Comprehensive Annotation of Various Types of Temporal Information on the Time Axis

Comprehensive Annotation of Various Types of Temporal Information on the Time Axis Tomohiro Sakaguchi, Daisuke Kawahara and Sadao Kurohashi Graduate School of Informatics, Kyoto University Yoshida-honmachi, Sakyo-ku, Kyoto, 606-8501, Japan sakaguchi@nlp.ist.i.kyoto-u.ac.jp, {dk, kuro}@i.kyoto-u.ac.jp Abstract In order to make the temporal interpretation of text, there have been many studies linking event and temporal information, such as temporal ordering of events and timeline generation. To train and evaluate models in these studies, many corpora that associate event information with time information have been developed. In this paper, we propose an annotation scheme that anchors expressions in text to the time axis comprehensively, extending the previous studies in the following two points. One of the points is to annotate not only expressions with strong temporality but also expressions with weak temporality, such as states and habits. The other point is that various types of temporal information, such as frequency and duration, can be anchored to the time axis. Using this annotation scheme, we annotated a subset of Kyoto University Text Corpus. Since the corpus has already been annotated predicate-argument structures and coreference relations, it can be utilized for integrated information analysis of events, entities and time. Keywords: Time annotation, Time anchoring, Time axis 1. Introduction Everyday many texts are generated on the Web, and a huge amount of texts have been accumulated so far. To extract knowledge about a certain topic from this large amount of texts, we need an information analysis technology to inte- grate, summarize and compare related texts. In order to analyze texts written at different times or texts referring to different times, it is necessary to interpret the temporal information implied in the texts. There have been many studies and tasks to understand the relationship between event information and time information in text. For exam- ple, temporal ordering of events that estimates the temporal relations of event-event and event-time was studied in Tem- pEval 1, 2, 3 (Verhagen et al., 2007; Verhagen et al., 2010; UzZaman et al., 2013), and the timeline generation task that links event and time in multiple documents was studied in SemEval 15 (Minard et al., 2015). In order to train models and evaluate results in these tasks, corpora in which event information is correlated with time information in text have been developed (Pustejovsky et al., 2003; Cassidy et al., 2014; Reimers et al., 2016). In these studies, expressions which have clear temporality were an- notated, but in order to know how people understand texts from the perspective of time, it is essential to know how the expressions with weak temporality are interpreted. To un- derstand temporal information in text exhaustively, we pro- pose an annotation scheme that anchors various expressions to the time axis, reflecting personal interpretation of text and common sense. Using this scheme, we annotate Ky- oto University Text Corpus (Kawahara et al., 2002), which is a Japanese newspaper corpus annotated with predicate- argument structures and coreference relations. The points of our annotation scheme are two-fold. One of the points is to …

Content truncated to 500 words