We present the results of feature engineeringand post-processing experiments conductedon a temporal expression recognitiontask. The former explores the use ofdifferent kinds of tagging schemes and ofexploiting a list of core temporal expressionsduring training. The latter is concernedwith the use of this list for postprocessingthe output of a system based onconditional random fields.We find that the incorporation of knowledgesources both for training and postprocessingimproves recall, while the useof extended tagging schemes may helpto offset the (mildly) negative impact onprecision. Each of these approaches addressesa different aspect of the overallrecognition performance. Taken separately,the impact on the overall performanceis low, but by combining the approacheswe achieve both high precisionand high recall scores.
展开▼
