Papers by Alban Goupil
When Annotators Disagree, Topology Explains: Mapper, a Topological Tool for Exploring Text Embedding Geometry and Ambiguity (2025.emnlp-main)
Copied to clipboard
| Challenge: | Language models are evaluated with scalar metrics like accuracy but do not capture how they encode ambiguity and more generally instances. |
| Approach: | They propose to analyze how fine-tuned models encode ambiguity and more generally instances. |
| Outcome: | The proposed tool uncovers decision regions, boundary collapses, and overconfident clusters in a RoBERTa-Large dataset. |
Beyond Black-Box Labels: Interpretable Criteria for Diagnosing Subjective NLP Tasks (2026.findings-acl)
Copied to clipboard
| Challenge: | Existing approaches to assess annotator judgments aggregate disagreement into a single gold label . et al., 2022) show disagreement is diffuse, but standard approaches are not rigorous . |
| Approach: | They propose a schema-level diagnostic for auditing expert-designed annotation schemas prior to gold-label commitment . they find disagreement is not diffuse: instability concentrates in a few criteria, while nearly half of covered sentences activate multiple categories. |
| Outcome: | The proposed diagnostic separates unstable criteria with hard-to-operationalize boundaries and systematic overlap that blurs the boundaries between mutually exclusive categories. |