Pivot-based word alignment

Word alignment is the task of, given two sentences that are translations of each other, determining which words correspond to each other across the two sentences. Word alignment is an important step in the pipeline of constructing a statistical machine translation system, but success at word alignment depends heavily on the quantity of training data available. The traditional methods for computational word alignment, proposed by Brown et al. (1993), require large quantities of training data. However, these methods fall short when such quantities of data are not available. To combat this problem, I propose a framework that fills in the data gap by using data from languages related to the one for which data are lacking. This technique is shown to improve significantly upon the baseline alignment error rate. In my proposed framework, aligning a sentence in a low-resource language with a sentence in a high-resource language follows a 3-step procedure. First, a pivot language is chosen such that it is both high-resource and as closely related to the low-resource language as possible. Edit distance is then used to create a correspondence between the low-resource language and the pivot language, while probabilities trained using the IBM Models are used to create a correspondence between the pivot language and the high-resource language. I test several different settings for these three basic components on the task of aligning Spanish and English, and I find that the most successful overall alignment system uses Portuguese as the pivot language, a cognate-based algorithm for calculating edit distance, and translation, distortion, and alignment probabilities from the IBM Models. With this framework settled on, I then conduct some sample translations to demonstrate the utility of this approach for machine translation. These translations are translating from Spanish to English, and despite being trained only on Portuguese the translator still manages to yield rough translations of Spanish text.