TermEnsembler: An enseble learning approach to bilingual term extraction and alignment

This paper describes TermEnsembler, a bilingual term extraction and alignment system utilizing a novel ensemble learning approach to bilingual term alignment. In the proposed system, the processing starts with monolingual term extraction from a language industry standard file type containing aligned English and Slovenian texts. The two separate term lists are then automatically aligned using an ensemble of seven bilingual alignment methods, which are first executed separately and then merged using the weights learned with an evolutionary algorithm. In the experiments, the weights were learned on one domain and tested on two other domains. When evaluated on the top 400 aligned term pairs, the precision of term alignment is over 96%, while the number of correctly aligned multi-word unit terms exceeds 30% when evaluated on the top 400 term pairs.