Abstract:
Despite ground breaking work in the academia, current state-of-the-art work in
goal-oriented conversational-agents has not been able to fulfil the demand of the industry
for multi-domain multi-lingual, adaptable, dialogue systems. Data-intensive nature of
deep learning models used in dialogue state tracking (DST) module, which is a core
component of the goal-oriented dialogue architecture, and the lack of large labelled
dialogue corpus for state tracking are two main factors which have hindered the progress.
We identified, modeling with separate natural language understanding (NLU) module
and joint modeling of dialogue state tracker with NLU as the two main approaches for
state tracking, and accordingly made two major contributions. First, we propose a novel
meta-learning algorithm for intent detection and slot-filling tasks, focusing on models
with separate NLU. Our work empirically demonstrates that the proposed meta-learning
approach is capable of learning a meta-parameter(prior) from similar, but different tasks.
Compared to the random initialization, which regular supervised learning algorithms
rely on, proposed method significantly improves the accuracy in both intent detection
and slot-filling tasks in few-shot (5-way 1-shot and 5-way 2-shot) settings. Further, our
effective use of meta-learning for intent detection and slot-filling opens up new line of
research for DST. Second, we systematically review the progression of joint NLU/DST
models with special emphasis on their ability to generalize and adapt to new domains
and languages.
