Adaptive Information Extraction for Complex Biomedical Tasks

Biomedical information extraction tasks are often more complex and contain uncertainty at each step during problem solving processes. We present an adaptive information extraction framework and demonstrate how to explore uncertainty using feedback integration. 1 Adaptive Information Extraction Biomedical information extraction (IE) tasks are often more complex and contain uncertainty at each step during problem solving processes. When in the first place the desired information is not easy to define and to annotate (even by humans), iterative IE cycles are to be expected. There might be gaps between the domain knowledge representation and computer processing ability. Domain knowledge might be hard to represent in a clear format easy for computers to process. Computer scientists may need time to understand the inherent characteristics of domain problems so as to find effective approaches to solve them. All these issues mandate a more expressive IE process. In these situations, the traditional, straightforward, and one-pass problem-solving procedure, consisting of definition-learning-testing, is no longer adequate for the solution. Figure 1. Adaptive information extraction. For more complex tasks requiring iterative cycles, an adaptive and extended IE framework has not yet been fully defined although variants have been explored. We describe an adaptive IE framework to characterize the activities involved in complex IE tasks. Figure 1 depicts the adaptive information extraction framework. This procedure emphasizes one important adaptive step between the learning and application phases. If the IE result is not adequate, some adaptations are required: Our study focuses on extracting tract-tracing experiments (Swanson, 2004) from neuroscience articles. The goal of tract-tracing experiment is to chart the interconnectivity of the brain by injecting tracer chemicals into a region of the brain and then identifying corresponding labeled regions where the tracer is transported to (Burns et al., 2007). Our work is performed in the context of NeuroScholar, a project that aims to develop a Knowledge Base Management System to benefit neuroscience research. We show how this new framework evolves to meet the demands of the more complex scenario of biomedical text mining. 2 Feedback Integration This task requires finding the knowledge describing one or more experiments within an article as well as identifying desired fields within individual sentences. Significant complexity arises from the presence of a variable number of records (experiments) in a single research article --anywhere from one to many. Table 1. An example tract-tracing experiment. Table 1 provides an example of a tract-tracing experiment. In this experiment, when the tracer was injected into the injection location “the contralateral AVCN”, “no labeled cells” was found in the labeling location “the DCN”. For sentence level fields labeling, the performance of F1 score is around 0.79 (Feng et al., 2008). 1 http://www.neuroscholar.org/