wordalignment

Word alignment is the task of identifying word-level correspondences between sentences in a bilingual or multilingual parallel corpus. For a given pair of sentences, the goal is to pair each word in the source with its translation(s) in the target, allowing one-to-one, one-to-many, or many-to-one relations, and sometimes leaving words unaligned. Alignments are a central component of statistical machine translation and are used to induce translation models, create bilingual lexicons, and provide alignment-based features for downstream tasks.

Classical approaches treat word alignment as a generative modeling problem. The IBM alignment models (Model 1 through Model 5) and hidden Markov model alignment approaches model the joint probability of a source sentence, a target sentence, and an alignment, typically estimated with expectation-maximization on parallel data. Practical implementations include GIZA++, fast_align, and related tools. Alignments are often represented as a sparse mapping with posterior probabilities and can be combined from bidirectional alignments via symmetrization.

Evaluation typically uses alignment error rate (AER) against a gold standard, or precision and recall of linked word pairs. In the neural era, attention mechanisms in neural machine translation provide soft alignments that correlate with word correspondence; explicit alignments are still extracted for lexicon induction, data pruning, and linguistic analysis.

Common challenges include handling non-literal translations, word order differences, polysemy, and many-to-many mappings. Word alignment relies on parallel corpora such as Europarl and News Commentary and remains a bridge between traditional feature-based statistical methods and modern neural approaches.