英文摘要 |
Conventional tagging models, with parameters estimated by the widely used maximum likelihood estimator, usually fail to achieve satisfactory performance in real applications. Since they achieve lexical disambiguation indirectly and implicitly via estin1ation, these models are usually unable to cover the statistical variation in the real text. In this paper, a discrimination oriented learning algorithm is proposed to directly pursue the goal of lexical disambiguation, so that the modeling error and the estimation error due to insufficient training data can be compensated. A 42% reduction in error rate, has been observed in the task of tagging Brown Corpus by using this. proposed method. |