Euler space-based emotion continuous space representation method

Abstract

The invention discloses an euler space-based emotion continuous space representation method. The euler space-based emotion continuous space representation method comprises the following steps of: converting emotion vocabularies: converting emotion related vocabularies into tokens which can be processed by a language model; vector embedding generation: converting emotional vocabularies into high-dimensional embedded vectors by using a pre-trained large language model, and representing distribution characteristics of emotions in a semantic space; vector dimension reduction processing: mapping the high-dimensional embedded vector to a two-dimensional plane space by using a dimension reduction algorithm; and euler space conversion and visualization: in the two-dimensional space after dimension reduction, defining the wake-up degree as a radius r and the pleasure degree as an angle theta, constructing euler space representation of emotions on the basis, and displaying emotional point distribution through a visualization tool, the wake-up degree is represented by the radius r, the pleasure degree is represented by the angle theta, and the emotional point distribution is displayed through a visual tool. And meanwhile, an emotion distribution model better conforming to human cognition is constructed by mining the relationship between emotion vectors generated by a large language model, and the correlation and distribution characteristics between different emotions can be reflected more accurately.

Description

Technical Field

[0001] The present invention relates to the technical field of emotion recognition and quantification, and in particular to an emotion continuous space representation method based on Euler space. Background Art

[0002] Existing emotion quantification methods usually use two-dimensional Cartesian coordinates to represent emotions, where valence and arousal are used as the horizontal and vertical axes, respectively, to construct a continuous space of emotions, such as Figure 1 However, this method has the following limitations:

[0003] 1. The representation of arousal is not intuitive: Representing arousal as the value of the coordinate axis cannot intuitively reflect the gradual change of arousal intensity, which is inconsistent with the way people are accustomed to perceiving intensity through length.

[0004] 2. The linearization of emotion distribution is unreasonable: the fixed radius and equal interval method are used to represent emotion points, ignoring the subjective correlation and nonlinear distribution relationship between emotions. Summary of the invention

[0005] The purpose of this section is to summarize some aspects of the embodiments of the present invention and briefly introduce some preferred embodiments. Some simplifications or omissions may be made in this section and the specification abstract and the invention title of this application to avoid blurring the purpose of this section, the specification abstract and the invention title, and such simplifications or omissions cannot be used to limit the scope of the present invention.

[0006] Therefore, the purpose of the present invention is to provide an emotion continuous space representation method based on Euler space, in which the arousal degree is represented by the radius r and the pleasantness is represented by the angle θ. At the same time, by mining the relationship between the emotion vectors generated by the large language model, an emotion distribution model that is more in line with human cognition is constructed, which can more accurately reflect the correlation and distribution characteristics between different emotions.

[0007] To solve the above technical problems, according to one aspect of the present invention, the present invention provides the following technical solutions:

[0008] A method for expressing emotion in continuous space based on Euler space comprises the following steps:

[0009] S1. Conversion of emotional vocabulary:

[0010] Convert sentiment-related words into tokens that can be processed by the language model;

[0011] S2. Vector embedding generation:

[0012] Use a pre-trained large language model to convert sentiment words into high-dimensional embedding vectors to represent the distribution characteristics of sentiment in the semantic space;

[0013] S3, vector dimensionality reduction processing:

[0014] Use dimensionality reduction algorithms to map high-dimensional embedding vectors to a two-dimensional plane space;

[0015] S4, Euler space transformation and visualization:

[0016] In the two-dimensional space after dimensionality reduction, the arousal degree is defined as the radius r and the pleasure degree is defined as the angle θ, and the Euler space representation of emotions is constructed based on this, and the distribution of emotion points is displayed through visualization tools.

