Two truths and a lie: Modern NLP vs ground truth
- Speaker
- Timothee Mickus (University of Helsinki)
- Date
- Room
- Salle 306, bâtiment IMAG
It is something of a trope to state that neural NLP has a problem of truthfulness. Hallucinations, for instance, are a problem so widespread in LLM technology that commercial providers often include disclaimers as to the truthfulness of what their systems can output. We often assume that such models, because they are derived from distributional data, have at best a loose relationship with truth. In this talk, I hope to bring some nuance to this perspective. On many levels, this is a more pervasive problem than we might think if we focus solely on hallucinations. While hallucinations are easy to come by, we can also highlight that annotators often do not agree as to what counts as a hallucination — and models routinely struggle with human label variation. Conversely, we can even highlight cases where LLMs are biased towards producing the expected true answer. In many instances, tackling this problem of truthfulness requires rethinking where the field of NLP is at: how do we train models that capture linguistic ambiguity? What are our metrics actually sensitive to? What counts as a true label?
