Learning to Refer from Estimated Listener Gaze
Through evaluation with human listeners, this work finds that speaker policies trained with gaze-estimating listeners result in significantly more pragmatically-optimal references than base models, reducing sequence length from 15.4 down to 4.0 words while increasing referential success from 75.2 up to 80.0%.
Téa Wright, Alane Suhr
· 0 citations