Analysis and Refinement of Temporal Relation Aggregation

Proceedings of COLING 2014, the 25th International Conference on Computational Linguistics: Technical Papers, pages 1556–1566, Dublin, Ireland, August 23-29 2014. Analysis and Refinement of Temporal Relation Aggregation Taylor Cassidy IBM Research Army Research Laboratory Adelphi, MD 20783, USA taylor.cassidy.ctr@mail.mil Heng Ji Computer Science Department Rensselaer Polytechnic Institute Troy, NY 12180, USA jih@rpi.edu Abstract To obtain a complete temporal picture of a relation it is necessary to aggregate fragments of tem- poral information across relation instances in text. This process is non-trivial even for humans because temporal information can be imprecise and inconsistent, and systems face the additional challenge that each of their classifications is potentially false. Even a small amount of incorrect proposed temporal information about a relation can severely affect the resulting aggregate tempo- ral knowledge. We motivate and evaluate three methods to modify temporal relation information prior to aggregation to address this challenge. 1 Introduction Temporal information about relations is conveyed in text at varying levels of completeness and speci- ficity. A sentence may indicate that a relation starts, ends, or that it is ongoing at a particular time. Furthermore, a time expression may be expressed at a variety of granularity levels (e.g., hour, day, or year). For instance, “Collins, ..., is a 61-year-old veteran who went 444-434 in six seasons as a man- ager, 1994-1996 with Houston” provides bounds on both the start and end date of the a relation but at a coarse granularity. Conversely, “Ivory Coast President Laurent Gbagbo on state television Friday dissolved parliament” conveys temporal information about an arbitrary part of Gbagbo’s presidency at a finer granularity: the relation simply holds true at the document creation time (DCT). Single instances in which a relation of interest is related to a time expression often fail to convey complete, fine-grained tem- poral information. Thus, it is necessary to aggregate information from multiple relation-time temporal relationship mentions to gain a complete temporal picture of a relation. We focus on the aggregation of temporal information about relations within the context of the Tem- poral Slot-Filling (TSF) Task (Ji et al., 2011; Surdeanu, 2013). TSF focusses on a class of relations called fluents (Russell and Norvig, 2010), which are properties of named entities whose values may vary over time. Systems must succinctly describe all temporal information about each query relation R – e.g., title(Gbagbo, President) – available in a source document collection by assigning it a single, final temporal four-tuple (Amigo et al., 2011). Given a relation mention r of R and a time expression γ, a four- tuple T r γ = � t(1), t(2), t(3), t(4)� characterizes their temporal relationship; namely, t(1) and t(2) represent the earliest and latest possible start date for R, while t(3) and t(4) represent the earliest and latest possible end dates, as inferred from the relation mention’s context (sec. 3). For instance, a sentence indicating that Gbagbo was President on 2010-02-12 yields � −∞, 2010-02-12, 2010-02-12, +∞ � , while the sen- tence “Gbagbo has been in power since 2000” yields …

Content truncated to 500 words