Search

The overarching objective of this thesis is to further the field of grounded semantics using a range of computational and empirical studies. Over the past thirty years, there have been many algorithmic advances in the modelling of semantic cognition. A commonality across these cognitive models is a reliance on hand-engineering "toy-models". Despite incorporating newer techniques (e.g. Long short-term memory), the model inputs remain unchanged. We argue that the inputs to these traditional semantic models have little resemblance with real human experiences. In this dissertation, we ground our neural network models by training them with real-world visual scenes using naturalistic photographs. Our approach is an alternative to both hand-coded features and embodied raw sensorimotor signals. We conceptually replicate the mutually reinforcing nature of hybrid (feature-based and grounded) representations using silhouettes of concrete concepts as model inputs. We next gradually develop a novel grounded cognitive semantic representation which we call scene2vec, starting with object co-occurrences and then adding emotions and language-based tags. Limitations of our scene-based representation are identified for more abstract concepts (e.g. freedom). We further present a large-scale human semantics study, which reveals small-world semantic network topologies are context-dependent and that scenes are the most dominant cognitive dimension. This finding leads us to conclude that there is no meaning without context. Lastly, scene2vec shows promising human-like context-sensitive stereotypes (e.g. gender role bias), and we explore how such stereotypes are reduced by targeted debiasing. In conclusion, this thesis provides support for a novel computational viewpoint on investigating meaning - scene-based grounded semantics. Future research scaling scene-based semantic models to human-levels through virtual grounding has the potential to unearth new insights into the human mind and concurrently lead to advancements in artificial general intelligence by enabling robots, embodied or otherwise, to acquire and represent meaning directly from the environment.