A device for classifying atypical arousal emotions of bullet screen text based on AeerBERT

Abstract

The present invention discloses a method for classifying atypical aroused emotions of bullet-screen texts based on AeerBERT, which belongs to the field of natural language processing and artificial intelligence technology. The method comprises the following steps: step 1, constructing a data set; step 2, constructing an input coding module, in which a bullet-screen group obtains feature vectors output by 12 output layers through a BERT encoder; step 3, constructing a feature enhancement module and a local perception module, which are used to optimize the feature vectors output by the 12 output layers to obtain comprehensive feature representation and local perception feature representation; step 4, constructing a feature fusion and emotion prediction module, which obtains emotion probability distribution and gives an emotion label with maximum probability; step 5, constructing a classification model for atypical aroused emotions of bullet-screen texts based on AeerBERT; and using the data set constructed in step 1 to train the classification model; and step 6, using the trained classification model to classify atypical aroused emotions of video bullet-screen texts.

Description

Technical Field

[0001] The present invention belongs to the field of natural language processing and artificial intelligence technology, and specifically relates to a method for classifying atypical aroused emotions of barrage texts based on AeerBERT. Background Art

[0002] In recent years, China's short video industry has experienced significant development and changes. The content types of short video platforms are becoming more and more diverse, covering various fields such as life, entertainment, education, food, travel, etc.

[0003] Traditionally, research in video sentiment analysis has focused on identifying the target emotions intended by video creators, known as "implied emotions." However, the actual dynamic emotional reactions viewers experience while watching—the "emotions evoked by video clips"—requires greater attention in the multimedia industry. In-depth analysis of the emotions evoked by video clips not only optimizes platforms' video recommendation algorithms but also provides users with warnings, such as "High-energy alert! Blood and violence ahead." While some platforms already offer content warnings, these are crude. For example, "This video may cause discomfort." They don't clearly indicate the specific clip or aspect of the discomfort. This discomfort stems from analysis of the video content, not user feedback. Some videos, in an effort to circumvent censorship mechanisms, employ homophonic characters or obscure alternative expressions. Analyzing the emotional content of videos cannot address this issue. Furthermore, in the field of internet governance, analyzing the emotions evoked by video clips can help assess abnormal fluctuations in user emotions, opening up new avenues for monitoring unusual internet events and assessing user mental health.

[0004] Danmu (bullet comments) provide instant feedback from users watching videos, and are timestamp-linked to the video playback time, making them ideal for studying the emotions elicited by video clips. Despite significant progress in Danmu sentiment analysis, existing research has largely focused on analyzing individual Danmu texts, while research on the dynamic connections and representations of Danmu groups within a specific timeframe remains insufficient. The interactive, sequential, and multi-layered nature of Danmu means that processing each comment in isolation can result in the loss of important contextual information and fail to fully capture the emotions elicited by the video within a specific timeframe.

[0005] Currently, sentiment classification for both bullet comments and videos is limited to emotion polarity (positive, negative, and neutral) or typical emotion classification. However, while typical emotion classification primarily reflects an individual's direct feelings, in video viewing scenarios, viewers often experience content from a third-party perspective. This discrepancy limits the application of traditional classification for bullet comment sentiment analysis. Furthermore, video content often incorporates common narrative techniques such as progression, transitions, and flashbacks. Therefore, viewers' emotions can continuously deepen or suddenly reverse. Consequently, when watching videos, users experience more nuanced and subtle emotions, such as resonance, speechlessness, and suffocation. Summary of the Invention

[0006] Purpose of the Invention: To address the current challenges of insufficient research on video emotion analysis and sentiment analysis of bullet comments, this paper proposes an AeerBERT-based method for classifying atypically aroused emotions in bullet comments. This method significantly improves classification accuracy by considering characteristics such as bullet comment interaction clustering, the incompatibility of traditional emotions, and the ambiguity between aroused emotions.

[0007] Technical solution: A method for classifying atypical evocative emotions in bullet comments based on AeerBERT, including the following steps:

[0008] Step 1: Build a dataset: Crawl bullet comments from video websites and annotate them. The annotations are the subjective emotion recognition results of the annotators. After obtaining the dataset, divide the data in the dataset into bullet comment groups using a dynamic sliding window and preprocess the dataset.

