Indexed on: 11 Sep '20Published on: 10 Sep '20Published in: arXiv - Computer Science - Computation and Language
Pre-trained Language Models (PrLMs) have been widely used as backbones in lots of Natural Language Processing (NLP) tasks. The common process of utilizing PrLMs is first pre-training on large-scale general corpora with task-independent LM training objectives, then fine-tuning on task datasets with task-specific training objectives. Pre-training in a task-independent way enables the models to learn language representations, which is universal to some extent, but fails to capture crucial task-specific features in the meantime. This will lead to an incompatibility between pre-training and fine-tuning. To address this issue, we introduce task-specific pre-training on in-domain task-related corpora with task-specific objectives. This procedure is placed between the original two stages to enhance the model understanding capacity of specific tasks. In this work, we focus on Dialogue-related Natural Language Processing (DrNLP) tasks and design a Dialogue-Adaptive Pre-training Objective (DAPO) based on some important qualities for assessing dialogues which are usually ignored by general LM pre-training objectives. PrLMs with DAPO on a large in-domain dialogue corpus are then fine-tuned for downstream DrNLP tasks. Experimental results show that models with DAPO surpass those with general LM pre-training objectives and other strong baselines on downstream DrNLP tasks.