Alexa Teacher Model: Pretraining and Distilling Multi-Billion-Parameter Encoders for Natural Language Understanding Systems
Open accessAbstract
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.
Integrity signals
Registry
Cites
Statistics
Claims the paper makes
Claims other papers make about this one
Each is checked against this paper’s own text: does it say what the citing paper says it does?
The gap between the high accuracy of NLP models and users' reluctance to use voice assistants for complex tasks may be explained by differences in how users and NLP practitioners evaluate model success.
“As voice assistants increasingly rely on large language models [24, 47], we believe the gap between the high accuracy of these models and users' reluctance to use voice assistants for complex tasks may be explained by differences in how users and NLP practitioners evaluate the success of a model.”
From A Mixed-Methods Approach to Understanding User Trust after Voice Assistant Failures (2023) · cites this paper as [24]
Cites
BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding
Devlin, J., Chang, M.-W., Lee, K., & Toutanova, K. (2019) · Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long and Short Papers) · doi:10.18653/v1/n19-1423 · Cited by 48,293 · see citing works
Exploring the Limits of Transfer Learning with a Unified Text-to-Text Transformer
Raffel, C., Shazeer, N., Roberts, A., Lee, K., Narang, S., Matena, M., Zhou, Y., Li, W., & Liu, P. J. (2019) · arXiv · doi:10.48550/arxiv.1910.10683 · Cited by 10,929 · see citing works
DistilBERT, a distilled version of BERT: smaller, faster, cheaper and lighter
Sanh, V., Debut, L., Chaumond, J., & Wolf, T. (2019) · arXiv · doi:10.48550/arxiv.1910.01108 · Cited by 4,653 · see citing works
Don’t Stop Pretraining: Adapt Language Models to Domains and Tasks
Gururangan, S., Marasović, A., Swayamdipta, S., Lo, K., Beltagy, I., Downey, D., & Smith, N. A. (2020) · Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics · doi:10.18653/v1/2020.acl-main.740 · Cited by 2,103 · see citing works
Scaling Laws for Neural Language Models
Kaplan, J., McCandlish, S., Henighan, T., Brown, T. B., Chess, B., Child, R., Gray, S., Radford, A., Wu, J., & Amodei, D. (2020) · arXiv · Cited by 1,749 · cited 1× · see citing works
“[19] proposed a power-law scaling relationship between parameter count and performance, and subsequent papers empirically confirmed this relationship for very large models, including [30] which trained models up to 1.”
BackgroundSmith, S., Patwary, M., Norick, B., LeGresley, P., Rajbhandari, S., Casper, J., Liu, Z., Prabhumoye, S., Zerveas, G., Korthikanti, V., et al. (2022) · arXiv · doi:10.48550/arxiv.2201.11990 · Cited by 286 · see citing works
Muppet: Massive Multi-task Representations with Pre-Finetuning
Aghajanyan, A., Gupta, A., Shrivastava, A., Chen, X., Zettlemoyer, L., & Gupta, S. (2021) · Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing · doi:10.18653/v1/2021.emnlp-main.468 · Cited by 183 · see citing works
A Study of BFLOAT16 for Deep Learning Training
Kalamkar, D., Mudigere, D., Mellempudi, N., Das, D., Banerjee, K., Avancha, S., Vooturi, D. T., Jammalamadaka, N., Huang, J., Yuen, H., et al. (2019) · arXiv · doi:10.48550/arxiv.1905.12322 · Cited by 169 · cited 1× · see citing works
“Another way to mitigate the stability issues is to use BFLOAT16 [18], which was not available in Deepspeed at the time of our experiments.”
MethodImproved Knowledge Distillation via Teacher Assistant
Mirzadeh, S. I., Farajtabar, M., Li, A., Levine, N., Matsukawa, A., & Ghasemzadeh, H. (2020) · Proceedings of the AAAI Conference on Artificial Intelligence · doi:10.1609/aaai.v34i04.5963 · Cited by 101 · cited 1× · see citing works
“Therefore, we distill the pretrained teacher models in two phases with a teacher assistant setup [26, 44].”
