Cleaning statistical language models
Résumé
In this paper, we describe how to decide a n-gram is actually impossible in a language. We use decision rules on a corpus tagged with POS. These rules are based on statistics and phonological criteria. In terms of statistical language modeling, deciding that a n-gram is impossible leads to assign to it a null probability.We defer on the possible n-grams the released mass of probabilities. To do this, we define a new formulation of P(w|h). We apply the principle of impossible events to bigrams. Then we use the list of impossible bigrams to build a list of impossible trigrams. The new trigram model exceeds the baseline model by 5.53% in terms of perplexity.