Saurabh Kumar Shashidhar


2020

pdf
Compressing Transformer-Based Semantic Parsing Models using Compositional Code Embeddings
Prafull Prakash | Saurabh Kumar Shashidhar | Wenlong Zhao | Subendhu Rongali | Haidar Khan | Michael Kayser
Findings of the Association for Computational Linguistics: EMNLP 2020

The current state-of-the-art task-oriented semantic parsing models use BERT or RoBERTa as pretrained encoders; these models have huge memory footprints. This poses a challenge to their deployment for voice assistants such as Amazon Alexa and Google Assistant on edge devices with limited memory budgets. We propose to learn compositional code embeddings to greatly reduce the sizes of BERT-base and RoBERTa-base. We also apply the technique to DistilBERT, ALBERT-base, and ALBERT-large, three already compressed BERT variants which attain similar state-of-the-art performances on semantic parsing with much smaller model sizes. We observe 95.15% 98.46% embedding compression rates and 20.47% 34.22% encoder compression rates, while preserving >97.5% semantic parsing performances. We provide the recipe for training and analyze the trade-off between code embedding sizes and downstream performances.