We first train word alignment models and will use them to evaluate the goodness of a phrase and a phrase pair.

Beginning with a flat lexicon, we train IBM Model-l word alignment model with 10 iterations for each translation direction.

We then train HMM word alignment models (Vogel et al., 1996) in two directions simultaneously by merging statistics collected in the

All these features are data-driven and defined based on models, such as statistical word alignment model or language model.

In a statistical generative word alignment model (Brown et al., 1993), it is assumed that (i) a random variable a specifies how each target word fj is generated by (therefore aligned to) a source 1 word eaj; and (ii) the likelihood function f (f, a|e) specifies a generative procedure from the source sentence to the target sentence.

This distribution is applicable to all word alignment models that follow assumptions (i) and (ii).

We employ features based on word alignment models and alignment matrix.

In this work we analyze a recently proposed agreement-constrained EM algorithm for unsupervised alignment models .

We propose and extensively evaluate a simple method for using alignment models to produce alignments better-suited for phrase-based MT systems, and show significant gains (as measured by BLEU score) in end-to-end translation systems for six languages pairs used in recent MT competitions.

They suggest how this framework can be used to encourage two word alignment models to agree during training.

Most MT systems train an alignment model in each direction and then heuristically combine their predictions.

In this work, we show that by changing the way the word alignment models are trained and

We present extensive experimental results evaluating a new training scheme for unsupervised word alignment models : an extension of the Expectation Maximization algorithm that allows effective injection of additional information about the desired alignments into the unsupervised training process.

We then train the competing alignment models and compute competing alignments using different decoding schemes.

2.1 Baseline word alignment models

Figure 1 illustrates the mapping between the usual HMM notation and the HMM alignment model .

All word alignment models we consider are normally trained using the Expectation Maximization

However, our best system does not apply VB to a single probability model, as we found an appreciable benefit from bootstrapping each model from simpler models, much as the IBM word alignment models are usually trained in succession.

As these word-level alignment models restrict the word alignment complexity by requiring each target word to align to zero or one source words, results are improved by aligning both source-to-target as well as target-to-source,

Ideally, such a procedure would remedy the deficiencies of word-level alignment models , including the strong restrictions on the form of the alignment, and the strong independence assumption between words.

Our second approach was to constrain the search space using simpler alignment models , which has the further benefit of significantly speeding up training.

First we train a lower level word alignment model , then we place hard constraints on the phrasal alignment space using confident word links from this simpler model.

alignment models is the EM algorithm (Brown et al., 1993) which iteratively updates parameters to maximize the likelihood of the data.