Illustrating Answers: An Evaluation of Automatically Retrieved Illustrations of Answers to Medical Questions

In this paper we discuss and evaluate a method for automatic text illustration, applied to answers to medical questions. Our method for selecting illustrations is based on the idea that similarities between the answers and picture-related text (the picture’s caption or the section/paragraph that includes the picture) can be used as evidence that the picture would be appropriate to illustrate the answer. In a user study, participants rated answer presentations consisting of a textual component and a picture. The textual component was a manually written reference answer; the picture was automatically retrieved by measuring the similarity between the text and either the picture’s caption or its section. The caption-based selection method resulted in more attractive presentations than the section-based method; the caption-based method was also more consistent in selecting informative pictures and showed a greater correlation between user-rated informativeness and the confidence of relevance of the system. When compared to manually selected pictures, we found that automatically selected pictures were rated similarly to decorative pictures, but worse than informative pictures.