Reproduce “Towards Automated Error Discovery: A Study in Conversational AI” 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
Dominic Petrak, Thy Thy Tran, Iryna Gurevych
SEEED treats conversational error analysis as generalized category discovery: it detects both predefined and previously unseen failure types, then uses an LLM to name and describe newly found categories. Combining dialogue summaries, soft clustering, margin-based contrastive learning, and label-based sample ranking, it achieves H-Scores of up to 0.64 on ABCEval and improves unknown-intent detection over prior clustering methods, while exposing remaining challenges from ambiguous labels, class imbalance, and long-context dialogues.
Key insights
- Dialogue summaries substantially improve discovery in long contexts: removing them lowered the FEDI-Error H-Score from 0.36 to 0.18 in ablation tests.
- Soft cluster membership is useful for overlapping conversational failures such as ignored requests versus ignored expectations; replacing NNK-Means with hard k-means reduced performance.
- Unknown-category discovery is more demanding than known-label classification, and generated definitions are generally useful but can focus on surface situations, duplicate concepts, or only part of a category’s meaning.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.