Presented video hosted on Youtube (with permission from presenter) at: https://youtu.be/BHQBkN4PyPc
Neural language models, which place probability distributions over sequences of words, produce vector representations of words and sentences that are useful for language processing tasks as diverse as machine translation, question answering, and image captioning. These models’ usefulness is partially explained by the fact that their representations robustly encode lexical and syntactic information. But the extent to which language model training also induces representations of meaning remains a topic of ongoing debate. I will describe recent work showing that language models—trained on text alone, without any kind of grounded supervision—build structured meaning representations that are used to simulate entities and situations as they evolve over the course of a discourse. These representations can be linearly decoded into logical representations of world state (e.g. discourse representation structures). They can also be directly manipulated to produce predictable changes in generated output. Together, these results suggest that (some) highly structured aspects of meaning can be recovered by relatively unstructured models trained on corpus data.