Compressing Transformer-Based Semantic Parsing Models using Compositional Code Embeddings
This addresses memory constraints for deploying semantic parsing models on edge devices, such as voice assistants, but is incremental as it builds on existing compressed variants like DistilBERT and ALBERT.
The paper tackles the problem of large memory footprints in transformer-based semantic parsing models for edge devices like voice assistants, proposing compositional code embeddings to compress models like BERT and RoBERTa, achieving 95.15% to 98.46% embedding compression and 20.47% to 34.22% encoder compression while preserving over 97.5% performance.
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 greater than 97.5% semantic parsing performances. We provide the recipe for training and analyze the trade-off between code embedding sizes and downstream performances.