The paper focuses on the issue of polarity bias in opinion summarization models. Previous summarization models tend to amplify the polarity bias, emphasizing the majority opinions while ignoring the minority opinions. To address this problem, the paper introduces the concept of polarity calibration, which aims to align the polarity of the output summary with that of the input text.
The authors develop a reinforcement learning approach for polarity calibration. Specifically, they design three reward models to assess the polarity distance between output and input, the content preservation, and the language fluency. The summarization model is then trained to minimize the polarity distance, while also maintaining the content semantics and language quality.
Experiments on two opinion summarization tasks, summarizing product reviews and political opinions articles, demonstrate the effectiveness of the proposed approach. The calibrated summarizer can significantly reduce the polarity distance between output and input, without compromising the content semantics and language quality, as shown by both automatic and human evaluation.
To Another Language
from source content
arxiv.org
Key Insights Distilled From
by Yuanyuan Lei... at arxiv.org 04-03-2024
https://arxiv.org/pdf/2404.01706.pdfDeeper Inquiries