Learning to Extract Entity Uniqueness from Web for Helping User Decision Making

Web entities are the building blocks of human knowledge and users are making decisions among vast varieties of entities. For example, recommendation systems generate lists of entities to users, but seldom show the reasons of recommendation such as the uniqueness of each item to assist user decision making. In this paper, we mathematically define Web entity uniqueness and uniqueness patterns, based on which we propose a novel unsupervised natural language learning algorithm for entity uniqueness extraction. We leverage the bootstrapping strategy to recognize uniqueness from the free-text Web corpus with assistance from semi-structured Web such as lists, tables and query logs. To avoid extracting the subjective entity uniqueness, which may bias user decision making, we propose the probabilistic likelihood of a uniqueness property using bipartite graph models over entities and properties. Experiments verify that our algorithms have higher accuracy and coverage of entity uniqueness extraction technique compared to other related algorithms. We also show by conducting a user study survey that entity uniqueness information indeed positively supports user decision making.