To resolve the problems of syntactic ambiguities, a unified probabilistic score function approach was proposed in our previous works. The parameters used in the score function were estimated by the maximum likelihood algorithm, and 53.1% accuracy rate was observed. To further improve the performance in the testing set, a discrimination and robustness oriented adaptive learning algorithm was derived from the score function, and the accuracy rate was pushed to 64.3%. However, the parameters corresponding to those rare events are usually quite unreliable and cannot be well tuned by the adaptive learning algorithm. For improving the parameter estimation from sparse training data, the effect of parameter smoothing is investigated in this paper first, and about 56% accuracy rate is achieved with the smoothing methods. Then, a hybrid approach, which incorporates the smoothing techniques and the robust learning algorithm, is proposed to further improve the performance. A very promising result, 69.8% accuracy rate, is obtained using this hybrid method. The significant error reduction rate of 35.6% (from 53.1% to 69.8%) shows that the smoothing technique not only reduce the estimation error but also lead the learning process to a better local optimal point.
Keywords: Adaptive Learning, Parameter Smoothing.
Bibliographic reference. Chiang, Tung-Hui / Su, Keh-Yih (1993): "The effects of parameter smoothing on robust learning in syntactic ambiguity resolution", In EUROSPEECH'93, 1183-1186.