[0017] As a preferred solution of the method for expressing emotion continuous space based on Euler space described in the present invention, in step S1, the step of converting emotion-related words into tokens that can be processed by the language model is as follows:

[0018] develop a vocabulary related to emotions;

[0019] Normalize the words in the sentiment vocabulary to ensure consistency of input;

[0020] Use a tokenizer to divide the text into individual words or subunits;

[0021] The sub-units after word segmentation are further converted into IDs acceptable to the model for input into the model for processing;

[0022] All tokens are filled and packaged into the input format of the model.

[0023] As a preferred solution of the method for expressing emotion continuous space based on Euler space described in the present invention, in step S2, the specific steps of training the large language model are as follows:

[0024] Collect large-scale text data from news, books, online articles, and social media content, remove noise from the text data, and process duplicate data, typos, and illegal content;

[0025] Use a specific word segmentation algorithm to convert text into word tokens, and map the word tokens to unique integer IDs;

[0026] Select a pre-trained architecture and use random initialization to set the initial values ​​of the model parameters;

[0027] Map tokens to unique integer IDs for easy model processing;

[0028] Design the objective function and use the autoregressive language modeling objective to predict the next word;

[0029] The word ID is input into the model, and the predicted value is obtained through the embedding layer, the Transformer layer, and the output layer. The error between the predicted result and the true label is calculated by the cross entropy loss function.

[0030] Use the back-propagation algorithm to calculate the gradient and use the optimizer to update the model parameters;

[0031] Use the validation set to evaluate the performance of the model and adjust the model hyperparameters as well as the model weights.

[0032] As a preferred solution of the method for expressing emotion continuous space based on Euler space described in the present invention, in step S3, the specific steps of using a dimensionality reduction algorithm to map a high-dimensional embedding vector to a two-dimensional plane space are as follows:

[0033] Prepare the embedding vectors, making sure you have a high-dimensional embedding vector matrix X, where the dimension of X is n×d, where n is the number of embedding vectors and d is the dimension of each vector;

[0034] Before dimensionality reduction, the embedding vectors were normalized by calculating the mean and standard deviation to center the data to zero mean and unit standard deviation;

[0035] Initialize the dimensionality reduction model, set the target dimension to 2, then train the model and transform the data to map the high-dimensional embedding vector to a two-dimensional plane space.

[0036] As a preferred solution of the method for expressing emotion continuous space based on Euler space described in the present invention, the dimension reduction model is t-SNE.

[0037] As a preferred solution of the method for expressing emotion continuous space based on Euler space described in the present invention, the pre-training architecture is a large language model based on Transformer architecture.

[0038] As a preferred solution of the method for expressing emotion continuous space based on Euler space described in the present invention, the words in the emotion vocabulary are normalized as follows:

[0039] Lowercase: convert all sentiment words into lowercase to avoid the impact of uppercase and lowercase differences on processing;

[0040] Remove extra characters: Delete special symbols and spaces in words.

[0041] Spelling standardization: dealing with vocabulary variations.

[0042] As a preferred solution of the method for expressing emotion continuous space based on Euler space described in the present invention, it is applied to image emotion annotation.

[0043] Compared with the prior art, the present invention has the following beneficial effects:

[0044] 1. The arousal degree is represented by the radius r, and the pleasure degree is represented by the angle θ. At the same time, by mining the relationship between the emotion vectors generated by the large language model, an emotion distribution model that is more in line with human cognition is constructed, which can more accurately reflect the correlation and distribution characteristics between different emotions.

[0045] 2. Enhanced intuitiveness: The radius is used to represent the degree of arousal, which is in line with human perception of intensity; the angle represents the degree of pleasure, showing the diversity of emotions.

[0046] 3. Nonlinear distribution: The distribution of emotional points no longer relies on fixed radius and equal-interval annotation, but is adaptively distributed through semantic relationships generated by the language model, which more realistically reflects the subjective correlation between emotions.

