Further develop our model representation and parameter sharing methods by automatically generating the smallest possible model without compromising accuracy, so that significant improvements in memory and speed are obtained, specifically working toward the goal of recognizing broadcast news in less than 10 times real-time.
Develop a named-entity tagger based on class-based statistical language modeling techniques so as to automatically learn the contexts in which named-entities occur in text.
Develop new training algorithms specifically tuned to our newly developed model representation and parameter sharing approaches so as to derive even larger improvements in accuracy, speed, and memory.