Low or no resource domain adaptation for task specific semantic parser

Semantic parsers are a critical component for voice agents (like Amazon's Alexa or Apple's Siri). Deep learning models can convert natural language to semantically parsed texts but requires large dataset to learn from individual topics (or domain) where they must do predictions. Since there are no limit to the possible number of topics for a conversational bot to operate, hence, developing a robust semantic parser becomes challenging. This bottleneck can be overcome by different domain adaptation techniques where the semantic parser trains on large source domain data and can easily adapt to a new target domain with very little target domain training data. But the effect of choosing a source domain has on the prediction capability of the model on target domain is very less explored and that is what we study in this research work. We concluded from this study that the choice of source domain is critical for domain adaptation tasks and can significantly increase or decrease the prediction accuracy of semantic parser models when operating on a specific target domain. We also identified a method to choose the best fit source domain for a specific target domain using cosine similarity score. Furthermore, we propose a novel method of designing a semantic parser model without any target domain training data and no target domain training and yet the model would be able to make good predictions in the target domain.