TweeTime : A Minimally Supervised Method for Recognizing and Normalizing Time Expressions in Twitter

Proceedings of the 2016 Conference on Empirical Methods in Natural Language Processing, pages 307–318, Austin, Texas, November 1-5, 2016. c⃝2016 Association for Computational Linguistics TweeTime: A Minimally Supervised Method for Recognizing and Normalizing Time Expressions in Twitter Jeniya Tabassum, Alan Ritter and Wei Xu Computer Science and Engineering Ohio State University {bintejafar.1,ritter.1492,xu.1265}@osu.edu Abstract We describe TweeTIME, a temporal tagger for recognizing and normalizing time expres- sions in Twitter. Most previous work in so- cial media analysis has to rely on tempo- ral resolvers that are designed for well-edited text, and therefore suffer from reduced perfor- mance due to domain mismatch. We present a minimally supervised method that learns from large quantities of unlabeled data and requires no hand-engineered rules or hand-annotated training corpora. TweeTIME achieves 0.68 F1 score on the end-to-end task of resolving date expressions, outperforming a broad range of state-of-the-art systems.1 1 Introduction Temporal expressions are words or phrases that re- fer to dates, times or durations. Resolving time ex- pressions is an important task in information ex- traction (IE) that enables downstream applications such as calendars or timelines of events (Derczyn- ski and Gaizauskas, 2013; Do et al., 2012; Rit- ter et al., 2012; Ling and Weld, 2010), knowledge base population (Ji et al., 2011), information re- trieval (Alonso et al., 2007), automatically schedul- ing meetings from email and more. Previous work in this area has applied rule-based systems (Mani and Wilson, 2000; Bethard, 2013b; Chambers, 2013) or supervised machine learning on small collections of hand-annotated news documents (Angeli et al., 2012; Lee et al., 2014). 1Our code and data are publicly available at https:// github.com/jeniyat/TweeTime. Figure 1: A tweet published on Friday 5/6/2016 that contains the temporal expression Monday referring to the date of the event (5/9/2016), which a generic temporal tagger failed to re- solve correctly. Figure 2: A tweet that contains a simple explicit time mention and an event (Mercury, 5/9/2016) that can be identified by an open-domain information extraction system. Social media especially contains time-sensitive information and requires accurate temporal anal- ysis, for example, for detecting real-time cyber- security events (Ritter et al., 2015; Chang et al., 2016), disease outbreaks (Kanhabua et al., 2012) and extracting personal information (Schwartz et al., 2015). However, most work on social media simply uses generic temporal resolvers and therefore suffers from suboptimal performance. Recent work on tem- poral resolution focuses primarily on news articles and clinical texts (UzZaman et al., 2013; Bethard and Savova, 2016). Resolving time expressions in social media is a non-trivial problem. Besides many spelling varia- tions, time expressions are more likely to refer to future dates than in newswire. For the example in 307 Figure 1, we need to recognize that Monday refers to the upcoming Monday and not the previous one to resolve to its correct normalized date (5/9/2016). We also need to identify that the word Sun is not re- ferring to a Sunday in this context. In this paper, we present a new minimally super- vised approach to temporal resolution …

Content truncated to 500 words