[0047] 4. Data independence: The emotion vector is derived from the semantic embedding generated by the language model, which avoids relying on faces or other data for emotion labeling, reducing subjective bias and data collection costs. BRIEF DESCRIPTION OF THE DRAWINGS

[0048] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the present invention will be described in detail below in combination with the accompanying drawings and detailed embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative labor. Among them:

[0049] Figure 1 A schematic diagram of the traditional Cartesian coordinates representing the emotion space provided by the present invention;

[0050] Figure 2 A schematic diagram of the emotion distribution based on Euler space provided by the present invention. DETAILED DESCRIPTION

[0051] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, the specific embodiments of the present invention are described in detail below with reference to the accompanying drawings.

[0052] Secondly, the present invention is described in detail with reference to schematic diagrams. When describing the embodiments of the present invention in detail, for the sake of convenience, the cross-sectional diagrams showing the device structure will not be partially enlarged according to the general scale, and the schematic diagrams are only examples, which should not limit the scope of protection of the present invention. In addition, in actual production, the three-dimensional dimensions of length, width and depth should be included.

[0053] In order to make the objectives, technical solutions and advantages of the present invention more clear, the embodiments of the present invention will be further described in detail below with reference to the accompanying drawings.

[0054] Example 1

[0055] The present invention provides an emotion continuous space representation method based on Euler space, in which the arousal degree is represented by a radius r and the pleasure degree is represented by an angle θ. At the same time, by mining the relationship between emotion vectors generated by a large language model, an emotion distribution model that is more in line with human cognition is constructed, which can more accurately reflect the correlation and distribution characteristics between different emotions.

[0056] The specific steps of this method of expressing emotion continuous space based on Euler space are as follows:

[0057] S1. Conversion of emotional vocabulary:

[0058] Convert sentiment-related words into tokens that can be processed by the language model;

[0059] S2. Vector embedding generation:

[0060] Use a pre-trained large language model to convert sentiment words into high-dimensional embedding vectors to represent the distribution characteristics of sentiment in the semantic space;

[0061] S3, vector dimensionality reduction processing:

[0062] Use dimensionality reduction algorithms to map high-dimensional embedding vectors to a two-dimensional plane space;

[0063] S4, Euler space transformation and visualization:

[0064] In the two-dimensional space after dimensionality reduction, the arousal degree is defined as the radius r and the pleasure degree is defined as the angle θ, and the Euler space representation of emotions is constructed based on this, and the distribution of emotion points is displayed through visualization tools.

[0065] In step S1, the specific steps for converting emotion-related words into tokens that can be processed by the language model are as follows:

[0066] 1) Identify emotion-related words

[0067] Before starting processing, you need to clarify the vocabulary related to emotions, which can be obtained in the following ways: Emotion dictionary: such as NRC emotion dictionary, WordNet-Affect, etc.; Manual screening: Manually collect common emotion words (such as "happy", "sad", "angry"); Domain-specific extension: Combined with actual application scenarios, add some specific emotion words (such as "sleepy", "tired").

[0068] 2) Text preprocessing

[0069] The words in the emotion vocabulary are normalized to ensure the consistency of the input, as follows: Lowercase: All emotion words are converted to lowercase to avoid the impact of uppercase and lowercase differences on processing; Remove redundant characters: Delete special symbols, spaces, etc. in words (such as "happy!" is converted to "happy"); Spelling standardization: Process vocabulary variants (such as "happier" is converted to "happy").

[0070] 3) Tokenization

[0071] Tokenization is the process of dividing text into individual words or sub-units (tokens). Tokenization of large language models usually relies on a specific tokenizer, and different models may use different tokenizers.

[0072] 3-1) Determine the word segmenter

[0073] Tokenizers of common language models: Tokenizers used by GPT / BERT: Byte Pair Encoding (BPE) or WordPiece; Tokenizers used by RoBERTa: an improved version based on BPE; Tokenizers used by OpenAI GPT: Clustering-based Subword Tokenizer;

[0074] 3-2) Call the word segmenter

[0075] Use the model's built-in word segmentation tool to segment sentiment words. For example, in Python, word segmentation can be implemented using the transformers library.

