Reproduce “DailyDialog: A Manually Labelled Multi-turn Dialogue Dataset” 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
Yanran Li, Hui Su, Xiaoyu Shen, Wenjie Li, Ziqiang Cao, Shuzi Niu
Researchers introduce DailyDialog, a manually labeled multi-turn dialogue dataset, to address the scarcity of high-quality, coherent, and richly annotated conversational data. The dataset comprises 13,118 human-written dialogues covering daily life topics, with explicit labels for communication intention and emotion, and establishes baselines by evaluating existing dialogue models.
Key insights
- DailyDialog effectively captures natural multi-turn conversational flows and specific human communication patterns, such as 'answer-then-question' sequences and multi-turn negotiations.
- Explicitly incorporating communication intention and emotion labels significantly improves the performance of generation-based dialogue models, leading to lower perplexity and better response quality.
- Pre-training on large, but domain-mismatched, datasets (e.g., OpenSubtitle) can negatively impact model performance when fine-tuned on daily conversation tasks, emphasizing the need for domain-aligned training data.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.