. Summary of the Invention
[0007] In view of this, the purpose of the present invention is to provide a graph neural network conversation emotion recognition method with noise reduction and error correction.
[0008] In order to achieve the above object, the present invention provides the following technical solutions:
[0009] A graph neural network conversation emotion recognition method with noise reduction and error correction, the method comprising the following steps:
[0010] Step 1: Through the preprocessing module, each sentence in the conversation is tokenized and special flags [CLS] and [SEP] are added to facilitate subsequent classification tasks. It is then processed into semantic features in the pre-trained language model.
[0011] Step 2: Use the context filtering module to evaluate the semantic relevance and information content of the utterance context respectively, discard some noise context based on the evaluation results, and establish the dependency relationship between utterances;
[0012] Step 3: Use the relational graph neural network to learn contextual information from the context and obtain context-embedded sentiment features;
[0013] Step 4: Use the feature error correction module to integrate semantic features and sentiment features, control the erroneous information in the features through the gating mechanism, and correct some of the errors;
[0014] Step 5: Predict the sentiment category of the speech through the fully connected network, and use the cross-entropy loss function to measure the difference between the label distribution and the model prediction distribution.
[0015] Optionally, the specific process of step 1 includes: first obtaining a discourse set {u1, u2, ..., u N}, use the word segmenter to segment the speech, and then add special flags [CLS] and [SEP] before and after the word segmentation results to facilitate subsequent classification tasks; then use the pre-trained language model to classify the speech u i Processed into a 1024-dimensional semantic feature vector H i In the form of:
[0016] X i ={[T1],[T2],...,[T N ]},
[0017] X i ={[CLS],[T1],[T2],...,[T N ],[SEP]},
[0018] H i =RoBERTa(X i ).last_hidden_layer[0].
[0019] Among them, X i For discourse u i The word segmentation after adding the special flag bit is represented as H i It is the feature vector represented by [CLS] of the last hidden layer in the pre-trained language model RoBERTa; [CLS] can represent the semantic feature vector of the entire sentence, which can be directly used for classification tasks.
[0020] Optionally, in step 2, the context screening process includes: using cosine similarity to measure the target speech feature H i The degree of semantic relevance between the context feature H is represented by the semantic similarity matrix s1;
[0021]
[0022]
[0023] At the same time, information entropy is used to measure the amount of information contained in the context. It is known that the higher the information entropy of a discourse, the higher the uncertainty of its emotion, and the model needs to focus on analyzing the discourse. By accumulating the information entropy of each word in the discourse, the information entropy matrix of the context can be obtained:
[0024]
[0025]
[0026]
[0027] Among them, p(T j ) represents the discourse u i The frequency of the jth word in the conversation; in order for the information entropy matrix to match the dimension of the semantic correlation matrix, stack N To construct the same information entropy matrix as s1, we can get
[0028] s=(1-α)·s1+α·s2
[0029] The weighted summation of semantic relevance matrix s1 and information entropy matrix s2 is performed, and each element in the summation matrix s is screened to remove some noise contexts with low scores;
[0030]
[0031] Among them, γ is a hyperparameter, which represents the threshold; s ij Represents the context utterance u j Relative to the target utterance u i The comprehensive score of mask ij Represents the discourse node u i and u j The connection relationship between them, 1 means there is a connection, 0 means there is no connection.
[0032] Optionally, in step 3, a graph processing module consisting of an N-layer relational graph neural network is defined, and a graph attention mechanism is used to focus on important nodes among neighboring nodes and assign weights;
[0033]
[0034] Discourse node u i The neighbor node information at the lth layer is aggregated to obtain the influence of the context on the discourse, while maintaining the sentiment information of the discourse itself at the lth layer to a certain extent:
[0035]
[0036]
[0037] in is α ij Then, the node features of each layer containing graph structure information and context information are All are stacked up to get a contextual embedding feature M from shallow to deep:
[0038] M=concat(H l ),l∈[0,L]
[0039] Finally, the emotional feature D of the speech is output through the fully connected layer and activation function, where the PReLU activation function can effectively alleviate the problem of overfitting:
[0040] D=PReLU(W d M+b d ).
