Multi-Intent Hierarchical Natural Language Understanding for Chatbots

Multi-intent natural language understanding aims at identifying multiple user intents (or goals) in a single natural language uttering (e.g., “I want to buy tomatoes and please, remove the 2 beer cans I've added to the cart before” - here intents are “add-product: tomatoes” and “remove-product: 2 beer cans”). The goal of this work is to explore precise tagging of multiple intents and other nonstandard chatbot problems in a newly created dataset, which was created in Polish language for a real-life domain of online shopping. We propose to achieve the goal by using a hierarchical model. The achieved results are on the scale of over 80 % F1 score.