[0009] Step 2: Build an input encoding module. In the input encoding module, the barrage group obtains the feature vectors of the 12 output layers through the BERT encoder.

[0010] Step 3: Construct a feature enhancement module and a local perception module to optimize the feature vectors output by the 12 output layers to obtain comprehensive feature representation and local perception feature representation;

[0011] Step 4: Construct a feature fusion and sentiment prediction module to combine the comprehensive feature representation and local perception feature representation obtained in step 3, perform weighted summation and fuse the features of these two parts, and then send them to the classification layer. Use the fully connected layer to obtain the sentiment probability distribution and give the sentiment label with the highest probability.

[0012] Step 5: Based on the constructed input encoding module, feature enhancement module, local perception module, feature fusion and sentiment prediction module, a barrage text atypical arousal sentiment classification model based on AeerBERT is constructed; and the classification model is trained using the dataset constructed in step 1;

[0013] Step 6: Use the trained classification model to classify atypical arousal emotions in video barrage text.

[0014] Furthermore, in step one, the atypical emotional categories include funny, sweet, refreshing, resonance, speechless, bloody and violent, and suffocating.

[0015] Furthermore, the specific steps for constructing the dataset in step 1 are as follows:

[0016] Step 1.1: Crawl some video commentary data and annotate each commentary with an emotion vector. Each value in the vector represents the subjective emotion recognition result of the commentary by different annotators, i.e., it labels the atypical arousing emotion category.

[0017] Use label i It represents the emotion recognition result annotated by the i-th annotator, d represents each barrage, t represents the timestamp corresponding to each barrage, and the aggregated data format can be expressed as a triple <t,d,[label1,label2,…,label 10 ]>;

[0018] Step 1.2, for [label1, label2, ..., label 10 ]Vote and get the label with the most votes * , identified as the final sentiment label of d; for voting results with multiple labels * In the case of , human arbitration is conducted to select the most appropriate label * ; After the annotation is completed, each data format is expressed as <t,d,label * >

[0019] Step 1.3, design a dynamic sliding window, which is based on the timestamp t i and emotional labels For data d i Combine and divide to form a barrage group D j =Conact(d i d i+1 …d k ), i represents the row number of a barrage d, and j represents the row number of a combined barrage group D;

[0020] Step 1.4: Data preprocessing: Clean the data, including removing duplicates, removing blanks, removing stop words, and removing noise information.

[0021] Furthermore, in step 2:

[0022] In the input encoding module, a text input vector is generated based on the text input data of the barrage group, and the text input vector is converted into the corresponding vector representation through the BERT embedding layer, and the feature vector H output by all BERT encoding layers is obtained. i , i<=12.

[0023] Furthermore, in step three:

[0024] In the feature enhancement module, the feature vector H is comprehensively applied by using multiple neurons i The multi-dimensional features of the first 11 output layers are used with a fully connected layer to obtain the output Y that combines phrase and syntactic information. i , the Y corresponding to the first 11 output layers i and the feature vector H of the 12th output layer 12 Combination, by setting the weight layer and taking the weighted average, we can get the comprehensive feature representation A.

[0025] Furthermore, in step three:

[0026] In the local perception module, the feature Y of the feature vectors of the first 11 output layers is combined with the 12th output layer, and the local information is captured through the CNN layer to obtain the local perception feature representation B.

[0027] Beneficial Effects: The AeerBERT-based barrage text atypical arousal sentiment classification method proposed in the present invention can effectively classify different sentiments with similar characteristics, such as sweetness and refreshing. The barrage atypical arousal sentiment analysis in the present invention is significantly different from the traditional barrage sentiment analysis and even the sentiment analysis of comment text. While improving the accuracy, the method of the present invention also surpasses the traditional sentiment classification model in terms of integrity and stability. This method also provides a new reference idea for video detection and recommendation. Specifically:

[0028] 1) This paper constructs a barrage group dataset and classification model for studying atypical emotional stimulation, which can identify 8 atypical emotion categories that are commonly applicable to videos: funny, sweet, refreshing, resonance, speechless, bloody and violent, and suffocating.

