Visual emotion adaptive explicit and implicit space self-supervised learning method

Abstract

The invention discloses a visual emotion adaptive explicit and implicit space self-supervised learning method, which comprises the following steps of: 1, separating redundant semantic features irrelevant to emotion from text embedding of CLIP, only retaining embedding relevant to emotion, and generating a pseudo tag through a tag propagation algorithm; 2, an image encoder of CLIP is used as an emotion encoder, the features are directly mapped to an emotion category space, and the definition and interpretability of the emotion features are enhanced; 3, a pseudo label generated by the implicit emotion space guides learning of the explicit emotion space, feedback of the explicit space optimizes the implicit space, and a cyclic feedback mechanism is formed; and 4, recording category prototypes of the source domain and the target domain in the emotion subspace by using a memory bank, and aligning different domains to reduce emotion differences by minimizing the distance between corresponding category features. According to the method, the problem of mismatching between the CLIP space and the emotion space and the problem of emotion difference between different emotion data sets are solved.

Description

Technical Field

[0001] The present invention belongs to the field of computer vision and emotion computing, and relates to a visual emotion recognition method, and specifically to a visual emotion recognition method based on unsupervised domain adaptation. Background Art

[0002] Research on visual emotion recognition (VER) has been ongoing for over two decades, evolving from early traditional methods to current deep learning approaches. Initial studies uniformly processed all regions of an image to extract global features. For example, Zhao et al. (2015) predicted the continuous probability distribution of emotion in the valence-arousal space, while Yuan et al. (2013) analyzed image emotion from a mid-level perspective.

[0003] As theory develops, research has found that emotional content often resides in specific key regions. More research has focused on extracting local features and exploring the relationships between them. For example, Yang et al. (2018) performed visual emotion prediction based on automatically discovered emotional regions, and Yao et al. (2019) proposed an attention-aware polarity-sensitive embedding method for emotional image retrieval.

[0004] Recent research has begun to incorporate CLIP into the VER task, such as Deng et al. (2022)'s approach to image sentiment classification based on language supervision, and Deng et al. (2024)'s approach to learning to construct diverse cues for image sentiment classification. However, these methods still require large amounts of accurately annotated training data. Therefore, domain adaptation methods are crucial to reducing this reliance on extensive annotations.

[0005] Considering the domain bias in the VER dataset, Zhao et al. (2019) proposed CycleEmotionGAN, the first unsupervised method for translating emotional images between source and target domains. Based on this, Zhao et al. (2022) improved the image translation effect through an improved CycleGAN, introducing multi-scale similarity and emotion consistency losses.

[0006] However, due to the ambiguity of sentiment features, these methods are difficult to capture the complete semantics of sentiment. Existing technologies have not yet solved the problem of separating the sentiment space from the traditional semantic space.

[0007] CLIP has shown impressive results in learning traditional visual representations and promoting zero-shot migration to diverse downstream classification tasks. Existing progress in CLIP for UDA can be divided into two categories: the first category reduces domain offset through prompt learning. Ge et al. first used CLIP for UDA tasks and introduced a method for designing specific domain contexts to promote the learning of representations in different domains. Singha et al. (2023) addressed the UDA challenge by leveraging the semantic richness of CLIP and learning domain-invariant and class-universal prompt tags. The second category directly utilizes CLIP's zero-shot capabilities instead of fine-tuning and prompts. For example, Hu et al. (2024) proposed the ReCLIP method to refine contrastive language image pre-training through passive domain adaptation. Although CLIP performs well in traditional UDA tasks, it still faces the following challenges in abstract tasks such as emotion recognition:

[0008] (1) Mismatch between CLIP space and sentiment space: CLIP encoders are trained to process millions of traditional semantic concepts, only a small fraction of which are crucial for VER. Both visual and textual embeddings retain a large amount of information irrelevant to sentiment, resulting in limited understanding of emotional images.

[0009] (2) Inter-domain sentiment differences: Due to differences in annotators and data collection scope, there are large sentiment differences between sentiment datasets, which seriously affects feature alignment and reduces prediction accuracy.

[0010] (3) Ambiguity of emotional features: Emotional images lack clear features, and the same semantic object (such as a dog) may express different emotions under different conditions, which increases the difficulty of cross-domain alignment.

