A Theory-Refinement Approach to Information Extraction

We investigate applying theory refinement to the task of extracting information from text. In theory refinement, partial domain knowledge (which may be incorrect) is given to a supervised learner, which is guided by this prior knowledge, but can refine or even discard it during training. Our supervised learner is a “knowledge-based” neural network that initially contains “compiled” prior knowledge about a particular information extraction (IE) task. The prior knowledge needs to specify the extraction slots for the specific IE task. Our approach uses generateand-test to address the IE task. In the generate step, we produce candidate extractions by intelligently searching the space of possible extractions. In the test step, we use the trained network to judge each candidate and output those that exceed a systemselected threshold. Experiments on the CMU seminar-announcements and the Yeast subcellular-localization domains demonstrate our approach’s value.