Associating Items with Scenes Identified in Social Q&A Data

Abstract

We discuss the problem of discovering associations between typical situations (scenes) in our daily lives and their characteristic items, which refer to anything from real objects to imaginary beings or abstract concepts. Once scenes are associated with items, the scenes can be further computationally analyzed (e.g., compared, tracked) on the basis of their associated items. In our approach for mining such associations, a list L of items and a set D of Web documents, in which scenes are identified, are first prepared. Next, D is divided using latent Dirichlet allocation (LDA) into clusters, each of which can be regarded as corresponding to a distinct characteristic scene. Then, the relevance between the scenes and items in L is estimated on the basis of the statistical significance of occurrence of items in the clusters. We developed two simple techniques for improving the quality (consistency) of the clustering result obtained using LDA with the expectation that the improved clustering result yields better performance in revealing item-scene associations. The most effective of the two techniques, PACA, purifies original clusters (i.e., eliminates unwanted elements in each cluster) created using a clustering algorithm by using the outcome from another clustering algorithm. Through an experiment using pages in a social Q&A site, we verified the effectiveness of the cluster purification techniques and the total effectiveness of our approach of associating items with scenes.