At Existor we are attempting to build algorithms to achieve a truly natural level of human like interaction. We are doing this by acquiring and assembling enormous conversational datasets.
Our work is focused on human-computer interaction and creating tools to facilitate bot creation. We are entering the age of bots where you will be holding a conversation rather than navigating a visual interface, right now, those conversations can feel limited and robotic. Our bots are probably the most popular bots globally, with over 6 million users a month, and hundreds of millions of views on You Tube.
Machine intelligence starts with the data computers analyze. We have acquired masses of the right data and the ability to clean it.
We are using state of the art advanced machine-learning techniques and creating our own highly accurate algorithms to process the data to extract structured information from the noise.
Our initial aim was to discover data patterns at a line level. In the last few years, great progress has been made in finding patterns at a word level. For example word2vec turns words into vectors (lists of numbers) which then allow mathematical operations such as queen - king = woman - man. However, doing this for entire sentences is much more challenging. So we constructed a machine learning pipeline involving word2vec (twice) and clustering, and to some extent we succeeded. We found a mathematical relationship between some basic conversational questions and answers, such as how old are you - I am 16 years old = who are you - I am Kate. Read more about this in our machine learning section.
Our next project was to try to generate conversational replies, with the ultimate aim of producing a machine learning version of Cleverbot. We started to work with RNNLMs (recurrent neural network language models). Language models estimate the probability of different words occurring in sequence. For example, after the sequence "Hello, what is your name?", there is a high probability that the next word is "My", followed by "name", then "is". RNNLMs implement language models using neural networks, one of the core machine learning techniques. We are currently researching different variants of RNNLMs and related algorithms like sequence-to-sequence networks to see which one gives the most logically consistent replies.
Our bots participate in around 5 million conversational interactions every day. We log all those interactions, and have now gathered about 3 billion of them. This is a substantial amount of data, one of the largest corpuses of machine-human chatting data in the world.
This allows us to engineer features, evaluate our learned models, serve predictions, incorporate changes to our system and so on. Our Sentiment and Emotion analysis model was built from our data.
Evie, the popular avatar which appears at www.eviebot.com. and her first artificial companion Boibot were joined by Chimbot in 2016. Our avatars now appear as fully moving faces on desktops and mobiles. They use Flash if your browser has support, but still work without it, thanks to our own Existor Avatar Player technology.
We've created a lot of new technology to give an entertaining chatting experience. There's the learning AI, the same software that runs the text-only chat at www.cleverbot.com on graphics cards. For our avatars the AI is extended to controlling the timing and degree of facial expressions. Their movements, reactions and emotions blend and vary in surprisingly complex ways. And voices are delivered to your browser along with lip synching information to really bring the avatars life!
Machine intelligence starts with the data computers analyze. We have assembled enormous, important new datasets.
There are only a few organisations that can amass a global dataset of significance because there are massive barriers to entry.
We created our own algorithms and user interface,things you say to our avatars today may influence what it says to others in future. The algorithm chooses how to respond to you fuzzily, and contextually, the whole of your conversation being compared to the millions that have taken place before. Many people say there is no bot - that it is connecting people together, live. The AI can seem human because it says things real people do say, but it is always software, imitating people.
Existor was founded in 2008. The active Cleverbot & Evie database now has over 300 million rows, so qualifies as Big Data and we use GPU programming to search it quickly. Our other main product is Cleverscript, a scripting engine which can be built on top of our data. It allows developers to harness the power and fun of Evie and Cleverbot to create Bots, Mobile apps, apps for the Internet of Things, or website, console and PC games.
Please contact us with any enquiries about using Cleverscript, Cleverbot or Evie commercially, or any questions related with artificial intelligence, machine learning and virtual assistants.
Call +44 20 8123 7525
Write to Existor Ltd, 32 Bartholomew Street West, Exeter EX4 3BN, United Kingdom