From Characters to Time Intervals: New Paradigms for Evaluation and Neural Parsing of Time Normalizations

From Characters to Time Intervals: New Paradigms for Evaluation and Neural Parsing of Time Normalizations Egoitz Laparra∗ Dongfang Xu∗ School of InformationSteven Bethard University of Arizona Tucson, AZ {laparra,dongfangxu9,bethard}@email.arizona.edu Abstract This paper presents the first model for time normalization trained on the SCATE corpus. In the SCATE schema, time expressions are annotated as a semantic composition of time entities. This novel schema favors machine learning approaches, as it can be viewed as a semantic parsing task. In this work, we propose a character level multi-output neural network that outperforms previous state-of-the-art built on the TimeML schema. To compare predic- tions of systems that follow both SCATE and TimeML, we present a new scoring metric for time intervals. We also apply this new metric to carry out a comparative analysis of the anno- tations of both schemes in the same corpus. 1 Introduction Time normalization is the task of translating natural language expressions of time to computer-readable forms. For example, the expression three days ago could be normalized to the formal representation 2017-08-28 in the ISO-8601 standard. As time nor- malization allows entities and events to be placed along a timeline, it is a crucial step for many informa- tion extraction tasks. Since the first shared tasks on time normalization (Verhagen et al., 2007), interest in the problem and the variety of applications have been growing. For example, Lin et al. (2015) use normal- ized timestamps from electronic medical records to contribute to patient monitoring and detect potential causes of disease. Vossen et al. (2016) identify multi- lingual occurrences of the same events in the news ∗These two authors contributed equally. by, among other steps, normalizing time-expressions in different languages with the same ISO standard. Fischer and Str¨otgen (2015) extract and normalize time-expressions from a large corpus of German fic- tion as the starting point of a deep study on trends and patterns of the use of dates in literature. A key consideration for time normalization sys- tems is what formal representation the time expres- sions should be normalized to. The most popular scheme for annotating normalized time expressions is ISO-TimeML (Pustejovsky et al., 2003a; Puste- jovsky et al., 2010), but it is unable to represent several important types of time expressions (e.g., a bounded set of intervals, like Saturdays since March 6) and it is not easily amenable to machine learning (the rule-based HeidelTime (Str¨otgen et al., 2013) still yields state-of-the-art performance). Bethard and Parker (2016) proposed an alternate scheme, Se- mantically Compositional Annotation of Time Ex- pressions (SCATE), in which times are annotated as compositional time entities (Figure 1), and suggested that this should be more amenable to machine learn- ing. However, while they constructed an annotated corpus, they did not train any automatic models on it. We present the first machine-learning models trained on the SCATE corpus of time normalizations. We make several contributions in the process: • We introduce a new evaluation metric for time normalization that can compare normalized times from different annotation schemes by mea- …

Content truncated to 500 words