Prenominal Modifier Ordering via Multiple Sequence Alignment

Human Language Technologies: The 2010 Annual Conference of the North American Chapter of the ACL, pages 600–608, Los Angeles, California, June 2010. c⃝2010 Association for Computational Linguistics Prenominal Modifier Ordering via Multiple Sequence Alignment Aaron Dunlop Oregon Health & Science University Portland, OR dunlopa@cslu.ogi.edu Margaret Mitchell University of Aberdeen Aberdeen, Scotland, U.K. m.mitchell@abdn.ac.uk Brian Roark Oregon Health & Science University Portland, OR roark@cslu.ogi.edu Abstract Producing a fluent ordering for a set of prenominal modifiers in a noun phrase (NP) is a problematic task for natural lan- guage generation and machine translation systems. We present a novel approach to this issue, adapting multiple sequence alignment techniques used in computa- tional biology to the alignment of modi- fiers. We describe two training techniques to create such alignments based on raw text, and demonstrate ordering accuracies superior to earlier reported approaches. 1 Introduction Natural language generation and machine trans- lation systems must produce text which not only conforms to a reasonable grammatical model, but which also sounds smooth and natural to a human consumer. Ordering prenominal mod- ifiers in noun phrases is particularly difficult in these applications, as the rules underlying these orderings are subtle and not well under- stood. For example, the phrase “big red ball” seems natural, while “red big ball” seems more marked, suitable only in specific contexts. There is some consensus that the order of prenom- inal modifiers in noun phrases is governed in part by semantic constraints, but there is no agreement on the exact constraints necessary to specify consistent orderings for any given set of modifiers. General principles of modifier order- ing based on semantic constraints also fall short on larger domains, where it is not always clear how to map prenominal modifiers to proposed semantic groups. With the recent advantages of large corpora and powerful computational resources, work on automatically ordering prenominal modifiers has moved away from approaches based on gen- eral principles, and towards learning ordering preferences empirically from existing corpora. Such approaches have several advantages: (1) The predicted orderings are based on prior evi- dence from ‘real-world’ texts, ensuring that they are therefore reasonably natural. (2) Many (if not all) prenominal modifiers can be ordered. (3) Expanding the training data with more and larger corpora often improves the system with- out requiring significant manual labor. In this paper, we introduce a novel approach to prenominal modifier ordering adapted from multiple sequence alignment (MSA) techniques used in computational biology. MSA is generally applied to DNA, RNA, and protein sequences, aligning three or more biological sequences in or- der to determine, for example, common ancestry (Durbin et al., 1999; Gusfield, 1997; Carrillo and Lipman, 1988). MSA techniques have not been widely applied in NLP, but have produced some promising results for building a generation map- ping dictionary (Barzilay and Lee, 2002), para- phrasing (Barzilay and Lee, 2003), and phone recognition (White et al., 2006). We believe that multiple sequence alignment is well-suited for aligning linguistic sequences, and that these alignments can be used to predict prenominal …

Content truncated to 500 words