Conversational agents on smart devices need to be efficient concerning latency in responding, for enhanced user experience and real-time utility. This demands on-device processing (as on-device processing is quicker), which limits the availability of resources such as memory and processing. Most state-of-the-art Dialog State Tracking (DST) systems make use of large pre-trained language models that require high resource computation, typically available on high-end servers. Whereas, on-device systems are memory efficient, have reduced time/latency, preserve privacy, and don’t rely on network. A recent approach tries to reduce the latency by splitting the task of slot prediction into two subtasks of State Operation Prediction (SOP) to select an action for each slot, and Slot Value Generation (SVG) responsible for producing values for the identified slots. SVG being computationally expensive, is performed only for a small subset of actions predicted in the SOP. Motivated from this optimization technique, we build a similar system and work on multi-task learning to achieve significant improvements in DST performance, while optimizing the resource consumption. We propose a quadruplet (Domain, Intent, Slot, and Slot Value) based DST, which significantly boosts the performance. We experiment with different techniques to fuse different layers of representations from intent and slot prediction tasks. We obtain the best joint accuracy of 53.3% on the publicly available MultiWOZ 2.2 dataset, using BERT-medium along with a gating mechanism. We also compare the cost efficiency of our system with other large models and find that our system is best suited for an on-device based production environment.
Sentence completion detection (SCD) is an important task for various downstream Natural Language Processing (NLP) based applications. For NLP based applications, which use the Automatic Speech Recognition (ASR) from third parties as a service, SCD is essential to prevent unnecessary processing. Conventional approaches for SCD operate within the confines of sentence boundary detection using language models or sentence end detection using speech and text features. These have limitations in terms of relevant available data for training, performance within the memory and latency constraints, and the generalizability across voice assistant domains. In this paper, we propose a novel sentence completion detection method with low memory footprint for On-Device applications. We explore various sequence-level and sentence-level experiments using state-of-the-art Bi-LSTM and BERT based models for English language.