[0076] 3-3) Special markings

[0077] The tokenizer usually adds special tags, such as: [CLS]: sentence start tag, used for classification tasks; [SEP]: sentence separator, used to distinguish different sentences; [PAD]: padding tag, used to fill the length.

[0078] 4) Convert to IDs

[0079] The tokens after word segmentation need to be further converted into IDs (integer indexes) acceptable to the model so that they can be input into the model for processing;

[0080] 5) Filling and alignment

[0081] For batch processing, in order to ensure the consistency of input length, all tokens need to be padded. The padding method is to fill all tokens to the length of the longest sequence. The padding mark is to use [PAD] or its corresponding ID for padding;

[0082] 6) Convert to model input format

[0083] The final result needs to be packaged into the input format of the model, including: input_ids: ID representation of tokens; attention_mask: marking which positions are filled and which are actual tokens; token_type_ids: used to distinguish sentences.

[0084] In step S2, the specific training steps of the large language model are as follows:

[0085] 1. Data preparation:

[0086] 1-1) Data collection: Collect large-scale text data, such as news, books, online articles, social media content, etc. The data volume is usually between hundreds of GB and several TB;

[0087] 1-2) Data cleaning: remove noise (such as garbled characters, HTML tags), process duplicate data, typos, and illegal content;

[0088] 1-3) Tokenization and encoding: Use a specific tokenization algorithm (such as Byte Pair Encoding or WordPiece) to convert text into tokens; map tokens to unique integer IDs for easy model processing.

[0089] 2. Model architecture design

[0090] 2-1) Choose an architecture: Common pre-trained architectures include: GPT (Generative Pre-trained Transformer): decoder-based; BERT (Bidirectional Encoder Representations from Transformers): encoder-based; T5 (Text-to-Text Transfer Transformer): encoder-decoder-based;

[0091] 2-2) Parameter initialization: Use a random initialization method (such as Xavier initialization) to set the initial values ​​of the model parameters;

[0092] 3. Pretraining

[0093] 3-1) Objective function design: GPT: uses autoregressive language modeling objectives to predict the next word; BERT: uses autoencoding language modeling objectives, including: MLM (Masked Language Model): randomly mask some words and predict the original values ​​of these words, NSP (Next Sentence Prediction): predict whether two sentences have a continuous relationship, T5: uniformly adopts a text-to-text approach and converts all tasks into generation problems.

[0094] 3-2) Forward propagation: Input the word unit ID into the model, pass through the embedding layer, Transformer layer and output layer, and get the predicted value;

[0095] 3-3) Loss calculation: The error between the predicted result and the true label is calculated through the cross-entropy loss function.

[0096] 3-4) Backpropagation and optimization: Use the backpropagation algorithm to calculate the gradient and use an optimizer (such as Adam or its variant AdamW) to update the model parameters.

[0097] 4) Verification and evaluation: 4-1) Use the validation set to evaluate the performance of the model (such as perplexity); 4-2) Adjust model hyperparameters, such as the number of layers, hidden unit size, etc.

[0098] 5) Fine-Tuning: Fine-tuning the model weights on specific tasks (such as sentiment analysis and machine translation) to adapt to the needs of specific tasks.

[0099] The advantages of the pre-trained large language model of the present invention over the traditional model are shown in Table 1

[0100] Table 1 The difference between the pre-trained large language model of the present invention and the traditional model training

[0101]

[0102] In step S3, the specific steps of using the dimensionality reduction algorithm to map the high-dimensional embedding vector to the two-dimensional plane space are as follows:

[0103] 1) Prepare embedding vectors: Make sure you have a high-dimensional embedding vector matrix X, with dimensions n×d, where n is the number of embedding vectors and d is the dimension of each vector;

[0104] 2) Select a dimensionality reduction algorithm. Currently, two commonly used dimensionality reduction algorithms are: PCA (Principal Component Analysis): used for linear dimensionality reduction, fast and efficient, t-SNE (t-distributed random neighbor embedding): suitable for nonlinear dimensionality reduction, often used for visualization of high-dimensional data. In the present invention, t-SNE is selected as the model of the dimensionality reduction algorithm.

