The use of higher level linguistic knowledge for spelling-to-pronunciation generation

We have previously designed a hierarchical lexical representation for reversible spelling/phonemics generation. In this paper, we demonstrate the advantages of using higher level linguistic knowledge in a hierarchical framework to provide constraints for spelling-to-pronunciation generation. We expect our current findings to be applicable to pronunciation-to-spelling generation also, since our approach casts the two tasks as directly symmetric problems. Comparison with an alternative, single-layer approach illustrates how the hierarchical framework provides a parsimonious description for English orthographic-phonological regularities, while simultaneously attaining competitive generation accuracy. The hierarchical approach achieves a top-choice word accuracy of 67.5% for spelling-to-sound generation (computed over the entire test set, including nonparsable words), and is capable of reversible generation using about 32,000 parameters. In comparison, the single-layer approach requires over 40 times the number of parameters (about 693,300) to achieve a word accuracy of 69.1%, and is incapable of reversible generation