[0041] Optionally, in step 4, the semantic features from the pre-processing language module and the sentiment features from the graph processing module are integrated through the feature error correction module, and the error information in the features is controlled by using a gating mechanism;
[0042] The input of the feature error correction module is divided into two parts. The upper branch input comes from the output of the preprocessing module, representing the semantic features of the speech, denoted by H; the lower branch input comes from the output of the graph processing module, representing the emotional features of the speech, denoted by D. Then, the weighted sum of the two branch inputs is used to obtain the fused feature C, which will pass through the fusion information branch alone. In most previous models, the classification module will directly classify the fused feature without controlling the erroneous information. The upper and lower branches each update the feature information through a memory gate, where the memory gate is composed of two activation functions and a multiplier. The fusion information branch uses a forget gate to forget part of the fused feature information, where the forget gate is composed of an activation function:
[0043] f=σ(W f H+Q f D+b f )
[0044] At the same time, the feature error correction module will evaluate the inputs of the upper and lower branches respectively:
[0045]
[0046]
[0047] Among them, σ is the activation function, W, Q and b are learnable parameters; the memory gate consists of a forget gate and a learning gate, which is used to remember the important feature information in the semantic feature; the semantic feature H passes through the forget gate z to determine which information in the feature needs to be updated, and learns new feature information in the learning gate c, and then the semantic feature H is updated to Similarly, the sentiment feature D is updated to For the forgotten parts in the fusion features, the updated semantic features and sentiment features in the upper and lower branches are used as supplements to control the error information in the features and obtain the output C:
[0048]
[0049] Among them, f, H, and D are the fusion features, semantic features, and emotional features after error correction, respectively.
[0050] Optionally, in step 5, the output after feature correction processing will be used as the final discourse classification feature, and the discourse u is obtained using the fully connected network i The probability P of each emotion category in i , and then take the category subscript with the highest probability as the final sentiment prediction label y i :
[0051] P i =softmax(W c C i +b c )
[0052]
[0053] in It is discourse i The predicted emotion, W C and b C are the trainable weights and the bias of the fully connected layer respectively; cross entropy is used as the loss function, and the expression of the function is:
[0054]
[0055] Among them, θ is the set of parameters that need to be trained in the model, y i represents the highest predicted probability of the i-th sentiment label, Represents the true value of the i-th sentiment label.
[0056] The beneficial effects of the present invention are as follows: the present invention proposes a graph neural network conversation emotion recognition method with noise reduction and error correction, comprising: (1) designing a context screening module that can simultaneously focus on the semantic relevance and information content of the context, and the utterance can be connected to the context with strong relevance and high information content through the module. (2) Using a multi-layer relational graph neural network to aggregate and transmit node information between neighboring nodes to obtain a context-embedded emotion feature vector. (3) Through a feature error correction module, the prediction errors of the model can be effectively reduced, thereby improving the performance of the model in emotion recognition. In conversation emotion recognition, errors in model emotion prediction are a common phenomenon. However, as far as is known, few studies on conversation emotion recognition in pure text modality focus on correcting these errors. Therefore, the present invention combines these modules at the same time to more accurately recognize the emotions expressed in the conversation.
[0057] Other advantages, objects, and features of the present invention will be described in part in the following description and, in part, will be apparent to those skilled in the art upon examination of the following description or may be learned from practice of the present invention. The objects and other advantages of the present invention may be realized and obtained through the following description. BRIEF DESCRIPTION OF THE DRAWINGS
[0058] In order to make the purpose, technical solutions and advantages of the present invention more clear, the present invention will be described in detail below with reference to the accompanying drawings, in which:
[0059] Figure 1 The process of emotion recognition in conversations using graph neural networks with noise reduction and error correction;
[0060] Figure 2 This is a graph neural network conversation emotion recognition model with noise reduction and error correction;
[0061] Figure 3 This is the feature error correction module diagram. DETAILED DESCRIPTION
[0062] The following describes the embodiments of the present invention by means of specific examples, and those skilled in the art can easily understand other advantages and effects of the present invention from the contents disclosed in this specification. The present invention can also be implemented or applied through other different specific embodiments, and the details in this specification can also be modified or changed in various ways based on different viewpoints and applications without departing from the spirit of the present invention. It should be noted that the illustrations provided in the following embodiments are only schematic illustrations of the basic concept of the present invention, and the following embodiments and features in the embodiments can be combined with each other without conflict.
