Generating Descriptions of Visible Objects

Опубликовано 8 августа 2016, 18:31

In this talk, I will be discussing the connection between vision and language, particularly focusing on how to generate human-like reference to visible objects. Some of the findings in this talk have been used to automatically generate descriptions of images (EACL 2012), order descriptive modifiers before a noun (ACL 11), and approximate human preferences for describing color and size (ENLG 2011, NAACL 2013). Evaluating automatically generated text offers interesting challenges, particularly when aiming to capture speaker variation, and I will spend some time discussing how best to measure the naturalness of generated descriptions against a corpus of human-produced descriptions.