Triple-Compressed BERT for Efficient Implementation on NLP Tasks
Bibliographic record
Abstract
Transformer-based models, exemplified by BERT, have significantly elevated the performance of Natural Language Processing (NLP) tasks, setting new benchmarks in accuracy, as demonstrated on GLUE and SQuAD. However, the large memory footprint and latency of BERT-based models present significant challenges for real-time applications. This paper introduces a novel compression strategy for BERT, synergistically combining knowledge distillation, layer pruning, and quantization. Focusing on the MNLI downstream task as a case study, the proposed approach not only maintains competitive performance but also significantly reduces model size and inference time. Specifically, our most optimized model, despite being approximately 8.06x smaller and around 9.5x faster than the original BERT, incurs only minor accuracy degradations of up to 6.5% and 6.6% on the matched and mismatched MNLI tasks, respectively. Our findings highlight the value of using a variety of compression techniques to enhance BERT-based models. The suggested method offers a potentially effective way to deal with the memory and latency issues brought on by BERT, creating opportunities for resource-effective language models for a range of NLP applications. By advancing the field of model compression through a synergistic approach to compressing BERT, this research provides a pragmatic strategy to enhance the deployability and effectiveness of language models, ultimately paving the way for more efficient and effective language models.
Fetched live from OpenAlex and de-inverted. Abstracts are not stored in this database: the inverted indexes are 8.6 GB of the frame’s 9.3 GB of text, and the host has 13 GB free.
How this classification was reachedexpand
Full frame machine prediction
Teacher imitationNot calibrated prevalence, not ground truth. Human validation pending. The Gemma side is a direct model label for every work in the frame, read from the title-only record. The Codex side is a classifier learned from the 10,348 direct Codex labels and calibrated to design-weighted sample rates; fields without enough sample support carry no Codex call. Candidate is the union of the two sides; consensus is their intersection. These outputs are machine_predicted_unvalidated and are not human labels.
Distilled classifier scores by category (both heads)
| Category | Codex | Gemma |
|---|---|---|
| Metaresearch | 0.001 | 0.005 |
| Meta-epidemiology (narrow) | 0.001 | 0.000 |
| Meta-epidemiology (broad) | 0.001 | 0.001 |
| Bibliometrics | 0.001 | 0.001 |
| Science and technology studies | 0.000 | 0.001 |
| Scholarly communication | 0.001 | 0.003 |
| Open science | 0.002 | 0.002 |
| Research integrity | 0.001 | 0.002 |
| Insufficient payload (model declined to judge) | 0.010 | 0.003 |
Machine scores (provisional)
The two teacher heads of the student model, read on this work. A score orders the frame for review; it never asserts a category, and the validation status ships verbatim with every row.
Baseline scores from an immature model (maturity gate not passed, 7 training rounds). Scores rank; they never assert a category.
score_only:v0-immature-baseline · verbatim from the scoring run: score_only means the number may rank works, and no category label ships from itClassification
machine, unvalidatedMachine predicted; a candidate call from one source (direct Gemma or distilled Codex), not a consensus.
How this classification was reached, model by model and score by score, is at the end of the page under "How this classification was reached".