Professional to Plain Language Neural Translation: A Path Toward Actionable Health Information

点击翻译按钮获取中文摘要

Health literacy is key to making well-informed health decisions that improve outcomes. However, while the peer- reviewed clinical literature contains valuable information to guide health decisions, it is generally written for an audience of healthcare professionals. Even in the context of good general literacy, medical jargon and the complex structure of professional language make this information especially hard to interpret. While efforts have been made to summarize some of this literature in plain language to make it accessible to the general public, these efforts depend on human expertise. This approach cannot scale to match the rapid pace at which new findings emerge in the literature. Thus, there is an urgent unmet need for automated methods to enhance the accessibility of the canonical biomedical literature to the general public. This problem can be framed as a type of translation problem, between the language of healthcare professionals, and that of healthcare consumers. The proposed research builds on recent advances in deep learning stemming from neural sequence- to-sequence models, which were originally evaluated in machine translation tasks. In our recent work, we showed these models can be effectively adapted to the task of translating between abstracts in the Cochrane Database of Systematic Reviews (CDSR) and corresponding professionally-authored plain language summaries. The resulting automatically-generated summaries outperformed those from other models in their alignment with professionally-authored summaries. Furthermore, in a pilot user evaluation in which participants were blinded as to summary provenance, they were generally judged favorably to their expert-authored counterparts. In the proposed research we will develop this line of research further, by evaluating the utility of additional pre-training and auxiliary fine-tuning tasks as a means to improve the quality of generated summaries. We will also customize the models concerned to enhance their factual accuracy and readability using novel auxiliary training objectives and post-processing procedures. We will evaluate our methods as compared with robust baseline models in system-centric evaluations of content alignment with reference summaries, readability and factual correctness. Using Mechanical Turk, we will conduct user-centric evaluations of the ease with which summaries from best-performing models can be understood, as compared with CDSR expert-authored plain language summaries. These evaluations will consider both perceived interpretability, and actual comprehension, with the latter evaluated using sets of multiple choice questions to probe comprehension, recall and learning. In doing so, the proposed research will advance the state-of-the-art in automated simplification and summarization of the biomedical literature for consumption by the general public.