Abstract
This thesis focuses on the linguistic phenomena of anaphora and their automatic resolution. The objective of this project is to study the contribution of the external knowledge in the performance of an automatic resolution system. To do so, we aim to design an explainable algorithm to list and explain the choices made. This makes it possible to identify the most influential mechanisms in the cases of successful resolutions. We define this objective in the setting of computer mediated communication presenting particularities making the task all the more complex. In order to overcome the lack of textual resources belonging to the context that interests us, we proceed to create a dataset composed of a collection of e-mails segmented into anonymized threads and manually annotated with anaphoric links. Then, we introduce the ARCS system (Anaphora Resolution using Common Sense). It is a model involving parameters from different theoretical frameworks, with an emphasis on semantic information from JeuxDeMots. The analysis of the parameters affecting the performance scores of our system shows a clear improvement allowed by the methods introducing semantic knowledge. It also allowed us to bring out a set of steps that should be taken in order to succeed in solving the most difficult cases.