Generalizing Hierarchical Phrase-based Translation using Rules with Adjacent Nonterminals
23838Authors
nan
Abstract
Hierarchical phrase-based translation (Hiero, (Chiang, 2005)) provides an attractive frame- work within which both short- and long- distance reorderings can be addressed consis- tently and ef�ciently. However, Hiero is gen- erally implemented with a constraint prevent- ing the creation of rules with adjacent nonter- minals, because such rules introduce compu- tational and modeling challenges. We intro- duce methods to address these challenges, and demonstrate that rules with adjacent nontermi- nals can improve Hiero's generalization power and lead to signi�cant performance gains in Chinese-English translation. 1
Keywords
Full Content
Human Language Technologies: The 2010 Annual Conference of the North American Chapter of the ACL, pages 349–352, Los Angeles, California, June 2010. c⃝2010 Association for Computational Linguistics Generalizing Hierarchical Phrase-based Translation using Rules with Adjacent Nonterminals Hendra Setiawan and Philip Resnik UMIACS Laboratory for Computational Linguistics and Information Processing University of Maryland, College Park, MD 20742, USA hendra, resnik @umd.edu Abstract Hierarchical phrase-based translation (Hiero, (Chiang, 2005)) provides an attractive frame- work within which both short- and long- distance reorderings can be addressed consis- tently and ef�ciently. However, Hiero is gen- erally implemented with a constraint prevent- ing the creation of rules with adjacent nonter- minals, because such rules introduce compu- tational and modeling challenges. We intro- duce methods to address these challenges, and demonstrate that rules with adjacent nontermi- nals can improve Hiero's generalization power and lead to signi�cant performance gains in Chinese-English translation. 1 Introduction Hierarchical phrase-based translation (Hiero, (Chi- ang, 2005)) has proven to be a very useful com- promise between syntactically informed and purely corpus-driven translation. By automatically learn- ing synchronous grammar rules from parallel text, Hiero captures short- and long-distance reorderings consistently and ef�ciently. However, implementa- tions of Hiero generally forbid adjacent nonterminal symbols on the source side of hierarchical rules, a practice we will refer to as the non-adjacent nonter- minals constraint. The main argument against such rules is that they cause the system to produce multi- ple derivations that all lead to the same translation � a form of redundancy known as spurious ambiguity. Spurious ambiguity can lead to drastic reductions in decoding ef�ciency, and the obvious solutions, such as reducing beam width, erode translation quality. In Section 2, we argue that the non-adjacent non- terminals constraints severely limits Hiero's gener- alization power, limiting its coverage of important reordering phenomena. In Section 3, we discuss the challenges that arise in relaxing this constraint. In Section 4 we introduce new methods to address those challenges, and Section 5 validates the ap- proach empirically. Improving Hiero via variations on rule prun- ing and �ltering is well explored, e.g., (Chiang, 2005; Chiang et al., 2008; Zollmann and Venugopal, 2006), to name just a few. These proposals dif- fer from each other mainly in the speci�c linguis- tic knowledge being used, and on which side the constraints are applied. In contrast, we complement previous work by showing that adding rules to Hiero can provide bene�ts if done judiciously. 2 Judicious Use of Adjacent Nonterminals Our motivations largely follow Menezes and Quirk's (2007) discussion of reorderings and generalization. As a speci�c example, we will use a Chinese to En- glish verb phrase (VP) translation (Fig. 1), which represents one of the most prominent phrase con- structions in Chinese. Here the construction of the Chinese VP involves joining a prepositional phrase (PP) and a smaller verbal phrase (VP-A), with the preposition at the beginning as a PP marker. In the translation, the VP-A precedes the PP, a shift from pre-verbal PP in Chinese to post-verbal in English. ���� ���� …
Content truncated to 500 words
Labeled Text
Hierarchical phrase-based translation (Hiero, (Chiang, 2005)) provides an attractive framework within which both short- and longdistance reorderings can be addressed consistently and ef�ciently. However, Hiero is generally implemented with a constraint preventing the creation of rules with adjacent nonterminals, because such rules introduce computational and modeling challenges. We introduce methods to address these challenges, and demonstrate that rules with adjacent nonterminals can improve Hiero's generalization power and lead to signi�cant performance gains in Chinese-English translation. 1
Publication Details
- Paper ID:
- 23838
- Year:
- 2010
- Publication:
- TIPSTER
- Publication Type:
- conf
- DOI:
- nan