AYO - Building an Engaging Chatbot at ASOS
We compare the performance of customer intent classifiers designed to predict the most popular intent received through ASOS.com Customer Care Team, namely order related issues (e.g “Where is my order?”). These queries are characterised by the use of colloquialism, label noise and short message length. We present the results of extensive experiments for two well established classification models: logistic regression via n-grams to account for sequences in the data and recurrent neural networks that perform the extraction of sequential patterns automatically. These models are calibrated to perform above Human performance (0.93 precision), revealing a small difference in Recall of 0.05 for the neural networks (training under 1hr), and of 0.07 for the linear n-grams (training under 10’).
• Simple linear models are a judicious choice of model architecture in modern AI production systems. We then leverage recent advances in pre-trained neural networks to offer an engaging experience to 20+ fashion lovers by learning to answer Frequently Asked Questions from small training sets and by learning to reply to Chit-Chat messages from fine-tuning Blender in the fashion domain.
• Large language models (e.g GPT3) used as synthetic data generators can improve the detection rate on small volume intents. When the language model can be fine-tuned, training from public conversations on Reddit will endow the chatbot with interesting chit-chat replies.
Fabon obtained his PhD in Computer Science in 2013 “on Learning and Representation from Texts for both Emotional and dynamical Information” at the University of Pierre et Marie Curie, DAPA department (LIP6). After graduating he held a post-doctoral position at LIP6, researching interpretable models for the classification of multivariate time series’ data. At this time he grew an interest in the analysis of time series’ data, and in the Fourier transform as a mean to extract meaningful features from data. He later joined the University of Bristol as a Research Associate in 2014 investigating efficient machine learning algorithms for data streams, and researching models to study our human behaviours at a collective level via the detection of periodic patterns in the social media or in large samples of press archives. Since 2018 Fabon is Lead Scientist ConversationalAI at ASOS.com working on the development of new NLP technology for Chatbot applications.