Sentiments analysis for intelligent customer service dialogue using hybrid word embedding and stacking ensemble

The intelligent customer service conversations between customer and customer service agent exhibit characteristics such as serious colloquialism, a high diversity of words, and short text length, making traditional sentiment analysis algorithms less effective. Besides, relying solely on coarse-grained sentiment analysis could not fully display the text information of users in the dialogue, resulting in poor performance of sentiment classification prediction. To address these issues, this paper introduces a hybrid word embedding method (HWE) based on Gaussian distribution to leverage the emotional syntactic and semantic richness of the two distributed word representations (Word2Vec, and GloVe). Additionally, a stacked ensemble method is employed, combining outputs from three deep learning models—Convolutional Neural Network (CNN), Long Short-Term Memory (LSTM), and Gated Recurrent Unit (GRU). This ensemble approach enables simultaneous prediction of both coarse-grained and fine-grained sentiment analysis for intelligent customer service conversation text. The results show that HWE method enhances the comprehension of word representations in the context of conversations, leading to more effective sentiment analysis. Furthermore, the proposed stacked ensemble model outperforms current state-of-the-art single-task models, showcasing a significant improvement in sentiment classification accuracy.