[0063] Among them, the accompanying drawings are only for illustrative purposes and represent only schematic diagrams rather than actual pictures, and should not be understood as limiting the present invention. In order to better illustrate the embodiments of the present invention, some parts of the accompanying drawings may be omitted, enlarged or reduced, and do not represent the dimensions of actual products. For those skilled in the art, it is understandable that some well-known structures and their descriptions may be omitted in the accompanying drawings.
[0064] The same or similar numbers in the drawings of the embodiments of the present invention correspond to the same or similar parts; in the description of the present invention, it should be understood that if there are terms such as "upper", "lower", "left", "right", "front", "back", etc. indicating directions or positional relationships, they are based on the directions or positional relationships shown in the drawings. They are only for the convenience of describing the present invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific direction, be constructed and operate in a specific direction. Therefore, the terms describing the positional relationship in the drawings are only used for illustrative purposes and cannot be understood as limiting the present invention. For ordinary technicians in this field, the specific meanings of the above terms can be understood according to specific circumstances.
[0065] like Figure 1 As shown in the figure, the present invention provides a graph neural network method for conversation emotion recognition with noise reduction and error correction. The utterance is first preprocessed into semantic features, then passes through a context filtering module, a graph processing module, a feature error correction module, and finally a classification module for emotion classification.
[0066] Figure 2 This is a system model diagram of the present invention. The following is an explanation with reference to the accompanying drawings, which includes the following modules:
[0067] Module 1: Through the preprocessing module, each sentence in the conversation will be segmented by the tokenizer, and then special flags [CLS] and [SEP] will be added before and after. The utterance will be processed into semantic features by the pre-trained language model;
[0068] Module 2: Through the context filtering module, the semantic relevance and information content of the utterance context are evaluated respectively. Based on the evaluation results, some noise context is discarded and the dependency relationship between utterances is established;
[0069] Module 3: Utilize a graph processing module composed of a relational graph neural network to learn contextual information from the context and obtain context-embedded sentiment features;
[0070] Module 4: The feature error correction module integrates semantic and sentiment features, controls erroneous information in the features through a gating mechanism, corrects some errors, and further improves model performance.
[0071] Module 5: Predict the emotional category of the speech through a classification module composed of a fully connected network, and use the cross-entropy loss function to measure the difference between the label distribution and the model prediction distribution.
[0072] Optionally, Module 1 specifically includes:
[0073] Preprocessing module. Define a conversation as a set of N sentences {u1,u2,...,u N Through the preprocessing module, these utterances are processed into N tokens, and special flags [CLS] and [SEP] are added before and after these tokens. The input processing process of the pre-trained language model is as follows:
[0074] X i ={[T1],[T2],...,[T N ]},
[0075] X i ={[CLS],[T1],[T2],...,[T N ],[SEP]},
[0076] H i =RoBERTa(X i ).last_hidden_layer[0].
[0077] where X i For discourse u i The word segmentation after adding the special flag bit is represented as H i It is the feature vector represented by [CLS] in the last hidden layer of the pre-trained language model RoBERTa. [CLS] can represent the semantic feature vector of the entire sentence, which can be directly used for classification tasks.
