Weighted linear bilingual word alignment model

1. A word alignment system, comprising:

a weighted linear word alignment model, linearly combining feature values for a plurality of different features of a given word alignment to assign a score to each of a plurality of word alignments for a pair of text fragments, each of the different feature values being weighted by a corresponding trained feature weight indicative of the relative importance of a given feature to the plurality of features, wherein one of the plurality of different features in the word alignment model is a word association metric indicative of a strength of association between words in the pair of text fragments, the word association metric being based on one of surface statistics from a training corpus, and a conditional link probability word association metric, the one of surface statistics from the training corpus including an association metric that is defined in a contingency table, the contingency table including a matrix of cells for each of the plurality of word alignments, each of the cells associated with a count and one of the plurality of word alignments, a first one of the cells corresponding to no words in the associated one of the plurality of word alignments being present, a second one of the cells corresponding to two words in the associated one of the plurality of word alignments being present, a third one of the cells corresponding to a first word in the associated one of the plurality of word alignments being present but not a second word in the associated one of the plurality of words alignments, and a fourth one of the cells corresponding to the second word in the associated one of the plurality of word alignments being present but not the first word in the associated one of the plurality of word alignments, and wherein one of the different features is something other than a word association metric indicative of a strength of association between words in the pair of text fragments that is summed with the word association metric;

an automatic training component training the feature weights for the individual feature values used by the weighted linear word alignment model, wherein the automatic training component uses annotated training data separate from the pair of text fragments to train the feature weights;

a word alignment component receiving the pair of text fragments and accessing the word alignment model to identify the best scoring word alignment for the pair of text fragments; and

a computer processor being a functional component of the word alignment system and facilitating training of the feature weights.

2. The word alignment system of claim 1 wherein the word association metric comprises a log-likelihood word association metric.

3. The word alignment system of claim 1 wherein the plurality of different features comprise:

a non-monotonicity feature indicative of a non-monotonicity of the given word alignment.

4. The word alignment system of claim 1 wherein the plurality of different features comprise:

an unlinked words feature indicative of a number of unlinked words in the given word alignment.

5. The word alignment system of claim 1 wherein the plurality of different features comprise:

a multiple links feature indicative of a number of words in the given alignment that are linked to multiple other words in the given alignment.

6. The word alignment system of claim 1 and further comprising:

an index of word association types coupled to the word alignment component.

7. The word alignment system of claim 1 wherein the automatic training component is a discriminative automatic training component.

8. The word alignment system of claim 7 wherein the discriminative automatic training component is based on a form of perceptron learning.

9. The word alignment system of claim 8 wherein the discriminative automatic training component is based on a form of averaged perceptron learning.