The UK arm of German supermarket company Lidl has launched a Facebook Messenger-based chatbot that is designed to help customers choose the right wine to compliment their food or occasion. The chatbot is named Margot and can interact with shoppers in four main ways: providing answers to frequently asked questions, pairing wines with food, finding a suitable wine, and taking a quiz (about wine, naturally). Lidl partnered with customer engagement company Aspect to go from conception to completion of the project in just twelve weeks.
Tobias Goebel, director of emerging technologies at Aspect, told InfoQ that they were able to achieve this velocity due to a combination of experience in the market and use of their proprietary platform, Aspect CXP, which includes a component based on a unique approach to the artificial intelligence topic of Natural Language Understanding (NLU) with multi-language support.
The Aspect CXP platform comprises full lifecycle tools such as a Developer IDE, a database with REST API, administration and analysis tools and a runtime Java application server. Embedded in all this is the Aspect NLU engine.
Goebel explained that the common approach to determining user intent from an incoming message is to use machine learning to train a neural network using huge amounts of manually tagged data. While effective for use cases such as the Lidl one, there is usually not enough data available to train the network effectively. Even when there is enough, once completed, the network is trained in only one language and must be retrained for each subsequent language.
Aspect’s NLU framework turns this around by maintaining a language-independent lexicon of hundreds of thousands of semantic concepts. These concepts are linked to their instances across 14 different languages (including complete dictionaries, morphology, semantics, and syntax for each language, including emojis). This means that a few abstractions of the key semantic elements of a user message are enough to build rules that can immediately understand hundreds of variations of any given question, and the learning can then be applied across all 14 languages at once without the need for retraining.
While its wine chatbot has been launched in the UK, the attraction of this approach to a pan-European trader like Lidl is apparent, and Goebel says that it was this approach to NLU that was an important factor in leading Lidl to choose them as partners.
To build the chatbot, Aspect partnered with application development firm 2-steps-ahead. Aspect provided project management, architecture and linguistic expertise and 2-steps-ahead added the programming skills to the effort.
Lidl has achieved much praise recently for their wine selections, winning awards both for red and sparkling wines. However, the knowledge and associated purchasing advice resides with their full-time sommelier who cannot be in all of the supermarket’s 800+ stores at once, nor available to everyone who connects with the contact centre. The chatbot represents the scaling of the sommelier’s expertise so customers can get instant access to wine advice by conversing or questioning the bot’s Conversational UI through Facebook messenger.
Margot mixes both menu-driven interactions and natural language dialogs, a style of interaction that was a deliberate design decision. Goebel stated that users still expect some guidance in the form of menus when conversing with a computer, however in the future, as chatbot experiences become more widespread, the need for such hybrid UIs go away.
If you are interested in trying the winebot yourself, either for research or recreational purposes, go to the Lidl UK Facebook page and click on 'send message'. Once on Facebook Messenger, select 'get started' from the menu and then 'start the wine bot.'