Representation and Learning of Temporal Relations

Proceedings of COLING 2016, the 26th International Conference on Computational Linguistics: Technical Papers, pages 1937–1948, Osaka, Japan, December 11-17 2016. Representation and Learning of Temporal Relations Leon Derczynski Department of Computer Science University of Sheffield S1 4DP, UK leon.d@shef.ac.uk Abstract Determining the relative order of events and times described in text is an important problem in natural language processing. It is also a difficult one: general state-of-the-art performance has been stuck at a relatively low ceiling for years. We investigate the representation of temporal re- lations, and empirically evaluate the effect that various temporal relation representations have on machine learning performance. While machine learning performance decreases with increased representational expressiveness, not all representation simplifications have equal impact. 1 Introduction Textual accounts often contain descriptions of events, times, and how they relate to one another tempo- rally. To connect events and times to each other, we need to know the kind of temporal ordering between them. This ordering can be modeled with temporal relations that hold between pairs of entities, each of which may be an event or time. Extracting these relations is critical to understanding the text: for example, given an almanac of presidents of the USA, there are likely to be many indications of different people being president – but only one will be factual at any given time. In order to reason about events and the applicability of information in a document, linguistic expressions of time need to be converted to a formal representation. This task is temporal relation annotation. To annotate temporal relations for reasoning or information extraction, one must to select a way of representing temporal relations. Such representations typically comprise a set of temporal relations, with each member describing a different temporal ordering. Building such representations is a key artificial intelligence task in reasoning and planning, and proposed solutions have amounted to major work in the field; e.g. Allen (1984), Freksa (1992). Temporal relation annotation has two key parts. One must decide which entity pairs to relate, and then determine the nature of their relation. These are referred to as temporal relation identification and temporal relation typing. These may be approached as a joint task, especially when it is possible for the type of one link to influence the type of another. For example, if event A is before event B and that event B is before event C, due to transitivity, the choice of relations between A and C may be constrained. Choosing how to represent the types of temporal relations is the focus of this paper. Machine learning of temporal relations is hard. Mani et al. (2007) detail experiments with features annotated in TimeML (Pustejovsky et al., 2004), and reach around 75% accuracy at overall relation typing,1 using gold-standard relation identification.2 This pattern repeats in subsequent literature. Many others reach a similar performance, or perhaps even exceed it by 1-2% (Mirroshandel et al., 2011; Do et al., 2012); some do well by focusing on specific sub-parts of the problem (e.g. relations that are …

Content truncated to 500 words