The Importance of Context for Sentiment Analysis in Dialogues

Abstract

Sentiment Analysis (SA) can be applied to dialogues to determine the emotional tone throughout the conversation. This is beneficial for dialogue systems because it may improve human-computer interaction. For instance, in case of negative sentiment, the system may switch to a human operator who can handle the situation more effectively. However, given that dialogues are a series of utterances, the context, including the previous text, plays a crucial role in analyzing the current sentiment. Our aim is to investigate the importance of context when monitoring the sentiment of every utterance during a conversation. To accomplish this goal, we assess sentiment analysis in dialogues with varying levels of context, specifically differing in the number and author of preceding utterances. We conduct experiments on Portuguese customer-support conversations, with each utterance manually labeled as having negative or non-negative sentiment. We test a wide range of text classification approaches, from traditional, as simplicity should not be overlooked, to more recent methods, as they are more likely to achieve better performances. Results indicate that the relevance of context varies. However, context assumes particular value in human-computer dialogues, when considering both speakers, and in shorter human-human conversations, when focusing on the client. Moreover, the best classifier for both scenarios, based on BERT, achieves the highest scores when considering the context.

Keywords

• Sentiment analysis
• dialogue analysis
• context awareness
• natural language processing
• deep learning
• machine learning