Second Languages as Meaning Annotation

In this talk, I will describe my recent work based on the idea that human translations into a second language can provide preexisting meaning annotations for natural language understanding tasks, and in particular for machine translation. In contrast to manually curated representations of word meaning, translated text is readily available in large amounts and avoids the need for costly annotation by human experts. For example, I will show that viewing translations as word meaning annotation can help (a) design better models for machine translation, such as discriminative translation lexicons that can select accurate translations in context, (b) improve automatic evaluation of machine translation, and (c) improve the portability of translation systems across domains. While my past work shows that the "translation as meaning" paradigm is effective for words and short phrases, my future work aims to use second language supervision to compose the meaning of words into complete sentences and coherent discourse. This includes investigating how to leverage more diverse sources of multilingual data as noisy supervision for natural language understanding.