We describe our efforts to generate alarge (100,000 instance) corpus of textualentailment pairs from the lead paragraphand headline of news articles. We manuallyinspected a small set of news storiesin order to locate the most productivesource of entailments, then built an annotationinterface for rapid manual evaluationof further exemplars. With thistraining data we built an SVM-baseddocument classifier, which we used forcorpus refinement purposes—we believethat roughly three-quarters of the resultingcorpus are genuine entailment pairs. Wealso discuss the difficulties inherent inmanual entailment judgment, and suggestways to ameliorate some of these.
展开▼