Learning Information Status of Discourse Entities

Proceedings of the 2006 Conference on Empirical Methods in Natural Language Processing (EMNLP 2006), pages 94–102, Sydney, July 2006. c⃝2006 Association for Computational Linguistics Learning Information Status of Discourse Entities Malvina Nissim∗ Laboratory for Applied Ontology Institute for Cognitive Science and Technology National Research Council (ISTC-CNR), Roma, Italy malvina.nissim@loa-cnr.it Abstract In this paper we address the issue of au- tomatically assigning information status to discourse entities. Using an annotated cor- pus of conversational English and exploit- ing morpho-syntactic and lexical features, we train a decision tree to classify entities introduced by noun phrases as old, medi- ated, or new. We compare its performance with hand-crafted rules that are mainly based on morpho-syntactic features and closely relate to the guidelines that had been used for the manual annotation. The decision tree model achieves an overall ac- curacy of 79.5%, significantly outperform- ing the hand-crafted algorithm (64.4%). We also experiment with binary classifica- tions by collapsing in turn two of the three target classes into one and retraining the model. The highest accuracy achieved on binary classification is 93.1%. 1 Introduction Information structure is the way a speaker or writer organises known and new information in text or dialogue. Information structure has been the subject of numerous and very diverse linguistic studies (Halliday, 1976; Prince, 1981; Hajiˇcov´a, 1984; Vallduv´ı, 1992; Lambrecht, 1994; Steed- man, 2000, for instance), thus also yielding a wide range of terms and definitions (see (Vallduv´ı, ∗The work reported in this paper was carried out while the author was a research fellow at the Institute for Com- municating and Collaborative Systems of the University of Edinburgh, United Kingdom, and was supported by a Scottish Enterprise Edinburgh-Stanford Link grant (265000- 3102-R36766). 1992; Kruijff-Korbayov´a and Steedman, 2003) for a discussion). In the present study, we adopt the term “Information Status”, following the defini- tion employed for the annotation of the corpus we use for our experiments (Nissim et al., 2004). In- formation status describes to which degree a dis- course entity is available to the hearer, in terms of the speaker’s assumptions about the hearer’s knowledge and beliefs. Although there is a fine line in the distinction between Information Sta- tus and Information Structure, it is fair to say that whereas the latter models wider discourse coher- ence, the former focuses mainly on the local level of discourse entities. Section 2 provides more de- tails on how this notion is encoded in our corpus. Information status has generated large interest among researchers because of its complex interac- tion with other linguistic phenomena, thus affect- ing several Natural Language Processing tasks. Since it correlates with word order and pitch ac- cent (Lambrecht, 1994; Hirschberg and Nakatani, 1996), for instance, incorporating knowledge on information status would be helpful for natural language generation, and in particular text-to- speech systems. St¨ober and colleagues, for ex- ample, ascribe to the lack of such information the lower performance of text-to-speech compared to concept-to-speech generation, where such knowl- edge could be made directly available to the sys- tem …

Content truncated to 500 words