Extraction and Normalization of Vague Time Expressions in German

Extraction and Normalization of Vague Time Expressions in German Ulrike May1, 2, Karolina Zaczynska1, Julián Moreno-Schneider1, and Georg Rehm1 1DFKI Gmbh, Berlin, Germany 2Linguistics Department, University of Potsdam, Germany ulrike.may@hotmail.de, {karolina.zaczynska, julian.moreno_schneider, georg.rehm}@dfki.de Abstract Existing datasets and methods that aim at the identification of time expressions in natural language text do not pay particular attention to expressions that are imprecise and that can- not be easily represented on a timeline. We call these vague time expressions (VTEs). We present an analysis of existing time extraction approaches and steps towards a novel scheme for the annotation of VTEs, developed using a corpus of German news articles. To the best of our knowledge, this work is the first to sug- gest an extension of the ISO standard TimeML with the goal of enabling the annotation of VTEs. In addition, we present a collection of 339 German VTEs as well as classification experiments on the news corpus with results from 60 up to 77 macro-avg. F1 score. 1 Introduction Time is critical to the meaning of language, for un- derstanding events, cause-effect relations and narra- tives, etc. In NLP, temporal expression analysis is a key issue which has been receiving a lot of attention in recent years, especially for English news texts (e. g., UzZaman et al., 2013; Caselli and Vossen, 2017; Strötgen et al., 2018). Existing approaches mostly focus on time expressions which can be more or less easily and specifically represented on a timeline with an accuracy of different granularity levels. Given an expression like at 6 ’o clock, the hour can be pinpointed, this Sunday or tomorrow refer to a specific day. Such time expressions can be annotated using standardized machine-readable expressions. This process is typically referred to as normalization. Nevertheless, there is a large proportion of time expressions that are inherently vague – they are neither exact nor precise and, i. e., they cannot be readily normalized. Examples for vague time ex- pressions (VTEs) are in the future or lately. VTEs are typically not taken into account by existing an- notation schemes; some simply normalize them as a reference to the past or future. According to Tissot et al. (2019), around 13% of time expres- sions in news articles can be considered vague. In our German corpus, almost 30% of all time ex- pressions are VTEs (Section 3). The annotation of time expressions in TimeML (Saurí et al., 2006), arguably the most well known scheme, is possible only if we can fully and precisely interpret the ex- pression (Mazur and Dale, 2011). As VTEs cannot be fully and precisely interpreted, we are unable to represent them using TimeML, which is why an an- notation scheme needs to be developed that is able to capture VTEs. Given that a large number of time expressions tend to be overlooked or oversimpli- fied, our goal is the annotation and normalization of VTEs by extending TimeML; we concentrate on German documents and the German extension for TimeML. According to our research, no corpora …

Content truncated to 500 words