Exploiting a lexical resource for discourse connective disambiguation in German

Proceedings of the 28th International Conference on Computational Linguistics, pages 5737–5748 Barcelona, Spain (Online), December 8-13, 2020 5737 Exploiting a lexical resource for discourse connective disambiguation in German Peter Bourgonje and Manfred Stede Applied Computational Linguistics Universit¨at Potsdam Potsdam, Germany {bourgonje|stede}@uni.potsdam.de Abstract In this paper we focus on connective identification and sense classification for explicit discourse relations in German, as two individual sub-tasks of the overarching Shallow Discourse Parsing task. We successively augment a purely-empirical approach based on contextualised embeddings with linguistic knowledge encoded in a connective lexicon. In this way, we improve over pub- lished results for connective identification, achieving a final F1-score of 87.93; and we introduce, to the best of our knowledge, first results for German sense classification, achieving an F1-score of 87.13. Our approach demonstrates that a connective lexicon can be a valuable resource for those languages that do not have a large PDTB-style-annotated coprus available. 1 Introduction An important difference between a text and a random collection of sentences is the amount of coherence it exhibits. In a text, sentences, or propositions therein, are connected through particular relations that can be, for example, causal, temporal or contrastive. Such relations can be obvious from the semantics of the involved text segments alone, as in (1), where the contrastive relation between male and female worker’s earnings is easily inferred. Or they can be explicitly signaled by (discourse) connectives. Connectives are usually understood to be a closed class of syntactically heterogeneous words and phrases, and are known to be ambiguous in two different ways. In (2), there are no two particular propositions that are related to each other by the potential connective once, thus it is said to have sentential reading, as opposed to its discourse reading in (3) and (4). Furthermore, once is ambiguous with regard to the particular relations it can express, and signals a temporal relation in (3) and a conditional relation in (4).1 (1) Earnings of year-round, full-time male workers fell 1.3% in 1988 after adjusting for higher prices. Earnings of female workers were unchanged. (wsj 1815) (2) Once again the company’s future looked less than rosy. (wsj 0564) (3) Once it gets there, a company can do with it what it wishes. (wsj 0989) (4) Normally, once the underlying investment is suspended from trading, the options on those invest- ments also don’t trade. (wsj 1962) Differentiating between sentential and discourse reading is often referred to as connective identifica- tion, and classifying the particular sense of a connective, or the relation it is involved in, is often referred to as sense classification. Both are sub-tasks of discourse parsing, which in turn has applications, for example, in text summarisation (Schilder, 2002; Yoshida et al., 2014), machine translation (Meyer and Popescu-Belis, 2012; Joty et al., 2014; Sim Smith, 2017) and argumentation mining (Eckle-Kohler et al., 2015). The availability of annotated data for the task of discourse parsing as a whole, and consequently its sub-tasks, is limited. With the PDTB being by far the largest corpus annotated …

Content truncated to 500 words