Continual Mixed-Language Pre-Training for Extremely Low-Resource Neural Machine Translation

Findings of the Association for Computational Linguistics: ACL-IJCNLP 2021, pages 269–280 August 1–6, 2021. ©2021 Association for Computational Linguistics 269 Contrastive Attention for Automatic Chest X-ray Report Generation Fenglin Liu1, Changchang Yin3, Xian Wu4, Shen Ge4, Ping Zhang3, Xu Sun2 1School of ECE, Peking University 2MOE Key Laboratory of Computational Linguistics, School of EECS, Peking University 3The Ohio State University 4Tencent Medical AI Lab, Beijing, China fenglinliu98@pku.edu.cn; {yin.731, zhang.10631}@osu.edu {kevinxwu, shenge}@tencent.com; xusun@pku.edu.cn Abstract Recently, chest X-ray report generation, which aims to automatically generate descriptions of given chest X-ray images, has received grow- ing research interests. The key challenge of chest X-ray report generation is to accurately capture and describe the abnormal regions. In most cases, the normal regions dominate the entire chest X-ray image, and the correspond- ing descriptions of these normal regions dom- inate the final report. Due to such data bias, learning-based models may fail to attend to ab- normal regions. In this work, to effectively capture and describe abnormal regions, we pro- pose the Contrastive Attention (CA) model. Instead of solely focusing on the current in- put image, the CA model compares the cur- rent input image with normal images to distill the contrastive information. The acquired con- trastive information can better represent the vi- sual features of abnormal regions. According to the experiments on the public IU-X-ray and MIMIC-CXR datasets, incorporating our CA into several existing models can boost their performance across most metrics. In addition, according to the analysis, the CA model can help existing models better attend to the abnor- mal regions and provide more accurate descrip- tions which are crucial for an interpretable di- agnosis. Specifically, we achieve the state-of- the-art results on the two public datasets. 1 Introduction A medical report is a paragraph containing multiple sentences that describe both the normal and abnor- mal regions in the chest X-ray image. Chest X-ray images and their corresponding reports are widely used in clinical diagnosis (Delrue et al., 2011). However, writing medical reports requires particu- lar domain knowledge (Goergen et al., 2013), and only experienced radiologists can accurately inter- pret chest X-ray images and note down correspond- Input Images Normal Image Contrastive Results Contrast Figure 1: By contrasting current input images and known normal images, it could be easier to capture the suspicious abnormal regions (Red bounding boxes). The images with Green boxes are normal. ing findings in a coherent manner. An automatic chest X-ray report generation system (Jing et al., 2018, 2019; Liu et al., 2021b,a, 2019c) can reduce the workload of radiologists and are in urgent need (Brady et al., 2012; Delrue et al., 2011). In recent years, several deep learning-based methods have been proposed (Jing et al., 2018, 2019; Li et al., 2018, 2019; Chen et al., 2020c; Liu et al., 2021b,a) for automatic chest X-ray re- port generation; however, there are serious data deviation problems in the medical report corpus. For example, 1) the normal images dominate the dataset over the abnormal ones (Shin et al., 2016); 2) given …

Content truncated to 500 words