Emanuela Boros, will defend her PhD Thesis entitled neural methods for event extraction on September 27, 2018 at 2pm. The defence will take place in conférence room, bâtiment 507, LIMSI, Orsay.


With the increasing amount of data and the exploding number data  sources, the extraction of information about events, whether from the  perspective of acquiring knowledge or from a more directly operational  perspective, becomes a more and more obvious need. From the point of  view of Natural Language Processing (NLP), the extraction of events from  texts is the most complex form of Information Extraction (IE)  techniques, which more generally encompasses the extraction of named  entities and relationships that bind them in the texts. The event  extraction task can be represented as a complex combination of relations  linked to a set of empirical observations from texts. In practice, an  event is described by a trigger (the word or phrase that evokes the  event) and a set of participants in that event (that is, arguments or  roles) whose values are text excerpts.

This thesis presents several strategies for improving the performance of  an Event Extraction (EE) system using neural-based approaches exploiting  morphological, syntactic, and semantic properties of word embeddings.  These have the advantage of not requiring a priori modeling domain  knowledge and automatically generate a much larger set of features to  learn a model. More specifically, we proposed different deep learning  models for two sub-tasks related to EE: event detection and argument  detection and classification. Event Detection (ED) is considered an  important sub-task of event extraction since the detection of arguments  is very directly dependent on its outcome.

As a preliminary to the introduction of our proposed models, we begin by  presenting in detail a state-of-the-art model which constitutes the  baseline. In-depth experiments are conducted on the use of different  types of word embeddings and the influence of the different  hyperparameters of the model using the ACE 2005 evaluation framework, a  standard evaluation for this task.

We then propose two new models to improve an event detection system. One  allows increasing the context taken into account when predicting an  event instance (event trigger) by using a sentential context, while the  other exploits the internal structure of words by taking advantage of  seemingly less obvious but essentially important morphological  knowledge. We also reconsider the detection of arguments as a high-order  relation extraction and we analyze the dependence of arguments on the ED  task.