Research

Event-centric narrative understanding

How can AI help us understand not only what happened, but also how the same event is explained, contextualized, and evaluated differently?

I study this through event-centric computational framing. Instead of labeling a whole article as biased, I identify the events that different sources share and model the narrative choices that set their coverage apart — which events are selected, how they are ordered, what causes and consequences are emphasized, who is held responsible, and how language and images shape interpretation. Two reports can describe the very same events and still tell different stories.

My goal is transparent, evidence-grounded systems that help researchers, journalists, and readers compare narratives without asking AI to decide which perspective is correct — surfacing how accounts differ and pointing to the evidence, rather than judging them.

What I have done

I began with multilingual semantic representations, factuality and modality, and annotation infrastructure — including Uniform Meaning Representation (UMR) and UMR Writer — which taught me how events, participants, time, and attribution can be represented in interpretable forms.

I then focused on cross-document event understanding: when different descriptions refer to the same real-world event. With collaborators I built Richer EventCorefBank via event decontextualization, and introduced framing-divergent event coreference through the FrECo task and dataset, showing that two descriptions can corefer while differing sharply in perspective and evaluation. More recently I have studied how causal relations between events reveal larger narrative structures — the foundation for comparing how sources construct explanations, assign responsibility, and communicate attitudes.

Where my research is going

My current research is developing a broader framework for event-centric narrative understanding. The central idea is to separate the events that sources share from the narrative structures through which each source interprets those events. I am pursuing four connected directions:

Event alignment — identifying equivalent events across documents, languages, perspectives, and modalities, even when their descriptions use very different language.
Narrative relation extraction — modeling causal explanation, temporal organization, discourse structure, attribution, certainty, evaluation, and responsibility.
Multilingual and multimodal framing — studying how narratives vary across linguistic communities and how images and videos foreground, contextualize, humanize, or evaluate events.
Transparent narrative-comparison systems — building interactive tools that allow users to inspect shared events, competing explanations, omitted context, and supporting evidence.

In the long term, I want to create computational tools that support narrative science: enabling communication scholars, computational social scientists, journalists, and the public to study how narratives emerge, diverge, and evolve across cultures and media systems.

Interested in working on these problems? See People & Join — I am recruiting Ph.D. students and research assistants.