MethodStanza: A Python Natural Language Processing Toolkit for Many Human Languages
Qi, P., Zhang, Y., Zhang, Y., Bolton, J., & Manning, C. D. (2020) · arXiv · doi:10.48550/arxiv.2003.07082 · cited 1×
“For each example, we use the Stanza tagger [29] to identify a noun word, then mask all subword tokens for that word.”
Method
Flagged first, then by citations
Cited by
Sorted by citations
A Comprehensive Overview of Large Language Models
Naveed, H., Khan, A. U., Qiu, S., Saqib, M., Anwar, S., Usman, M., Akhtar, N., Barnes, N., & Mian, A. (2025) · ACM Transactions on Intelligent Systems and Technology · doi:10.1145/3744746 · Cited by 1,257 · see citing works
A Mixed-Methods Approach to Understanding User Trust after Voice Assistant Failures
Baughan, A., Wang, X., Liu, A., Mercurio, A., Chen, J., & Ma, X. (2023) · Proceedings of the 2023 CHI Conference on Human Factors in Computing Systems · doi:10.1145/3544548.3581152 · Cited by 32 · see citing works
Efficient Fine-Tuning Large Language Models for Knowledge-Aware Response Planning
Nguyen, M., Kishan, K. C., Nguyen, T., Chadha, A., & Vu, T. (2023) · Lecture Notes in Computer Science · doi:10.1007/978-3-031-43415-0_35 · Cited by 8 · see citing works
- Authors
- FitzGerald, J., Ananthakrishnan, S., Arkoudas, K., Bernardi, D., Bhagia, A., Delli Bovi, C., Cao, J., Chada, R., Chauhan, A., Chen, L., Dwarakanath, A., Dwivedi, S., Gojayev, T., Gopalakrishnan, K., Gueudre, T., Hakkani-Tur, D., Hamza, W., Hüser, J. J., Jose, K. M., Khan, H., Liu, B., Lu, J., Manzotti, A., Natarajan, P., Owczarzak, K., Oz, G., Palumbo, E., Peris, C., Prakash, C. S., Rawls, S., Rosenbaum, A., Shenoy, A., Soltan, S., Sridhar, M. H., Tan, L., Triefenbach, F., Wei, P., Yu, H., Zheng, S., Tur, G., & Natarajan, P.
- Affiliations
- Amazon (United States) (US)
- Amazon, Aachen, Germany
- Amazon (Germany) (DE)
- Amazon, Sunnyvale, CA, USA
- Spotify, Turin, Italy
- ORCID
- Venue
- Proceedings of the 28th ACM SIGKDD Conference on Knowledge Discovery and Data Mining
- Publisher
- ACM
- Published
- 2022-08-14
- Pages
- 2893-2902
- OpenAlex
- W4283070885
- Cited by
- 68
- References
- 10
- Citations by year
- 2026: 1 · 2025: 3 · 2024: 5 · 2023: 55 · 2022: 4
- FWCI
- 9.26
- Type
- conference paper
- Language
- en
- Open access
- hybrid
- doi.org
- Licence
- Locations
- ACM SIGKDD Conference on Knowledge Discovery and Data Mining (KDD) (publishedVersion, cc-by-nc-nd)
- arXiv (Cornell University) (submittedVersion)
- arXiv (Cornell University)
- Topics
- Physical Sciences › Computer Science › Artificial Intelligence › Topic Modeling
- Physical Sciences › Computer Science › Artificial Intelligence › Domain Adaptation and Few-Shot Learning
- Physical Sciences › Computer Science › Artificial Intelligence › Natural Language Processing Techniques
- Keywords
- knowledge distillation, teacher model, pre-training, virtual assistants, natural language understanding, XNLI, intent classification, slot filling, user dissatisfaction
About these checks
Registry, retraction, Cites and statistics results are computed, not judged, with one exception: a printed reference that no index lists is read once by a model before it counts as a citation error. Claims are extracted by a model and kept only when their quote appears verbatim in the text; each cited source is then checked, with a model’s help, for whether it says what the claim relies on. Those verdicts can be wrong — read the claim’s quote and the cited source before relying on one.
Statistics and arithmetic are recomputed, not judged.
Confidence bands are shown only for fields calibrated on our annotated test set. None is calibrated yet, so no band is shown.
What counts as a claim ›How we measure confidence ›How we check this ›
Are you an author, or do you see a problem with this page? Request a review
