🧑‍🔬 Aida Mostafazadeh Davani

Aida Mostafazadeh Davani works on natural language processing for tasks where human judgement is subjective, such as hate speech detection, emotion and sentiment classification, and the detection of moral sentiment. A recurring question in the papers here is what gets lost when the differing labels of several annotators are collapsed into a single answer, and how classifiers absorb social bias from the data they learn from. The methods include building annotated corpora, designing multi-task neural models, applying post-hoc explanation techniques, and analysing annotator agreement across socio-demographic groups.

The most cited work, Dealing with Disagreements, argues against merging labels before training. It proposes multi-annotator models in which each annotator's judgements are predicted by a separate output head built on a shared representation, and evaluates them on hate speech and emotion tasks. Contextualizing Hate Speech Classifiers with Post-hoc Explanation shows that classifiers lean heavily on group identifiers as signals of hate. It then adds a training penalty based on explanation scores, which pushes the models to attend to the surrounding context instead. On Releasing Annotator-Level Labels and Information in Datasets compares how closely individual annotators match the majority label and asks whether that match differs across demographic groups.

The newest paper, The Moral Foundations Reddit Corpus, turns to moral sentiment. It offers a new collection of Reddit comments, each coded by several trained annotators. The coding follows the revised Moral Foundations Theory, which separates fairness into equality and proportionality. The corpus extends an earlier Twitter-only resource to a different platform and to the updated set of categories.

Davani is currently at Google, and earlier papers list the University of Southern California. Frequent co-authors include Vinodkumar Prabhakaran, Brendan Kennedy and Mark Díaz. The papers here fall under Sentiment & emotion and Bias, fairness & safety.

Aida Mostafazadeh Davani has 4 papers in our library, in 2020–2026, most of them on Sentiment & emotion, Bias, fairness & safety; often with Brendan Kennedy, Mark Díaz, Vinodkumar Prabhakaran.