Skip to content

Research

Research philosophy Link to heading

Understanding the world requires a system of concepts that supports computation and logical reasoning.

How AI and science deepen understanding
AI proposes revisions

Suggest new concepts, models, and rules to explain the evidence.

Conceptual, computational, and logical systems

Define what to check, compute expected results, and specify what follows from the rules.

Verify AI outputs

Compare predictions with observations and check conclusions against premises and rules.

Verification results

Supporting evidence and counterexamples guide AI’s next revisions.

Deeper scientific understanding

Revisions supported by evidence guide the next round of inquiry and verification.

Concepts, computation, and logic Link to heading

Concepts organize observations into objects, states, and relations. Computation lets us model and predict their behavior, while logic specifies what follows from our facts and assumptions.

AI and scientific discovery Link to heading

AI can help identify missing variables, propose new relations, or suggest a better way to describe a phenomenon. Scientific progress comes when those revisions explain observations and yield new testable predictions.

Building verification systems for AI Link to heading

A conceptual system defines what an output refers to. Computational models can produce independent reference results, and logic supplies the rules and constraints to check. These checks help establish where an AI result is reliable and when more evidence is needed.

From verification back to discovery Link to heading

Experiments and counterexamples reveal where concepts, models, or rules need revision. Changes that survive testing improve our understanding of the world and give AI a stronger basis for the next round of reasoning.

Research directions Link to heading

My work connects three steps, from representing what language expresses to learning structured knowledge and reasoning with it. The results of inference can point back to distinctions, facts, or rules that need revision.

Structured language representation

Make the structure of language available to computation so that different tasks can reuse the same analysis.

  • Lingua-Graph (ACL 2026). Represents entities, relations, and nested facts in a single graph so that several language understanding tasks can share the same structure.
  • Open Information Annotation (EMNLP 2020). Introduces a predicate–function–argument annotation and a labeled dataset as a shared intermediate step for open information extraction.

Learning between language and knowledge

Learn to map text to structured facts and express those facts in language, using each direction to improve the other.

  • Logician (WSDM 2018). Introduces the SAOKE knowledge format and trains a neural model to extract a sentence’s facts jointly.
  • Logician and Orator (EMNLP 2018). Pairs open-domain information extraction with its reverse task, information narration, and uses feedback from the two directions to train both models.

Logic and reasoning systems

Combine explicit rules with large, incomplete knowledge bases and learned representations to make inference practical.

  • MLN4KB (ACM Web Conference 2023). Exploits the sparsity of a knowledge base to scale Markov logic inference in the open-world setting.
  • DiffLogic (NeurIPS 2023). Uses probabilistic soft logic to train knowledge graph embeddings alongside weighted rules, with a filter that limits inference to relevant triples.