Machine Compilation of Scientific Knowledge

Papers are written for humans

Scientific knowledge is still primarily published as prose. Papers are written for humans, and much of what makes them scientifically meaningful is carried implicitly through language: what the authors observed, what they infer, what they believe the evidence supports, what they merely hypothesize, and what they attribute to prior work. I am exploring whether some of that structure can be made explicit.

Why this is newly practical

Modern language models make this question newly practical. They can already extract entities, relationships, claims, experimental conditions, and other structured information from scientific text. But extracting information is not the same as correctly interpreting scientific discourse. A system may reproduce the content of a sentence while misrepresenting its epistemic role.

An author may be proposing a possibility rather than asserting a fact, describing an observation rather than explaining it, reporting someone else’s conclusion rather than endorsing it, or presenting evidence that supports a claim without establishing it.

Research agenda

I am approaching this problem incrementally. The current research agenda investigates five related problems:

1. Epistemic acts — identifying what an author is doing with a proposition: asserting it, reporting it, hypothesizing it, inferring it, challenging it, or otherwise qualifying its status. This is the subject of the program’s first project, Epistemic Act Recognition in Scientific Literature.
2. Grounded extraction — testing whether structured representations become more reliable when every extracted object must remain anchored to the exact source text from which it was derived.
3. Verified compilation — combining probabilistic language-model interpretation with deterministic validation and independent semantic verification, including generate–verify–repair workflows.
4. Selective compilation — determining when a system should abstain from formalizing ambiguous language rather than silently producing an unreliable representation.
5. Scientific intermediate representation — using the results of these experiments to develop a minimal representation for propositions, evidence, epistemic acts, context, provenance, and the relationships between them.

The longer arc

The longer-term idea is to treat the scientific paper as a human-readable serialization of a deeper scientific object rather than as the canonical container of scientific knowledge.

If such an object can be represented reliably, the same underlying research could eventually support many interfaces: traditional papers, evidence graphs, structured literature reviews, interactive scientific browsers, machine reasoning systems, and new forms of scientific publishing.