[0011] Existing technologies have not yet effectively solved the problem of redundant information in CLIP embedding, and are unable to make features better focus on VER tasks. Summary of the Invention

[0012] The present invention provides an explicit and implicit space self-supervised learning method for visual emotion adaptation. The method utilizes the CLIP (Contrastive Language-Image Pretraining) model to construct explicit and implicit emotion spaces, thereby solving the mismatch problem between the CLIP space and the emotion space and the emotion difference problem between different emotion datasets, thereby achieving efficient cross-domain emotion recognition.

[0013] The purpose of the present invention is achieved through the following technical solutions:

[0014] A method for explicit and implicit spatial self-supervised learning of visual emotion adaptation, including the following steps:

[0015] Step 1: Construct an implicit sentiment space: Separate redundant semantic features irrelevant to sentiment from the text embeddings of CLIP, retain only the sentiment-related embeddings, and generate pseudo labels through the label propagation algorithm. The specific steps are as follows:

[0016] Step 1: Given an input image x and the corresponding prompt text p, pass the text encoder E of CLIP t and image encoder E I Get text embedding t and image embedding v:

[0017] t=E t (p) v=E I (x)

[0018] Step 1 and 2: Apply singular value decomposition to the text embedding t and map the image embedding to the implicit sentiment space through the mapping matrix P:

[0019]

[0020] Where I represents the embedding set of all target domain images, P is the mapping matrix obtained from SVD decomposition, and V proj is the image feature after mapping, d represents the feature dimension, N T is the number of samples in the target domain;

[0021] Step 1 and 3: In the implicit emotion space, use the label propagation algorithm to generate pseudo labels. The specific steps are as follows:

[0022] (1) The converted text features are used as the propagation center point C and embedded with all the converted target images V proj Aggregate together to form feature set X;

[0023] (2) By calculating the similarity matrix between features and using an iterative optimization algorithm, the label information is propagated from the labeled center point to the unlabeled target sample, and finally a reliable pseudo label is generated for each target sample.

[0024] Step 2: Construct an explicit emotion space: Use the CLIP image encoder as the emotion encoder to directly map the features to the emotion category space to enhance the clarity and interpretability of the emotion features. The specific steps are as follows:

[0025] Use CLIP's image encoder as an emotion encoder to generate emotion feature representation for the input image:

[0026] e=E m (x)

[0027] Among them E mrepresents the emotion encoder, x is the input image, and e represents the feature representation in the generated emotion subspace;

[0028] Step 3: Spatial Interaction: The pseudo-labels generated by the implicit emotion space guide the learning of the explicit emotion space. The feedback of the explicit space optimizes the implicit space, forming a circular feedback mechanism. The specific steps are as follows:

[0029] Use the collaborative attention mechanism to fuse the prediction results of the two spaces to obtain more accurate classification predictions:

[0030]

[0031] Where ⊙ represents element-by-element multiplication, o i and o e are the predicted probability values of the implicit space and the explicit space respectively. The softmax function converts the probability value into a probability distribution. Represents the final predicted label after fusion;

[0032] Step 4: Domain alignment: Use the memory library to record the category prototypes of the source domain and the target domain in the sentiment subspace. By minimizing the distance between the corresponding category features, different domains are aligned to reduce the sentiment difference. The specific steps are as follows:

[0033] Step 4.1: Propose a memory mechanism to construct feature prototypes for each domain category:

[0034]

[0035] in and Represent the feature prototypes of the kth class in the source domain and the target domain respectively, and denotes the number of samples of the kth class in the source domain and the target domain respectively, and Represents all feature representations of the kth class in the source domain and the target domain;

[0036] Step 4.2: Align different domains by minimizing the distance between corresponding category prototypes:

[0037]

[0038] Where D represents the cosine similarity metric function, E k Represents the expected value of all categories k, loss function It encourages the feature distributions of the same sentiment category in the source and target domains to be consistent.

[0039] In this paper, "explicit space" refers to the space that directly maps features to emotion categories, so that each point clearly corresponds to an emotion category; "implicit space" refers to the emotion-related feature space obtained by decomposing CLIP text embeddings and removing redundant semantic information; "self-supervised learning" refers to the method of automatically generating pseudo-labels for training using unlabeled target domain data; and "visual emotion adaptation" refers to the process of adapting the source domain emotion recognition model to the target domain to address the differences between different visual emotion datasets.

