Unsupervised extraction of semantic relations using discourse cues

Proceedings of COLING 2014, the 25th International Conference on Computational Linguistics: Technical Papers, pages 2184–2194, Dublin, Ireland, August 23-29 2014. Unsupervised extraction of semantic relations using discourse cues Juliette Conrath Stergos Afantenos Nicholas Asher Philippe Muller IRIT, Université Toulouse & CNRS, Univ. Paul Sabatier, 118 Route de Narbonne, 31062 Toulouse {firstname.lastname@irit.fr} Abstract This paper presents a knowledge base containing triples involving pairs of verbs associated with semantic or discourse relations. The relations in these triples are marked by discourse connectors between two adjacent instances of the verbs in the triple in the large French corpus, frWaC. We detail several measures that evaluate the relevance of the triples and the strength of their association. We use manual annotations to evaluate our method, and also study the coverage of our resource with respect to the discourse annotated corpus Annodis. Our positive results show the potential impact of our resource for discourse analysis tasks as well as other semantically oriented tasks like temporal and causal information extraction. 1 Introduction Relational lexical resources, which describe semantic relations between lexical items, have tradition- ally focused on relations like synonymy or similarity in thesauri, perhaps including some hierarchical semantic relations like hyperonymy or hyponomy or part-whole relations as in the resource Wordnet (Fel- baum, 1998). Some distributional thesauri contain more varied relations, see e.g. (Grefenstette, 1994), however these relations are not typed. The lexical semantics given by FrameNet (Baker et al., 1998) does include causal and temporal relations, as does Verbocean (Chklovski and Pantel, 2004), but coverage is limited and empirical validation of these resources is partial and still largely remains to be done. Lexical relations, in particular between verbs, are nevertheless crucial for understanding natural lan- guage and for many information processing tasks. They are needed for textual inference, in which one has to infer certain relations between eventualities (Hashimoto et al., 2009; Tremper and Frank, 2013), for information extraction tasks, like finding temporal relations between eventualities mentioned in a text (UzZaman et al., 2013), for automatic summarization (Liu et al., 2007), and for discourse parsing in the absence of explicit discourse markers (Sporleder and Lascarides, 2008). In this paper we report on our efforts to extract semantic relations essential to the analysis of discourse and its interpretation, in which links are made between units of text or rather their semantic representa- tions as in (1) in virtue of semantic information about the two main verbs of those clauses. (1) The candidate demonstrated his expertise during the interview. The committee was completely convinced. We follow similar work on the extraction of causal, temporal, entailment and presuppositional relations from corpora (Do et al., 2011; Chambers and Jurafsky, 2008; Hashimoto et al., 2009; Tremper and Frank, 2013), though our goals and validation methods are different. While one of our goals is to use this information to improve performance in predicting discourse relations between clauses, we believe that such a lexical resource will have other uses in other tasks in which semantic information is needed. Discourse analysis is …

Content truncated to 500 words