A Concurrent Intelligent Natural Language Understanding Model for an Automated Inquiry System

Abstract

The work is intended to tackle a vital field that lies at the intersection of speech processing and natural language processing: Spoken Language Understanding (SLU). Its idea is to understand the essence of machine-directed human speech in order to facilitate its further processing and take on board its cognitive impact. The proposed system is CIDIS -Concurrent Intelligent Model for Dialogue Act Classification, Intent Detection and Slot Filling, that uses a deep concurrent multi-task paradigm to perform the three fundamental tasks of the SLU domain: Dialogue Act Classification, Intent Detection and Slot Filling. Since the model is orchestrated in a multi-task fashion, every task interacts with the other to have a global understanding of the input query. It follows an intelligent encoding strategy involving concatenation of the query’s BERT and CharCNN embedding to handle all possible edge cases and ambiguities involved in human speech queries. This intelligent encoding is passed through a Stacked BiLSTM architecture followed by task-specific attention layers. The three supplementary outputs are in turn fed to the final module that generates the expected query response in real-time based on the dialogue act, intent and slot. The developed models are evaluated against standard benchmark datasets like ATIS, TRAINS and FRAMES and the achieved state-of-the-art performances are eventually tabulated.