Lifelong Chatbots for Customer Support

Abstract

With advances in machine learning, chatbots are gaining increasing popularity in different domains, specifically bots for customer support. The benefit of these bots is that they provide customers instant responses anytime and are able to save cost and time while enhancing customer service. However, one of the limitations of existing work on customer support bots (CSB) is that models developed for CSBs are either static or updated infrequently as more training data becomes available. Additionally, updating the learned models often needs human intervention. To address the limitations of discontinuous CSB learning and the required manual intervention, we propose the design of an automated Lifelong Learning CSB (LL-CSB) that can continuously learn and adapt its answers based on new knowledge acquired from different sources like support forums and discussions on the web. The proposed design of the LL-CSB addresses several challenges for lifelong learning (LL) including continuous tasks: extraction of new knowledge, update to existing knowledge, integration, and updates to LL-CSB response model. We propose to setup the CSB as an Information Retrieval (IR) system where the user asks for the solution of a technical problem and the response is a potential solution from the support knowledge base. To facilitate continuous knowledge updates, we design knowledge base of the LL-CSB as a knowledge graph (KG) consisting of <problem, solution> pairs. For continuous knowledge update, we propose a lifelong learning algorithm capturing rules for extraction of new knowledge from the web and checking whether a problem already exists in the knowledge base before adding it with the corresponding solution and linking it to similar problems in the KG. The similarity matching is based on a computationally low-cost and fast method using hashing TF-IDF vectorizer. For continuous updates of the LL-CSB response model and instead of retraining with the whole dataset with new knowledge, a TF-IDF hashing solution is used to enable fast additions of vectors representing new problems. Finally, we implemented the LL-CSB with real data from CISCO corporation’s network support. In addition to the LL-CSB design and implementation, we proposed a CSB-specific simulator for evaluation of the LL strategy. The simulator models real-time updates of knowledge over time and evaluates the performance of customer queries over simulated time. To evaluate the proposed design, we created a validation dataset comprised of real question-answer discussions crawled from CISCO’s online support forum. Using the simulator, our experimental results demonstrated the superiority of LL-CSB with up to 3.18X improvement in F1 score compared to CSB without LL. The simulator also showed how the baseline CSB suffered from a drop in recall and precision with more queries when it does not take advantage of lifelong learning.