Uncertainty Detection in Hungarian Texts

Proceedings of COLING 2014, the 25th International Conference on Computational Linguistics: Technical Papers, pages 1844–1853, Dublin, Ireland, August 23-29 2014. Uncertainty Detection in Hungarian Texts Veronika Vincze1,2 1University of Szeged Department of Informatics 2MTA-SZTE Research Group on Artificial Intelligence vinczev@inf.u-szeged.hu Abstract Uncertainty detection is essential for many NLP applications. For instance, in information re- trieval, it is of primary importance to distinguish among factual, negated and uncertain informa- tion. Current research on uncertainty detection has mostly focused on the English language, in contrast, here we present the first machine learning algorithm that aims at identifying linguistic markers of uncertainty in Hungarian texts from two domains: Wikipedia and news media. The system is based on sequence labeling and makes use of a rich feature set including orthographic, lexical, morphological, syntactic and semantic features as well. Having access to annotated data from two domains, we also focus on the domain specificities of uncertainty detection by compar- ing results obtained in indomain and cross-domain settings. Our results show that the domain of the text has significant influence on uncertainty detection. 1 Introduction Uncertainty detection has become one of the most intensively studied problems of natural language pro- cessing (NLP) in these days (Morante and Sporleder, 2012). For several NLP applications, it is essential to distinguish between factual and nonfactual, i.e. negated or uncertain information: for instance, in medical information retrieval, it must be known whether the patient definitely suffers, probably suffers or does not suffer from an illness. This type of information can only be revealed from the texts of the documents if reliable uncertainty detectors are available, which are able to identify linguistic markers of uncertainty, i.e. cues within the text. To the best of our knowledge, uncertainty detectors have been mostly developed for the English language (Morante and Sporleder, 2012; Farkas et al., 2010). Here, we present our machine learning based uncertainty detector developed for Hungarian, a morphologically rich language, and report our results on a manually annotated uncertainty corpus, which contains texts from two domains: first, Hungarian Wikipedia texts and second, pieces of news from a Hungarian news portal. The main contributions of this paper are the following: • it presents the first uncertainty corpus for Hungarian; • it reports the first results on uncertainty detection in Hungarian texts; • it introduces new features in the machine learning setting like semantic and pragmatic features; • we show that there are domain specificities in the distribution of uncertainty cues in Hungarian texts; • we show that domain specificities have a considerable effect on the efficiency of machine learning. The structure of the paper is the following. First, related work on uncertainty detection is presented. Then our corpus is described in detail, which is followed by the elaboration of machine learning methods and results on uncertainty detection. The paper concludes with a discussion of results and possible ways for future work are also outlined. This work is licenced under a Creative Commons Attribution 4.0 International License. Page numbers and proceedings footer are …

Content truncated to 500 words