Lifelong Pretraining: Continually Adapting Language Models to Emerging Corpora
This work addresses the challenge of maintaining language model performance in real-world deployments where data distributions shift over time, offering incremental improvements for continual learning in NLP.
The paper tackles the problem of adapting pretrained language models to evolving data distributions by proposing lifelong pretraining, where models are continually updated on emerging corpora. The results show that distillation-based approaches effectively retain performance on earlier domains while improving knowledge transfer and temporal generalization on new data.
Pretrained language models (PTLMs) are typically learned over a large, static corpus and further fine-tuned for various downstream tasks. However, when deployed in the real world, a PTLM-based model must deal with data distributions that deviate from what the PTLM was initially trained on. In this paper, we study a lifelong language model pretraining challenge where a PTLM is continually updated so as to adapt to emerging data. Over a domain-incremental research paper stream and a chronologically-ordered tweet stream, we incrementally pretrain a PTLM with different continual learning algorithms, and keep track of the downstream task performance (after fine-tuning). We evaluate PTLM's ability to adapt to new corpora while retaining learned knowledge in earlier corpora. Our experiments show distillation-based approaches to be most effective in retaining downstream performance in earlier domains. The algorithms also improve knowledge transfer, allowing models to achieve better downstream performance over the latest data, and improve temporal generalization when distribution gaps exist between training and evaluation because of time. We believe our problem formulation, methods, and analysis will inspire future studies towards continual pretraining of language models.