Identification of Temporal Event Relationships in Biographical Accounts

Proceedings of NAACL-HLT 2013, pages 529–533, Atlanta, Georgia, 9–14 June 2013. c⃝2013 Association for Computational Linguistics Identification of Temporal Event Relationships in Biographical Accounts Lucian Silcox University of Texas PanAm 1201 W. University Dr. Edinburg, Tx 78539 lucian.silcox@gmail.com Emmett Tomai University of Texas PanAm 1201 W. University Dr. Edinburg, Tx 78539 tomaie@utpa.edu Abstract This paper examines the efficacy of the appli- cation of a pre-existing technique in the area of event-event temporal relationship identifi- cation. We attempt to both reproduce the re- sults of said technique, as well as extend the previous work with application to a newly- created domain of biographical data. We find that initially the simpler feature sets perform as expected, but that the final improvement to the feature set underperforms. In response, we provide an analysis of the individual features and identify differences existing between two corpora. 1 Introduction As natural language systems continue to grow, so too does the importance of extracting temporal information from text. Narratives often contain a wealth of temporal information, linking specific events to each other and to individual named enti- ties of importance, but such information is often implicitly conveyed, rather than explicitly stated. The continued interest in Question Answering and other data extraction systems has emphasized the need to better understand these relations to move past superficial understanding to a level of deeper comprehension. For native speakers, the temporal clues hidden in the text are relatively simple to comprehend. However, even for human annotators, the task of identifying and classifying the specific relationship between two events can be problemat- ic. This complexity, of course, only exacerbates the problem of trying to automate the process for any information extraction system. The creation of the TimeBank Corpus (Pustejovsky et al, 2003a), a fully-annotated newswire domain, opened up the possibility of ap- plying machine learning techniques to the task of automatically extracting temporal relations. We look to the standards of the TimeBank Corpus to create a corpus of biographical accounts, and apply techniques that have been shown to work on TimeBank to the new domain. 2 Related Work Domain-independent approaches have often focused on events that can be bound to a global timeline (Mani et al, 2003). This includes dates and times, but often neglects phrases that indicate events occurring in relative time (e.g. “during school,” “before the crash,” or “recently”). Re- search conducted on news articles attempted to identify the specific temporal relationships be- tween two events, as seen in (Mani et al, 2006). Further work in that domain extended this start by identifying additional features that better predicted those temporal relations. (Lapata & Lascarides 2007; Chambers et al, 2007). In this work, we are primarily interested in applying event ordering techniques to documents less structured than news articles, specifically bi- ographies. It is the intention of our work to validate the efficacy of previous techniques in a different domain, and thus we attempt to extend the work completed by Chambers et al through application to a newly created corpus of …

Content truncated to 500 words