Using artificial neural networks to separate the cream from the crop in systematic reviews

In a recent publication in Conservation Biology, Uri, Oded and Ricardo Correia from the University of Aveiro, developed a new tool, based on machine learning, which enables separating related from unrelated content in bibliometric searches.

There are a lot of challenges in conducting systematic reviews. When initiating a systematic review we have to separate content relevant to the question in hand from irrelevant content, usually in large textual corpora. However a technical challenge that can severely hinder the feasibility of this separation is the presence of homonyms. Homonyms are terms that share spelling but differ in meaning - for example, the word 'orange' is both a fruit and a color. Homonyms add a lot of noise to literature search results and cannot be easily identified and removed. In the current work we developed a new semi-automated approach, using artificial neural networks, which aids in in the classification of homonyms in search results into relevant and irrelevant contexts. We tested their new approach and found it to be more than 99% accurate when compared to a classification conducted manually. The results of this can be easily used with any homonym term, simplifying the production of systematic reviews, or any similar cases where homonyms make the interpretation of large databases difficult.