Annotation and Data Mining of the Penn Discourse TreeBank

Annotation and Data Mining of the Penn Discourse TreeBank Rashmi Prasad University of Pennsylvania Philadelphia, PA 19104 USA rjprasad@linc.cis.upenn.edu Eleni Miltsakaki University of Pennsylvania Philadelphia, PA 19104 USA elenimi@linc.cis.upenn.edu Aravind Joshi University of Pennsylvania Philadelphia, PA 19104 USA joshi@linc.cis.upenn.edu Bonnie Webber University of Edinburgh Edinburgh, EH8 9LW Scotland bonnie@inf.ed.ac.uk Abstract The Penn Discourse TreeBank (PDTB) is a new re- source built on top of the Penn Wall Street Journal corpus, in which discourse connectives are anno- tated along with their arguments. Its use of stand- off annotation allows integration with a stand-off version of the Penn TreeBank (syntactic structure) and PropBank (verbs and their arguments), which adds value for both linguistic discovery and dis- course modeling. Here we describe the PDTB and some experiments in linguistic discovery based on the PDTB alone, as well as on the linked PTB and PDTB corpora. 1 Introduction Large scale annotated corpora such as the Penn TreeBank (Marcus et al., 1993) have played a cen- tral role in speech and natural language research. However, with the demand for more powerful NLP applications comes a need for greater richness in annotation – hence, the development of PropBank (Kingsbury and Palmer, 2002), which adds basic se- mantics to the PTB in the form of verb predicate- argument annotation and eventually similar annota- tion of nominalizations. We have been developing yet another annotation layer above these both. The Penn Discourse TreeBank (PDTB) adds low-level discourse structure and semantics through the anno- tation of discourse connectives and their arguments, using connective-specific semantic role labels. With this added knowledge, the PDTB (together with the PTB and PropBank) should support more in-depth NLP research and more powerful applications. Work on the PDTB is grounded in a lexical- ized approach to discourse – DLTAG (Webber and Joshi, 1998; Webber et al., 1999a; Webber et al., 2000; Webber et al., 2003). Here, low-level dis- course structure and semantics are taken to re- sult (in part) from composing elementary predicate- argument relations whose predicates come mainly from discourse connectives1 and whose arguments 1Despite this, we have deliberately adopted a policy of hav- come from units of discourse – clausal, sentential or multi-sentential units. The PDTB therefore dif- fers from the RST-annotated corpus (Carlson et al., 2003) which starts with (abstract) rhetorical rela- tions (Mann and Thompson, 1988) and annotates a subset of the Penn WSJ corpus with those relations that can be taken to hold between (primarily) pairs of discourse spans identified in the corpus. The current paper focuses on what can be dis- covered through analyzing PDTB annotation, both on its own and together with the Penn TreeBank. Section 2 of the paper briefly reviews the theo- retical background of the project, its current state, the guidelines given to annotators, the annotation tool they used (WordFreak), and the extent of inter- annotator agreement. Section 3 shows how we have used PDTB annotation, along with the PTB, to ex- tract several features pertaining to discourse con- nectives and their arguments, and discusses the …

Content truncated to 500 words