Transformer Architecture-Based Transfer Learning for Politeness Prediction in Conversation

Abstract

Politeness is an essential part of a conversation. Like verbal communication, politeness in textual conversation and social media posts is also stimulating. Therefore, the automatic detection of politeness is a significant and relevant problem. The existing literature generally employs classical machine learning-based models like naive Bayes and Support Vector-based trained models for politeness prediction. This paper exploits the state-of-the-art (SOTA) transformer architecture and transfer learning for respectability prediction. The proposed model employs the strengths of context-incorporating large language models, a feed-forward neural network, and an attention mechanism for representation learning of natural language requests. The trained representation is further classified using a softmax function into polite, impolite, and neutral classes. We evaluate the presented model employing two SOTA pre-trained large language models on two benchmark datasets. Our model outperformed the two SOTA and six baseline models, including two domain-specific transformer-based models using both the BERT and RoBERTa language models. The ablation investigation shows that the exclusion of the feed-forward layer displays the highest impact on the presented model. The analysis reveals the batch size and optimization algorithms as effective parameters affecting the model performance.