Example-based error recovery method for speech translation: repairing sub-trees according to the semantic distance

In speech translation, recognition errors produced by the speech recognition process can cause parsing and translation errors. Because of this, the development of a robust error handling framework is quite essential to improve the performance of the speech translation system. Previously, a robust translation method was proposed by Wakita, which translates only reliable parts in utterances. In this method, however, the recall of translated parts for a whole utterance is low, and sometimes no translation is output. In this paper, we propose an example-based error recovery method to solve the low recall problem of Wakita’s method. The proposed method recovers an unreliable utterance, by repairing the parse-tree of the utterance based on similar example parse-trees in the treebank. A recovered translation is generated from the recovered tree.