Modeling Dialogue Acts with Content Word Filtering and Speaker Preferences

Area of Science:

• Computational Linguistics
• Natural Language Processing
• Artificial Intelligence

Background:

• Dialogue act recognition is crucial for understanding conversational structure.
• Existing models often struggle to balance content and function in predicting dialogue acts.
• Modeling both temporal dependencies and speaker-specific behaviors is key for robust dialogue analysis.

Purpose of the Study:

• To develop an unsupervised model for dialogue act sequences.
• To investigate the impact of de-emphasizing content words and incorporating speaker tendencies.
• To improve the accuracy of dialogue act prediction in diverse conversational corpora.

Main Methods:

• An unsupervised model was developed to analyze dialogue act sequences.
• The model de-emphasizes content-related words, focusing on conversational function words.
• Speaker-specific tendencies were incorporated as an additional predictive factor.

Main Results:

• The model's effectiveness varied across corpora (CNET forum, NPS Chat).
• De-emphasizing content words improved performance on the CNET corpus.
• Utilizing speaker tendencies was advantageous for the NPS corpus.

Conclusions:

• The model components effectively complement each other for robust performance.
• The proposed model outperforms state-of-the-art baseline models on both tested corpora.
• The study highlights the context-dependent effectiveness of different modeling assumptions in dialogue act recognition.