Improving Thai-English word alignment for interrogative sentences in SMT by grammatical knowledge

This paper presents a method to improve Thai-English word alignment in statistical machine translation (SMT) for interrogative sentences in a parallel corpus. We utilize the Thai and English grammatical knowledge i.e. tense, part of speech (POS), and question inversion pattern. The proposed method handles the difference of Thai and English interrogative sentences using sentence transformation, interrogative grammatical attribute extraction, and interrogative grammatical attribute annotation. This method works as a pre-processing of GIZA, a standard word co-occurrence alignment tool in SMT. We hypothesize that using grammatical knowledge as a pre-processing of GIZA can provide higher accuracy. We experiment by using 43,500 interrogative sentences to compare alignment result between interrogative sentences attached an interrogative grammatical label and interrogative sentences unattached interrogative grammatical label. The experimental results yield 95% of accuracy with significant improvement than the conventional one. With the increasing accuracy of word alignment, the translation accuracy is consequently improved.