An Experiment in Word Alignment with a Parallel Corpus

This report documents an experiment done on word alignment using a parallel , sentence aligned corpus. The languages are English and Japanese and the corpus is derived from the Asahi Shinbun daily newspaper editorials. The aims of the experiment are To nd out how accurate word alignment is with simple pattern matching. To nd out how useful a conventional English-Japanese lexicon is. To observe how well English and Japanese newspaper editorials correspond at the lexical level of transfer. The results show approximately 80% word alignment success using a small bilingual lexicon of 81 English words and a mean 17.4 translations per word. The reason for error is analysed and it is concluded that gaps in the senses of words in the lexicon and stylistic diierences in the English and Japanese are responsible. The shallow level of analysis is not a major problem although it is expected to become more important as coverage is increased.