Aspects of salience in natural language generation
Nick J. Cercone, T. Pattabhiraman · 1992
This dissertation examines the role of salience in natural language generation (NLG). The salience of an entity, in intuitive terms, refers to its prominence, and is interpreted as a measure of how well an entity stands out from other entities and biases the preference of the generator in selecting words and complex constructs. Through an analysis of previous work in diverse disciplines, we show the variety of salience effects in NLG. Next, we classify several important determinants of salience, corresponding to different factors contributing to salience. We then delineate two theoretically-significant categories: canonical salience and instantial salience. The former is characterized as a built-in preference in the general conceptual- and linguistic knowledge of the speaker. The latter refers to the salience of specific objects in the context of NLG, and may accrue through such determinants as vividness and recency of mention. Psycholinguistic results of Osgood and Bock are highlighted to suggest the multiplicative interaction between canonical salience and instantial salience. This interaction is captured in a decision-theoretic formalization which models canonical salience as probabilities, instantial salience as utilities, and the selection criterion as maximization of expected utility. Next we consider the phenomena of basic level- and entry level preference in naming objects. We argue that basic level preference in taxonomic concept knowledge is a form of canonical salience. By establishing the stark similarity of the conclusions of the psychological experiments of Jolicoeur et al. to those of Osgood and Bock, the dynamics of preferring names for objects is conceptualized as an interaction between canonical salience and instantial salience. We demonstrate the suitability of decision theory to model this interaction. Our model of salience interactions is further reinforced in a study of simile generation. We model property salience as information-theoretic redundancy and equivalently, as expected utility. To avert miscommunication and anomaly, we propose two different cost measures (antipodal to utility) using probabilistic knowledge and intrinsicness, and develop net expected utility as a decision criterion for selecting objects of comparison in similes. We conclude in this thesis that the multi-aspect notion of salience provides subtle quantitative control in language generation decisions, and captures interesting and significant effects such as graded judgments and relative frequencies of linguistic constructs.