Iterative decoding: A novel re-scoring framework for confusion networks

Recently there has been a lot of interest in confusion network re-scoring using sophisticated and complex knowledge sources. Traditionally, re-scoring has been carried out by the N-best list method or by the lattices or confusion network dynamic programming method. Although the dynamic programming method is optimal, it allows for the incorporation of only Markov knowledge sources. N-best lists, on the other hand, can incorporate sentence level knowledge sources, but with increasing N, the re-scoring becomes computationally very intensive. In this paper, we present an elegant framework for confusion network re-scoring called `iterative decoding´. In it, integration of multiple and complex knowledge sources is not only easier but it also allows for much faster re-scoring as compared to the N-best list method. Experiments with language model re-scoring show that for comparable performance (in terms of word error rate (WER)) of iterative decoding and N-best list re-scoring, the search effort required by our method is 22 times less than that of the N-best list method.