It’s no secret that today’s conversational assistants leave a lot to be desired. We spend a lot of time talking about how bad they are at hearing what we say and even worse at interpreting it. But one thing that doesn’t get nearly enough attention is the one-directional nature of conversations we have with Siri, Cortana and the rest. These systems don’t know how to ask questions — real inquiries that reflect a nuanced understanding of what information is known and what information is desired.
Microsoft Maluuba, the research team of PhDs that Microsoft acquired back in January, has been hard at work exploring the nexus of machine learning and question generation with the aim of delivering an intelligent personal assistant that surpasses the limited capabilities of today’s market leaders. In a recent paper, the team outlined its approach of using recurrent networks to generate questions from a body of text.
The team expanded upon existing question generators by combining supervised learning with reinforcement learning. Reinforcement learning makes use of the economic concept of utility to assign “points” for performing what are deemed to be optimal actions. The model is built to maximize this value and over many iterations it improves by continuously adjusting its strategy.
The Maluuba team prioritized accuracy and grammar. The aim is to be able to generate questions in proper English and to be able to generate questions that can be answered within the confines of the source text. Measuring the quality of a generated question based on how accurately a machine learning model can answer it is an interesting proposition.
If the model was generating questions based on this story, it would be reasonable to ask what company authored the research we are discussing. It would not be reasonable to ask a generic question about why the team used a recurrent network or to ask “team recurrent net not use?”
The team explained to me that the ability to ask questions improves the ability to answer them. This is because questions typically have only a single answer but it is possible to ask multiple questions that all lead to the same answer.
In practice, Maluuba’s question generator could be turned into a rather meta system for training other machine learning models by intentionally collecting new information. More concretely, question generation also has obvious uses in education — automatically generating questions for students to answer based on course material.