Representation Deficiency in Masked Language Modeling
This addresses a fundamental limitation in pretraining for NLP, offering a method to enhance model performance on tasks like GLUE and SQuAD, though it is incremental as it builds on existing MLM and Masked Autoencoder approaches.
The paper identifies that Masked Language Modeling (MLM) allocates model dimensions for [MASK] tokens, causing representation deficiency for real tokens and limiting expressiveness in downstream tasks. It proposes MAE-LM, which excludes [MASK] tokens from the encoder, improving dimension utilization and outperforming MLM-pretrained models on GLUE and SQuAD benchmarks.
Masked Language Modeling (MLM) has been one of the most prominent approaches for pretraining bidirectional text encoders due to its simplicity and effectiveness. One notable concern about MLM is that the special $\texttt{[MASK]}$ symbol causes a discrepancy between pretraining data and downstream data as it is present only in pretraining but not in fine-tuning. In this work, we offer a new perspective on the consequence of such a discrepancy: We demonstrate empirically and theoretically that MLM pretraining allocates some model dimensions exclusively for representing $\texttt{[MASK]}$ tokens, resulting in a representation deficiency for real tokens and limiting the pretrained model's expressiveness when it is adapted to downstream data without $\texttt{[MASK]}$ tokens. Motivated by the identified issue, we propose MAE-LM, which pretrains the Masked Autoencoder architecture with MLM where $\texttt{[MASK]}$ tokens are excluded from the encoder. Empirically, we show that MAE-LM improves the utilization of model dimensions for real token representations, and MAE-LM consistently outperforms MLM-pretrained models across different pretraining settings and model sizes when fine-tuned on the GLUE and SQuAD benchmarks.