Alexa Teacher Model: Pretraining and Distilling Multi-Billion-Parameter Encoders for Natural Language Understanding Systems

TL;DR

This paper explores large-scale pretraining and distillation of multi-billion-parameter encoders for natural language understanding, demonstrating improved performance on NLU tasks and virtual assistant systems through in-domain data and model compression.

Contribution

It introduces a large-scale pretraining and distillation pipeline for multi-billion-parameter encoders tailored for NLU, with significant performance gains from in-domain data and model size reduction.

Findings

Teacher models perform comparably to XLM-R and mT5 on XNLI.

In-domain pretraining improves intent and slot filling error rates.

Distilled models outperform smaller baselines on NLU tasks and user dissatisfaction metrics.

Abstract

We present results from a large-scale experiment on pretraining encoders with non-embedding parameter counts ranging from 700M to 9.3B, their subsequent distillation into smaller models ranging from 17M-170M parameters, and their application to the Natural Language Understanding (NLU) component of a virtual assistant system. Though we train using 70% spoken-form data, our teacher models perform comparably to XLM-R and mT5 when evaluated on the written-form Cross-lingual Natural Language Inference (XNLI) corpus. We perform a second stage of pretraining on our teacher models using in-domain data from our system, improving error rates by 3.86% relative for intent classification and 7.01% relative for slot filling. We find that even a 170M-parameter model distilled from our Stage 2 teacher model has 2.88% better intent classification and 7.69% better slot filling error rates when compared to the 2.3B-parameter teacher trained only on public data (Stage 1), emphasizing the importance of in-domain data for pretraining. When evaluated offline using labeled NLU data, our 17M-parameter Stage 2 distilled model outperforms both XLM-R Base (85M params) and DistillBERT (42M params) by 4.23% to 6.14%, respectively. Finally, we present results from a full virtual assistant experimentation platform, where we find that models trained using our pretraining and distillation pipeline outperform models distilled from 85M-parameter teachers by 3.74%-4.91% on an automatic measurement of full-system user dissatisfaction.