Entity Contrastive Learning in a Large-Scale Virtual Assistant System

Abstract

Conversational agents are typ of domain (DC) and intent cla identify the general subject a longs to and the specifc action achieve. In addition, named en (NER) performs per token lab specifc entities of interest in a s We investigate improving joi models using entity contrasti attempts to cluster similar ent a learned representation space full virtual assistant system trai contrastive learning to a baseline system that does not use contrastive learning. We present both offine results, using retrospective test sets, as well as online results from an A/B test that compared the two systems. In both the offine and online settings, entity contrastive training improved overall performance against baseline systems. Furthermore, we provide a detailed analysis of learned entity embeddings, including both qualitative analysis via dimensionality-reduced visualizations and quantitative analysis by computing alignment and uniformity metrics. We show that entity contrastive learning improves alignment metrics and produces well-formed embedding clusters in representation space.