Building Advanced Dialogue Managers for Goal-Oriented Dialogue Systems
This addresses the data scarcity issue for developers of goal-oriented chatbots, though it is incremental as it builds on existing transfer learning techniques.
The paper tackled the problem of low data availability for training dialogue policies in goal-oriented dialogue systems by introducing a transfer learning method, which improved the bot's success rate by 20% for distant domains and more than doubled it for close domains, while also speeding up policy learning by 5 to 10 times.
Goal-Oriented (GO) Dialogue Systems, colloquially known as goal oriented chatbots, help users achieve a predefined goal (e.g. book a movie ticket) within a closed domain. A first step is to understand the user's goal by using natural language understanding techniques. Once the goal is known, the bot must manage a dialogue to achieve that goal, which is conducted with respect to a learnt policy. The success of the dialogue system depends on the quality of the policy, which is in turn reliant on the availability of high-quality training data for the policy learning method, for instance Deep Reinforcement Learning. Due to the domain specificity, the amount of available data is typically too low to allow the training of good dialogue policies. In this master thesis we introduce a transfer learning method to mitigate the effects of the low in-domain data availability. Our transfer learning based approach improves the bot's success rate by $20\%$ in relative terms for distant domains and we more than double it for close domains, compared to the model without transfer learning. Moreover, the transfer learning chatbots learn the policy up to 5 to 10 times faster. Finally, as the transfer learning approach is complementary to additional processing such as warm-starting, we show that their joint application gives the best outcomes.