Method and apparatus for improving statistical word alignment models using smoothing

1. A method of estimating translation probabilities, the method comprising:

determining initial values for the translation probabilities;

iteratively re-estimating the values of the translation probabilities while utilizing a smoothing technique by performing a division wherein the denominator of the division comprises a summation of counts summed with an additional value wherein each count in the summation of counts represents an estimated number of times a source word is translated as a target word in a corpus comprising a plurality of bilingual sentence pairs and wherein the summation is performed across all target words observed in the corpus; and

determining that the iteration is complete and that the translation probabilities form a trained model.

2. The method of claim 1 wherein utilizing a smoothing technique comprises utilizing add-n smoothing.

3. The method of claim 1 wherein the additional value comprises the product of a hypothesized number of words in a target language vocabulary and a virtual count, wherein the hypothesized number of words is larger than the number of target words observed in the corpus.

4. The method of claim 3 wherein re-estimating the value of a translation probability further comprises:

determining a count for the source word and the target word, the count representing an estimated number of occurrences of the source word translating as the target word in the corpus;

adding the virtual count to the count to form a sum; and

using the sum as the numerator in the division.

5. The method of claim 1 wherein re-estimating the values of the translation probabilities further comprises estimating a translation probability for translating a null token into a target word.

6. The method of claim 5 further comprising increasing the translation probability for translating a null token into a target word to produce an adjusted translation probability.

7. A method of estimating a translation probability, the method comprising:

forming a numerator through steps comprising adding a smoothing value to an expected count of the number of times a source word will appear as the translation of a target word in a plurality of sentence pairs of a bilingual corpus;

forming a denominator through steps comprising multiplying the smoothing value by a number that at least approximates the number of words in a target language and adding the product to an expected count for the number of times a source word will appear as the translation of any target word in the plurality of sentence pairs of the bilingual corpus;

dividing the numerator by the denominator to estimate a translation probability; and

storing the translation probability as part of a trained model.