[0105] 3) Standardized embedding vector: Before dimensionality reduction, the embedding vector is standardized: the mean and standard deviation are calculated, and the data is centralized to zero mean and unit standard deviation.

[0106] 4) Dimensionality reduction: Initialize the t-SNE model, set the target dimension to 2, train the model and transform the data.

[0107] Figure 2 This is a schematic diagram of emotion distribution generated by a method for expressing emotion in continuous space based on Euler space in the present invention, showing a spatial model in which arousal is represented by radius r and pleasure is represented by angle θ. Figure 2 The following points can be seen:

[0108] (1) According to intensity, emotions can be divided into four ring areas from the inside out: relaxed, depressed, bored, tired, sleepy; pleased, excited, happy; sad, miserable; surprised, contempt, disgust, angry, fear. The more peripheral, the higher the arousal, which is consistent with our intuitive feelings.

[0109] (2) According to the perspective, we can find that the emotions in the first quadrant are sad, miserable, surprised, contempt, disgust, angry, and fear, which are negative emotions. The emotions in the fourth quadrant are pleased, excited, and happy, which are positive emotions. The second and third quadrants are relaxed, depressed, bored, tired, and sleepy, which are relatively neutral emotions.

[0110] According to the degree of aggregation, disgust, contempt, fear, surprise, and anger are close to each other, which reflects the correlation of negative emotions. Disgust and contempt are surrounded by fear, surprise, and anger, which reflects the complex relationship between these emotions. This is also the reason why the recognition accuracy of disgust and contempt in machine vision is not high.

[0111] In order to further verify the consistency between the Euler space distribution generated by the emotion continuous space representation method based on Euler space of the present invention and the actual research data, and prove its scientificity and rationality in emotion quantification, the emotion subjective correlation matrix in psychological research is compared, and the specific analysis method is as follows:

[0112] 1) Definition of the Emotional Subjective Correlation Matrix: The emotional subjective correlation matrix is ​​constructed based on psychological research and describes the subjective similarities or relationships between different emotions. Its core is based on human experimental data (such as questionnaires, psychological tests, etc.) and is obtained through statistical analysis.

[0113] Each element of the matrix represents the similarity or correlation between two emotions, usually in the range ([-1,1]):

[0114] o Positive correlation (close to 1): The emotions are similar (such as "happy" and "pleased").

[0115] o Negative correlation (close to -1): Emotional opposition (such as "happy" and "sad").

[0116] o Zero correlation (close to 0): no significant relationship (such as "excited" and "bored").

[0117] The sample matrix (partial) is shown in Table 2:

[0118] Table 2 - Example matrix (partial):

[0119]

[0120] This matrix reflects human's subjective perception of emotions and is an important basis for studying the spatial distribution of emotions.

[0121] 2) Euler space generated by the present invention

[0122] The emotional Euler space generated by the method of the present invention uses the embedding vector of the pre-trained large language model and is mapped to a two-dimensional plane through a dimensionality reduction method to obtain a polar coordinate representation (radius r and angle) of each emotion. This representation reflects the arousal and pleasure of the emotion.

[0123] Calculated similarity matrix: The similarity between emotions is calculated by the cosine similarity or Euclidean distance between two points in Euler space:

[0124] or:

[0125]

[0126] 3) Comparison method

[0127] The comparison method includes the following steps:

[0128] 3-1) Obtaining the subjective correlation matrix in psychological research: Collect existing psychological research data, such as Russell's emotion circle model or other relevant experimental data, and convert them into a standardized emotion correlation matrix.

[0129] 3-2) Calculating the emotion correlation matrix of the Euler space in the present invention: Using the two-dimensional coordinates generated by the embedded vector to calculate the similarity or distance between emotions, and constructing the corresponding similarity matrix.

