Artificial intelligence is a much used (and hyped) term, as is the concept of machine learning. How can a machine learn? Doesn’t it have to be programmed to follow a decision tree and do exactly as it is told? That is true of ‘expert systems’ but machine learning is designed to produce outcomes that it has been ‘taught’, sometimes reinforced many times, until it reaches a point where it deduces from its learning experience what is required in a new situation.
There, I’ve managed to get this far without referring to algorithms, which require an explanation of their own. So pretend I haven’t mentioned them. Instead, let me share with you a non-human learning experience, which I’ve had the pleasure to observe at close quarters in recent times. In order to do that, I would like to introduce you to my dog Gus – see his photo above – otherwise known as the firm’s Chief Welfare Officer, because he calms us down when we’re feeling stressed (at least when he’s not the cause of it).
Gus came into our lives as a puppy, with no knowledge of the world outside the litter he was born in. The early days were testing for him and for me, and one of the most fascinating aspects of this was the way in which he learned what various commands mean.
Machine learning vs dog training… Sit!
One of the most common demands a dog needs to understand is to ‘sit’, usually to stop him getting over excited or to avoid danger. This is achieved largely by offering treats (notwithstanding anti-bribery legislation) for a job well done. Usually, this is perfected in the home, which Gus achieved, so it comes as a surprise when, immune system being given the all clear, the dog is taken out of the home environment for the first time, no heed at all is taken of the command.
Imagine my consternation, when reaching the roadside and confidently telling Gus to sit, he completely ignored me. I put more emphasis on the command (others might call this shouting), which instead produce a perplexed look. It was only on going back to the books that I discovered that dogs understand commands in context. So he had learned that ‘sit’ when commanded in the home, meant he had to sit, but it had no meaning anywhere else.
Back out again, I spent more time on training and after a while, and a spike in the treats budget, Gus worked out that ‘sit’ means the same out and about as it does at home. He was first taught the basic principle and the learning was then reinforced by further training.
Testing, testing, testing!
There then came a further test by way of taking Gus to the local pub for the first time (he likes the water there). It was quiz night and very busy, so I asked him to sit. This time, he didn’t ignore me, but he didn’t immediately oblige either. Instead of just a vacant look in his eyes, there appeared something to suggest that some cognition was occurring. Still looking at me for reassurance, he lowered himself to the floor in the manner of a helicopter landing carefully onto a ship at sea.
Sure enough, Gus was lavished with praise and more treats and he looked very pleased with himself. I was delighted too, because he had provided me with the content for this article: Just like Gus, machine learning software can be taught. That teaching can be reinforced, and, by comparing previous commands with similar commands given in new situations, the software works it out for itself without being expressly told what is required.
We’re putting these principles of machine learning through their paces with the legal chatbots that we’re developing. The chatbots can answer legal questions and then act as a lead generator. We have to feed a great many questions on a particular field of law (there is chatbot for each type), and the software decides on a balance of probabilities what it is being asked, and then matches the stored answer to that question. As it is being built, it is corrected each time it gets it wrong, until it has had enough tuition to perfect the probability assessment. Machine learning in action!
So, thank you Gus, now let’s try ‘drop’ again…
To find out more – or if you’d like to discuss chatbots for your own particular field of law – contact us on 01243 859605 or firstname.lastname@example.org. To stay up to date, why not Follow Us on LinkedIn, or signup to receive Legal Workflow Decoded, our regular news service direct to your mailbox.