Annotation and Automatic Classification of Aspectual Categories

Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, pages 3335–3341 Florence, Italy, July 28 - August 2, 2019. c⃝2019 Association for Computational Linguistics 3335 Annotation and Automatic Classification of Aspectual Categories Markus Egg Helena Prepens Humboldt-Universität zu Berlin Berlin, Germany {markus.egg,helena.prepens,will.roberts}@hu-berlin.de Will Roberts Abstract We present the first annotated resource for the aspectual classification of German verb to- kens in their clausal context. We use aspec- tual features compatible with the plurality of aspectual classifications in previous work and treat aspectual ambiguity systematically. We evaluate our corpus by using it to train super- vised classifiers to automatically assign aspec- tual categories to verbs in context, permitting favourable comparisons to previous work. 1 Introduction The universal linguistic category of aspect de- scribes how a verb or a verbal projection (including sentences, ‘predicates’ for short) characterises the temporal course of a state of affairs or ‘eventuality’. Such information is relevant for tasks that extract temporal information from texts, such as informa- tion extraction, question answering, and document summarisation (Costa and Branco, 2012). Further tasks in which aspectual information plays a cru- cial role include computational semantic analysis (Caselli and Quochi, 2007), zoning (analysing the argumentative and rhetorical structure of texts; Ba- iamonte et al. 2016), and the analysis of specific textual elements, e.g., captions (Alikhani and Stone, 2019). Aspect must also be considered in event annota- tion (Pustejovsky et al., 2010; Bittar et al., 2011; Caselli et al., 2011). Aspect is a universal semantic category; thus, the same aspectual patterns reap- pear across languages. We created the first resource of German verbs annotated for aspectual class in context. We use aspectual features compatible with various differ- ent previously published aspectual classifications, and model the pervasive phenomenon of aspectual ambiguity. We evaluate the resource by using it in supervised aspectual classifiers for verbs in context. 2 Aspectual classes and ambiguity Aspectual classes are established by feature di- chotomies (Vendler, 1967; Moens and Steedman, 1988; Egg, 2005). First, stative predicates describe purely static situations (e.g., be happy or love); dynamic ones introduce eventualities with develop- ment (e.g., continuous change of place in move). Dynamic predicates can be either unbounded (in- troduce eventualities without inherent boundaries, e.g., move or play the piano), or bounded (e.g., run a mile or build a house). Bounded predicates (also called ‘telic’) have four subgroups that are cross- classified by the features change - no change and punctual - extended: The first pair distinguishes predicates that express an explicit change of state (e.g., leave as change from being present to be- ing away) from predicates that do not (e.g., play a sonata).1 The second pair distinguishes e.g. the no-change predicates cough and play a sonata or the change predicates explode and build a house. The punctual - extended distinction is gradual (while the others are binary). This will tend to aggravate both the annotation and the automatic classification of aspect. These features define six aspectual classes: Only dynamic predicates can be bounded or not, and only …

Content truncated to 500 words