Research
Fact-checking & Claim Detection
Building systems that assist human fact-checkers — identifying check-worthy claims, matching them to previously fact-checked evidence, and the shared-task ecosystem around them.
Human fact-checkers are overwhelmed. Every hour, a flood of political claims, viral posts, and ambiguous half-truths arrives, and only a small fraction of them can be investigated in depth. My work in this area focused on assisting fact-checkers rather than replacing them: surfacing which claims deserve a closer look, and — when a claim has already been checked somewhere — finding that prior verdict quickly.
Check-worthiness estimation
Given a stream of tweets or debate transcripts, predict which sentences a fact-checker would actually choose to verify. The signal here is subtle: many claims are technically falsifiable but pragmatically unimportant, and prioritization is a core bottleneck in real-world fact-checking operations.
Previously fact-checked claim matching
Given a new claim, retrieve the most relevant already-fact-checked claim from a database of past verdicts — the “known lie” problem. Our ACL 2020 paper That Is a Known Lie introduced the task; follow-ups explored the role of context (the surrounding document) and the document-level variant of the problem, where multiple claims need to be matched jointly.
Shared-task infrastructure
The CLEF CheckThat! lab has run yearly tasks on these problems since 2018, driving reproducible benchmarks for the community. I co-organized the English tracks in 2020 and 2021, and contributed to lab overviews through 2022. These shared tasks shaped how the community measures progress on the assistive-fact-checking problem.
Why I stepped back
The pandemic-era work (below) was intense and largely convergent — the same core methods applied across many languages and platforms. After 2022, I shifted my focus to long-form generation and evaluation, which I felt had more open questions and room to develop a distinctive research agenda. The fact-checking work continues without me; the infrastructure is in good hands.