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.

Shuyang Li, Jin Cao, Mukund Sridhar, Henghui Zhu, Shang-Wen Li, Wael Hamza, Julian McAuley. Proceedings of the 16th Conference of the European Chapter of the Association for Computational Linguistics: Main Volume. 2021.

Neural models have yielded state-of-the-art results in deciphering spoken language understanding (SLU) problems; however, these models require a significant amount of domain-specific labeled examples for training, which is prohibitively expensive. While pre-trained language models like BERT have been shown to capture a massive amount of knowledge by learning from unlabeled corpora and solve SLU using fewer labeled examples for adaption, the encoding of knowledge is implicit and agnostic to downstream tasks. Such encoding results in model inefficiencies in parameter usage: an entirely new model is required for every domain. To address these challenges, we introduce a novel SLU framework, comprising a conversational language modeling (CLM) pre-training task and a light encoder architecture. The CLM pre-training enables networks to capture the representation of the language in conversation style with the presence of ASR errors. The light encoder architecture separates the shared pre-trained networks from the mappings of generally encoded knowledge to specific domains of SLU, allowing for the domain adaptation to be performed solely at the light encoder and thus increasing efficiency. With the framework, we match the performance of state-of-the-art SLU results on Alexa internal datasets and on two public ones (ATIS, SNIPS), adding only 4.4% parameters per task.