Controlling the Reading Level of Machine Translation Output

Today’s machine translation systems output the same translation for a given input, despite important differences between users. In practice, translations should be customized for each reader, for instance when translating for children versus in a business setting. In this paper, we introduce the task of reading level control to machine translation, and provide the first results. Our methods can be used to raise or lower the reading level of output translations. In our first approach, sourceside sentences in the training corpus are tagged based on the reading level (readability) of the matching target sentences. Our second approach alters the traditional encoder-decoder architecture by specifying a joint encoder and separate decoders for simple and complex decoding modes, with training data partitioned by reading level. We demonstrate control over output readability score on three test sets in the Spanish–English language direction.