Combination of mixed encoding and deep learning for email security sentiment classification

Abstract Email sentiment classification plays an important role in enhancing email content understanding and automated processing. To better model semantic and emotional features and achieve more robust classification results, this paper proposes CSME-DLMES (Combination of Sequence Mixed Encoding and Deep Learning Models for Email Sentiment Analysis), a method specifically designed for email sentiment classification. It adopts a sequence mixed encoding strategy that integrates word normalization with sentence weighting mechanisms to capture key information and structural features in email texts. In addition, sentiment feature encoding is introduced by leveraging weighted SentiWordNet scores to extract fine-grained sentiment weights. The enriched features are further processed by bidirectional gated recurrent unit and bidirectional long short-term memory networks to model long-range dependencies. Meanwhile, a deep convolutional neural network (CNN) is incorporated to extract local sentiment-related features. CNN demonstrates significant advantages in learning local contextual patterns and fine-grained characteristics, enabling the model to capture subtle emotional differences in text. To ensure comprehensive coverage of word vectors, a sliding window mechanism is also introduced to enhance the detection of fine-grained sentiment cues. Experimental results show that CSME-DLMES achieves superior performance on email sentiment classification tasks, validating its effectiveness in recognizing complex sentiment patterns.