The structure of unseen trigrams and its application to language models: A first investigation

In a series of preparatory experiments in 4 languages on subsets of the Europarl corpus, we show that a large number of unseen trigrams can be reconstructed by proportional analogy with trigrams having the lowest frequencies. We derive a very simple smoothing scheme from this empirical result and show that it outperforms Good-Turing and Kneser-Ney smoothing schemes on trigrams models in all 11 languages on the common multilingual part of the Europarl corpus, except Finnish.