Event Extraction as Frame-Semantic Parsing

Proceedings of the Fourth Joint Conference on Lexical and Computational Semantics (*SEM 2015), pages 159–164, Denver, Colorado, June 4–5, 2015. Event Extraction as Frame-Semantic Parsing Alex Judea and Michael Strube Heidelberg Institute for Theoretical Studies gGmbH Schloss-Wolfsbrunnenweg 35 69118 Heidelberg, Germany (alex.judea|michael.strube)@h-its.org Abstract Based on the hypothesis that frame-semantic parsing and event extraction are structurally identical tasks, we retrain SEMAFOR, a state- of-the-art frame-semantic parsing system to predict event triggers and arguments. We de- scribe how we change SEMAFOR to be better suited for the new task and show that it per- forms comparable to one of the best systems in event extraction. We also describe a bias in one of its models and propose a feature factor- ization which is better suited for this model. 1 Introduction Event Extraction is a task in information extraction where mentions of predefined events are extracted from texts. We follow the task definition of the Au- tomatic Content Extraction (ACE) program of 2005. It defines 33 event types, organized in eight cate- gories. Each event type has associated roles, e.g., ATTACK has the roles attacker, target, and instru- ment, whereas DIE has the roles agent, victim, and instrument. The roles place and time are shared by all event types. ACE events occur only within sentences. Each event is indicated by a word, the trigger. The roles associated with the respective event type are filled by zero or more arguments. Most arguments are mentions of entities, e.g. persons, locations, or or- ganizations. Some arguments are mentions of points in time, amounts of money, etc. Arguments may be shared by multiple events and may play different roles in each of them. Figure 1 illustrates an example. The sentence contains two events, DIE and ATTACK, triggered by “died” and “fired”, respectively. For DIE, the roles victim, instrument, and place are filled with the argu- ments “cameraman”, “American tank”, and “Bagh- dad”, respectively. For ATTACK, the role target has two arguments, namely “cameraman” and “Palestine hotel”, the roles instrument, and place have the ar- guments, “American tank”, and “Baghdad”, respec- tively. Three arguments are shared. One of them, “cameraman”, plays different roles in the events, namely victim of DIE and target of ATTACK. Frame-semantic parsing is the task of extracting semantic predicate-argument structures from texts. It is built on the theory of frame semantics and FrameNet (Fillmore et al., 2003; Das et al., 2014). As in event extraction, frames occur within sen- tences and have triggers and roles (called lexical units and frame elements). Our hypothesis is that the two tasks are struc- turally identical. From a computational point of view, they differ only in feature types. We can use the same approach and infrastructure to tackle both. Based on this hypothesis, we retrain a frame- semantic parsing system, SEMAFOR, for event ex- traction. We describe differences between frame-semantic parsing and event extraction and the adaptions needed to better prepare SEMAFOR for the new task. We also describe a bias in the trigger classifi- cation model which …

Content truncated to 500 words