Masked language modeling has become a standard pretraining objective for training encoder-based language models. In this approach, certain tokens in the input are masked, and the model learns to predict them using the surrounding context. This process enables the model to capture both syntactic and semantic properties of language. Conventionally, the tokens selected for masking are chosen at random, which may not always yield the most effective learning signals. In this work, we examine a token masking strategy based on entropy distribution. We use the model’s entropy over token predictions to identify which tokens should be masked. This method aims to target tokens that are more informative and uncertain to improve the training efficacy. We also propose a novel self-masking approach that enhances training efficiency without relying on an external reference model. Experimental results demonstrate that our method achieves an average performance improvement of 5% in GLUE scores compared to the baseline. Further, we experiment with combining knowledge distillation with entropy masking, resulting in the best overall results.

Pretrained transformer-based language models have produced state-of-the-art performance in most natural language understanding tasks. These models undergo two stages of training: pretraining on a huge corpus of data and fine-tuning on a specific downstream task. The pretraining phase is extremely compute-intensive and requires several high-performance computing devices like GPUs and several days or even months of training, but it is crucial for the model to capture global knowledge and also has a significant impact on the fine-tuning task. This is a major roadblock for researchers without access to sophisticated computing resources. To overcome this challenge, we propose two novel hybrid architectures called HybridBERT (HBERT), which combine self-attention and additive attention mechanisms together with sub-layer normalization. We introduce a computing budget to the pretraining phase, limiting the training time and usage to a single GPU. We show that HBERT attains twice the pretraining accuracy of a vanilla-BERT baseline. We also evaluate our proposed models on two downstream tasks, where we outperform BERT-base while accelerating inference. Moreover, we study the effect of weight initialization with a limited pretraining budget. The code and models are publicly available at: www.github.com/gokulsg/HBERT/.