How do humans analyze the sentiments embedded in text? When attempting to analyze a text, humans construct a “scene” in their minds through imagination based on the text, generating a vague image. They then synthesize the text and the mental image to derive the final analysis result. However, current sentiment analysis models lack such imagination; they can only analyze based on existing information in the text, which limits their classification accuracy. To address this issue, we propose the SceEmoNet model. This model endows text classification models with imagination through Stable diffusion, enabling the model to generate corresponding visual scenes from input text, thus introducing a new modality of visual information. We then use the Contrastive Language-Image Pre-training (CLIP) model, a multimodal feature extraction model, to extract aligned features from different modalities, preventing significant feature differences caused by data heterogeneity. Finally, we fuse information from different modalities using late fusion to obtain the final classification result. Experiments on six datasets with different classification tasks show improvements of 9.57%, 3.87%, 3.63%, 3.14%, 0.77%, and 0.28%, respectively. Additionally, we set up experiments to deeply analyze the model’s advantages and limitations, providing a new technical path for follow-up research.

Due to their extensive common-sense knowledge and linguistic understanding, large language models (LLMs) have demonstrated remarkable capabilities in text comprehension and logical reasoning for natural language processing tasks. Traditional prompt-based learning methods, which rely on contextual pattern matching, have proven to be effective in extracting knowledge from LLMs. However, these approaches are constrained by training data pattern matching, overlook reasoning processes, and consequently suffer from suboptimal prompt performance and limited interpretability. Moreover, considering that the intermediate steps generated by single-chain reasoning may not effectively assist LLMs in identifying the sentiment polarity of aspect terms, and that multiple reasoning paths often exist for complex reasoning tasks to reach correct conclusions, this paper proposes a Multi-Chain Thought Prompt Learning framework (MT-CPL). Starting from fundamental concepts, this method simulates human multi-path reasoning patterns to progressively construct comprehensive thought processes and deeply explore sentiment cues. Based on syntactic structures and the semantic logic of text, the framework incorporates four distinct perspectives of text comprehension: hierarchical reading, experiential reading, keyword-based reading, and analogical reading. It establishes a multi-chain prompt template and employs voting mechanisms to select correct reasoning path outcomes. The MT-CPL approach aims to guide LLMs in mining multi-dimensional textual information from different perspectives, gradually uncovering hidden contextual sentiment clues, while mitigating issues caused by irrelevant sentiment cues in intermediate reasoning steps. By decomposing main tasks incrementally, the method achieves progressive reasoning, effectively reduces the difficulty of direct analysis, and further enhances model interpretability through the integration of inherent common-sense knowledge.

Fine-grained toxic language detection aims to assess the harmfulness, expression style, and attack target of textual content, supporting refined content moderation on online platforms. Most existing methods model each subtask independently, failing to exploit semantic correlations among tasks, while conventional contrastive learning is prone to introducing false negative samples under multi-task, multi-label settings. To address these limitations, we propose a Prompt-Enhanced Multi-Task Contrastive Learning framework (PEMTCL). The framework employs a dynamic prompt generation mechanism to construct instance-level soft prompts, enabling the encoder to adaptively capture fine-grained semantic cues required by different subtasks. It then jointly optimizes four subtasks—toxicity identification, toxic type discrimination, expression type detection, and targeted group detection—achieving inter-task knowledge complementarity within a shared representation space. Furthermore, a label-aware multi-head contrastive learning module leverages multi-task label relationships for reliable negative sample selection and incorporates a sample reweighting mechanism, effectively mitigating false negatives and class imbalance. Experiments on two Chinese datasets, ToxiCN and FG-COLD, demonstrate that PEMTCL achieves the best performance on most subtasks, with notable improvements on fine-grained tasks. Ablation studies, training data scale analysis, and case studies further validate the effectiveness and robustness of the proposed model.