Multilabel Tagging of Discourse Relations in Ambiguous Temporal Connectives

Proceedings of Recent Advances in Natural Language Processing, pages 154–161, Hissar, Bulgaria, 12-14 September 2011. Multilabel Tagging of Discourse Relations in Ambiguous Temporal Connectives Yannick Versley Collaborative Research Centre (SFB) 833 University of T¨ubingen versley@sfs.uni-tuebingen.de Abstract Many annotation schemes for discourse relations allow combinations such as tem- poral+cause (for events that are tempo- rally and causally related to each other) and temporal+contrast (for contrasts be- tween subsequent time spans, or between events that are temporally coextensive). However, current approaches for the auto- matic classification of discourse relations are limited to producing only one relation and disregard the others. We argue that the information contained in these ‘additional’ relations is indeed use- ful and present an approach to tag mul- tiple fine-grained discourse relations in ambiguous connectives from the German T¨uBa-D/Z corpus. Using a rich feature set, we show that good accuracy is possi- ble even for inferred relations that are not part of the connective’s ‘core’ meaning. 1 Introduction In order to account for the structure of text beyond the level of single clauses, it is common to pos- tulate discourse relations holding between clauses or groups of clauses. Discourse relations are fre- quently marked by connectives such as because, as or while, which give an indication both of (syn- tactic or anaphoric) linking possibilities for the spans and of the possible relations. Many connectives (such as because or for in- stance) always signal one specific discourse rela- tion. This fact has, after initial successes in purely structural discourse parsing (Soricut and Marcu, 2003), led to decreased attention from researchers. Other connectives, however, are ambiguous be- tween multiple readings and their disambiguation necessitates similar semantic information as im- plicit (connective-less) discourse relations. Ambiguous temporal markers such as after, as or while usually occur with a purely temporal reading, but also with additional non-temporal dis- course relations, such as causal and contrastive readings. When these non-temporal relations oc- cur instead of, or in addition to, the temporal read- ing, they require similar similar inferences from the reader as in connective-less discourse rela- tions, but may be easier to detect automatically. For our goal of accurate classification, multilabel classification becomes necessary when the non- temporal discourse relations co-occur with the temporal ones: (1) a. As [arg2 individual investors have turned away from the stock market over the years], [arg1 securities firms have scrambled to find new products that brokers find easy to sell]. b. [arg1 “Forget it,” he said] as [arg2he handed her a paper]. c. But as [arg2 the French embody a Zen- like state of blase when it comes to athletics] (try finding a Nautilus ma- chine in Paris), [arg1 my fellow con- ventioners were having none of it]. In the examples from (1), the sentence in (b) is clearly temporal (and non-causal), and the one in (c) is clearly causal (and non-temporal), whereas in (a) the connective contributes both a causal and a temporal aspect to the coherence of the text. In the Penn Discourse Treebank (Prasad et al., …

Content truncated to 500 words