ACL2021
Elaborative Simplification: Content Addition and Explanation Generation in Text Simplification
Neha Srikanth, Junyi Jessy Li
摘要
Much of modern-day text simplification research focuses on sentence-level simplification, transforming original, more complex sentences into simplified versions. However, adding content can often be useful when difficult concepts and reasoning need to be explained. In this work, we present the first datadriven study of content addition in text simplification, which we call elaborative simplification. We introduce a new annotated dataset of 1.3K instances of elaborative simplification in the Newsela corpus, and analyze how entities, ideas, and concepts are elaborated through the lens of contextual specificity. We establish baselines for elaboration generation using large-scale pre-trained language models, and demonstrate that considering contextual specificity during generation can improve performance. Our results illustrate the complexities of elaborative simplification, suggesting many interesting directions for future work. Simplified Text Generated Elaborations It also has a high amount of smoking for a big city, according to a health department report. In Philadelphia, smoking is more common among the poor than people who aren't poor, said Giridhar Mallya. He works for the Philadelphia Department of Public Health. Being poor can be very hard on a person's mind. It is not easy to relax, or to feel positive. Contextual Specificity: High Greedy: "It's not just the money," he said. Top-k: Mallya said the study showed that the poorest people in the city smoke a lot more than the richest people. Contextual: It can make them feel guilty. Workers at the Danish zoo said they had no choice but to kill the giraffe. The reason why is complicated. They said its genes were too much like the genes of other giraffes. All plants and animals have genes. They play a big part in what animals and plants look and act like. Genes are passed down from parents. They determine eye color, hair color, athletic ability, health and more. Contextual Specificity: Low Greedy: The scientists found that the giraffe's genes were very similar to those of other giraffes. Top-k: Scientists have known for years that the giraffe's genes were more like those of other giraffes. Contextual: The genes are passed down from generations. Because they look like a poisonous butterfly, they are left alone by predators. Looking like another species works just fine for them. While they avoid getting eaten, the male and the masculine-colored female are still at risk. For a long time, biologists have studied how animals try to look like other animals. It teaches them a lot about natural selection. Natural selection is sometimes called "survival of the fittest." It means that animals that can change will survive. Contextual Specificity: Low Greedy: It is a way of thinking about how animals live. Top-k: It means that the more an animal looks like another animal, the more likely it is to survive. Contextual: It's a way of thinking about how organisms adapt to change. The newly found cat is probably part of a separate branch of the cat family, said Tseng. The modern snow leopard is most likely part of the same branch. Both are able to live high in the Tibetan mountains. Big cats present serious problems for paleontologists. They have existed millions of years longer than people have. Contextual Specificity: High Greedy: They are slow and slow to change. Top-k: Paleontologists study ancient animals. Contextual: They are difficult to study. Reporters and photographers watched. The big cat crept out of the cage. He looked around, and then ran off down a dirt road. The panther disappeared into the forest. Wildlife workers were setting him free. Contextual Specificity: High Greedy: The news was not good. Top-k: Then, it disappeared into the darkness. Contextual: The news of the escape made people angry.