Research
8 peer-reviewed publications, with first-author papers at EMNLP and COLING. Open models and datasets released on Hugging Face.
Interpretable Topic Modeling of Spontaneous Speech in Depression Using Large Language Models
We combine multilingual representations, clustering, and language-model-generated descriptions to identify topics in spontaneous speech. Across four cohorts, we examine how these topics relate to clinical measures, diagnoses, interview questions, and demographics.
Formalizing Style in Personal Narratives
We represent narrative style through recurring patterns in how people describe subjective experiences. Language models extract processes, participants, and circumstances, which we analyze using sequential pattern mining and relate to psychological observations.
Textualized and Feature-based Models for Compound Multimodal Emotion Recognition in the Wild
We compare two approaches to multimodal emotion recognition: using visual and acoustic features directly, and translating these cues into text for a language model. Our submission placed third in the ABAW compound expression recognition challenge.
Improving Language Models for Emotion Analysis: Insights from Cognitive Science
We examine emotion analysis in NLP through theories of emotion and emotional communication from psychology, cognitive science, and pragmatics. We identify limitations of emotion-label prediction and propose directions for annotation, modeling, and evaluation.
Sequence-to-Sequence Language Models for Character and Emotion Detection in Dream Narratives
I formulate character and emotion coding in dream narratives as a sequence-to-sequence task. In these experiments, supervised models outperform in-context learning with a substantially larger language model. I also examine the effects of model size, character representation, and prediction order, and apply the models to annotate DreamBank.
Emotion Recognition based on Psychological Components in Guided Narratives for Emotion Regulation
We introduce a French corpus of autobiographical narratives annotated for behavior, feeling, thinking, and context. By using these components separately and together, we study which information contributes to emotion recognition.