RASwDA: Re-Aligned Switchboard Dialog Act Corpus for Dialog Act Prediction in Conversations

Run Chen, Eleanor M Lin, Shayan Hooshmand, Mariam Mustafa, Rose Sloan, Ritika Nandi, Alicia Yang, Andrea Lopez, Ansh Kothary, Isaac Suh, Catherine Lyu, Eric Chen, Sophia Horng, Julia Hirschberg

Abstract: The Switchboard Dialog Act (SwDA) corpus has been widely used for dialog act prediction and generation tasks. However, due to misalignment between the text and speech data in this corpus, models incorporating prosodic information have shown poor performance. In this paper, we report the misalignment issues present in the SwDA corpus caused by previous automatic alignment methods and introduce a re-aligned, improved version called RASwDA (Re-Aligned Switchboard Dialog Act Corpus). Our goal is to create the largest publicly available two-speaker dialogue act corpus which has correctly aligned transcripts and speech. Through manual realignment and validation of 537.5 conversations completed so far, we have exceeded the state-of-the-art dialog act recognition results trained on SwDA. As we continue to expand RASwDA by re-aligning the remaining conversations from SwDA, we anticipate further improvements in model performance, facilitated by a larger and more accurate dataset.