Identification of subject shareness for Korean-English machine translation

One of the most critical issues in translating Korean into other languages is the common use of empty arguments. Since even mandatory elements in Korean are often dropped unlike English, the missing elements should be resolved during translation to obtain grammatical sentences. In this paper, we focus on missing subjects in intra-sentential level, which can be regarded as the identification of subject sharing between clauses. In order to reflect syntactic information in resolving missing subjects, we use a parse tree kernel, a specialized convolution kernel. In experimental evaluation, syntactic information turns out to be positively related to the identification of subject shareness. Our method achieves an accuracy of 81.39% and outperforms the baseline system assuming that two adjacent clauses share a subject.