A Semi-Supervised Word Alignment Algorithm with Partial Manual Alignments

Proceedings of the Joint 5th Workshop on Statistical Machine Translation and MetricsMATR, pages 1–10, Uppsala, Sweden, 15-16 July 2010. c⃝2010 Association for Computational Linguistics A Semi-supervised Word Alignment Algorithm with Partial Manual Alignments Qin Gao, Nguyen Bach and Stephan Vogel Language Technologies Institute Carnegie Mellon University 5000 Forbes Avenue, Pittsburgh PA, 15213 {qing, nbach, stephan.vogel}@cs.cmu.edu Abstract We present a word alignment framework that can incorporate partial manual align- ments. The core of the approach is a novel semi-supervised algorithm extend- ing the widely used IBM Models with a constrained EM algorithm. The par- tial manual alignments can be obtained by human labelling or automatically by high-precision-low-recall heuristics. We demonstrate the usages of both methods by selecting alignment links from manu- ally aligned corpus and apply links gen- erated from bilingual dictionary on unla- belled data. For the first method, we con- duct controlled experiments on Chinese- English and Arabic-English translation tasks to compare the quality of word align- ment, and to measure effects of two differ- ent methods in selecting alignment links from manually aligned corpus. For the second method, we experimented with moderate-scale Chinese-English transla- tion task. The experiment results show an average improvement of 0.33 BLEU point across 8 test sets. 1 Introduction Word alignment is used in various natural lan- guage processing applications, and most statistical machine translation systems rely on word align- ment as a preprocessing step. Traditionally the word alignment model is trained in an unsuper- vised manner, e.g. the most widely used tool GIZA++ (Och and Ney, 2003), which implements the IBM Models (Brown et. al., 1993) and the HMM model (Vogel et al., 1996). However, for language pairs such as Chinese-English, the word alignment quality is often unsatisfactory (Guzman et al., 2009). There has been increasing interest on using manual alignments in word alignment tasks. Ittycheriah and Roukos (2005) proposed to use only manual alignment links in a maximum en- tropy model. A number of semi-supervised word aligners are proposed (Blunsom and Cohn, 2006; Niehues and Vogel, 2008; Taskar et al., 2005; Liu et al., 2005; Moore, 2005). These approaches use held-out manual alignments to tune the weights for discriminative models, with the model param- eters, model scores or alignment links from un- supervised word aligners as features. Also, sev- eral models are proposed to address the prob- lem of improving generative models with small amount of manual data, including Model 6 (Och and Ney, 2003) and the model proposed by Fraser and Marcu (2006) and its extension called LEAF aligner (Fraser and Marcu, 2007). The approaches use labelled data to tune parameters to combine different components of the IBM Models. 2005� � �� 2005nian de xiatian The summer of 2005 Figure 1: Partial and full alignments An interesting question is, if we only have par- tial alignments of sentences, can we make use of them? Figure 1 shows the comparison of par- tial alignments (the bold link) and full alignments (both of the dashed and the bold links). A partial alignment of …

Content truncated to 500 words