Master thesis : Extending Joint Entity and Event Coreference Resolution across Documents

Abstract

Detecting corefering events and entities in texts is an important task in NLP, where it plays a role in many other tasks and applications. In this work, we build on a joint approach of entity and event coreference resolution, pioneered by H. Lee, Recasens, et al. 2012 and matured by Barhom et al. 2019 using a neural architecture. In particular we look at coreference resolution across documents, more complicated and less researched than coreference resolution within documents. Using the Barhom et al. 2019’s model, we propose a series of extensions to improve its results. This is done by increasing the amount of information provided to the model, in particular the joint nature of the modelling and by improving entity and event representation with the use of document embedding. As a secondary problem, we investigate ways to improve the model’s time performance through compressing the mention representations. Our results are compared with other works tackling the problem of cross document coreference resolution on the ECB+ dataset, the standard dataset for cross document entity and event coreference resolution.