Epistemic Act Recognition in Scientific Literature

More than propositions

Scientific sentences do more than communicate propositions. They also indicate what relationship an author has to those propositions. Consider the difference between:

• an observation reported from an experiment,
• a conclusion inferred from those observations,
• a hypothesis proposed for future testing,
• a claim attributed to another researcher, and
• an established result that the authors treat as background knowledge.

The propositional content may be similar, while the scientific meaning is substantially different. This project investigates whether current language models can reliably recover that distinction.

Working premise

Before attempting to construct complete scientific knowledge graphs or formal representations of papers, a compiler should first be able to determine what role each proposition plays in the author’s reasoning.

Questions under study

• Which epistemic acts can be distinguished reliably?
• Which distinctions are consistently ambiguous, even for humans?
• How much surrounding context does a model require?
• How sensitive are classifications to model choice, prompting, or terminology?
• Can models identify uncertainty rather than forcing every sentence into a category?
• Does grounding classifications in exact textual evidence improve reliability?
• Can disagreement between independent models help identify difficult cases?

Scope and first study

The immediate objective is not to build a universal taxonomy of scientific discourse. Instead, the project begins with a small operational vocabulary of epistemic acts and evaluates how reliably current models can distinguish them under controlled conditions.

Results from this project will inform later work on grounded extraction, abstention, verification, and the design of a scientific intermediate representation. The first study, Can Language Models Distinguish Scientific Epistemic Acts?, is structured as a research proposal and is currently in progress.