Improving Word Alignment Quality Using Linguistic Knowledge

Word alignment of bilingual parallel corpora is usually generated using only statistical information. External linguistic information like e.g. a dictionary or linguistic structural annotation of the texts is used rarely, despite its usefulness. Additionally, it has to our knowledge never been examined systematically how linguistic information can be employed for word alignment improvement. In this paper, we present our experiments on finding out which linguistic information has which effect on word alignment quality, and we evaluate our experiments using precision and recall calculated for dictionaries that were generated after word alignment. The experiments show that information on e.g. lemmas and word category is useful to increase recall without lowering precision. Additionally, we discuss whether linguistic information can be used to compensate weak points of standard word alignment systems, and which features an ideal procedure should possess.