Friday, February 1, 2019

The Alquist Technology Presentaion

We hosted an Alquist meeting for our friends and partners this Wednesday (1/30/19). Our goal was to share the Conversational AI technology of the twice in row second (Alexa Prize) Alquist. Here is a copy if you have any question about technology, or if you want to use it, or if you want to join us do not hesitate and drop us an email. Let me give you a short summary.

In the introduction, we pointed out the recent advances in AI. Deep learning improved the speech recognition accuracy. Google is one percentage point better than humans in 2017. Advances in speech accuracy enabled the new type of devices the intelligent speakers. The best examples are Alexa and Google Home. They are the target platforms for Alquist. The initial presentation slides explained the underlying architecture and how the intelligent speakers carry on the conversation. 

In the following part of the presentation, Jan explained the very basics of the Alquist Natural Language Understanding (NLU). He has shown some of the details of our keyword extraction and named entity recognition implementation. Next, he has very quickly mentioned the dialog acts recognition and the profanity detector. Profanity detection is a challenging problem essential to keep the bot pleasant to talk to. See what kind of language we need to fight.

When people meet they start chatting. There is an endless number of conversational topics. We tried to teach Alquist the most frequent ones. The current version has 30 topics, such as sport, politics, travel, etc. The Dialog Designer (DD) plays the most important role in making the topic entertaining and interesting. He invents the dialogs. Usually, the bot starts the conversation uttering an initial sentence. The user replies. We call the human bot exchange a turn. Each topic is then a set of a large number of turns. The DD prepares the turns for Alquist in the form of templates.

Petr is responsible for a dialog manager (DM) ordering the turns in a sensible dialog. In his MSc dissertation, he has developed an original hybrid DM, which we use in Alquist. The offline processing DM has two parts a graphical UI and the training set generation. The DD generates a dialog tree (composition of turns) from graphical dialog objects for a topic. The second part of the offline DM processes the topic dialog turns, and it automatically creates a rich set of sentences, the training set. We use the set to train an LSTM neural net, which predicts the next sentence on a topic. The production DM runtime is a mixture of several rules and an NN for each dialog topic. Overall such a system handles a large number of different sentences. The hybrid design makes the Alquist conversation much more robust.

We also wanted to show how to use the Alquist technology for other modality. With the Rebel & Glory, we have put together an interactive movie about Alquist. You will learn how we have chosen the name Alquist, and you can learn about Karel Capek, watch it. We also have prepared a short interactive invitation for our presentation. The team has first shot the clip and wrote the dialogs. Ondrej even acted as a guide. This movie may inspire you how to author interactive ads or simple product introductions. We are looking forward to working with creative teams on the conversational part of entertaining clips. If you have an interesting idea, let us know. In the presentation, we have pointed out some of the unexpected questions. Try them! 

The last speaker was Ondrej who is our DD. He is inventing and designing the dialogs. The dialogs have two parts the Alquist messages, and the user replies. The DD needs to work on both. He predicts what will be the user interested in and he generates the templates to react on the user's requests. Language is very complicated. The user does not want to hear repeated or tiresome replies. The conversation must be interesting. Therefor Alquist creates responses form templates as well as using generative algorithms. Ondrej has also overviewed some of the underlying problems and linguistic peculiarities he had encountered.

The Alquist presentation has attracted almost eighty people to our beautiful top floor presentation room of the CIIRC building. We were very excited by a fruitful discussion with plenty of questions at the end of our meeting. If you are interested, you may go through the presentation too. Let us know if you like it. Enjoy and if you have any questions let us know. We would also be happy to hear your ideas for new applications.