Cross-Domain Training for Goal-Oriented Conversational Agents

Proceedings of Recent Advances in Natural Language Processing, pages 142–150, Varna, Bulgaria, Sep 2–4, 2019. https://doi.org/10.26615/978-954-452-056-4_017 142 Cross-Domain Training for Goal-Oriented Conversational Agents Alexandra Bodirlau, Stefania Budulan, Traian Rebedea University Politehnica of Bucharest, Romania alexandra.bodirlau@stud.acs.upb.ro {stefania.budulan, traian.rebedea}@cs.pub.ro Abstract Goal-Oriented Chatbots in fields such as customer support, providing specific infor- mation or general help with bookings or reservations, suffer from low performance partly due to the difficulty of obtaining large domain-specific annotated datasets. Given that the problem is closely related to the domain of the conversational agent and that data belonging to a specific domain is difficult to annotate, there have been some attempts at surpassing these chal- lenges such as unsupervised pre-training or transfer learning between different do- mains. A more thorough analysis of the transfer learning mechanism is justified by the significant boost of the results demon- strated in the results section. We de- scribe extensive experiments using trans- fer learning and warm-starting techniques with improvements of more than 5% in relative percentage of success rate in the majority of cases, and up to 10x faster con- vergence as opposed to training the system without them. 1 Introduction Goal-Oriented Conversational Agents (GO Chat- bots) are seeing increased use to help users to achieve predetermined goals, but they can han- dle only very simple tasks, such as playing songs, searching information, set alarms or reminders. Building a dialogue agent to fulfill complex tasks remains one of the fundamental challenges for the Natural Language Processing (NLP) community and Artificial Intelligence (AI) in general. There are two dominant approaches for solv- ing this problem. The first one relies on (fully) supervised learning, e.g. using sequence-to- sequence (Sutskever et al., 2014) models, encod- ing a user’s utterance and its context to decode the answer provided by the chatbot. However, this method does not explicitly allow to locate and make use of specific information such as entity recognition (e.g. a person’s workplace) and re- quires large amounts of data in order to flawlessly extract and process particular pieces of informa- tion relevant for the task at hand, usually a manda- tory requirement for GO Chatbots. The second category entails partitioning the di- alog system into smaller subsystems, usually im- plemented and trained separately. An example of such a system (Li et al., 2017) consists of several components: Natural Language Under- standing, Dialog Manager, and Natural Language Generation. The Dialog Manager is often im- plemented with the aid of reinforcement learning (RL) based techniques, for instance using Deep Q- Nets (DQN) (Mnih et al., 2015) and having the main goal of learning the policy on account of which the agent will be able to provide answers. The first approach is used with more favourable outcomes in the case of open-domain dialogue systems (Serban et al., 2016) than in closed- domain dialogue systems (Peng et al., 2017), be- cause it does not require a method to reward the accomplishment of the task. Instead, the suc- cess of the conversation resides in the engage- ment of the …

Content truncated to 500 words