Meta-analysis is a systematic approach for understanding a phenomenon by analyzing the results of many previously published experimental studies. It is central to deriving conclusions about the summary effect of treatments and interventions in medicine, poverty alleviation, and other applications with social impact. Unfortunately, meta-analysis involves great human effort, rendering a process that is extremely inefficient and vulnerable to human bias. To overcome these issues, we work toward automating meta-analysis with a focus on controlling for risks of bias. In particular, we first extract information from scientific publications written in natural language. From a novel causal learning perspective, we then propose to frame automated meta-analysis -- based on the input of the first step -- as a multiple-causal-inference problem where the summary effect is obtained through intervention. Built upon existing efforts for automating the initial steps of meta-analysis, the proposed approach achieves the goal of automated meta-analysis and largely reduces the human effort involved. Evaluations on synthetic and semi-synthetic datasets show that this approach can yield promising results.
翻译:元分析是一种系统的方法,通过分析许多以前发表的实验研究的结果来理解一种现象,对于就医学、减贫和其他具有社会影响的应用的治疗和干预的总结效果得出结论至关重要。不幸的是,元分析涉及巨大的人类努力,使一个效率极低和容易受人类偏见影响的过程。为了克服这些问题,我们致力于使元分析自动化,重点是控制偏见的风险。特别是,我们首先从自然语言的科学出版物中提取信息。从新的因果学习角度出发,我们然后提议将自动化元分析 -- -- 以第一步的投入为基础 -- -- 作为一种通过干预取得摘要效果的多重因果推论问题来设置。在现有努力使元分析初步步骤自动化的基础上,拟议方法实现了自动化元分析的目标,并在很大程度上减少了人类的工作。对合成和半合成数据集的评价表明,这种方法可以产生有希望的结果。