Discourse Annotation and Semantic Annotation in the GNOME corpus

Discourse Annotation and Semantic Annotation in the GNOME Corpus Massimo Poesio University of Essex, Department of Computer Science and Centre for Cognitive Science, United Kingdom Abstract The GNOME corpus was created to study the dis- course and semantic properties of discourse entities that affect their realization and interpretation, and particularly salience. We discuss what information was annotated and the methods we followed. 1 Introduction The GNOME corpus was created to study the as- pects of discourse that appear to affect generation, especially salience (Pearson et al., 2000; Poesio and Di Eugenio, 2001; Poesio and Nissim, 2001; Poesio et al., 2004b). Particular attention was paid to the factors affecting the generation of pronouns (Pear- son et al., 2000; Henschel et al., 2000), demon- stratives (Poesio and Nygren-Modjeska, To appear) possessives (Poesio and Nissim, 2001) and definites in general (Poesio, 2004a). These results, and the annotated corpus, were used in the development of both symbolic and statistical natural language gen- eration algorithms for sentence planning (Poesio, 2000a; Henschel et al., 2000; Cheng et al., 2001), aggregation (Cheng, 2001) and text planning (Kara- manis, 2003). The empirical side of the project in- volved both psychological experiments and corpus annotation, based on a scheme based on the MATE proposals, as well as on a detailed annotation man- ual (Poesio, 2000b), the reliability of whose instruc- tions was tested by extensive experiments (Poe- sio, 2000a). More recently, the corpus has also been used to develop and evaluate anaphora resolu- tion systems, with a special focus on the resolution of bridging references (Poesio, 2003; Poesio and Alexandrov-Kabadjov, 2004; Poesio et al., 2004a) Although the results of the studies using the GNOME corpus mentioned above have been pub- lished in a number of papers, and although a de- tailed annotation manual was written and has been available on the Web for a few years (Poesio, 2000b), none of the previously published papers dis- cusses in detail the goals of the annotation and the methodology that was followed, especially for the non-anaphoric aspects. In this paper we discuss the methods used to identify possible ‘utterances,’ the properties of NPs and discourse entities that were annotated, and (very briefly) anaphoric information. . 2 The Data Texts from three domains were (partially) anno- tated. The museum subcorpus consists of descrip- tions of museum objects and brief texts about the artists that produced them.1 The pharmaceutical subcorpus is a selection of leaflets providing the patients with legally mandatory information about their medicine.2 The GNOME corpus also includes tutorial dialogues from the Sherlock corpus col- lected at the University of Pittsburgh. Each sub- corpus contains about 6,000 NPs, but not all types of annotation have been completed for all domains. All sentences, units and NPs have been identified, and all ‘syntactic’ properties of NPs (agreement fea- ture and grammatical function). Anaphoric rela- tions have been annotated in about half of the texts in each domain; and the more complex semantic properties (taxonomic properties, genericity, etc.) in about 25% of these texts. The …

Content truncated to 500 words