ManMachine: Building an Efficient Chat Tool Prototype
Artificial Intelligence (AI) has been a major theme in the last decade and numerous big companies have invested a lot of effort into the technology. Within the scope of our last ti&m garage project, we too developed a small but efficient chat tool prototype for a big company in Switzerland.
It is common knowledge that most questions asked by customers are standard inquiries about the products or the company itself. This is why we wanted to automate the process and our idea for the chat was born: A customer should be able to ask basic, product-related questions through a chat included in the website. Or in other words: If a customer on the website wants to ask a common question, a bot should be able to answer it. Of course, the customer is still provided with the option to directly speak to a call center agent.
In this context, we all remember old chat bots with inept or unrelated answers that would have failed the Turing test miserably. Therefore, to achieve the best user experience possible, we decided to solve this issue by applying the newest developments in the field of Artificial Intelligence.
Because of the nature of a garage project – it has a really short development period – it was not possible to implement the results of our research ourselves. The other option was to find a product on the market that would be a match for our problem. However, the market for AI solutions is constantly growing and can be confusing. This is why we formulated three basic criteria for our required system:
- It should be possible to implement in six weeks.
- It should be based on the newest findings in AI research.
- It should provide the best results possible, in more than one language.
One of our first guesses was IBM Watson, one of the top cognitive systems that has thought processes similar to a human. Unfortunately, Watson did not pass our criteria. Firstly, it would not have been possible to implement the technology into our system within only six weeks – it already took us three weeks to get in touch with IBM. Secondly, the technology Watson is based on wasn’t an exact fit to what we had in mind and thirdly, we did not have the possibility to test Watson on our own. All in all, Watson was out of the race.
The next promising product we looked at was FreeHAL, an open source self-learning dialog bot based on a semantic net. After a first look and a rough implementation, FreeHAL was looking really good. The technology was simple to implement into our product and it could communicate in German as well as English. However, after playing around with FreeHAL, we were disappointed as the AI was not quite as good as we had initially expected. Most of the answers turned out to be unrelated to the initial question.
We had to find another solution. Fortunately, we were able to find an interesting product called wit.ai, a cloud based AI for natural language processing grounded in the latest research. However, the first impression was not promising as the examples on the website where not exactly within the scope of our problem. Nevertheless, by digging a little deeper, we managed to find exactly what we needed and the implementation into our system turned out to be simple.
We only had to call the API of wit.ai with the question asked by the user and the cloud service would then return an answer. On the service page of wit.ai, we were able to define a basic catalogue of topics and answers. After testing the product, we got some really nice results and decided to go with wit.ai for our project.
All in all, it was a pleasure to implement and play with all these new products in the field of artificial intelligence and we noticed that in most cases, the implementation of complex systems are not necessary to proof a concept. Working in a ti&m garage was fun for all the participants and lead to interesting insights, cool products and a good time all around.
Are millennial men the right target group for a Pampers marketing campaign? Of course - if they have children. Anticipating what the customer wants is like mind reading. This is now possible by analyzing customer behavior and situations.Mehr erfahren
Die Kreativität stand im Mittelpunkt der fünften Ausgabe des ti&m code camp in den Flumserbergen. Über 30 Entwickler und 1 Designer stellten sich verschiedenen Challenges und versuchten diese innerhalb von 30 Stunden zu lösen.Mehr erfahren
Imagine a world without projects, project leaders and product owners. Imagine a broader product definition, a larger backlog and multiple feature teams all working on the same effort. In this article, ti&m's CTO Martin Fabini explores an organizational design based on LeSS, the framework for scaling agile software development to multiple teams.Mehr erfahren