[0040] Compared with the prior art, the present invention has the following advantages:

[0041] 1. Solve the mismatch problem between CLIP space and sentiment space: Through the explicit-implicit interactive sentiment space, the redundant semantic information in CLIP space is eliminated, making the features more focused on sentiment extraction.

[0042] 2. Effectively eliminates inter-domain sentiment differences: The memory-based domain alignment method aligns category features in the sentiment subspace, narrowing the inter-domain differences.

[0043] 3. Superior performance: On the FI dataset, the proposed method is 38.99% better than the previous state-of-the-art emotion UDA method and 4.64% better than the CLIP-based UDA method.

[0044] 4. Strong versatility: In experiments conducted on different backbone networks (such as ResNet101), the EIDA method of the present invention consistently outperforms all competing methods, demonstrating the versatility of the method on various models.

[0045] 5. Good parameter stability: Sensitivity analysis of loss weight and learning rate shows that the performance of EIDA is relatively stable under different hyperparameters and learning rates. BRIEF DESCRIPTION OF THE DRAWINGS

[0046] Figure 1 Schematic diagram of the EIDA system architecture of the present invention. DETAILED DESCRIPTION

[0047] The technical solution of the present invention is further described below with reference to the accompanying drawings, but is not limited thereto. Any modification or equivalent replacement of the technical solution of the present invention that does not depart from the spirit and scope of the technical solution of the present invention should be included in the scope of protection of the present invention.

[0048] This paper proposes an EIDA (Explicit and Implicit Domain Adaptation) method. Based on the CLIP model, it designs an explicit and implicit interactive emotion space and uses a memory library to record and align category prototypes from different domains. The overall framework of EIDA is as follows:

[0049] Implicit Sentiment Space: We separate redundant semantic features irrelevant to sentiment from CLIP’s text embeddings, retain only sentiment-related embeddings, and generate pseudo labels through a label propagation algorithm.

[0050] Explicit emotion space: CLIP's image encoder is used as the emotion encoder to directly map features to the emotion category space, enhancing the clarity and interpretability of emotion features.

[0051] Spatial interaction: The pseudo-labels generated by the implicit emotion space guide the learning of the explicit emotion space, while the feedback of the explicit space optimizes the implicit space, forming a circular feedback mechanism.

[0052] Domain alignment: A memory bank is used to record the category prototypes of the source and target domains in the sentiment subspace, and different domains are aligned to reduce sentiment differences by minimizing the distance between corresponding category features.

[0053] The EIDA architecture diagram is as follows Figure 1 As shown in Figure 1, the left side of the figure shows the input, which includes two types of textual cues (emotional cues for artistic photos and emotional cues for website photos) and two types of image data (labeled source data and unlabeled target data). The upper middle section shows the implicit sentiment space structure. The text encoder processes the sentiment cues to generate text features t, which are decomposed using SVD to obtain the matrix U and the mapping matrix P. The image encoder processes the input image to generate image features v, which are then propagated through interaction with the text features and the mapping matrix to generate pseudo-labels. The lower middle section shows the explicit sentiment space structure. The sentiment encoder maps the source and target domain images to a shared subspace feature e. The memory bank records and updates the category prototypes of the source and target domain features, achieving domain alignment by reducing the distance between corresponding category prototypes. The right side shows the output logits of the two sentiment spaces being fused through a co-attention mechanism to produce the final prediction. Solid arrows indicate paths used in both training and inference, while dashed arrows indicate paths used only in training. The entire framework captures sentiment features from two complementary perspectives, implicit and explicit, effectively resolving the mismatch between the CLIP space and the sentiment space. The following is a detailed description of the EIDA architecture:

[0054] 1. Implicit Emotional Space

[0055] In the unsupervised domain adaptation task of visual emotion recognition, constructing informative and clear features to generate reliable pseudo-labels is a key challenge. This paper develops an implicit emotion space method to improve the quality of pseudo-label generation by reducing redundant semantic features in the embedding.

[0056] While the CLIP model excels at aligning image and text modalities, it has limitations in capturing abstract emotional relationships. When CLIP processes an input image and the corresponding emotionally suggestive text, it generates image and text embeddings. However, these embeddings contain a large amount of semantic information that is irrelevant to emotion. This redundant information interferes with the model's understanding of emotion.