[0078] Module 2 (optional) specifically includes:
[0079] Context filtering module. To reduce the noise information introduced by the context, a context filtering module is designed to establish dependencies between utterances. The module can simultaneously consider the semantic relevance and information content of the context, avoiding the loss of distant context information and reducing the interference of noise on the model. Through the context filtering module, cosine similarity is used to measure the target utterance feature H i The degree of semantic relevance between the context feature H is represented by the semantic similarity matrix s1:
[0080]
[0081]
[0082] Information entropy is used to measure the amount of information contained in the context. The higher the information entropy of an utterance, the greater the uncertainty in its sentiment, and the more focused the model analysis should be on that utterance. The information entropy of an utterance is calculated by summing the information entropies of all words in the utterance. If prepositions or pronouns frequently appear in an utterance, the information entropy may be low, acting as a transition between conversations. These low-information sentences are not conducive to the model's use as context for sentiment analysis and may even introduce noise during model training. The information entropy matrix is calculated as follows:
[0083]
[0084]
[0085]
[0086] Among them, p(T j ) represents the discourse u i The word frequency of the jth participle in .
[0087] In order to make full use of the semantic correlation matrix and the information entropy matrix, the two matrices are weighted and summed to obtain the final comprehensive score matrix:
[0088] s=(1-α)·s1+α·s2
[0089] Here, α is a hyperparameter representing the score weight. By adjusting the weight, the comprehensive score matrix can be adjusted to be dominated by the information entropy matrix or the semantic relevance matrix. The threshold function then processes each element of the comprehensive score matrix s to construct the adjacency matrix of the target utterance node. Because the constructed graph structure does not contain self-connections, the diagonal elements of the adjacency matrix are set to 0:
[0090]
[0091] Among them, γ is a hyperparameter, which represents the threshold. ij Represents the context utterance u j Relative to the target utterance u i The comprehensive score of mask. ij Represents the discourse node u i and u j The connection relationship between them, 1 means there is a connection, 0 means there is no connection.
[0092] Module 3 is optional and includes:
[0093] Graph processing module. A graph processing module is constructed by stacking L layers of relational graph neural networks to simulate the interaction and information propagation between discourses. First, the edge weights between the target discourse node and the context discourse node in the lth layer are initialized using the similarity-based attention mechanism, which represents u j To u i Degree of impact:
[0094]
[0095] in, Represents node u i The adjacency matrix of is a learnable parameter. According to the information propagation formula in the relational graph neural network, for the l+1 layer of speech u i , through the aggregation node u i The influence of context on the discourse is obtained by using the neighbor node information at the lth layer, while maintaining the sentiment information of the discourse itself at the lth layer to a certain extent:
[0096]
[0097] in, is a trainable parameter, α ii It is the edge weight of the self-connection between the same node in different layers, which means the semantic similarity of the same sentence. The default value is 1, so the formula can be rewritten as:
[0098]
[0099] in is α ij Then, the node features of each layer containing graph structure information and context information are All are stacked up to get a contextual embedding feature M from shallow to deep:
[0100] M=concat(H l ),l∈[0,L]
[0101] Finally, the emotional feature D of the speech is output through the fully connected layer and activation function, where the PReLU activation function can effectively alleviate the problem of overfitting:
[0102] D=PReLU(W d M+b d )
[0103] Module 4 is optional and includes:
[0104] Feature Correction Module. To reduce model prediction errors, a feature correction module was proposed, inspired by an enhanced LSTM network proposed in a named entity recognition study. This module utilizes a gating mechanism to separately control semantic features, context-embedded sentiment features, and the fusion of the two, enabling the model to focus on the semantic information of the utterance while also considering contextual information. Unlike the enhanced LSTM, because the graph processing module already enables the acquisition and propagation of long-distance information, the feature correction module no longer uses a chain structure to obtain information from the previous layer. Instead, each feature correction module acts as an independent unit to process the input at the current moment.
