Temporal Relation Classification using a Model of Tense and Aspect

Proceedings of Recent Advances in Natural Language Processing, pages 118–122, Hissar, Bulgaria, Sep 7–9 2015. Temporal Relation Classification using a Model of Tense and Aspect Leon Derczynski University of Sheffield leon@dcs.shef.ac.uk Robert Gaizauskas University of Sheffield robertg@dcs.shef.ac.uk Abstract Determining the temporal order of events in a text is difficult. However, it is crucial to the extraction of narratives, plans, and context. We suggest that a simple, estab- lished framework of tense and aspect pro- vides a viable model for ordering a sub- set of events and times in a given text. Using this framework, we investigate ex- tracting features that represent temporal information and integrate these in a ma- chine learning approach. These features improve event-event ordering. 1 Introduction It is important to understand time in language. The ability to express and comprehend expressions of time enables us to plan, to tell stories, and to dis- cuss change in the world around us. When we automatically extract temporal infor- mation, we are often concerned with events and times – referred to collectively as temporal inter- vals. We might ask, for example, “Who is the current President of the USA?.” In order to ex- tract a single contemporary answer to this ques- tion, we need to identify events related to persons becoming president and the times of those events. Crucially, however, we also need to identify the ordering between these events and times, by as- signing a temporal relation type (from e.g. Allen (1983)). This last task, temporal relation typing, is challenging (UzZaman et al., 2013; Bethard et al., 2015), and is the focus of this paper. When events are expressed as verbs, tense and aspect are used to convey temporal features of these events. Thus, it is intuitive that tense and aspect will be of value in determining the type of temporal relation that holds between two verb events, and evidence in human-annotated corpora supports this intuition. Event-event relations are often the hardest to la- bel (Derczynski, 2015). Around 45% of links in TempEval-3 (UzZaman et al., 2013) event-event ordering tasks cannot reliably be labelled automat- ically. Temporal relations involving at least one argu- ment with tense or aspect information are preva- lent. Verb-verb links make up around a third of TimeBank’s temporal relations,1 and tensed verb- verb links the largest share of that set, so of all verb-verb relations, the majority are between two tensed verbs. Data-driven approaches to the relation typing task are hampered in two ways. Firstly, there is a shortage of ground truth training data. This leads to low volumes of instances for many combina- tions of tense and aspect values for pairs of events, hampering automatic hypothesis learning (Lapata and Lascarides, 2006). Secondly, the range of tense and aspect expression in TimeML is rela- tively limited, describing three “tenses”2 (past and past participle, present and present participle, and future) and three “aspects” (none, perfective and progressive). This markup language may be insuf- ficiently descriptive to capture relations implied by variations in linguistic use of tense …

Content truncated to 500 words