Ordering Prenominal Modifiers with a Reranking Approach

Proceedings of the 49th Annual Meeting of the Association for Computational Linguistics, pages 1109–1116, Portland, Oregon, June 19-24, 2011. c⃝2011 Association for Computational Linguistics Ordering Prenominal Modifiers with a Reranking Approach Jenny Liu MIT CSAIL jyliu@csail.mit.edu Aria Haghighi MIT CSAIL me@aria42.com Abstract In this work, we present a novel approach to the generation task of ordering prenomi- nal modifiers. We take a maximum entropy reranking approach to the problem which ad- mits arbitrary features on a permutation of modifiers, exploiting hundreds of thousands of features in total. We compare our error rates to the state-of-the-art and to a strong Google n- gram count baseline. We attain a maximum error reduction of 69.8% and average error re- duction across all test sets of 59.1% compared to the state-of-the-art and a maximum error re- duction of 68.4% and average error reduction across all test sets of 41.8% compared to our Google n-gram count baseline. 1 Introduction Speakers rarely have difficulty correctly ordering modifiers such as adjectives, adverbs, or gerunds when describing some noun. The phrase “beau- tiful blue Macedonian vase” sounds very natural, whereas changing the modifier ordering to “blue Macedonian beautiful vase” is awkward (see Table 1 for more examples). In this work, we consider the task of ordering an unordered set of prenomi- nal modifiers so that they sound fluent to native lan- guage speakers. This is an important task for natural language generation systems. Much linguistic research has investigated the se- mantic constraints behind prenominal modifier or- derings. One common line of research suggests that modifiers can be organized by the underlying semantic property they describe and that there is a. the vegetarian French lawyer b. the French vegetarian lawyer a. the beautiful small black purse b. the beautiful black small purse c. the small beautiful black purse d. the small black beautiful purse Table 1: Examples of restrictions on modifier orderings from Teodorescu (2006). The most natural sounding or- dering is in bold, followed by other possibilities that may only be appropriate in certain situations. an ordering on semantic properties which in turn restricts modifier orderings. For instance, Sproat and Shih (1991) contend that the size property pre- cedes the color property and thus “small black cat” sounds more fluent than “black small cat”. Using > to denote precedence of semantic groups, some commonly proposed orderings are: quality > size > shape > color > provenance (Sproat and Shih, 1991), age > color > participle > provenance > noun > denominal (Quirk et al., 1974), and value > dimension > physical property > speed > human propensity > age > color (Dixon, 1977). However, correctly classifying modifiers into these groups can be difficult and may be domain dependent or con- strained by the context in which the modifier is being used. In addition, these methods do not specify how to order modifiers within the same class or modifiers that do not fit into any of the specified groups. There have also been a variety of corpus-based, computational approaches. …

Content truncated to 500 words