Inducing Discourse Marker Inventories from Lexical Knowledge Graphs

Proceedings of the 13th Conference on Language Resources and Evaluation (LREC 2022), pages 2401–2412 Marseille, 20-25 June 2022 © European Language Resources Association (ELRA), licensed under CC-BY-NC-4.0 2401 Inducing Discourse Marker Inventories from Lexical Knowledge Graphs Christian Chiarcos Applied Computational Linguistics (ACoLi), Goethe University Frankfurt, Germany Institute for Digital Humanities (IDH), University of Cologne, Germany chiarcos@cs.uni-frankfurt.de Abstract Discourse marker inventories are lexical resources that define the meaning of discourse cues (discourse markers) in terms of associated discourse relation types. They are thus important tools for the development of both discourse parsers and corpora with discourse annotations. This paper explores the potential of massively multilingual lexical knowledge graphs to induce multilingual discourse marker lexicons by means of propagation methods. Given one or multiple source language discourse marker inventories and a large number of bilingual dictionaries to link them – directly or indirectly – with the target language, we study to what extent discourse marker induction can benefit from the integration of information from different sources, the impact of sense granularity and what limiting factors may need to be considered. Our study uses discourse marker inventories from nine European languages normalized against the discourse relation inventory of the Penn Discourse Treebank (PDTB), as well as three collections of machine-readable dictionaries with different characteristics, so that the interplay of a large number of factors can be studied. Keywords: discourse marker, lexical knowledge graphs, lexical induction, OntoLex 1. Background Discourse parsing has been a topic that received con- siderable renewed attention in the last years. The area does, however, still suffer from a general sparsity of resources. At the moment, we are aware of discourse corpora for little more than a dozen languages, only, including English (Carlson et al., 2003; Wolf and Gib- son, 2005; Prasad et al., 2008), Chinese (Huang and Chen, 2011; Long et al., 2020), Czech (Pol´akov´a et al., 2013), Dutch (Van Der Vliet et al., 2011), French (Dan- los et al., 2015), German (Stede, 2004; Gastel et al., 2011), Hindi (Oza et al., 2009), Italian (Raymond et al., 2007), Portuguese (Pardo and Nunes, 2004), Spanish (Da Cunha et al., 2011), and a small number of cross- linguistic discourse annotations, e.g. Buch-Kromann and Korzen (2010) and Zeyrek et al. (2019). While for most languages, full-fledged discourse pars- ing is thus beyond reach, it is well-known that al- ready more shallow techniques can be employed to fa- cilitate the development of discourse-aware technolo- gies. One such technique is the application of dis- course marker inventories, i.e., machine-readable dic- tionaries of discourse cues (‘markers’) such as adverbs, conjunctions and particular phrases, classified accord- ing to fine-grained sense inventories that represent their discourse functions in terms of associated discourse re- lations. These can then be used as gazetteers to de- termine possible discourse functions of the associated marker in a text. A plain lookup may thus provide in- formation about the discourse relation an utterance is subject to, or, at least, to narrow their possible band- width, and this information can either be directly …

Content truncated to 500 words