[0130] 3-3) Compare the consistency between matrices: Use statistical methods to quantify the similarity between two matrices, such as: Pearson correlation coefficient: measures the linear correlation of corresponding elements of two matrices, mean square error (MSE): measures the overall error between two matrices.

[0131] 4) Visualization Analysis

[0132] 4-1) Heat map of subjective correlation matrix: Heat map can be used to visually display the subjective correlation of psychological research:

[0133] 4-2) Heat map of the matrix generated in Euler space: The heat map of the matrix of the present invention is also generated and compared.

[0134] 4-3) Distribution diagram: Plot the distribution of emotions in two-dimensional Euler space and observe whether the relative position relationship between emotions is consistent with psychological theory.

[0135] 5) Conclusion: By comparing the consistency of the two matrices, combined with the heat map and distribution map, it can be proved whether the Euler space generated by the present invention is consistent with the psychological research data, and then prove its scientificity and rationality. This process can provide strong mathematical and experimental support for emotion quantification.

[0136] Example 2

[0137] The present invention also provides an application of an emotion continuous space representation method based on Euler space in image emotion annotation, which is as follows:

[0138] In machine vision tasks, image emotion annotation is an important part of emotion computing model training. However, since emotions are highly subjective, different annotators may have significant differences in their emotion judgments on the same image. This difference will lead to a decrease in the consistency and reliability of model training data, thereby affecting model performance. In order to overcome this problem, the Euler space emotion distribution model proposed in this paper is introduced into image emotion annotation. The specific steps are as follows:

[0139] 1. Image emotion feature extraction: Use the pre-trained emotion detection model to extract the preliminary emotion features of the image to be annotated, and generate a set of emotion candidates for each image. These features are generated based on the facial expressions, scenes, and contextual semantics contained in the image;

[0140] 2. Emotion mapping to Euler space: Using the emotion Euler space representation method of the present invention, the candidate emotion features are mapped to the two-dimensional Euler space, and their specific coordinates in the space are calculated, where arousal represents the intensity of the emotion, which is the radius r of the Euler space, and valence represents the positive and negative polarity of the emotion, which is the angle θ of the Euler space.

[0141] 3. Annotator auxiliary interface: In the annotation interface, the preliminary emotional distribution of the image is displayed in a visual form. Annotators can see the distribution of the image emotions in the Euler space, as well as the most likely emotion categories recommended by the system (such as "happy", "sad"). In order to reduce subjective bias, annotators can complete the annotation by adjusting the position of the emotion point or selecting the emotion category recommended by the system.

[0142] 4. Calculation of consistency of multi-annotation: After multiple annotators have completed their annotations, calculate their correlation and consistency in the Euler space: use cosine similarity to calculate the similarity of each pair of annotations. If the consistency is lower than the set threshold, the system will prompt the annotator to re-annotate.

[0143] 5. Optimization and fusion of annotation results: Based on the sentiment distribution model of Euler space, the annotation results of multiple annotators are weighted and fused to generate final sentiment labels with higher consistency for training the sentiment computing model.

[0144] The experimental results are as follows: The emotion labeling auxiliary tool of the present invention is applied to a set of public emotion picture datasets and compared with the traditional direct emotion classification and labeling method. The experiment shows that:

[0145] - Annotators’ sentiment label consistency improved by 25%.

[0146] -After the final labeled data was used for sentiment model training, the classification accuracy of the model increased by 15%.

[0147] -Annotation efficiency has been improved by 30%, allowing annotators to complete tasks more quickly.

[0148] The present invention introduces the emotion representation method of Euler space to make image emotion annotation more intuitive and scientific, greatly improving the quality of annotation results and the training performance of the model.

