Improving the Factual Consistency of Abstractive Summarization: Model Self-Improvement Contrastive Learning
Abstract: Abstractive summarization models often produce summaries that are inconsistent with the content of the original text. Contrastive learning is an effective strategy for improving the factual consistency of generative text summarization. However, the success of contrastive learning depends largely on the construction of the dataset. Existing contrastive learning methods usually directly consider the gold summaries of summary datasets as factual summaries, and ignore the factual consistency problem of the gold summaries. They mainly focus on the generation of hallucinated summaries, i.e., negative samples, and the construction for negative samples is usually based on the gold summaries rather than from the perspective of the model itself. The quality of the positive and negative samples of these methods is not high enough, which will affect the effect of contrastive learning. Therefore, this paper proposes Model Self-Improvement Contrastive Learning: a method to improve the factual consistency of abstractive summarization. This method begins with fine-tuning the model itself, considering its already acquired knowledge of generating summaries. It focuses on the inference aspect of the generation phase, and delves deeper into the content that may cause factual errors. At the same time, it takes into account the factual consistency of both the positive and negative samples, constructs the negative samples in a targeted manner and improves the positive samples. It then further improves the factual consistency of the model through contrastive learning.
Paper Type: Long
Research Area: Summarization
Research Area Keywords: abstractive summarization, factual consistency,data filtering,contrastive learning
Contribution Types: NLP engineering experiment
Languages Studied: English
Submission Number: 3571
Loading