Concept discovery and argument bundles in the web of experiences

Millions of people interact and share interesting information every day in the Social Web. From daily conversations to comments about products in e-commerce sites, the content generated by people in these sites is huge and diverse. Among the wide diversity of user-contributed content on the web, there is a particular kind that has the potential of being put to good use by intelligent systems: human experiences. People very often use other people's experiences before making decisions, and when these kind of human experiences are expressed and recorded on the web, they can be shared with by large number of people. Nevertheless sometimes this content is not easily accessible, so a person trying to book a hotel may read a few reviews over a few hotels - but cannot possibly read them all. There is a clear need for an in-depth analysis of this kind of information, based on textual expressions of human particular experiences. Our approach, in the framework of the Web of Experiences, aims at acquiring practical knowledge from individual experiences with entities in the real world expressed in textual form. Moreover, this knowledge has to be represented in a way that facilitates the reuse of the experiential knowledge by other individuals with different preferences. Our approach has three stages: First, we extract the most salient set of aspects used by the individuals to describe their experiences with the entities in a domain. Second, using the set of extracted aspects, we group them in concepts to create a concept vocabulary that models the set of issues addressed in the reviews. Third, using the vocabulary of concepts, we create a bundle of arguments for each entity. An argument bundle characterizes the pros and cons of an entity, aggregating practical knowledge from judgments written by individuals with different biases and preferences. Moreover, we show how argument bundles allow us to define the notions of user query and the satisfaction degree of a bundle by a user query, proving that argument bundles are not only capable of representing practical knowledge but they are also useful to perform inference given a set of user preferences specified in a query. We evaluate the argument bundles of our approach with the Amazon score ratings and the camera characterizations of Dpreview. We show that pro and con arguments are very close to those listed in Dpreview. Evaluating entity rankings, we show that Dpreview and our approach give congruent rankings, while Amazon's is not congruent neither with Dpreview's or ours.