Recognising and Interpreting Named Temporal Expressions

Proceedings of Recent Advances in Natural Language Processing, pages 113–121, Hissar, Bulgaria, 7-13 September 2013. Recognising and Interpreting Named Temporal Expressions Matteo Brucato University of Bologna matteo.brucato@studio.unibo.it Leon Derczynski University of Sheffield leon@dcs.shef.ac.uk Hector Llorens Nuance Communications hector.llorens@nuance.com Kalina Bontcheva University of Sheffield kalina@dcs.shef.ac.uk Christian S. Jensen Aarhus University csj@cs.au.dk Abstract This paper introduces a new class of tem- poral expression – named temporal ex- pressions – and methods for recognis- ing and interpreting its members. The commonest temporal expressions typically contain date and time words, like April or hours. Research into recognising and in- terpreting these typical expressions is ma- ture in many languages. However, there is a class of expressions that are less typical, very varied, and difficult to automatically interpret. These indicate dates and times, but are harder to detect because they of- ten do not contain time words and are not used frequently enough to appear in con- ventional temporally-annotated corpora – for example Michaelmas or Vasant Pan- chami. Using Wikipedia and linked data, we auto- matically construct a resource of English named temporal expressions, and use it to extract training examples from a large corpus. These examples are then used to train and evaluate a named temporal ex- pression recogniser. We also introduce and evaluate rules for automatically interpret- ing these expressions, and we observe that use of the rules improves temporal annota- tion performance over existing corpora. 1 Introduction The ability to express time in language is critical. We require this ability in order to communicate plans, to tell stories, and to describe change in the world around us. Phrases that explicitly describe certain periods of time, or temporal expressions, are particularly useful. They may be calendar dates, mentions of months, relative expressions like “tomorrow”, and so on. In-depth accounts of temporal expres- sions – timexes – are given by Ferro et al. (2005) and Llorens et al. (2012a). In this paper, we discuss a new class of timexes that signify a date or range of dates, but that do not explicitly include information about which dates these are (e.g., October 31 vs. Halloween). Fol- lowing the description of expressions that clearly identify one entity from a set of others by use of a proper noun as named entities, we call these named temporal expressions (or NTEs). As with many linguistic phenomena, the phrases used as timexes have a power law-like fre- quency distribution in text. A few forms of ex- pression make up for the bulk of occurrences of temporal expressions. However, existing research has been typically evaluated on only a small cor- pus of hand-annotated temporal expressions. With such resources, it is difficult to build or evalu- ate tools for recognising or interpreting the less- frequent temporal expressions, and this is reflected in the performance plateau of recent TempEval ex- ercises (Verhagen et al., 2010; UzZaman et al., 2013). Existing temporal expression recognition tools are typically rule-based (Str¨otgen and Gertz, 2010). These perform reasonably well on exist- ing datasets, achieving F-scores of …

Content truncated to 500 words