Translating Government Agencies' Tweet Feeds: Specificities, Problems and (a few) Solutions

摘要

While the automatic translation of tweets has already been investigated in different scenarios, we are not aware of any attempt to translate tweets created by government agencies. In this study, we report the experimental results we obtained when translating 12 Twitter feeds published by agencies and organizations of the government of Canada, using a state-of the art Statistical Machine Translation (SMT) engine as a black box translation device. We mine parallel web pages linked from the URLs contained in English-French pairs of tweets in order to create tuning and training material. For a Twitter feed that would have been otherwise difficult to translate, we report significant gains in translation quality using this strategy. Furthermore, we give a detailed account of the problems we still face, such as hashtag translation as well as the generation of tweets of legal length.