Generating example contexts to illustrate a target word sense

Abstract

Learning a vocabulary word requires seeing it in multiple informative contexts. We describe a system to generate such contexts for a given word sense. Rather than attempt to do word sense disambiguation on example contexts already generated or selected from a corpus, we compile information about the word sense into the context generation process. To evaluate the sense-appropriateness of the generated contexts compared to WordNet examples, three human judges chose which word sense (s) fit each example, blind to its source and intended sense. On average, one judge rated the generated examples as sense-appropriate, compared to two judges for the WordNet examples. Although the system's precision was only half of WordNet's, its recall was actually higher than WordNet's, thanks to covering many senses for which WordNet lacks examples.