TimeML Events Recognition and Classification: Learning CRF Models with Semantic Roles

Proceedings of the 23rd International Conference on Computational Linguistics (Coling 2010), pages 725–733, Beijing, August 2010 TimeML Events Recognition and Classification: Learning CRF Models with Semantic Roles Hector Llorens, Estela Saquete, Borja Navarro-Colorado Natural Language Processing Group University of Alicante {hllorens,stela,borja}@dlsi.ua.es Abstract This paper analyzes the contribution of se- mantic roles to TimeML event recognition and classification. For that purpose, an approach using conditional random fields with a variety of morphosyntactic features plus semantic roles features is developed and evaluated. Our system achieves an F1 of 81.4% in recognition and a 64.2% in classification. We demonstrate that the application of semantic roles improves the performance of the presented system, es- pecially for nominal events. 1 Introduction Event recognition and classification has been pointed out to be very important to improve com- plex natural language processing (NLP) applica- tions such as automatic summarization (Daniel et al., 2003) and question answering (QA) (Puste- jovsky, 2002). Natural language (NL) texts often describe sequences of events in a time line. In the context of summarization, extracting such events may aid in obtaining better summaries when these have to be focused on specific happenings. In the same manner, the access to such information is crucial for QA systems attempting to address questions about events. The analysis of events as well as the classifica- tion of the different forms they adopt in NL text is not a new issue (Vendler, 1967). It relates not only to linguistics but different scientific areas such as philosophy, psychology, etc. In NLP, different definitions of event can be found regarding the target application. On the one hand, in topic detection and track- ing (Allan, 2002), event is defined as an instance of a topic identified at document level describing something that happen (e.g., “wars”). The aim of this task is to cluster documents on the same topic, that is to say, the same event. On the other hand, information extraction (IE) provides finer granularity event definitions. IE proposes standard schemes to annotate the indi- vidual events within the scope of a document. STAG scheme (2000) was aimed to identify events in news and their relationship with points in a tem- poral line. More recently, TimeML (Pustejovsky et al., 2003a) presented a rich specification for an- notating events in NL text extending the features of the previous one. This paper is focused on the TimeML view of events. TimeML defines events as situations that happen or occur, or elements describing states or circumstances in which something obtains or holds the truth. These events are generally ex- pressed by tensed or untensed verbs, nominaliza- tions, adjectives, predicative clauses or preposi- tional phrases. TimeML guidelines define seven classes of events: • Reporting. Action of a person or organization declar- ing or narrating an event (e.g., “say”) • Perception. Physical perception of another event (e.g., “see”, “hear”) • Aspectual. Aspectual predication of another event (e.g., “start”, “continue”) • I Action. Intensional action (e.g., “try”) • I State. Intensional state (e.g., “feel”, “hope”) • State. …

Content truncated to 500 words