[0105] The input of the feature error correction module is divided into two parts. The upper branch input comes from the output of the preprocessing module, which represents the semantic features of the discourse and is represented by H; the lower branch input comes from the output of the graph processing module, which represents the emotional features of the discourse and is represented by D. Then, the weighted sum of the two branch inputs is obtained to obtain the fused feature C, and the fused feature will pass through the fusion information branch alone. In most previous models, the classification module will directly classify the fused features without controlling the erroneous information therein. The upper and lower branches each update the feature information through a memory gate, where the memory gate is composed of two activation functions and a multiplier. The fusion information branch uses a forget gate to forget part of the fused feature information, where the forget gate is composed of an activation function:
[0106] f=σ(W f H+Q f D+b f )
[0107] At the same time, the feature error correction module will evaluate the inputs of the upper and lower branches respectively:
[0108]
[0109]
[0110] Where σ is the activation function, W, Q, and b are learnable parameters. The memory gate consists of a forget gate and a learning gate, which are used to remember the important feature information in the semantic feature. The semantic feature H passes through the forget gate z to determine which information in the feature needs to be updated, and learns new feature information in the learning gate c, and then the semantic feature H is updated to Similarly, the sentiment feature D is updated to For the forgotten parts in the fusion features, the updated semantic features and sentiment features in the upper and lower branches are used as supplements to control the error information in the features and obtain the output C:
[0111]
[0112] Figure 3 This is the structure of the feature error correction module. The upper branch input is the output of the preprocessing module, which represents the original semantic features of the utterance. The lower branch input is the output of the graph processing module, which represents the sentiment features embedded in the context. After the error correction module, the final output is the fused features of the utterance.
[0113] Module 5 is optional and includes:
[0114] The output after feature correction processing will be used as the final discourse feature, and the discourse u will be obtained using the fully connected network. i The probability P of each emotion category in i , and then take the category subscript with the highest probability as the final sentiment label:
[0115] P i =softmax(W c C i +b c )
[0116]
[0117] in It is discourse i The predicted emotion, W C and b C are the trainable weights and the bias of the fully connected layer respectively; cross entropy is used as the loss function, and the expression of the function is:
[0118]
[0119] Among them, θ is the set of parameters that need to be trained in the model, y i represents the highest predicted probability of the i-th sentiment label, Represents the true value of the i-th sentiment label.
[0120] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not limiting. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solutions of the present invention can be modified or replaced by equivalents without departing from the purpose and scope of the technical solutions, which should all be included in the scope of the claims of the present invention.
Claims
1. A graph neural network conversation emotion recognition method with noise reduction and error correction, characterized by: The method comprises the following steps: Step 1: Through the preprocessing module, each sentence in the conversation is tokenized and special flags [CLS] and [SEP] are added to facilitate subsequent classification tasks. It is then processed into semantic features in the pre-trained language model. Step 2: Use the context filtering module to evaluate the semantic relevance and information content of the utterance context respectively, discard some noise context based on the evaluation results, and establish the dependency relationship between utterances; Step 3: Use the relational graph neural network to learn contextual information from the context and obtain context-embedded sentiment features; Step 4: Use the feature error correction module to integrate semantic features and sentiment features, control the erroneous information in the features through the gating mechanism, and correct some of the errors; Step 5: Predict the sentiment category of the speech through the fully connected network, and use the cross-entropy loss function to measure the difference between the label distribution and the model prediction distribution.
2. The method for conversation emotion recognition using a graph neural network with noise reduction and error correction according to claim 1, characterized in that: The specific process of step 1 includes: first obtaining the discourse set {u1, u2, ..., u N }, use the word segmenter to segment the speech, and then add special flags [CLS] and [SEP] before and after the word segmentation results to facilitate subsequent classification tasks; then use the pre-trained language model to classify the speech u i Processed into a 1024-dimensional semantic feature vector H i In the form of: X i ={[CLS],[T1],[T2],...,[T N ],[SEP]} H i =RoBERTa(X i ).last_hidden_layer[0] Among them, X i For discourse u i The word segmentation after adding the special flag bit is represented as H i It is the feature vector represented by [CLS] of the last hidden layer in the pre-trained language model RoBERTa; [CLS] can represent the semantic feature vector of the entire sentence, which can be directly used for classification tasks.
