EM Decipherment for Large Vocabularies
Malte Nuhn, Hermann Ney · 2014
This paper addresses the problem of EMbased decipherment for large vocabularies.Here, decipherment is essentially a tagging problem: Every cipher token is tagged with some plaintext type.As with other tagging problems, this one can be treated as a Hidden Markov Model (HMM), only here, the vocabularies are large, so the usual O(N V 2 ) exact EM approach is infeasible.When faced with this situation, many people turn to sampling.However, we propose to use a type of approximate EM and show that it works well.The basic idea is to collect fractional counts only over a small subset of links in the forward-backward lattice.The subset is different for each iteration of EM.One option is to use beam search to do the subsetting.The second method restricts the successor words that are looked at, for each hypothesis.It does this by consulting pre-computed tables of likely n-grams and likely substitutions.