From Models of Language Understanding to Agents of Language Use

Опубликовано 12 июля 2016, 18:17

Neural network-style algorithms and architectures have recently been applied to achieve notable improvements to various language applications. However, when it comes to general language understanding, the performance of these models is still far from human-like. One possible limiting factor is the clear differences between the training regimes of artificial and human language learners. Current models are trained to minimise a formal cost function on fixed, prescribed datasets, whereas humans learn language in an interactive, multimodal environment, motivated by a desire to achieve communicative goals. This talk focuses on my PhD work, in which I aimed to adapt the training data or architecture of (generally supervised) neural language models to better reflect the human language learning environment. I finish by mapping out a programme of work in which these function-approximation methods are integrated into a more dynamic, interactive learning environment. The programme is ambitious, and may require expertise in various learning and AI subfields, such as games and even robotics, as well as deep learning and computational linguistics. Ultimately, I hope it will lead to both more robust and realistic language understanding applications and important scientific insights into how humans acquire and process language.