Learning interaction dynamics with an interactive LSTM for conversational sentiment analysis.

Area of Science:

• Natural Language Processing
• Artificial Intelligence
• Computational Linguistics

Background:

• Conversational sentiment analysis is challenging due to the lack of suitable datasets and inability to model interpersonal dynamics.
• Existing methods fail to capture the rich interaction information influencing sentiment in conversations.

Purpose of the Study:

• To introduce a novel conversational sentiment analysis dataset, ScenarioSA.
• To investigate the multidimensional nature of interaction dynamics (understandability, credibility, influence) in conversations.
• To propose an advanced model for conversational sentiment analysis that effectively incorporates interaction dynamics.

Main Methods:

• Development and release of the ScenarioSA dataset for conversational sentiment analysis.
• Investigation of interaction dynamics, including understandability, credibility, and influence.
• Proposal of an interactive Long Short-Term Memory (LSTM) network with confidence gates and learned influence scores to model speaker interactions.

Main Results:

• The proposed interactive LSTM network significantly outperforms existing baselines on the ScenarioSA dataset.
• The model achieves competitive results with state-of-the-art approaches on both ScenarioSA and IEMOCAP datasets.
• Experimental results validate the effectiveness of modeling interaction dynamics for improved conversational sentiment analysis.

Conclusions:

• The developed ScenarioSA dataset and the interactive LSTM network provide valuable resources for advancing conversational sentiment analysis.
• Modeling interaction dynamics, such as credibility and influence, is crucial for accurately understanding affective states in conversations.
• The proposed approach offers a promising direction for future research in sentiment analysis within conversational contexts.