To advance AI reasoning, study computational argumentation

I argue that AI development could benefit by learning from the body of work in computational argumentation.

In the core of the so-called intelligence lies reasoning. Types of reasoning can be broadly classified into two classes: formal and informal. Formal reasoning is the driving tool of maths and the natural sciences. Informal reasoning, on the other hand, facilitates fuzzy interactions among humans. In formal reasoning, you are concerned with proofs and the like. So what are you concerned in informal reasoning? Argumentation.

The study of argumentation can be traced back to the Greek philosophers like Aristotle and Socrates, one of which is attributed for the famous syllogy that “All men are mortal. Socrates are a man. Therefore, Socrates are mortal.” Modern argumentation seems to have converged on the theory by Stephen Toulmin in his 1958 book “The Uses of Argument”. The book seems to have proposed the vocabulary to talk about an argument, including its components and their relations.

With the invention language technologies, computers have been more capable of processing human arguments. Like the processing of anything, argumentation processing contains two main problem classes: argument understanding and argument generation.

• For understanding, given an argumentative document, a computer can restate the topic, the stance, and even recover the entire argument’s structure in a graph following Toulmin’s model.
• For generation, a computer can sometimes convince humans completely in a debate. Such a system was developed by IBM. On the importance of this technology, Geofrey Hinton once said in a panel discussion¹ roughly that, a more productive standard for AGI is whether an AI can convincingly wins a human in a debate.

Reasoning and argumentation are closely related, if not identical. Reasoning is usually to answer a question, and is typically a solitary activity. Argumentation, on the other hand, is usually about defending a standpoint and attacking another, and typically done between at least 2 parties. However, those differences are superficial and requires the same underlying skill! Answering a question is essentially exploring the possible stances and arguing for all of them alternatingly until the reasoning body sees a stance to be clearly more convincing. Argumentation, if done right, is just a team effort to reason on a question. Therefore, those who work on AI reasoning could benefit by learning from the body of work in computational argumentation.

What do Computational Argument’s literature has to offer? Via argument understanding², an AI can verify its own natural-language reasoning tokens by looking at the underlying argument structure in a graphical form. Using the metric of reasonableness (which is close to natural language entailment assessment, another well-developed technology), a computer can rigorously reflects on its own thoughts without hallucinating. And overtime, with more evidence, it can updates it beliefs. It is a good exercise to compare this model of beliefs with Bayesian belief network and combine ideas. And eventually, an AI well-versed in argumentation will be well-versed in self-understanding, reasoning, and truth seeking.

When an AI says things that you can’t deny, it is very hard to ignore that it is more intelligent than you.

1. It is the panel among laureats of the Queen Elizabeth Prize. ↩

2. This field is currently called Argument Mining for some weird historical reasons. I think this naming has unfortunately undermined its applications to other areas. ↩