Generating Scientific Definitions with Controllable Complexity

Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics Volume 1: Long Papers, pages 8298 - 8317 May 22-27, 2022 c⃝2022 Association for Computational Linguistics Generating Scientific Definitions with Controllable Complexity Tal August† Katharina Reinecke† Noah A. Smith†♦ †Paul G. Allen School of Computer Science & Engineering, University of Washington, Seattle, WA, USA ♦Allen Institute for Artificial Intelligence, Seattle, WA, USA {taugust,reinecke,nasmith}@cs.washington.edu Abstract Unfamiliar terminology and complex lan- guage can present barriers to understanding science. Natural language processing stands to help address these issues by automatically defining unfamiliar terms. We introduce a new task and dataset for defining scientific terms and controlling the complexity of gen- erated definitions as a way of adapting to a specific reader’s background knowledge. We test four definition generation methods for this new task, finding that a sequence-to-sequence approach is most successful. We then explore the version of the task in which definitions are generated at a target complexity level. We in- troduce a novel reranking approach and find in human evaluations that it offers superior flu- ency while also controlling complexity, com- pared to several controllable generation base- lines. 1 Introduction Unfamiliar concepts and complex language can make understanding scientific information difficult for readers (Brossard and Shanahan, 2006; Shea, 2015; Martínez and Mammola, 2021), especially because understanding such terms is highly de- pendent on their domain knowledge. Given the wide variation in such knowledge, providing a one- size-fits-all definition may not be sufficiently un- derstandable for all readers. We envision a software tool designed to aid read- ers with varying domain knowledge by automati- cally defining scientific terms. Such a tool would afford readers control over generated definitions, including their complexity. This hypothetical sys- tem motivates research on automated generation of scientific definitions and generation-time control of definition complexity. Prior work in generating definitions and person- alizing generations to a reader falls short of these goals. Most definition generation has focused on common, general-usage words in English (Noraset A molecule that binds to a hydrophobic surface. A substance that dissolves water in a liquid or solute. Journal Abstracts Question: What is (are) surfactants? Input complex definition simple definition Output Figure 1: Example of our task. Definitions are gener- ated with a controlled amount of complexity based on the question, “What is (are) X?” et al., 2017; Balachandran et al., 2018); however, these approaches and models may not be suitable for generating scientific definitions (Beltagy et al., 2019). Scientific terms rarely reach common usage (Shea, 2015; Britt et al., 2014) and the contexts in which their definitions might appear (e.g., a re- search paper) are often much more complex than general-purpose resources for definitions (e.g., dic- tionaries or standard word embeddings). Previous methods focused on reader personalization have aimed at generating based on a reader’s prior knowl- edge and interests (Acharya et al., 2018; Murthy et al., 2021). These approaches work well when models can leverage a reader profile (Murthy et al., 2021) or incorporate reader feedback over time. However, in many …

Content truncated to 500 words