Using a Random Forest Classifier to recognise translations of biomedical terms across languages
Georgios Kontonatsios, Ioannis Korkontzelos, Sophia Ananiadou, J Tsujii · 2013
We present a novel method to recognise semantic equivalents of biomedical terms in language pairs. We hypothesise that biomedical term are formed by semantically similar textual units across languages. Based on this hypothesis, we employ a Random Forest (RF) classifier that is able to automatically mine higher order associations between textual units of the source and target language when trained on a corpus of both positive and negative examples. We apply our method on two language pairs: one that uses the same character set and another with a different script, English-French and English-Chinese, respectively. We show that English-French pairs of terms are highly transliterated in contrast to the English-Chinese pairs. Nonetheless, our method performs robustly on both cases. We evaluate RF against a state-of-the-art alignment method, GIZA++, and we report a statistically significant improvement. Finally, we compare RF against Support Vector Machines and analyse our results. 1