Distributional semantics for situated spatial language? Functional, geometric and perceptual perspectives

Abstract

Distributional semantics has been at the core of recent developments in deep learning work for natural language processing. This distributional semantics plus neural processing paradigm has resulted in significant improvements in state of the art results across a large number of tasks, including parsing, text classification, and machine translation. However, there are a number of areas of natural language processing research where this shift in paradigm has not resulted in significant improvements in system performance. One such area is in situated dialogue systems (such as those studied in the field of human-robot interaction), and in particular with respect to the processing of spatial references. This chapter examines why this lack of progress has occurred, through a review of existing research on grounding language in perception that is structured around three forms of semantic information available in situated dialogue: functional, geometric and perceptual. Through this review we identify which aspects of perceptual grounding distributional semantics naturally accommodates and which aspects it does not. Building on this insight we suggest avenues for future work that attempt to integrate distributional and non-distributional information in order to progress research in perceptual grounding of language, and discuss the broader implications of our findings for computational representations of natural language semantics.