Reproduce “Approaching Dialogue State Tracking via Aligning Speech Encoders and LLMs” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Šimon Sedláček, Bolaji Yusuf, Ján Švec, Pradyoth Hegde, Santosh Kesiraju, Oldřich Plchot, Jan Černocký
An end-to-end system for spoken Dialogue State Tracking directly processes speech input to output structured dialogue states, aligning pre-trained speech encoders with Large Language Models. This approach using Gemma-2-9B achieved 42.17% Joint Goal Accuracy on the SpokenWOZ benchmark, outperforming cascaded ASR-DST systems while ensuring transparency with open-source components.
Key insights
- Two-stage training, starting with ASR pre-training, is crucial for effectively aligning speech and language model representation spaces.
- Low-Rank Adaptation (LoRA) is more effective than full fine-tuning for adapting the LLM component in this end-to-end spoken DST setup.
- Including full dialogue history (both user and agent turns) and applying fuzzy matching for slot values significantly improves dialogue state tracking accuracy.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.