Improved Word Alignment with Statistics and Linguistic Heuristics

Proceedings of the 2009 Conference on Empirical Methods in Natural Language Processing, pages 229–237, Singapore, 6-7 August 2009. c⃝2009 ACL and AFNLP Improved Word Alignment with Statistics and Linguistic Heuristics Ulf Hermjakob University of Southern California Information Sciences Institute 4676 Admiralty Way Marina del Rey, CA 90292, USA ulf@isi.edu Abstract We present a method to align words in a bitext that combines elements of a tra- ditional statistical approach with linguis- tic knowledge. We demonstrate this ap- proach for Arabic-English, using an align- ment lexicon produced by a statistical word aligner, as well as linguistic re- sources ranging from an English parser to heuristic alignment rules for function words. These linguistic heuristics have been generalized from a development cor- pus of 100 parallel sentences. Our aligner, UALIGN, outperforms both the commonly used GIZA++ aligner and the state-of-the- art LEAF aligner on F-measure and pro- duces superior scores in end-to-end sta- tistical machine translation, +1.3 BLEU points over GIZA++, and +0.7 over LEAF. 1 Introduction Word alignment is a critical component in training statistical machine translation systems and has re- ceived a significant amount of research, for exam- ple, (Brown et al., 1993; Ittycheriah and Roukos, 2005; Fraser and Marcu, 2007), including work leveraging syntactic parse trees, e.g., (Cherry and Lin, 2006; DeNero and Klein, 2007; Fossum et al., 2008). Word alignment is also a required first step in other algorithms such as for learning sub-sentential phrase pairs (Lavie et al., 2008) or the generation of parallel treebanks (Zhechev and Way, 2002). Yet word alignment precision remains surpris- ingly low, under 80% for state-of-the-art aligners on not closely related language pairs. Consider the following Arabic/English sen- tence pair with alignments built by the statistical word aligner LEAF: Bitext Arabic: � � A � � A � � K � Q � � � P � X @ P A K . � Y � K A � K � A � J � @ � P A � � � ��J � � � � J� @ � , 6 - 4 � 6 - 4 � � Q J . � J � J � � � J � � � � � K � A k . � � @ Q � J� A � @ � � � 6 - 4 � 6 - 4 Q � � � � � � P � K . � � P A � � � K A � � A� @ � � � � J � � K A � � � Q J � K � Gloss: Won(1) Thai Paradorn Srichaphan(1) on/to(2) Australian Jason(2) Stoltenberg(3) 6(4) - 4(5) and(3) 6(4) - 4(5), and Czech Jiˇr´ı(7) Vanˇek(7) on/to German(6) Lars Burgsm¨uller 6(4) - 4 and(3) 6(4) - 4 Bitext English: Thailand ’s(1) Baradorn Srich- fan(1) beat(2) Australian Gayson(1) Stultenberg(3) 6(4) - 6(4) 6(4) - 4(5) , Czech player(1) Pierre(1) Vanic(7) beat(6) Germany(6) ’s Lars Burgsmuller 6 - 4 6 - 4 In the example above, words with the same …

Content truncated to 500 words