Short text sentiment classification algorithm based on iAFF and memory network improvements

With the development of deep learning, short text sentiment classification technology has made significant progress. However, accurately identifying sentiment in short texts remains a major challenge. Addressing issues such as insufficient semantic representation of traditional word embeddings, inadequate semantic structure extraction, and insufficient contextual sentiment feature extraction in short text sentiment classification, this paper proposes an improved method based on iAFF and memory networks. We use the iAFF module to fuse different granular word vectors generated by RoBERTa + CNN, enhancing the semantic expression and contextual relationships of features through multi-scale and attention mechanisms, thereby addressing the limitations of traditional word embeddings. At the same time, we employ ON-LSTM to process the fused features, enabling the model to better extract sentiment features from the context. Inspired by end-to-end memory networks, the entire network structure is designed as an end-to-end memory network, combined with a multi-head attention mechanism to repeatedly process the features, thereby enhancing the model's ability to represent long-distance dependencies. Finally, we perform classification of short texts through max pooling module and Softmax module. With these improvements, our method significantly enhances model performance in text sentiment recognition and classification tasks, and comparative experiments on the SST-1, SST-2, and Yelp-5 datasets showed precision rates of 0.911, 0.924, and 0.891, respectively.