[0057] Specifically, given an input image x and the corresponding prompt text p, the present invention can be implemented by using the text encoder E of CLIP. t and image encoder E I Get text embedding t and image embedding v:

[0058] t=E t (p) v=E I (x) (1)

[0059] To solve the problem of redundant semantic information in text embedding, the present invention applies singular value decomposition (SVD) to the text embedding t, decomposing it into a matrix U containing linearly independent vectors:

[0060] U ，_，_ =svd(t) (2)

[0061] where U∈E d×c Contains a set of linearly independent vectors, d represents the feature dimension, and c represents the number of categories.

[0062] Due to the nonlinear relationship between the redundant vector and the emotional space, we first delete the most redundant features from the matrix U and retain the remaining emotional features to obtain a submatrix. This submatrix is then multiplied by its own transpose to form a mapping matrix P. The image embedding is mapped to the implicit emotional space through the mapping matrix P:

[0063]

[0064] Where I represents the embedding set of all target domain images, P is the mapping matrix obtained from SVD decomposition, and V proj is the image feature after mapping, which retains the information related to emotion and filters the redundant information. T is the number of target domain samples.

[0065] In the implicit sentiment space, the present invention uses the label propagation algorithm to generate pseudo labels. First, the present invention uses the converted text features as the propagation center point C and embeds all the converted target images into V proj Aggregate together to form a feature set X. By calculating the similarity matrix between features and using an iterative optimization algorithm, the label information is propagated from the labeled center point to the unlabeled target sample, and finally a reliable pseudo label is generated for each target sample.

[0066] The advantage of this approach is that it not only effectively filters out information irrelevant to emotion recognition, but also leverages the underlying sentiment semantics in the CLIP model to generate high-quality pseudo-labels, providing a reliable guidance signal for unsupervised domain adaptation. The label propagation process fully exploits the inherent similarity structure of the data, enabling more accurate category predictions even in unsupervised situations.

[0067] Explicit Emotional Space

[0068] To solve the problems of differences and unclear features between emotion domains, this paper proposes an explicit emotion space and memory library method to extract representative category features.

[0069] A major challenge with sentiment data is the large inter-domain emotional disparity. Different datasets may express the same emotion category in very different ways. To minimize this disparity, this paper projects the source and target domain images into a shared sentiment subspace for alignment.

[0070] Specifically, this paper uses CLIP's image encoder as an emotion encoder to generate emotion feature representation for the input image:

[0071] e=E m (x) (4)

[0072] Among them E m represents the emotion encoder, x is the input image, and e represents the feature representation in the generated emotion subspace.

[0073] Considering that the ambiguity of sentiment features will introduce noise into sentiment embedding, this paper proposes a memory bank mechanism to construct feature prototypes for each domain category:

[0074]

[0075] in and Represent the feature prototypes of the kth class in the source domain and the target domain respectively, and denotes the number of samples of the kth class in the source domain and the target domain respectively, and represents all feature representations of the kth class in the source and target domains.

[0076] The memory library continuously updates the feature prototypes of each emotion category during training to ensure that they accurately represent the center of each class. To bridge the emotional gap and enhance the consistency of feature extraction, the present invention aligns different domains by minimizing the distance between the corresponding category prototypes between domains:

[0077]

[0078] Where D represents the cosine similarity metric function, E k Represents the expected value of all categories k, the loss function It encourages the feature distributions of the same sentiment category in the source and target domains to be consistent.

[0079] Furthermore, the present invention constructs an explicit label space, directly embedding and projecting the subspace into specific emotion categories. Through a multi-layer perceptron network, the present invention achieves a direct mapping from feature space to emotion categories, enhancing the model's ability to discriminate emotion categories.

[0080] 3. Spatial Interaction and Domain Alignment

[0081] The implicit and explicit spaces capture different aspects of sentiment information. The implicit space focuses on the deep features and contextual information of sentiment, while the explicit space focuses on learning clear classification boundaries between sentiment categories. To fully leverage the complementary advantages of the two spaces, this paper designs a spatial interaction mechanism.

[0082] Specifically, the present invention uses a collaborative attention mechanism to fuse the prediction results of the two spaces to obtain more accurate classification predictions:

[0083]

[0084] Where ⊙ represents element-by-element multiplication, o i and o e are the predicted logits in implicit space and explicit space respectively. The softmax function converts logits into probability distribution. Represents the final predicted label after fusion.

