Birch et al, StatMT 2006

Citation

Alexandra Birch, Chris Callison-Burch, and Miles Osborne. 2006. Constraining the phrase-based, joint probability statistical translation model. In The Conference for the Association for Machine Translation in the Americas.

Online version

pdf

Summary

The model proposed in Marcus and Wong, EMNLP 2002 provides a strong framework for phrase-to-phrase alignments, but its applicability is hamstrung by the computational complexity of the running EM in the large space of latent variables generated from all possible phrases and alignments.

This work describes a phrase-to-phrase alignment model, uses word-to-phrase alignments to constrain the space of phrasal alignments, improving the scalability of the model and also the performance in the Machine Translation task.

Description of the method

The joint model proposed in Marcus and Wong, EMNLP 2002 searches the space of all possible latent variables (phrases and alignments between phrases) during the EM algorithm, which is computationally expensive. The goal of this method is to define hard constraints on the possible latent variables using a high confidence set of alignments.

The high confidence alignments are built using the interception of two unidirectional word-to-phrase alignments, which generally generates a set of alignments with high precision and low recall. Furthermore, more alignment points are added by aligning identical words in the two language and entries in a dictionary that match both sides of a sentence are also aligned.

Using this high confidence alignment, the space of possible phrase pairs is limited to phrase pairs that contain at least one alignment between the source words and target words, similarly to the alignment template defined in Och et al, 2004. To further constrict the search space, phrase pairs must occur at least a given number of times in the training corpora to be considered. These constraints are applied during the initialization of the parameters of the model and during the E-step of the EM algorithm.

Another improvement in the model is the introduction of the lexical weighting information which are calculated for phrase pairs using the high confidence alignments between words in the phrase pair.

Experiments

Tests were conducted by evaluating the translation quality using BLEU. The EUROPARL German-English data was used, which contains around 1.6 millions training sentences. However, due to the limitations of the model, only up to 40000 training sentences were used. The test set was composed by 1755 with length between 5 and 15 characters.

Model	10000	20000	40000
Joint model	21.69	23.61	25.52
+ lex	22.79	24.33	25.99
+ lex + ident	23.30	24.90	26.12
+ lex + ident + dict	23.20	24.96	26.13