Discriminative word alignment by learning the alignment structure and syntactic divergence between a language pair
Sriram Venkatapathy, Aravind K. Joshi · 2007
Discriminative approaches for word alignment have gained popularity in recent years because of the flexibility that they offer for using a large variety of features and combining information from various sources. But, the models proposed in the past have not been able to make much use of features that capture the likelihood of an alignment structure (the set of alignment links) and the syntactic divergence between sentences in the parallel text. This is primarily because of the limitation of their search techniques. In this paper, we propose a generic discriminative re-ranking approach for word alignment which allows us to make use of structural features effectively. These features are particularly useful for language pairs with high structural divergence (like English-Hindi, English-Japanese). We have shown that by using the structural features, we have obtained a decrease of 2.3% in the absolute value of alignment error rate (AER). When we add the cooccurence probabilities obtained from IBM model-4 to our features, we achieved the best AER (50.50) for the English-Hindi parallel corpus.