[0029] 2) The present invention adopts the BERT coding layer to obtain the vector representation of the barrage group, and designs a model based on the uniqueness of the barrage group and atypical emotional stimulation. By enhancing the boundaries of fuzzy emotional features and combining multi-dimensional comprehensive features and local features, the accuracy is improved while surpassing the traditional sentiment classification model in terms of integrity and stability. BRIEF DESCRIPTION OF THE DRAWINGS

[0030] Figure 12 is a schematic diagram of a processing flow framework of a method for classifying atypical aroused emotions of bullet-screen texts based on AeerBERT in an embodiment of the present invention;

[0031] Figure 2 Schematic diagram of data set construction in the method for classifying atypical aroused emotions of bullet-screen text based on AeerBERT in an embodiment of the present invention;

[0032] Figure 3 Schematic diagram of the input encoding module of the AeerBERT-based barrage text atypical arousal emotion classification model in an embodiment of the present invention.

[0033] Figure 4 2. It is a schematic diagram of a feature enhancement module of a barrage text atypical arousal emotion classification model based on AeerBERT in an embodiment of the present invention.

[0034] Figure 5 It is the intention of the local perception module of the barrage text atypical stimulation emotion classification model based on AeerBERT in an embodiment of the present invention.

[0035] Figure 6 Schematic diagram of the feature fusion and emotion prediction module of the barrage text atypical arousal emotion classification model based on AeerBERT in an embodiment of the present invention. DETAILED DESCRIPTION

[0036] The technical solution of the present invention is described in detail below with reference to the accompanying drawings, but the protection scope of the present invention is not limited to the embodiments.

[0037] like Figure 1 As shown, the present invention provides a method for classifying atypical arousal emotions of bullet screen text based on AeerBERT, comprising the following steps:

[0038] Step 1: Build a dataset: Crawl bullet comments from video websites and annotate them. The annotations are the subjective emotion recognition results of the annotators. After obtaining the dataset, divide the data in the dataset into bullet comment groups using a dynamic sliding window and preprocess the dataset.

[0039] Existing public datasets cannot meet the requirements of this invention, so a new dataset needs to be constructed. Regarding the processing of barrage, existing research focuses on the analysis of single barrage text, while the research on the dynamic connection and characterization of barrage groups within a specific time period is still insufficient. Processing each barrage in isolation may lead to the loss of important contextual information and fail to fully capture the user emotions aroused by the video within a certain time period. Therefore, the present invention adopts the form of barrage groups.

[0040] like Figure 2 As shown, the specific operations of step 1 are as follows:

[0041] Step 1.1: In this embodiment, we select and crawl several barrage data from multiple types of videos on Bilibili, such as social, legal, and psychological. We crawl several barrage data from the video barrage, and label each barrage data with an emotion vector. Each value in the vector is the subjective emotion recognition result of different annotators for the barrage; assuming that label i It represents the sentiment result of the i-th annotator, d represents each comment, t represents the timestamp of each comment, and the aggregated data format can be expressed as a triple <t,d,[label1,label2,…,label 10 ]>.

[0042] The subjective emotion recognition result is an atypical arousal emotion category, including funny, sweet, refreshing, sympathetic, speechless, bloody violence, and stifling.

[0043] Step 1.2, then [label1, label2, ..., label 10 ]Vote and get the label with the most votes * , identified as the final sentiment label of d. For voting results with multiple labels * In the case of , human arbitration is conducted to select the most appropriate label * After the annotation is completed, each data format can be expressed as <t,d,label * >.

[0044] Step 1.3, after that, design a dynamic sliding window that is based on the timestamp t i and emotional labels For data d i Combine and divide to form a barrage group D j =Conact(d i d i+1 …d k ), i represents the line number of a barrage d, and j represents the line number of a combined barrage group D.

[0045] Finally, in step 1.4, data preprocessing is performed. The data is cleaned, including removing duplicates, blanks, stop words, and noise. Regular expressions are used to filter out noise in the text.

[0046] Step 2: Build an input encoding module and pass the barrage group through the BERT encoder to obtain a 12-layer feature representation;

[0047] Generate text input vector based on the text input data of the bullet screen group, such as Figure 3 As shown, the text is converted into the corresponding vector representation through the BERT embedding layer, and the feature vectors H1, H2, ..., H output by all BERT encoding layers are obtained. 12 .

[0048] In step 2 of the present invention, all hidden layer codes in the encoder are extracted, and different fully connected layers are used to integrate and reorganize their surface phrases, middle-level syntactic information, and high-level semantic feature representations, attempting to mine atypical emotions from semantic information at all levels to distinguish emotions with blurred boundaries.

