Weasels, Hedges and Peacocks: Discourse-level Uncertainty in Wikipedia Articles

International Joint Conference on Natural Language Processing, pages 383–391, Nagoya, Japan, 14-18 October 2013. Weasels, Hedges and Peacocks: Discourse-level Uncertainty in Wikipedia Articles Veronika Vincze1,2 1Hungarian Academy of Sciences, Research Group on Artificial Intelligence 2Department of Informatics, University of Szeged vinczev@inf.u-szeged.hu Abstract Uncertainty is an important linguistic phenomenon that is relevant in many areas of language processing. While earlier research mostly concentrated on the semantic aspects of uncertainty, here we focus on discourse- and pragmatics- related aspects of uncertainty. We present a classification of such linguistic phenom- ena and introduce a corpus of Wikipedia articles in which the presented types of discourse-level uncertainty – weasel, hedge and peacock – have been manually annotated. We also discuss some exper- imental results on discourse-level uncer- tainty detection. 1 Introduction In many areas of natural language processing, it is essential to distinguish between factual and non- factual information. Thus, depending on the pre- cise task, negated or uncertain propositions should be treated separately by e.g. information extrac- tion systems or they should be neglected. For instance, in medicine, if it is uncertain whether the patient suffers from an illness, the doctor should undertake further examinations to determine the final diagnosis. In another case, only those pieces of news are relevant in news media that are true and come from a reliable source. Uncertain infor- mation or unreliable sources should not be part of the news. In order to be able to find uncertain propositions in a huge amount of texts, a reliable uncertainty detector is needed, which can be only developed if annotated resources are at hand. Previous studies on uncertainty detection con- centrated mostly on the semantic dimensions. Indeed, in many cases it is the lexical content (meaning) of the uncertainty marker (cue) that is responsible for uncertainty, i.e. it can be identified in texts with the help of semantic tools. However, there are other types of uncertainty which can- not be described by just concentrating on seman- tics. For instance, many may denote quite differ- ent approximations: in the sentence Many of the students did not read the book, many may signal about 60-70% of the students (or at least more than 50%), while in This airline loses many suitcases, many may be only 20% but this number is still high enough for passengers to call it many. Here, the context and world knowledge determine how the quantifier many should be interpreted. Here, we will focus on pragmatics- and discourse-related aspects of uncertainty. We will examine the concepts of source, fuzziness and subjectivity and their connection with uncertainty. As a first contribution, we will present a language- independent classification of such linguistic phe- nomena. As another contribution, we will intro- duce a corpus of Wikipedia articles in which lin- guistic cues of the presented types of discourse- level uncertainty have been manually annotated, hence empirical data on the frequency of such phe- nomena can also be provided. We will report the results of our experiments and we will also com- pare …

Content truncated to 500 words