[0085] During training, the present invention simultaneously optimizes two spaces using a multi-task learning framework. For labeled source domain data, the present invention uses the true label y for supervision; for unlabeled target domain data, the present invention uses pseudo-labels generated in the implicit space for self-supervision. This multi-task learning framework enables the model to simultaneously focus on classification accuracy in the source domain, self-supervised consistency in the target domain, and feature alignment between domains, thereby achieving effective cross-domain emotion recognition.

[0086] Example:

[0087] 1. System environment and data set

[0088] In this example, the present invention applies the EIDA method to the domain adaptation task between two classic visual emotion datasets: migrating ArtPhoto to FI (Flickr and Instagram). The system operating environment is a server equipped with an NVIDIA A100 GPU and implemented using the PyTorch deep learning framework.

[0089] 2. Dataset Introduction

[0090] ArtPhoto dataset: Contains 806 art photographs categorized by artists using Mikels' emotion categories. Each photo is assigned to one of eight emotion categories: amusement, anger, awe, contentment, disgust, excitement, fear, and sadness. The images were captured and uploaded to the website by the artists, who then assigned the emotion labels to each photo.

[0091] FI dataset: Contains 23,308 images obtained from Flickr and Instagram, annotated by 225 Amazon Mechanical Turk (AMT) workers using the same eight emotion categories as ArtPhoto.

[0092] There are obvious differences in the style and content of the two datasets: ArtPhoto contains artistic photos created by professional artists, while FI contains everyday photos from social media. This difference makes them ideal candidate datasets for testing cross-domain emotion recognition.

[0093] 3. Model implementation details

[0094] The EIDA method of the present invention is implemented based on the CLIP model, and the specific parameters are set as follows:

[0095] Model architecture: Use CLIP's ViT-B / 32 as the backbone network, fix the CLIP encoder parameters, and only fine-tune the layer normalization weights.

[0096] Input processing: The input image is resized and center-cropped to 224×224 pixels, the batch size is set to 64, and training is performed for 30 epochs.

[0097] Optimizer configuration:

[0098] Weight decay: 1e-4;

[0099] Learning rate for image, text, and sentiment encoders: 1e-3;

[0100] Learning rate for MLP classifier: 1e-3.

[0101] Key hyperparameters:

[0102] The number of features discarded in the implicit space is p = 1;

[0103] Label propagation smoothing parameter α = 0.99;

[0104] Loss weights λ = 1.1, γ = 0.8.

[0105] 4. Specific implementation steps

[0106] (1) Load the pre-trained CLIP model (ViT-B / 32), freeze the main parameters, and only keep the layer normalization weights trainable.

[0107] (2) Construct text prompts for eight emotion categories and obtain initial text embeddings through CLIP’s text encoder.

[0108] (4) Initialize the MLP projection layer of the explicit emotion space: a two-layer fully connected network with a ReLU activation function in the middle.

[0109] (4) Initialize the memory bank to store the prototype features of each emotion category in the source domain and the target domain.

[0110] 5. Experimental results and analysis

[0111] This paper compares the EIDA method with existing methods on the domain adaptation task of ArtPhoto→FI. Experimental results show that EIDA achieves an average accuracy of 71.00% on this task, a 4.64% improvement over the state-of-the-art CLIP-based UDA method, UniMoS, and a 38.99% improvement over the previous state-of-the-art emotional UDA method, CEGAN++. EIDA particularly demonstrates a significant advantage in emotion categories such as "Awe," "Excitement," and "Sadness."

[0112] These results demonstrate that EIDA successfully addresses the mismatch between the CLIP space and the sentiment space, effectively narrowing the inter-domain sentiment gap. In particular, our approach achieves significant improvements on the "Disgust" and "Excitement" categories, which are typically difficult to identify in cross-domain scenarios and are nearly impossible to identify with traditional methods.

[0113] 6. Alternative Implementation

[0114] In addition to the above main implementation methods, the EIDA method can also have the following alternative implementations:

[0115] Different Backbone Networks: In addition to using CLIP's ViT-B / 32 as the backbone network, we also tested ResNet101 as the encoder backbone for CLIP. The results show that the EIDA method maintains superior performance across different backbone networks. With the ResNet101 backbone, we achieved an average accuracy of 63.89% on the ArtPhoto→FI task, which still outperforms existing methods.

