Determinants of Adjective-Noun Plausibility

Proceedings of EACL '99 Determinants of Adjective-Noun Plausibility Maria Lapata and Scott McDonald and Frank Keller School of Cognitive Science Division of Informatics, University of Edinburgh 2 Buccleuch Place, Edinburgh EH8 9LW, UK {mlap, scottm, keller} @cogsci.ed.ac.uk Abstract This paper explores the determinants of adjective-noun plausibility by using cor- relation analysis to compare judgements elicited from human subjects with five corpus-based variables: co-occurrence fre- quency of the adjective-noun pair, noun fre- quency, conditional probability of the noun given the adjective, the log-likelihood ra- tio, and Resnik's (1993) selectional asso- ciation measure. The highest correlation is obtained with the co-occurrence frequency, which points to the strongly lexicalist and collocational nature of adjective-noun com- binations. 1 Introduction Research on linguistic plausibility has focused mainly on the effects of argument plausibility during the pro- cessing of locally ambiguous sentences. Psycholin- guists have investigated whether the plausibility of the direct object affects reading times for sentences like (1). Here, argument plausibility refers to "prag- matic plausibility" or "local semantic fit" (Holmes et al., 1989), and judgements of plausibility are typi- cally obtained by asking subjects to rate sentence frag- ments containing verb-argument combinations (as an example consider the bracketed parts of the sentences in (1)). Such experiments typically use an ordinal scale for plausibility (e.g., from 1 to 7). (1) a. [The senior senator regretted the decision] had ever been made public. b. [The senior senator regretted the reporter] had ever seen the report. The majority of research has focussed on investigating the effect of rated plausibility for verb-object combi- nations in human sentence processing (Garnsey et al., 1997; Pickering and Traxler, 1998). However, plausi- bility effects have also been observed for adjective- noun combinations in a head-modifier relationship. Murphy (1990) has shown that typical adjective- noun phrases (e.g., salty olives) are easier to in- terpret in comparison to atypical ones (e.g., sweet olives). Murphy provides a schema-based explana- tion for this finding by postulating that in typical adjective-noun phrases, the adjective modifies part of the noun's schema and consequently it is understood more quickly, whereas in atypical combinations, the adjective modifies non-schematic aspects of the noun, which leads to interpretation difficulties. Smadja (1991) argues that the reason people prefer strong tea to powerful tea and powerful car to strong car is neither purely syntactic nor purely semantic, but rather lexical. A similar argument is put forward by Cruse (1986), who observes that the adjective spotless collocates well with the noun kitchen, relatively worse with the noun complexion and not all with the noun taste. Ac- cording to Cruse, words like spotless have idiosyn- cratic collocational restrictions: differences in the de- gree of acceptability of the adjective and its collocates do not seem to depend on the meaning of the individ- ual words. 1.1 Motivation Acquiring plausibility ratings for word combinations (e.g., adjective-noun, verb-object, noun-noun) can be useful in particular for language generation. Consider a generator which has to make a choice between spot- less kitchen and flawless kitchen. An empirical model of plausibility …

Content truncated to 500 words