SED-GPT: A Non-Invasive Method for Long-Sequence Fine-Grained Semantics and Emotions Decoding

Traditional emotion decoding methods typically rely on short sequences with limited context and coarse-grained emotion categories. To address these limitations, we proposed the Semantic and Emotion Decoding Generative Pre-trained Transformer (SED-GPT), a non-invasive method for long-sequence fine-grained semantics and emotions decoding on extended narrative stimuli. In the encoding stage, we employed a Semantic to Brain Re-sponse Conversion Module (SBRCM) to align neural signals with semantic representa-tions, which modeled the relationship between stimulus features and brain responses. To enhance generalization, we constructed a word rate model and estimated noise covari-ance. In the decoding stage, new semantic sequences were generated according to previ-ous candidate semantic sequences and the prior probabilities of large language models (LLMs). These sequences were converted into brain response sequences through SBRCM, which were then compared with actual brain responses. By combining prior probabilities with likelihood probabilities, the optimal sequence was iteratively derived, generating long-sequence semantic vectors. Finally, the GoEmotions framework was applied to quantify the multiclass emotional distribution of the long-sequence semantic representa-tions. SED-GPT achieves a BERTScore-F1 of 0.650 on semantic decoding and attains a co-sine similarity (CS) of 0.504 and a Jensen–Shannon similarity (JSS) of 0.469 for emotion decoding (p< 0.05). Functional connectivity analyses reveal persistent coupling between the language network and the emotion network, which provides neural evidence for the language-emotion interaction mechanism in Chinese.