Trainable Speaker-Based Referring Expression Generation

CoNLL 2008: Proceedings of the 12th Conference on Computational Natural Language Learning, pages 151–158 Manchester, August 2008 Trainable Speaker-Based Referring Expression Generation Giuseppe Di Fabbrizio and Amanda J. Stent and Srinivas Bangalore AT&T Labs - Research, Inc. 180 Park Avenue Florham Park, NJ 07932, USA {pino,stent,srini}@research.att.com Abstract Previous work in referring expression gen- eration has explored general purpose tech- niques for attribute selection and surface realization. However, most of this work did not take into account: a) stylistic dif- ferences between speakers; or b) trainable surface realization approaches that com- bine semantic and word order information. In this paper we describe and evaluate sev- eral end-to-end referring expression gener- ation algorithms that take into considera- tion speaker style and use data-driven sur- face realization techniques. 1 Introduction Natural language generation (NLG) systems have typically decomposed the problem of generating a linguistic expression from a conceptual specifi- cation into three major steps: content planning, text planning and surface realization (Reiter and Dale, 2000). The task in content planning is to select the information that is to be conveyed to maximize communication efficiency. The task in text planning and surface realization is to use the available linguistic resources (words and syntax) to convey the selected information using well-formed linguistic expressions. During a discourse (whether written or spoken, monolog or dialog), a number of entities are in- troduced into the discourse context shared by the reader/hearer and the writer/speaker. Construct- ing linguistic references to these entities efficiently and effectively is a problem that touches on all c⃝ 2008. Licensed under the Creative Commons Attribution-Noncommercial-Share Alike 3.0 Unported li- cense (http://creativecommons.org/licenses/by-nc-sa/3.0/). Some rights reserved. parts of an NLG system. Traditionally, this prob- lem is split into two parts. The task of selecting the attributes to use in referring to an entity is the attribute selection task, performed during content planning or sentence planning. The actual con- struction of the referring expression is part of sur- face realization. There now exist numerous general-purpose al- gorithms for attribute selection (e.g., (Dale and Re- iter, 1995; Krahmer et al., 2003; Belz and Gatt, 2007; Siddharthan and Copestake, 2004)). How- ever, these algorithms by-and-large focus on the algorithmic aspects of referring expression gener- ation rather than on psycholinguistic factors that influence language production. For example, we know that humans exhibit individual differences in language production that can be quite pronounced (e.g. (Belz, 2007)). We also know that the language production process is subject to lexical priming, which means that words and concepts that have been used recently are likely to appear again (Levelt, 1989). In this paper, we look at attribute selection and surface realization for referring expression gener- ation using the TUNA corpus 1, an annotated cor- pus of human-produced referring expressions that describe furniture and people. We first explore the impact of individual style and priming on at- tribute selection for referring expression genera- tion. To get an idea of the potential improvement when modeling these factors, we implemented a version of full brevity search that …

Content truncated to 500 words