Aspect Extraction and Sentiment Classification (AESC) is designed to conduct Aspect Term Extraction (ATE) and Aspect Sentiment Classification (ASC) from the sentence. Recent works have employed interaction networks and local context focus methods to mitigate the sentiment noise caused by multiple sentiment polarities within a sentence. However, the above methods lack fine-grained modeling between aspect and opinion terms, and they overlook the semantic correlation and mutual enhancement between them. To address the above issues, we propose a bidirectional enhancement contrastive learning framework (BiAESC). Specifically, we design the Aspect-to-Opinion Enhancement (A2O) module and the Opinion-to-Aspect Enhancement (O2A) module to leverage semantic correlations for bidirectional enhancement of aspect and opinion term representations. In addition, we construct the Dependency Contrastive Learning (DCL) module, which leverages dependency graph structures and contrastive learning to achieve fine-grained modeling between aspect and opinion terms, thereby mitigating the impact of sentiment noise. Experimental results on three benchmark datasets demonstrate that BiAESC significantly outperforms baseline models and exhibits improved robustness.

Sarcasm detection remains a major challenge in natural language processing due to its reliance on identifying semantic incongruity between literal expressions and contextual implications. Existing approaches fail to adequately model semantic conflicts between pivotal lexical units and underutilize contextual information. To address these gaps, we propose the Gated Contrast-Attention Contextualized-Understanding (GCACU) network, a dual-path architecture that integrates contrastive semantic learning with deep contextual comprehension via gated mechanisms. The contrast-attention pathway quantifies conflicts between key lexical pairs, while the contextualized-understanding pathway leverages ELMo-enhanced bidirectional GRU networks for contextual understanding. A hierarchical gating system dynamically fuses these representations to capture both lexical-level incongruity and discourse-level context. Experimental results show that GCACU achieves an F1-score of 77.0 % on Twitter SemEval-2018, outperforming the best baseline by 0.7 %; on the IAC-V1 and IAC-V2 datasets, it reaches 67.64 % and 82.66 % in F1-score respectively, with improvements of 1.24 % and 0.56 % over the baselines. Ablation studies further validate the gating system as a critical contributor to performance gains, underscoring the effectiveness of explicit incongruity modeling for sarcasm analysis.

As multimodal systems increasingly process sensitive personal data, the ability to selectively revoke specific data modalities has become a critical requirement for privacy compliance and user autonomy. We present Missing-by-Design (MBD), a unified framework for revocable multimodal sentiment analysis that combines structured representation learning with a certifiable parameter-modification pipeline. Revocability is critical in privacy-sensitive applications where users or regulators may request removal of modality-specific information. MBD learns property-aware embeddings and employs generator-based reconstruction to recover missing channels while preserving task-relevant signals. For deletion requests, the framework applies saliency-driven candidate selection and a calibrated Gaussian update to produce a machine-verifiable Modality Deletion Certificate. Experiments on benchmark datasets show that MBD achieves strong predictive performance under incomplete inputs and delivers a practical privacy-utility trade-off, positioning surgical unlearning as an efficient alternative to full retraining.

Recent advancements in emotion recognition research based on physiological data have been notable. However, existing multimodal methods often overlook the interrelations between various modalities, such as video and Electroencephalography data, in emotion recognition. In this paper, a feature fusion-based hierarchical cross-modal spatial fusion network is proposed that effectively integrates EEG and video features. By designing an Electroencephalography feature extraction network based on 1D convolution and a video feature extraction network based on 3D convolution, corresponding modality features are thoroughly extracted. To promote sufficient interaction between the two modalities, a hierarchical cross-modal coordinated attention module is proposed in this paper. Additionally, to enhance the network's perceptual ability for emotion-related features, a multiscale spatial pyramid pooling module is also designed. Meanwhile, a self-distillation method is introduced, which enhances the performance while reducing the number of parameters in the network. The hierarchical cross-modal spatial fusion network achieved an accuracy of 97.78% on the valence-arousal dimension of the DEAP dataset, and it also obtained an accuracy of 60.59% on the MAHNOB-HCI dataset, reaching the state-of-the-art level.

Interactive fusion methods have been successfully applied to multimodal sentiment analysis, due to their ability to achieve data complementarity via interaction of different modalities. However, previous methods treat the information of each modality as a whole and usually treat them equally, failing to distinguish the contribution of different semantic regions in non-textual features towards textual features. It caused that the public regions fail to be captured and private regions are hard to be predicted only with textual. Meanwhile, these methods use sentiment-independent encoder to encode textual features, which may mistakenly identify syntactically irrelevant contextual words as clues for predicting sentiment. In this paper, we propose a coordinated-joint translation fusion framework with sentiment-interactive graph to solve these problems. Specifically, we generate a novel sentiment-interactive graph to incorporate sentiment associations between different words into the syntactic adjacency matrix. The relationships between nodes are no longer limited to the sole existence of syntactic associations but fully consider the interaction of sentiment between different words. Then, we design a coordinated-joint translation fusion module. This module utilizes a cross-modal masked attention mechanism to determine whether there is a correlation between the text and non-text inputs, thereby identifying the most relevant public semantic features in the visual and acoustic modalities corresponding to the text modality. Subsequently, a cross-modal translation-aware mechanism is used to calculate the differences between the visual and acoustic modalities features transformed into the text modality and the text modality itself, which allows us to reconstruct the visual and acoustic modalities towards text modality to obtain private semantic features. In addition, we construct a multimodal fusion layer to fuse textual features and non-textual public and private features to improve multimodal interaction effects. Experimental results on publicly available datasets CMU-MOSI and CMU-MOSEI illustrate that our proposed model achieve a best accuracy of 86.5% and 86.1%, and best F1 of 86.4% and 86.1%. A series of further analyses also indicate the proposed framework effectively improve the sentiment identification capability.

