Modeling Human Referring Expression Generation
Sahar Kazemzadeh · Carolina Digital Repository (University of North Carolina at Chapel Hill) · 2019
Referring generation expression is a natural language processing task that involves creating noun phrases that identify a referent object to a listener. We evaluate the state-of-the-art Visible Objects Algorithm for referring expression generation presented by Mitchell et. al (2012), and find that it does not perform as well with our natural image set than with the computer-generated image set that was originally used. Further, we analyze over 7,000 referring expressions generated by players of ReferIt Game, an online game that we developed, and by Amazon Mechanical Turk workers to identify metrics with which to create an improved stochastic model that can be coupled with computer vision to mimic human referring expression generation from visual input.