Demoting the Lead Bias in News Summarization via Alternating Adversarial Learning
This addresses a specific issue in news summarization for researchers and practitioners, offering an incremental improvement by mitigating bias without harming in-distribution performance.
The paper tackled the problem of lead bias dominating neural extractive summarizers in news articles, which limits performance on data with different biases, and introduced a technique that effectively demotes this bias, improving generality on out-of-distribution data with little to no loss on in-distribution data, as shown in experiments on two news corpora.
In news articles the lead bias is a common phenomenon that usually dominates the learning signals for neural extractive summarizers, severely limiting their performance on data with different or even no bias. In this paper, we introduce a novel technique to demote lead bias and make the summarizer focus more on the content semantics. Experiments on two news corpora with different degrees of lead bias show that our method can effectively demote the model's learned lead bias and improve its generality on out-of-distribution data, with little to no performance loss on in-distribution data.