Extracting Comparative Commonsense from the Web

Commonsense acquisition is one of the most important and challenging topics in Artificial Intelligence. Comparative commonsense, such as "In general, a man is stronger than a woman", denotes that one entity has a property or quality greater or less in extent than that of another. This paper presents an automatic method for acquiring comparative commonsense from the World Wide Web. We firstly extract potential comparative statements from related texts based on multiple lexico-syntactic patterns. Then, we assess the candidates using Web-scale statistical features. To evaluate this approach, we use three measures: coverage of the web corpora, precision and recall which achieved 79.2%, 76.4% and 83.3%, respectively in our experiments. And the experimental results show that this approach profits significantly when the semantic similarity relationships are involved in the commonsense assessment.