[0049] Step 3: Optimize and reorganize the feature vectors output by the 12 output layers to obtain comprehensive feature representation and local perception feature representation;

[0050] Step 3.1: Construct a feature enhancement module to enhance the feature vectors of the first 11 output layers and the feature vector of the 12th output layer, and perform weighted averaging to obtain a comprehensive feature representation.

[0051] Specifically, such as Figure 4 As shown, by using multiple neurons to comprehensively apply the feature vector H i The multi-dimensional features of the first 11 output layers are used with a fully connected layer to obtain the output Y that combines phrase and syntactic information. i , the Y corresponding to the first 11 output layers i and the feature vector H of the 12th output layer 12 Combination, by setting the weight layer and taking the weighted average, we get the comprehensive feature representation A;

[0052] In the present invention, by using multiple neurons to comprehensively utilize the multidimensional features of Hi, a fully connected layer is used for each Hi (i <= 11) to obtain the output Y that integrates the phrase and syntactic information. i , can enhance the feature representation of each sample. 12 With Y i Splicing, using weighted averaging to deal with the problem of discrete features.

[0053] Step 3.2: Construct a local perception module, combine the features of the feature vectors of the first 11 output layers with the feature vector of the 12th output layer, capture local information through the CNN layer, and obtain the local perception feature representation B;

[0054] Specifically, the feature Y of the feature vectors of the first 11 output layers is combined with the 12th output layer, and the local information is captured through the CNN layer to obtain the local perception feature representation B.

[0055] In step 2, the reorganized phrases, grammar and semantic representations of the barrage group text are obtained. Because of the differences in structure and content between the barrage group text and the comment text, phrases and grammar have a great influence on the analysis of the sentiment of the barrage group, not just from semantics. However, the emphasis of each encoding layer output of BERT on these three aspects is different, and the features are relatively discrete. Therefore, it is necessary to use a weighted average method for it. Therefore, in step 3, a weight layer is set to allow the model to learn the optimal weight combination. Since the barrage group contains multiple sentences, and the close dependency between word phrases in each sentence helps to enhance boundary clarity, the present invention uses a convolutional neural network (CNN) to capture the close dependency in the text.

[0056] Step 4: Feature fusion and sentiment prediction module, such as Figure 5 As shown in the figure, the comprehensive feature representation A and the local perception feature representation B strengthened in step three are weighted and fused, and then sent to the classification layer. The emotion probability distribution is obtained using the fully connected layer to give the emotion label with the maximum probability.

[0057] Step 5: Based on the modules constructed in Steps 2-4, construct a classification model for the AeerBERT-based method for classifying atypical evocative emotions in bullet comments. The classification model includes an input encoding module, a feature enhancement module, a local perception module, and an output module. The classification model is trained using the dataset constructed in Step 1.

[0058] Input encoding module: used to obtain vector representations of different dimensions for the barrage group text through the BERT encoding layer.

[0059] Feature Enhancement Module: The fully connected layer reorganizes the feature vector. Each encoding layer output of BERT contains different levels of multi-dimensional information such as phrases, syntax, and semantics. Through multiple neurons, the multi-dimensional features of the text are reintegrated to obtain a vector representation of comprehensive phrase and syntactic information. The final layer output can incorporate rich semantic information, ultimately obtaining a comprehensive feature representation.

[0060] Local perception module: Convolutional neural network (CNN) can capture the close dependencies in the text, enhance the boundary clarity between fuzzy features, and obtain local perception feature representation.

[0061] Feature fusion and sentiment prediction module: It fuses the features obtained by feature enhancement and the information obtained by local perception, and then uses the fully connected layer to perform sentiment classification.

[0062] Use the dataset constructed in step 1 to train the classification model;

[0063] The feature enhancement module and the feature fusion and sentiment prediction module use weight layers. The model training process determines the distribution of hidden states in different layers. The resulting weight layer parameters for the feature enhancement module are [0.005, 0.005, 0.005, 0.055, 0.03, 0.05, 0.05, 0.1, 0.1, 0.15, 0.4], and the weight parameters for the feature fusion and sentiment prediction module are [0.05, 0.95].