[0149] Although the present invention has been described above with reference to the embodiments, various modifications may be made thereto and parts thereof may be replaced by equivalents without departing from the scope of the present invention. In particular, as long as there is no structural conflict, the various features in the embodiments disclosed in the present invention may be used in combination with each other in any manner, and the fact that these combinations are not exhaustively described in this specification is only for the sake of omitting space and saving resources. Therefore, the present invention is not limited to the specific embodiments disclosed herein, but includes all technical solutions falling within the scope of the claims.

Claims

1. A method for expressing emotion in continuous space based on Euler space, characterized in that: The steps include: S1. Conversion of emotional vocabulary: Convert sentiment-related words into tokens that can be processed by the language model; S2. Vector embedding generation: Use a pre-trained large language model to convert sentiment words into high-dimensional embedding vectors to represent the distribution characteristics of sentiment in the semantic space; S3, vector dimensionality reduction processing: Use dimensionality reduction algorithms to map high-dimensional embedding vectors to a two-dimensional plane space; S4, Euler space transformation and visualization: In the two-dimensional space after dimensionality reduction, the arousal degree is defined as the radius r and the pleasure degree is defined as the angle θ, and the Euler space representation of emotions is constructed based on this, and the distribution of emotion points is displayed through visualization tools.

2. The method for expressing emotion continuous space based on Euler space according to claim 1, characterized in that: In step S1, the steps for converting emotion-related words into tokens that can be processed by the language model are as follows: develop a vocabulary related to emotions; Normalize the words in the sentiment vocabulary to ensure consistency of input; Use a tokenizer to divide the text into individual words or subunits; The sub-units after word segmentation are further converted into IDs acceptable to the model for input into the model for processing; All tokens are filled and packaged into the input format of the model.

3. The method for expressing emotion continuous space based on Euler space according to claim 2, characterized in that: In step S2, the specific steps of training the large language model are as follows: Collect large-scale text data from news, books, online articles, and social media content, remove noise from the text data, and process duplicate data, typos, and illegal content; Use a specific word segmentation algorithm to convert text into word tokens, and map the word tokens to unique integer IDs; Select a pre-trained architecture and use random initialization to set the initial values ​​of the model parameters; Map tokens to unique integer IDs for easy model processing; Design the objective function and use the autoregressive language modeling objective to predict the next word; The word ID is input into the model, and the predicted value is obtained through the embedding layer, the Transformer layer, and the output layer. The error between the predicted result and the true label is calculated by the cross entropy loss function. Use the back-propagation algorithm to calculate the gradient and use the optimizer to update the model parameters; Use the validation set to evaluate the performance of the model and adjust the model hyperparameters as well as the model weights.

4. The method for expressing emotion in continuous space based on Euler space according to claim 1, characterized in that: In step S3, the specific steps of using the dimensionality reduction algorithm to map the high-dimensional embedding vector to the two-dimensional plane space are as follows: Prepare the embedding vectors, making sure you have a high-dimensional embedding vector matrix X, where the dimension of X is n×d, where n is the number of embedding vectors and d is the dimension of each vector; Before dimensionality reduction, the embedding vectors were normalized by calculating the mean and standard deviation to center the data to zero mean and unit standard deviation; Initialize the dimensionality reduction model, set the target dimension to 2, then train the model and transform the data to map the high-dimensional embedding vector to a two-dimensional plane space.

5. The method for expressing emotion in continuous space based on Euler space according to claim 4, characterized in that: The dimensionality reduction model is t-SNE.

6. The method for expressing emotion continuous space based on Euler space according to claim 1, characterized in that: The pre-training architecture is a large language model based on the Transformer architecture.

7. The method for expressing emotion in continuous space based on Euler space according to claim 1, characterized in that: The words in the sentiment vocabulary are normalized as follows: Lowercase: convert all sentiment words into lowercase to avoid the impact of uppercase and lowercase differences on processing; Remove extra characters: Delete special symbols and spaces in words. Spelling standardization: dealing with vocabulary variations.

8. The method for expressing emotion continuous space based on Euler space according to claim 1, characterized in that: Applied to image emotion annotation.

Images