Cross-lingual Argumentation Mining: Machine Translation (and a bit of Projection) is All You Need!

Proceedings of the 27th International Conference on Computational Linguistics, pages 831–844 Santa Fe, New Mexico, USA, August 20-26, 2018. 831 Cross-lingual Argumentation Mining: Machine Translation (and a bit of Projection) is All You Need! Steffen Eger, Johannes Daxenberger, Christian Stab, and Iryna Gurevych Ubiquitous Knowledge Processing Lab (UKP-TUDA) Department of Computer Science, Technische Universit¨at Darmstadt http://www.ukp.tu-darmstadt.de Abstract Argumentation mining (AM) requires the identification of complex discourse structures and has lately been applied with success monolingually. In this work, we show that the existing re- sources are, however, not adequate for assessing cross-lingual AM, due to their heterogeneity or lack of complexity. We therefore create suitable parallel corpora by (human and machine) translating a popular AM dataset consisting of persuasive student essays into German, French, Spanish, and Chinese. We then compare (i) annotation projection and (ii) bilingual word em- beddings based direct transfer strategies for cross-lingual AM, finding that the former performs considerably better and almost eliminates the loss from cross-lingual transfer. Moreover, we find that annotation projection works equally well when using either costly human or cheap machine translations. Our code and data are available at http://github.com/UKPLab/ coling2018-xling_argument_mining. 1 Introduction Argumentation mining (AM) is a fast-growing research field with applications in discourse analysis, summarization, debate modeling, and law, among others (Peldszus and Stede, 2013a). Recent studies have successfully applied computational methods to analyze monological argumentation (Wachsmuth et al., 2017; Eger et al., 2017). Most of these studies view arguments as consisting of (at least) claims and premises—and so do we in this work. Thereby, our focus is on token-level argument component extraction, that is, the segmentation and typing of argument components. AM has thus far almost exclusively been performed monolingually, e.g. in English (Mochales-Palau and Moens, 2009), German (Eckle-Kohler et al., 2015), or Chinese (Li et al., 2017). Working only mono- lingually is problematic, however, because AM is a difficult task even for humans due to its dependence on background knowledge and parsing of complex pragmatic relations (Moens, 2017). As a result, ac- quiring (high-quality) datasets for new languages comes at a high cost—be it in terms of training and/or hiring expert annotators or querying large crowds in crowd-sourcing experiments. It is thus of utmost importance to train NLP systems in AM that are capable of going cross-language, so that annotation efforts do not have to be multiplied by the number of languages of interest. This is in line with current trends in NLP, which increasingly recognize the possibility and the necessity to work cross-lingually, be it in part-of-speech tagging (Zhang et al., 2016), dependency parsing (Agic et al., 2016), sentiment mining (Chen et al., 2016; Zhou et al., 2016), or other fields. In this work, we address the problem of cross-lingual (token-level) AM for the first time. We initially experiment with available resources in English, German, and Chinese. We show that the existing datasets for analyzing argumentation are not suitable for assessing cross-lingual component extraction due to their heterogeneity or lack of complexity. Given this scarcity of homogeneously annotated …

Content truncated to 500 words