Graph-Based Ranking on Chinese Product Features with a General Structure for Noun Phrases

For Chinese product feature extraction, we use a general structure (BaseNP) for noun phrases, aiming to cover more product features. To rank the extracted feature candidates, we use a modified co-ranking model to rank feature candidates and opinion candidates simultaneously. For two types of nodes (\emph{feature candidates} and \emph{opinion candidates}), the model integrates mutual reinforcement between heterogeneous nodes and intra-class propagation within homogenous nodes.Experimental results show that that BaseNP structures cover more product features, and the modified co-ranking model obtains encouraging performance through mutual reinforcement and intra-class propagation.