A Semantically Compositional Annotation Scheme for Time Normalization

A Semantically Compositional Annotation Scheme for Time Normalization Steven Bethard, Jonathan L Parker University of Alabama at Birmingham Birmingham, AL, USA bethard@uab.edu, aescgar@uab.edu Abstract We present a new annotation scheme for normalizing time expressions, such as three days ago, to computer-readable forms, such as 2016-03-07. The annotation scheme addresses several weaknesses of the existing TimeML standard, allowing the representation of time expressions that align to more than one calendar unit (e.g., the past three summers), that are defined relative to events (e.g., three weeks postoperative), and that are unions or intersections of smaller time expressions (e.g., Tuesdays and Thursdays). It achieves this by modeling time expression interpretation as the semantic composition of temporal operators like UNION, NEXT, and AFTER. We have applied the annotation scheme to 34 documents so far, producing 1104 annotations, and achieving inter-annotator agreement of 0.821. Keywords: time expressions, normalization, compositionality 1. Introduction Time normalization is the task of translating natural lan- guage expressions of time, such as three days ago, to computer-readable forms, such as 2015-10-12. Accu- rate time normalization is critical for enabling temporally- constrained search over free text language resources. Ap- plications range from the study of literary texts (Fischer and Str¨otgen, 2015) to mining patient records for potential causes of disease (Lin et al., 2014). The most popular scheme for annotating such normalized forms is ISO-TimeML (Pustejovsky et al., 2010), an exten- sion of the TIDES annotation guidelines (Ferro et al., 2005). Figure 1 shows a sample of such annotations. Time ex- pressions are annotated as phrases and the VALUE attribute indicates the normalized form. This type of annotation has formed the basis for a wide variety of annotated corpora, in- cluding the TimeBank (Pustejovsky et al., 2003), WikiWars (Mazur and Dale, 2010), TimeN (Llorens et al., 2012), and the various TempEval shared tasks (Verhagen et al., 2007; Verhagen et al., 2010; UzZaman et al., 2013). On TIMEX3 TYPE=DATE VALUE=1995-05-22 May 22, 1995 ...and TIMEX3 TYPE=DATE VALUE=1996 ANCHORTIME the following year Figure 1: ISO-TimeML annotation of times in a sentence from the TimeBank article APW19980418.0210 However, there are a few drawbacks of the ISO-TimeML ap- proach. First, it struggles to represent times that do not align to a single calendar unit (day, week, month, etc.), such as the past three summers, since this cannot be described with some prefix of a YYYY-MM-DDTHH:MM:SS date-time. Second, it does not permit times to be defined relative to events (only relative to times), as in three weeks postopera- tive, where the three-week period is defined relative to some operation event. Finally, the flat nature of the annotations does not reflect the compositional semantics of time expres- sions. For example, following can be applied not only to year as in Figure 1, but also to week, Tuesday, or December, and the semantics is identical: find the first such calendar unit in the future relative to the anchor time (May 22, 1995). We thus propose a new annotation scheme for time nor- malization that can faithfully represent a wider …

Content truncated to 500 words