In recent years, we have witnessed a revolution regarding how data can be retrieved, used, or shared. News articles and social media have provided unique access to vast amounts of information, and a significant portion of this data describes events: who did what to whom, when, and where. The automated collection and structuring of such information have long been a central focus in NLP and is referred to as event extraction, aiming to create structured representations of events described in unstructured text.
Advancement in this important information extraction task can offer great potential for several other academic fields. One example is the subfield of social science known as peace and conflict studies, focusing on armed conflicts to study broader patterns in how conflicts evolve and to develop theories on how conflicts can be resolved.
One of the most used event extraction datasets in this domain is the MUC-4 dataset (named after the Message Understanding Conference). It consists of newswire articles covering political conflict in Latin American countries, with terrorism-focused event types. Although the dataset was released in the 1990s already, it has now re-emerged in research focusing on document-level event extraction. The events in the MUC data are associated with so-called templates, each annotated with a set of 24 roles (event arguments). The corresponding event extraction task is to detect the event types and predict the correct value for each slot in the template.
Most recent research, however, has narrowed its focus to just five of the 24 possible arguments: perpetrator, perpetrator organisation, target, victim’s name, and weapon. These arguments are explicitly mentioned in the text, leaving the remaining arguments, which often need to be inferred (rather than "extracted") and thereby require more inference and contextual understanding, largely unexplored.
In this master project, you will explore how we can expand the scope of event extraction to a more comprehensive set of arguments. One example is to include temporal and geospatial information (when and where the event occurred). An important part of this project will also be conducting an error analysis to better understand the limitations of existing models and domain specific challenges with the dataset, and identify areas for improvement.
Some relevant links:
- Overview of MUC-4: https://aclanthology.org/M92-1001.pdf
- Automatic error analysis on MUC: https://aclanthology.org/2022.acl-long.274.pdf