Discriminative Learning of Feature Functions of Generative Type in Speech Translation

 The speech translation (ST) problem can be formulated as a log-linear model with multiple features that capture different levels of dependency between the input voice observation and the output translations. However, while the log-linear model itself is of discriminative nature, many of the feature functions are derived from generative models, which are usually estimated by conventional maximum likelihood estimation. In this paper, we first present the formulation of the ST problem as a log-linear model with a plurality of feature functions. We then describe a general discriminative learning framework for training these generative features based on a technique called growth transformation (GT). The proposed approach is evaluated on a spoken language translation benchmark test of IWSLT. Our experimental results show that the proposed method leads to significant improvement of translation quality. Fast and stable convergence can also be achieved by the proposed method. 1. Electronic Submission Speech translation (ST) takes the source speech signal as input and produces as output the translated text of that utterance in another language. It can be viewed as automatic speech recognition (ASR) and machine translation (MT) in tandem. Like many other machine learning problems, the speech translation (ST) problem can be modeled by a log-linear model with multiple features that capture different dependencies between the input voice observation and the output translations. Although the log-linear model itself is a discriminative model, many of the feature functions, such as scores of ASR outputs, are still derived from generative models. Further, these features are usually trained by conventional maximum likelihood estimation. In this paper, we propose a general framework of discriminative training for these generative features based on a technique called growth transformation (GT). The proposed approach is evaluated on a spoken language translation benchmark test called IWSLT. Our experimental results show that the proposed method leads to significant translation performance improvement. It is also shown that fast and stable convergence can be achieved by the proposed GT based optimization method.