Multimodal sarcasm detection aims to identify whether the literal expression is contrary to the authentic attitude within multimodal data. Sarcasm incongruity method has been successfully applied to multimodal sarcasm detection, due to its ability to flexibly capture the intrinsic differences between modalities. However, previous incongruity methods primarily focused on the semantic level, often overlooking more specific forms of sarcasm incongruity. Sarcasm incongruity, in particular, encompasses fact incongruity, sentiment incongruity, and combination incongruity. Therefore, we propose a fact-sentiment incongruity combination network from a novel perspective, which draws the multimodal sarcastic relations by exploring the multimodal factual disparities, sentiment incongruity, and combination fusion. Specifically, we design a dynamic connecting component calculating dynamic routing probability weights via graph attention and mask routing matrices, which selects the most suitable image-text pairs to capture fact incongruity between images and text. Then, we retrieve sentiment relations between text tokens and image objects using external sentiment knowledge to reconstruct edge weights in the cross-modal graph matrix to capture sentiment incongruity. Furthermore, we introduce a combination incongruity fusion layer and cross-modal contrastive loss to fuse fact incongruity and sentiment incongruity for further enhancing the incongruity representations. Extensive experiments and further analyses on publicly available datasets demonstrate the superiority of our proposed model.

Graph convolutional networks have been successfully applied to aspect-based sentiment analysis, due to their ability to flexibly capture syntactic information and word dependencies. However, most existing graph network-based models only consider the syntactic dependencies between specific aspects and contexts. These cannot capture the internal semantic correlations within aspect-specific phrases and ignore the sentiment interaction relations between different aspects of a sentence. In this paper, we propose a novel graph convolutional network with sentiment interaction and multi-graph perception for aspect-based sentiment analysis. The proposed model considers the complementarity of semantic dependencies and sentiment interactions simultaneously. Specifically, we generate four types of adjacency graphs by integrating the internal semantic correlations between aspect phrases and linking the sentiment interaction relations among different aspects. Adjacency graphs are used to construct graph convolutional neural networks to enrich aspect-centric dependencies and enhance the capability of context-awareness. In addition, we construct a multi-graph perception mechanism to capture the specific dependency information that cannot be captured between different graphs and hence reduce the amount of overlapping information. Experimental results on five publicly-available datasets demonstrate that our proposed model outperforms state-of-the-art methods and achieves the best performance in terms of accuracy and macro-F1 score.

In recent years, Deep Neural Networks (DNNs) have achieved excellent performance on many tasks, but it is very difficult to train good models from imbalanced datasets. Creating balanced batches either by majority data down-sampling or by minority data up-sampling can solve the problem in certain cases. However, it may lead to learning process instability and overfitting. In this paper, we propose the Batch Balance Wrapper (BBW), a novel framework which can adapt a general DNN to be well trained from extremely imbalanced datasets with few minority samples. In BBW, two extra network layers are added to the start of a DNN. The layers prevent overfitting of minority samples and improve the expressiveness of the sample distribution of minority samples. Furthermore, Batch Balance (BB), a class-based sampling algorithm, is proposed to make sure the samples in each batch are always balanced during the learning process. We test BBW on three well-known extremely imbalanced datasets with few minority samples. The maximum imbalance ratio reaches 1167:1 with only 16 positive samples. Compared with existing approaches, BBW achieves better classification performance. In addition, BBW-wrapped DNNs are 16.39 times faster, relative to unwrapped DNNs. Moreover, BBW does not require data preprocessing or additional hyper-parameter tuning, operations that may require additional processing time. The experiments prove that BBW can be applied to common applications of extremely imbalanced data with few minority samples, such as the classification of EEG signals, medical images and so on.

In comparison with the single mode, the utilization of multi-mode information of text, video and audio could lead to more accurate sentiment analysis. GME-Dialogue-NET, a gated multi-modal sentiment analysis model, is raised for the multi-modal emotion prediction and sentiment analysis. The model judges whether the audio or video modal is the noise through GME (Gated Multi-modal Embedding, GME) and then accepts or refuses the modal information based on the judgement. The model uses the Attention Mechanism of context vector to allocate more attention to the context with greater relevance to the current sentence. GME-Dialogue-NET divides participants of the dialogue into speaker and listener to better capture the dependence between emotion and state. It raises that the fusion mechanism CPA (Circulant-Pairwise Attention, CPA) could pay effective attention with different degrees on different modals to attain more helpful emotional and sentimental representation and thus make emotion prediction and sentiment analysis. Compared with the current model, both the weighted accuracy and the F1 score of emotion prediction were improved, especially for the three emotions of sadness, anger and excitement. In the sentiment regression task, the comparison between GME-Dialogue-NET with current advanced model Multilogue-Net shows that MAE (Mean absolute error, MAE) of GME-Dialogue-NET reduces by 0.1 percentage and the Pearson Correlation Coefficient (R) of GME-Dialogue-NET rises by 0.11 percentage.