International Conference on Grammatical Inference edition:13 location:Delft, The Netherlands date:5-7 October 2016
Language learning has been studied for decades. For a long time, the focus was on learning the grammatical structure of a language from sentences, or learning the semantics of sentences from examples of sentence/meaning pairs. More recently, there has been increasing interest in grounded language learning, where the language is learned by observing sentences used in a particular context, and trying to link elements of these sentences to elements of the context.
This talk is about an approach called relational grounded language learning. In this approach, the semantics of a sentence is a relational structure, and this structure is learned from sentence/context pairs in which the context is represented in a relational format. Once a model of the link between sentences and semantic structures is in place, it can be used for a variety of purposes: generating sentences describing a given scene, identifying the elements in a scene that a sentence refers to, translating a sentence from one language to another through its semantic representation, and more. The potential of this approach for all these uses has been demonstrated on some simple problems. Although the approach is clearly still in its infancy, we believe it has much potential in terms of helping us understand how humans learn their first language, as well as improving natural language processing technology.