A cross-lingual approach for opinion holder extraction

Opinion holder extraction is an important subtask in opinion analysis. However, the lack of fine annotated corpus influences the supervised machine learning based approach. The cross-lingual approach which leverages the opinion annotation information on other languages is considered as a feasible solution. In this paper, we propose a new cross-lingual opinion analysis approach to leverage fine-annotated opinion corpus. We firstly generate the translations and corresponding annotations of MPQA, which is most important opinion corpus in English, through cross-lingual projection. The obtained transferred corpus is used as the supplementary training data to train a classifier based on Tree Kennel Support Vector Machine (TK-SVM). The experiments on NTCIR-7 MOAT dataset (Simplified Chinese side) show that the proposed cross-lingual approach achieves a higher performance compared to CRFs models.