HOW DO YOU measure progress? That is the question Kyle Van Houtan, an ecol-ogist at the Monterey Bay Aquarium, in California, found himself asking when he faced the task of working out whether methods of boosting the populations of endangered species in the wild have improved over the years. In normal circumstances, those keen on studying the effectiveness of research write reviews of the scientific literature. In a flourishing field, though, this may involve reading and extracting information from hundreds, possibly thousands, of papers. That requires a large team, and brings problems of co-ordination. Dr Van Houtan therefore wondered whether getting computers to do the heavy lifting might help. The answer is that it does. His study on the matter, published this week in Patterns, tapped into a branch of machine learning called natural-language processing. This is a way of analysing large volumes of text with minimal human supervision. He and his colleagues identified five existing natural-language-processing systems and borrowed them. They used them to search the abstracts of 4,313 papers on species-conservation projects published over the course of the past four decades. The software's task was to look for words associated with success, such as "protect", "support", "help", "benefit" and "growth", and also words associated with failure, like "threaten", "loss", "kill", "problem" and "risk". Different words had different values attached to them, depending on how positive or negative they were felt to be by the original model-makers. The result was that each abstract could be assigned a sentiment score, averaged from the five different inputs.
展开▼
机译：您如何衡量进度？这就是这个问题，加利福尼亚州蒙特利湾水族馆的生态学家凯尔·范·霍坦（Kyle Van Houtan）发现，自己问自己何时面临着一项任务，即如何解决多年来增加野生濒危物种种群的方法是否有所改善。在通常情况下，那些热衷于研究有效性的人会对科学文献发表评论。但是，在蓬勃发展的领域中，这可能涉及读取和提取数百篇（可能数千篇）论文中的信息。这需要庞大的团队，并带来协调问题。因此，范霍坦博士想知道让计算机来做繁重的工作是否会有所帮助。答案是确实如此。他关于这个问题的研究发表在本周的“模式”中，涉及了机器学习的一个分支，即自然语言处理。这是一种在无需人工监督的情况下分析大量文本的方法。他和他的同事们确定了五个现有的自然语言处理系统，并从中借鉴了它们。他们使用它们搜索了过去40年中发表的4,313篇有关物种保护项目的论文摘要。该软件的任务是寻找与成功相关的词，例如“保护”，“支持”，“帮助”，“收益”和“增长”，以及与失败相关的词，例如“威胁”，“损失”， “杀死”，“问题”和“风险”。不同的词具有不同的附加值，这取决于原始模型制作者对它们的正面或负面看法。结果是，可以为每个摘要分配一个情感分数，该分数由五个不同的输入取平均值。
展开▼
