Inner-Outer Bracket Models for Word Alignment using Hidden Blocks

Proceedings of Human Language Technology Conference and Conference on Empirical Methods in Natural Language Processing (HLT/EMNLP), pages 177–184, Vancouver, October 2005. c⃝2005 Association for Computational Linguistics Inner-Outer Bracket Models for Word Alignment using Hidden Blocks Bing Zhao School of Computer Science Carnegie Mellon University {bzhao}@cs.cmu.edu Niyu Ge and Kishore Papineni IBM T. J. Watson Research Center Yorktown Heights, NY 10598, USA {niyuge, papineni}@us.ibm.com Abstract Most statistical translation systems are based on phrase translation pairs, or “blocks”, which are obtained mainly from word alignment. We use blocks to infer better word alignment and improved word alignment which, in turn, leads to better inference of blocks. We propose two new probabilistic models based on the inner- outer segmentations and use EM algorithms for estimating the models’ parameters. The first model recovers IBM Model-1 as a spe- cial case. Both models outperform bi- directional IBM Model-4 in terms of word alignment accuracy by 10% absolute on the F-measure. Using blocks obtained from the models in actual translation systems yields statistically significant improvements in Chinese-English SMT evaluation. 1 Introduction Today’s statistical machine translation systems rely on high quality phrase translation pairs to acquire state-of-the-art performance, see (Koehn et al., 2003; Zens and Ney, 2004; Och and Ney, 2003). Here, phrase pairs, or “blocks” are obtained automati- cally from parallel sentence pairs via the underlying word alignments. Word alignments traditionally are based on IBM Models 1-5 (Brown et al., 1993) or on HMMs (Vogel et al., 1996). Automatic word align- ment is challenging in that its accuracy is not yet close to inter-annotator agreement in some language pairs: for Chinese-English, inter-annotator agree- ment exceeds 90 on F-measure whereas IBM Model- 4 or HMM accuracy is typically below 80s. HMMs assume that words “close-in-source” are aligned to words “close-in-target”. While this locality assump- tion is generally sound, HMMs do have limitations: the self-transition probability of a state (word) is the only control on the duration in the state, the length of the phrase aligned to the word. Also there is no natural way to control repeated non-contiguous vis- its to a state. Despite these problems, HMMs remain attractive for their speed and reasonable accuracy. We propose a new method for localizing word alignments. We use blocks to achieve locality in the following manner: a block in a sentence pair is a source phrase aligned to a target phrase. We assume that words inside the source phrase cannot align to words outside the target phrase and that words out- side the source phrase cannot align to words inside the target phrase. Furthermore, a block divides the sentence pair into two smaller regions: the inner part of the block, which corresponds to the source and target phrase in the block, and the outer part of the block, which corresponds to the remaining source and target words in the parallel sentence pair. The two regions are non-overlapping; and each of them is shorter than the original parallel sentence pair. The regions are thus easier to align than the original …

Content truncated to 500 words