Minimal Dependency Length in Realization Ranking

Proceedings of the 2012 Joint Conference on Empirical Methods in Natural Language Processing and Computational Natural Language Learning, pages 244–255, Jeju Island, Korea, 12–14 July 2012. c⃝2012 Association for Computational Linguistics Minimal Dependency Length in Realization Ranking Michael White and Rajakrishnan Rajkumar Department of Linguistics The Ohio State University Columbus, OH, USA {mwhite,raja}@ling.osu.edu Abstract Comprehension and corpus studies have found that the tendency to minimize dependency length has a strong influence on constituent or- dering choices. In this paper, we investigate dependency length minimization in the con- text of discriminative realization ranking, fo- cusing on its potential to eliminate egregious ordering errors as well as better match the dis- tributional characteristics of sentence order- ings in news text. We find that with a state- of-the-art, comprehensive realization rank- ing model, dependency length minimization yields statistically significant improvements in BLEU scores and significantly reduces the number of heavy/light ordering errors. Through distributional analyses, we also show that with simpler ranking models, dependency length minimization can go overboard, too of- ten sacrificing canonical word order to shorten dependencies, while richer models manage to better counterbalance the dependency length minimization preference against (sometimes) competing canonical word order preferences. 1 Introduction In this paper, we show that for the constituent or- dering problem in surface realization, incorporating insights from the minimal dependency length the- ory of language production (Temperley, 2007) into a discriminative realization ranking model yields sig- nificant improvements upon a state-of-the-art base- line. We demonstrate empirically using OpenCCG, our CCG-based (Steedman, 2000) surface realiza- tion system, the utility of a global feature encoding the total dependency length of a given derivation. Although other works in the realization literature have used phrase length or head-dependent distances in their models (Filippova and Strube, 2009; Velldal and Oepen, 2005; White and Rajkumar, 2009, i.a.), to the best of our knowledge, this paper is the first to use insights from the minimal dependency length theory directly and study their effects, both qualita- tively and quantitatively. The impetus for this paper was the discovery that despite incorporating a sophisticated syntac- tic model borrowed from the parsing literature— including features with head-dependent distances at various scales—White & Rajkumar’s (2009) real- ization ranking model still often performed poorly on weight-related decisions such as when to em- ploy heavy-NP shift. Table 1 illustrates this point. In wsj 0034.9, the full model (incorporating numer- ous syntactic features) succeeds in reproducing the reference sentence, which is clearly preferable to the rather awkward variant selected by the base- line model (using various n-gram models). How- ever, in wsj 0013.16, the full model fails to shift the temporal modifier for now next to the phrasal verb turned down, leaving it at the end of its very long verb phrase where it is highly ambiguous (with mul- tiple intervening attachment sites). Conversely, in wsj 0044.3, the full model shifts before next to the verb, despite the NP cheating being very light, yield- ing a very confusing ordering given that before is meant to be intransitive. …

Content truncated to 500 words