Zero Shot Intent Classification Using Long-Short Term Memory Networks

We describe a zero shot approach to intent classification that allows for the identification of intents that were not present during training. Our approach makes use of a Long-short Term Memory neural network to encode user queries and intents and uses these encodings to score previously unseen intents based on their semantic similarity to the queries. We test our model on intent classification in a personal digital assistant and show an improvement of 15% over a strong baseline. We also investigate the effect of adding a few training samples for the previously unseen intents in a few shot learning setting and show improvements of up to 16% over the baseline method.