基于多尺度特征与混合注意力的社交媒体文本情感分类研究OA

Text in social media often exhibits fragmented and colloquial characteristics with sparse effective emo-tional information,which poses challenges for emotion classification tasks in capturing effective emotional features.To address the deficiencies of existing methods in the extraction and integration of effective emotional feature,a Multi-scale Feature and Hybrid Attention mechanism-based Emotion Classification Model(MFHA-ECM)is proposed.The model leverages Chinese-bert-wwm-ext to extract global semantic features of text,employs Convolutional Neural Net-works(CNN)to capture character-level surface features of pinyin and Chinese character radicals.It also utilizes Bidi-rectional Gated Recurrent Units(BiGRU)to model deep semantic dependencies of HowNet sememes,thereby forming a multi-scale feature system.The aim is to effectively discover the scarce emotional features in social media text and improve the accuracy and robustness of emotion classification.Meanwhile,the model introduces a hybrid attention mechanism,including BERT self-attention,channel attention,sequence attention,and gated attention,achieving adaptive enhancement and fusion of emotional features and improving comprehensive perception capability of emotional cues.On the SMP2020 Weibo General Sentiment Classification dataset,the model achieved a macro-averaged F1 score of 77.06%,validating its effectiveness.