How AI can achieve deep understanding of text and speech: Structure-exploiting strategies for knowledge extraction

Humans are good at engaging in conversation and understanding written text, so much so that these sorts of natural language processing (NLP) tasks are widely considered essential requirements for any truly intelligent artificial agent. The NLP community has made a great deal of progress over the last few years on certain challenges (for example, the SuperGLUE benchmarks and Winograd Schemas), thanks to the power of neural networks and pre-trained models like BERT and GPT-3. Nevertheless, existing systems still don’t approach human performance for useful tasks such as answering general, open-ended questions about stories or participating in dialogue.

In this talk, I analyze why NLP’s recent achievements do not generalize to true language understanding. I suggest a broad remedy – automated large-scale extraction of knowledge from text and reasoning with that knowledge, together with existing background information, to achieve deeper understanding – and propose methods to achieve this. I focus on texts that have exploitable structure, including bulleted lists and tables, and discuss the techniques I’ve used for automating understanding of regulatory and molecular biology texts. Finally, I discuss how comprehending the goal-action-effect structure implicit in narrative texts, including news articles and business case studies, can help extract useful general knowledge about actions and plans, and consider how these techniques could contribute to AI succeeding at the difficult tasks of story understanding and analysis.