[0116] Progressive pseudo-label update strategy: To further improve pseudo-label quality, a progressive update strategy can be employed. Initially, only high-confidence pseudo-labels are used. As training progresses, the confidence threshold is gradually lowered to include more samples. This strategy can reduce the negative impact of erroneous pseudo-labels and improve model adaptability.

[0117] Multi-level emotion feature fusion: In addition to using the final output layer features of CLIP, we can also fuse the intermediate layer features to construct a multi-level emotion representation. Low-level features may contain more visual details, while high-level features contain more abstract semantic information. Fusion of these features may help capture more comprehensive emotion information.

Claims

1. An explicit and implicit spatial self-supervised learning method for visual emotion adaptation, characterized by The method comprises the following steps: Step 1: Construct an implicit sentiment space: Separate redundant semantic features irrelevant to sentiment from the text embeddings of CLIP, retain only the sentiment-related embeddings, and generate pseudo labels through the label propagation algorithm; Step 2: Construct an explicit emotion space: Use CLIP’s image encoder as an emotion encoder to directly map features to the emotion category space, enhancing the clarity and interpretability of emotion features. Step 3: Spatial Interaction: The pseudo-labels generated by the implicit emotion space guide the learning of the explicit emotion space, and the feedback of the explicit space optimizes the implicit space, forming a circular feedback mechanism. Step 4: Domain alignment: Use the memory library to record the category prototypes of the source domain and the target domain in the sentiment subspace, and align different domains to reduce sentiment differences by minimizing the distance between corresponding category features.

2. The method for explicit and implicit spatial self-supervised learning of visual emotion adaptation according to claim 1, characterized in that The specific steps of step one are as follows: Step 1: Given an input image x and the corresponding prompt text p, pass the text encoder E of CLIP t and image encoder E I Get text embedding t and image embedding v: t=E t (p)v=E I (x) Step 1 and 2: Apply singular value decomposition to the text embedding t and map the image embedding to the implicit sentiment space through the mapping matrix P: Where I represents the embedding set of all target domain images, P is the mapping matrix obtained from SVD decomposition, and V proj is the image feature after mapping, d represents the feature dimension, N T is the number of samples in the target domain; Step 1: In the implicit sentiment space, a label propagation algorithm is used to generate pseudo labels.

3. The explicit and implicit spatial self-supervised learning method for visual emotion adaptation according to claim 2 is characterized in that The specific steps of steps 1 and 3 are as follows: (1) The converted text features are used as the propagation center point C and embedded with all the converted target images V proj Aggregate together to form feature set X; (2) By calculating the similarity matrix between features and using an iterative optimization algorithm, the label information is propagated from the labeled center point to the unlabeled target sample, and finally a reliable pseudo label is generated for each target sample.

4. The method for explicit and implicit spatial self-supervised learning of visual emotion adaptation according to claim 1, characterized in that The specific steps of step 2 are as follows: Use CLIP's image encoder as an emotion encoder to generate emotion feature representation for the input image: and=And m (x) Among them E m represents the emotion encoder, x is the input image, and e represents the feature representation in the generated emotion subspace.

5. The method for explicit and implicit spatial self-supervised learning of visual emotion adaptation according to claim 1, characterized in that The specific steps of step three are as follows: Use the collaborative attention mechanism to fuse the prediction results of the two spaces to obtain more accurate classification predictions: Where ⊙ represents element-by-element multiplication, o i and o e are the predicted probability values of the implicit space and the explicit space respectively. The softmax function converts the probability value into a probability distribution. Represents the final predicted label after fusion.

6. The method for explicit and implicit spatial self-supervised learning of visual emotion adaptation according to claim 1, characterized in that The specific steps of step 4 are as follows: Step 4.1: Propose a memory mechanism to construct feature prototypes for each domain category: in and Represent the feature prototypes of the kth class in the source domain and the target domain respectively, and denotes the number of samples of the kth class in the source domain and the target domain respectively, and Represents all feature representations of the kth class in the source domain and the target domain; Step 4.2: Align different domains by minimizing the distance between corresponding category prototypes: Where D represents the cosine similarity metric function, E k Represents the expected value of all categories k, loss function It encourages the feature distributions of the same sentiment category in the source and target domains to be consistent.

Images