Incorporating diverse information sources in handwriting recognition postprocessing

This article describes the proposed implementation of a new model for the linguistic postprocessing component of the Human Language Technology (HLT) project. The model was designed for handwriting recognition applications but can be used for other text recognition problems and speech recognition. We demonstrate here that the current implementation (the POS model) fails to incorporate new sources of information such as word n-grams, and further handles the recognizer's scores incorrectly. We propose an alternative approach (the SSS model) which remedies these shortcomings. We also show that the SSS algorithm has a direct interpretation as a Hidden Markov Model whose states correspond to words that have been tagged with their parts of speech, and whose observations are discretized recognizer confidences. The HMM interpretation has the added advantage that the approach can be naturally extended to handle error recovery of the recognizer. Preliminary results indicate that the SSS model is successful in selecting the truth path over alternate paths.