A discriminative latent model for Chinese multiword expression extraction

Abstract

In this paper, we propose a discriminative latent model to extract Chinese multiword expressions from corpus resources as part of a larger research effort to improve machine translation(MT) system such as Super Function based Machine Translation(SFBMT). For existing MT systems, especially Statistic based MT(SBMT), the issue of multiword expressions (MWEs) detection and accurate correspondence from source to target language remains an unsolved problem. Template or Super Function based machine translation system suffer less from the existance of MWEs, but MWEs are still main holdback of MT systems. Our initial experiments on the Chinese-Japanese MT systems reveal that, where MWEs exist, SFBMT system suffer in terms of both efficiency ,comprehensibility and adequacy of finding the translation functions. Statistic based Machine Translation suffers more than SFBMT. For SFBMT systems to become of further practical use, they need to be enhanced with MWEs processing capability. As part of our study towards this goal, we proposed a discriminative latent model, which was developed for sequence labeling task, for identifying and extracting Chinese MWEs. In our evaluation, the tool achieved precisions ranging from 71.46% to 95.87% for different types of MWEs. Such results demonstrate that it is feasible to automatically identify many Chinese MWEs using our tool, Super Function based MT will be further improvement after the Chinese MWEs have been detected.