Context Memory Networks for Multi-objective Semantic Parsing in Conversational Understanding

The end-to-end multi-domain and multi-task learning of the full semantic frame of user utterances (i.e., domain and intent classes and slots in utterances) have recently emerged as a new paradigm in spoken language understanding. An advantage of the joint optimization of these semantic frames is that the data and feature representations learnt by the model are shared across different tasks (e.g., domain/intent classification and slot filling tasks use the same feature sets). It’s important that the model should learn to pay attention to global and local aspects of the utterances while learning to map the entire utterance to an intent class and tag each word with a slot tag. We introduce the Context Memory Network (CMN), a neural network architecture which specifically focuses on learning better representations as attention vectors from past memory to be reasoned with for the end task of jointly learning the intent class and slot tags. The utterances trigger a dynamic memory network, which learns attention based representation for each word by allowing the model to condition on the list of related phrases in the form of memory networks. These representations are then provided to a new multi-objective long short term memory network (LSTM) to infer the intent class and slot tags. Our empirical investigations on CMN show impressive gains over the end-to-end LSTM baselines on ATIS dataset as well as two other humanto-machine conversational datasets.