Humanizing Artificial Intelligence. Another step closer to our vision.

Over the past 10 years, we’ve developed a unique capability in Artificial Intelligence by teaching language to machines sequentially, as a human learns a language.

Pat.ai has reached another milestone in achieving our vision to humanize conversation with machines, with the announcement of our final benchmarking against Facebook’s proposed ‘Intelligent Dialogue’ tests: an impressive success rate of 100% on eight of nine tasks.

I’ll get to that ninth task later. But first I want to share why our approach, based on a modern linguistic framework, Role and Reference Grammar (RRG), and the brain-based Patom theory, is creating truly conversational AI.

The Pat approach: how it’s different

Traditional approaches go down the parsing road or statistical analysis, or combine the two. The problem is that both parsing and statistics are blunt instruments.

Google’s Parsey McParseface opts for the parsing route. However, it’s possible for robots to generate perfectly parsed nonsense — and they often do.

The statistical approach can do the same, with the co-location algorithm leading to answers that may be statistically correct, but don’t give the full picture.

Both parsing and statistical methods will suffer from diminishing returns as they require ever-greater data sets to feed them. Instead, we’re teaching language to machines using ‘matched patterns’ or ‘link sets’ — the Patom model.

Pattern matching is not a banal set of five symbols in sequence, but rather a match of a list of sets. In short, we break down language by meaning and then can generate responses in natural language rather than keywords.

The key advantage is that we don’t need massive data sets to scale up to conversational AI. And the potential for Natural Language application is staggering. As Pat’s knowledge increases, its potential for API applications grows. One of the most obvious starting points for us, given our natural language-based approach, is English language learning. We’ve now reached another milestone in achieving our vision to humanize conversation with machines, with the announcement of our final benchmarking against Facebook’s proposed ‘Intelligent Dialogue’ tests: an impressive success rate of 100% on eight of nine tasks.

Pat for language learning and education

If you’ve ever tried to learn a second language, you’ll know that learning the words is the easy part. It’s how you string them together that matters, and the confidence you have in using that language.

If a language student is shy, or afraid of using incorrect grammar, they won’t get enough practice. Pat can overcome this as a ‘non-judgemental robot’. By modelling the correct conversational interaction, it can guide you to learn using a similar sequence to a child learning their first language.

With its next-generation Natural Language Understanding (NLU) API, Pat can deliver ‘Meaning-as-a-Service’ by processing natural language and human conversations into structured output about its meaning. This allows real human-like conversation.

Our scorecard to date

Pat is not learning patterns about language meaning from big data. Rather, Pat builds knowledge on language (just like a human) by progressively learning the way words are combined, regarding real objects, people, processes and events in context.

So how is it doing? We’ve now attempted nine of the Facebook FAIR (bAbI) tasks, including Single Supporting Facts, Three Argument Relations, Yes/No Questions, Counting, Simple Negation, Conjunction, Basic Coreference, Conjunction and Compound Coreference.

We passed with 100% accuracy on eight of nine tasks. For Task 5, where we achieved a score of 99.5%, we aren’t concerned as the ‘failing’ tests represent an error in the test result due to an error in the input file, and an error in interpretation of a logical structure. We saw no point in recoding to produce the ‘wrong’ result simply to pass the test.

You can read a full breakdown of our results in this paper.

bAbI tests can be limiting, given our unique approach. With a few modifications — such as replacing a keyword response with an English language response — I believe they can more closely approximate true human-like interaction and responses.

Here’s an example of more natural responses to a given question, rather than stored text answers.

INPUT — The woman who went to the kitchen went to the garden. Who is in the garden?

Answer:

a. The woman who went to the kitchen, or It is the woman who went to the kitchen.

b. The woman who went to the kitchen is.

c. The woman who went to the kitchen is in the garden.

The challenge of NLU is in understanding the true meaning of sentences and conversations, not simply translating words or guessing the intent of a question.

For Pat, the only learning step is definition of patterns. Pat then has that knowledge forever, then builds on the acquired knowledge.

This has an important additional benefit. We need to be able to trust AI responses and decisions to realise the full potential of machine learning — but if you build a system entirely with deep learning and something goes wrong, it’s hard to know why and that makes it hard to debug.

However, with our meaning-based system every matched pattern and promoted phrase can be seen and traced back.

What’s next for Pat?

Our meaning- and semantic-based approach will have enormous ramifications for developers, allowing them to build applications with a true conversational interface.

Once we see how Pat performs in a language learning environment, we can extend its capabilities to manage home automation apps beyond today’s stilted commands.

With Pat, your chatbot could be a more intelligent virtual assistant, able to understand complex instructions or queries. For example, you could ask Pat to “Turn on the lights and the microwave, no, I mean the oven, and have music playing.”

It’s a big leap from today’s pre-programmed, scripted responses — and another step closer to our vision of human conversations with machines.

If you’re interested in waitlisting for beta access to our API please register here or sign up for further updates.