[0064] Step 6: The trained classification model is used to classify the atypical arousal emotions of video barrage text.

[0065] The above description is only a preferred embodiment of the present invention, and the protection scope of the present invention is not limited to the above embodiment. Any equivalent modifications or changes made by ordinary technicians in this field based on the contents disclosed in the present invention should be included in the protection scope recorded in the claims.

Claims

1. A method for classifying atypical arousal emotions in bullet comments based on AeerBERT, characterized by: The steps include: Step 1: Build a dataset: Crawl bullet comments from video websites and annotate them. The annotations are the subjective emotion recognition results of the annotators. After obtaining the dataset, divide the data in the dataset into bullet comment groups using a dynamic sliding window and preprocess the dataset. Step 2: Build an input encoding module. In the input encoding module, the barrage group obtains the feature vectors of the 12 output layers through the BERT encoder. Step 3: Construct a feature enhancement module and a local perception module to optimize the feature vectors output by the 12 output layers to obtain comprehensive feature representation and local perception feature representation; Step 4: Construct a feature fusion and sentiment prediction module to combine the comprehensive feature representation and local perception feature representation obtained in step 3, perform weighted summation and fuse the features of these two parts, and then send them to the classification layer. Use the fully connected layer to obtain the sentiment probability distribution and give the sentiment label with the highest probability. Step 5: Based on the constructed input encoding module, feature enhancement module, local perception module, feature fusion and sentiment prediction module, a barrage text atypical arousal sentiment classification model based on AeerBERT is constructed; and the classification model is trained using the dataset constructed in step 1; Step 6: Use the trained classification model to classify atypical arousal emotions in video barrage text.

2. The method for classifying atypical arousal emotions of bullet comments based on AeerBERT according to claim 1 is characterized in that: In step one, atypical emotional categories include funny, sweet, refreshing, resonance, speechless, bloody and violent, and suffocating.

3. The method for classifying atypical sentiments of bullet comments based on AeerBERT according to claim 1 is characterized in that: The specific steps for constructing the dataset in step 1 are as follows: Step 1.1: Crawl some video commentary data and annotate each commentary with an emotion vector. Each value in the vector represents the subjective emotion recognition result of the commentary by different annotators, i.e., it labels the atypical arousing emotion category. Use label i It represents the emotion recognition result annotated by the i-th annotator, d represents each barrage, t represents the timestamp corresponding to each barrage, and the aggregated data format can be expressed as a triple <t,d,[label1,label2,…,label 10 ]>; Step 1.2, for [label1, label2, ..., label 10 ]Vote and get the label with the most votes * , identified as the final sentiment label of d; for voting results with multiple labels * In the case of , human arbitration is conducted to select the most appropriate label * ; After the annotation is completed, each data format is expressed as <t,d,label * > Step 1.3, design a dynamic sliding window, which is based on the timestamp t i and sentiment label *i For data d i Combine and divide to form a barrage group D j =Conact(d i d i+1 …d k ), i represents the row number of a barrage d, and j represents the row number of a combined barrage group D; Step 1.4: Data preprocessing: Clean the data, including removing duplicates, removing blanks, removing stop words, and removing noise information.

4. The method for classifying atypical emotional stimulation of bullet comments based on AeerBERT according to claim 3 is characterized in that: In step 2: In the input encoding module, a text input vector is generated based on the text input data of the barrage group, and the text input vector is converted into the corresponding vector representation through the BERT embedding layer, and the feature vector H output by all BERT encoding layers is obtained. i , i<=12.

5. The method for classifying atypical sentiments of bullet comments based on AeerBERT according to claim 4 is characterized in that: In step three: In the feature enhancement module, the feature vector H is comprehensively applied by using multiple neurons i The multi-dimensional features of the first 11 output layers are used with a fully connected layer to obtain the output Y that combines phrase and syntactic information. i , the Y corresponding to the first 11 output layers i and the feature vector H of the 12th output layer 12 Combination, by setting the weight layer and taking the weighted average, we can get the comprehensive feature representation A.

6. The method for classifying atypical emotional stimulation of bullet comments based on AeerBERT according to claim 5 is characterized in that: In step three: In the local perception module, the feature Y of the feature vectors of the first 11 output layers is combined with the 12th output layer, and the local information is captured through the CNN layer to obtain the local perception feature representation B.

Images