Neural coreference resolution

Coreference resolution aims at detecting chains of coreference mentions in a text, that is mentions in the text that refer to the same entity.
While at first coreference resolution was split into two separated sub-problems, i.e. mention detections and resolution of coreferent mentions [1], thanks to the development of sophisticated neural models [2,3,4], end-to-end coreference resolution system can be based on a whole single model.
The aim of this stage is to study Sequence-to-Sequence [5] and Transformer [6] neural models for coreference resolution, integrating different types of attention mechanisms and possibly arbitrarily-long context [8], with the goal of understanding their impact in dealing with this complex NLP problem.

In this internship the student will implement parts of the systems for coreference resolution with Sequence-to-Sequence and Transformer neural models.
The student will run experiments on his own using GPUs, and the systems will be tested on the CoNLL Semeval 2012 benchmark [7].

Profile:
- Student for internship level stage (Master 2) in computer science, or from engineering school
- Computer science skills:
    Python programming with good knowledge of deep learning libraries (TensorFlow or PyTorch)
    Textual data manipulation (xml format, tabular format, CoNLL format)
- Interested in Natural Language Processing
- Skills in machine learning for probabilistic models

Context: