Identification and Translation of Significant Patterns for Cross-Domain SMT Applications

Adaptation of statistical machine translation (SMT) systems from generic to specific domains is challenging due to the lack of training data. In this paper we propose a framework for domain adaptation by exploiting a large monolingual in-domain corpus. We identify the significant patterns to capture the domain specific writing styles. The patterns are then translated with the involvements of domain experts. The major issue of our framework is to reduce the cost of the experts and better allocate their efforts. The experimental results show the proposed methods are effective, in terms of the significance and diversity of the patterns. The approaches to integrate the mined patterns into background SMT are also discussed.