ChemBrowser: A Flexible Framework for Mining Chemical Documents

Abstract

The ability to extract chemical and biological entities and relations from text documents automatically has great value to biochemical research and development activities. The growing maturity of text mining and artificial intelligence technologies shows promise in enabling such automatic chemical entity extraction capabilities (called “Chemical Annotation” in this paper). Many techniques have been reported in the literature, ranging from dictionary and rule-based techniques to machine learning approaches. In practice, we found that no single technique works well in all cases. A combinatorial approach that allows one to quickly compose different annotation techniques together for a given situation is most effective. In this paper, we describe the key challenges we face in real-world chemical annotation scenarios. We then present a solution called ChemBrowser which has a flexible framework for chemical annotation. ChemBrowser includes a suite of customizable processing units that might be utilized in a chemical annotator, a high-level language that describes the composition of various processing units that would form a chemical annotator, and an execution engine that translates the composition language to an actual annotator that can generate annotation results for a given set of documents. We demonstrate the impact of this approach by tailoring an annotator for extracting chemical names from patent documents and show how this annotator can be easily modified with simple configuration alone.