Reproduce “UniSLU: Unified Spoken Language Understanding from Heterogeneous Cross-Task Datasets” 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
Zhichao Sheng, Shilin Zhou, Chen Gong, Zhenghua Li
UniSLU, developed at Soochow University, introduces a unified generative framework for spoken language understanding that jointly models ASR, Named Entity Recognition, and Sentiment Analysis. It effectively leverages heterogeneous datasets through a consistent output representation and dynamic loss weighting, outperforming separate models and achieving an overall SLUE score of 73.81 with a Whisper-medium backbone.
Key insights
- A unified generative framework effectively leverages heterogeneous data, enabling cross-task learning between ASR, NER, and SA.
- Dynamic loss weighting is crucial for balancing training signals across tasks with varying output lengths, preventing longer outputs from dominating training.
- Integrating Large Language Models (LLMs) as decoders significantly enhances semantic understanding for NER and SA tasks, leveraging their pre-trained knowledge.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.