Disambiguating Explicit Discourse Connectives without Oracles

International Joint Conference on Natural Language Processing, pages 997–1001, Nagoya, Japan, 14-18 October 2013. Disambiguating explicit discourse connectives without oracles Anders Johannsen Anders Søgaard University of Copenhagen {ajohannsen,soegaard}@hum.ku.dk Abstract Deciding whether a word serves a dis- course function in context is a prerequi- site for discourse processing, and the per- formance of this subtask bounds perfor- mance on subsequent tasks. Pitler and Nenkova (2009) report 96.29% accuracy (F1 94.19%) relying on features extracted from gold-standard parse trees. This fig- ure is an average over several connectives, some of which are extremely hard to clas- sify. More importantly, performance drops considerably in the absence of an ora- cle providing gold-standard features. We show that a very simple model using only lexical and predicted part-of-speech fea- tures actually performs slightly better than Pitler and Nenkova (2009) and not sig- nificantly different from a state-of-the-art model, which combines lexical, part-of- speech, and parse features. 1 Introduction Discourse relations structure text by linking seg- ments together in functional relationships. For in- stance, someone might say ”Saber-toothed tigers are harmless because they’re extinct”, making the second part of the sentence serve as an explana- tion for the first part. In the example the discourse connective because functions as a lexical anchor for the discourse relation. Whenever an anchor is present we say that the discourse connective is ex- plicit. Complicating the matter, phrases used as dis- course connectives sometimes appear in a non- discourse function. For instance, ”and” may be either a simple conjunction, as in ”sugar and salt”, or a discourse relation suggesting a temporal re- lationship between events, for instance ”he struck the match and went away”. The Penn Discourse 0 20 40 60 80 100 No. of markers 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1.0 Cumulative occurences Figure 1: A picture of the problem. 10% of con- nectives account for roughly 75% of occurrences Treebank (PDTB) (Prasad et al., 2008) distinguish 100 types of explicit connectives—a subset of these are listed in Table 2. The type of relation- ship is selected from a hierachial structure where the four top-level categories are Comparison, Con- tingency, Temporal, and Expansion. Discourse relations are important for many ap- plications and, since the PDTB was released, much effort has gone into developing tools for recreating the annotations of the resource automat- ically. Recently two ambitious end-to-end parsers have appeared which transform plain text to full PDTB-style annotations (Lin et al., 2010; Ghosh, 2012). Both systems share a pipelined architecture in which the output of one component becomes the input to the next. A crucial first step in their pro- cessing is correctly identifying explicit discourse connectives; when unsuccessful subsequent steps fail. An accuracy in the high ninetees seems to sug- gest that the problem is almost solved. For the task of discourse connective disambiguation this unfor- tunately does not hold true, because, as we argue here, the task benefits from being seen and eval- uated as a number of smaller tasks, one for each …

Content truncated to 500 words