PhD defence by Rahul Rajendra Arailkatte

Rahul PhD Defense Banner 1100x600

The PhD defense will be conducted as an online event via Zoomlink:


Towards Better Neural Coreference Resolution


This thesis describes novel methods of improving state-of-the-art neural coreference resolvers without the need for additional annotated data. The three newly introduced methods can be summarized as follows:

(1) Augmenting external knowledge: knowledge bases contain enormous amounts of real-world knowledge which can be used to resolve ambiguities that arise from grounding assumptions. We introduce a reinforcement learning-based approach that improves performance by verifying model decisions against external knowledge bases and rewarding them based on their validity.

(2) Remodeling of tasks: performance of some tasks can be improved if they are recast into a different form that is more suitable for learning. Since the availability of training data varies drastically across tasks, we remodel a low-resource task to take the form of a high-resource task, and use models pre-trained for the latter and finetune them to get significant improvements on the former.

(3) Encouraging coherence in MTL: in standard multitask learning setups, strongly correlated tasks result in better overall performance. Taking this a step further, we build simple meaning representations from the outputs of the model to explicitly quantify the coherence between them and use this coherence value as a reward to further finetune the models.
We thoroughly experiment with and analyze these methods and report performance improvements across the board.

Assessment Committee

  • Chair: Assistant professor, Desmond Elliott, Department of Computer Science
  • Senior Research Scientist, Ivan Vulic, Cambridge University, UK
  • Research Scientist, Adina Williams, Facebook AI Research

Academic supervisor

Professor, Anders Søgaard, Department of Computer Science, UCPH

Moderator at this defense

Assistant professor, Daniel Hershcovich, Department of Computer Science, University of Copenhagen

For an electronic copy of the thesis, please go to: