The long-term goal is to be able to "teach" a language-model-based question-answering system by interacting with it, so as to understand the reasoning behind its answers and correct it when it makes mistakes. Three things are needed for this to happen:
- language models (LMs) need be able to reason systematically with what they know and explain that reasoning.
- users need to be able to interact with those explanations to identify mistakes the system made.
- and the system needs to be able to react appropriately to corrections from the user, so as to update its knowledge and not make similar mistakes in future.

In this talk I will present our current work towards this goal. I will describe:
(a) how we can train language models to produce answers supported by a faithful chain of reasoning, revealing their latent (hidden) knowledge that supports the answer, hence allowing users to diagnose failures.
(b) how we augment the LM with a dynamic memory of user-corrected facts, allowing those corrected facts to influence future answers and help the system avoid similar mistakes in future.

We find that on two datasets, using simulated user feedback, the resulting system is able to continuously improve with time, without requiring model retraining. This suggests new opportunities for using LMs in an interactive setting where users can inspect, debug, correct, and improve a system performance over time.

Lucky draw prizes

We will raffle 5 winners for the book during the event. To enter the lucky draw, please complete one of the two steps:
- Twitter the event with hashtag #nlpaicamp and tag @aicampai . for example:

#nlpaicamp online AI/ML tech talk series by @aicampai: Peter Clark (Research Manager at @ai2_aristo @allen_ai) will discuss their latest research on NLP teachable systems. Free RSVP: https://www.aicamp.ai

This tech talk event is sponsored by Packt.