3. The method for conversation emotion recognition using a graph neural network with noise reduction and error correction according to claim 2, characterized in that: In the second step, the context screening process includes: using cosine similarity to measure the target speech feature H i The degree of semantic relevance between the context feature H is represented by the semantic similarity matrix s1; At the same time, information entropy is used to measure the amount of information contained in the context. It is known that the higher the information entropy of a discourse, the higher the uncertainty of its emotion, and the model needs to focus on analyzing the discourse. By accumulating the information entropy of each word in the discourse, the information entropy matrix of the context can be obtained: Among them, p(T j ) represents the discourse u i The frequency of the jth word in the conversation; in order for the information entropy matrix to match the dimension of the semantic correlation matrix, stack N To construct the same information entropy matrix as s1, we can get s=(1-α)·s1+α·s2 The weighted summation of semantic relevance matrix s1 and information entropy matrix s2 is performed, and each element in the summation matrix s is screened to remove some noise contexts with low scores; Among them, γ is a hyperparameter, which represents the threshold; s ij Represents the context utterance u j Relative to the target utterance u i The comprehensive score of mask ij Represents the discourse node u i and u j The connection relationship between them, 1 means there is a connection, 0 means there is no connection.
4. The method for conversation emotion recognition using a graph neural network with noise reduction and error correction according to claim 3, characterized in that: In step 3, a graph processing module consisting of an N-layer relational graph neural network is defined, and a graph attention mechanism is used to focus on important nodes among neighboring nodes and assign weights; Discourse node u i The neighbor node information at the lth layer is aggregated to obtain the influence of the context on the discourse, while maintaining the sentiment information of the discourse itself at the lth layer to a certain extent: in is α ij Normalized representation of ; The node features of each layer containing graph structure information and context information All are stacked up to get a contextual embedding feature M from shallow to deep: M=concat(H l ),l∈[0,L] Finally, the emotional feature D of the speech is output through the fully connected layer and activation function, where the PReLU activation function can effectively alleviate the problem of overfitting: D(PReLU(W). d M+b d )。 。 5. The method for conversation emotion recognition using a graph neural network with noise reduction and error correction according to claim 4, characterized in that: In step 4, the semantic features from the pre-processing language module and the sentiment features from the graph processing module are integrated through the feature correction module, and the error information in the features is controlled by using the gating mechanism; The input of the feature correction module is divided into two parts. The upper branch input comes from the output of the preprocessing module, which represents the semantic features of the speech and is represented by H. The lower branch input comes from the output of the graph processing module, which represents the emotional features of the speech and is represented by D. Then, the weighted sum of the two branch inputs is used to obtain the fusion feature C. The fused features will pass through the fusion information branch alone. In most previous models, the classification module will directly classify the fused features without controlling the erroneous information. The upper and lower branches each update the feature information through a memory gate, where the memory gate is composed of two activation functions and a multiplier. The fusion information branch uses a forget gate to forget part of the fused feature information, where the forget gate is composed of an activation function: f=σ(W f H+Q f D+b f ) At the same time, the feature error correction module will evaluate the inputs of the upper and lower branches respectively: Among them, σ is the activation function, W, Q and b are learnable parameters; the memory gate consists of a forget gate and a learning gate, which is used to remember the important feature information in the semantic feature; the semantic feature H passes through the forget gate z to determine which information in the feature needs to be updated, and learns new feature information in the learning gate c, and then the semantic feature H is updated to Similarly, the sentiment feature D is updated to For the forgotten parts in the fusion features, the updated semantic features and sentiment features in the upper and lower branches are used as supplements to control the error information in the features and obtain the output C: Among them, f They are the fusion features, semantic features and emotional features after error correction.
6. The method for conversation emotion recognition using a graph neural network with noise reduction and error correction according to claim 5, characterized in that: In step 5, the output after feature correction processing will be used as the final discourse classification feature, and the discourse u is obtained using the fully connected network. i The probability P of each emotion category in i , and then take the category subscript with the highest probability as the final sentiment prediction label y i : P i =softmax(W c C i +b c ) in It is discourse i The predicted emotion, W C and b C are the trainable weights and the bias of the fully connected layer respectively; cross entropy is used as the loss function, and the expression of the function is: Among them, θ is the set of parameters that need to be trained in the model, y i represents the highest predicted probability of the i-th sentiment label, Represents the true value of the i-th sentiment label.