Predicting Entity Relationships Using Hidden Markov Random Fields. An Application to Conversion of Greeklish Text

Abstract

We present results of a new optimal selection algorithm, applied to the transliteration of greeklish text. Our method uses existing greek text samples to “learn” greek word correlations by applying a Hidden Markov Random Field model to the underlying word relationships. The model “learns” the semantic correlations between word pairs in the training samples. When presented with trial greeklish text, the model uses this knowledge to disambiguate between all the possible transliterations of each greeklish word, by selecting the transliterations that have the highest learned correlation. Trials demonstrate that a model trained in text focused in a single topic can achieve up to 100% accuracy in sample sentence transliteration.