Concept Discovery and Argument Bundles in the Experience Web

Abstract

Abstract. In this paper we focus on a particular interesting web usergenerated content: people's experiences. We extend our previous work on aspect extraction and sentiment analysis and propose a novel approach to create a vocabulary of basic level concepts with the appropriate granularity to characterize a set of products. This concept vocabulary is created by analyzing the usage of the aspects over a set of reviews, and allows us to find those features with a clear positive and negative polarity to create the bundles of arguments. The argument bundles allow us to define a concept-wise satisfaction degree of a user query over a set of bundles using the notion of fuzzy implication, allowing the reuse experiences of other people to the needs a specific user.

On argument bundles in the Web of Experiences

“…The difference comes not only from the name of the concept but from the aspects grouped under a concept in different vocabularies. For instance DSLR and Compact differentiate some picture-related concepts such as 'iso', 'noise' or 'resolution', while in Point & Shoot they are conflated in the concept 'picture' [4]. A similar situation is observed in concept 'button': this concept groups aspects 'button shutter', 'button' and 'shutter' in the three concept vocabularies C D , C C and C P .…”

“…The three aspect vocabularies consist of set of salient aspects used to define important characteristics of photographic digital cameras. Moreover, as we show in [4], the three aspect vocabularies contain different sets of aspects.…”

“…Taxonomic Similarity (L): PhotoDict is a small taxonomy of camera-related terms, where similarity is measured as the shortest path between two terms. We automatically generated PhotoDict from a camera related website [4].…”

“…As a final step in order to create a collection of bundles (B G , B s , B F ) for each product in P , we first need, considering the whole P in a corpus K , to rescale the sentiment values of the arguments that form the bundles of the collection of bundles in a way that the most positive argument sentiment about a concept has a sentiment 1, and the most negative a sentiment -1. We rescale the rest of the sentiment values accordingly (see [4] for more details). This way, considering a collection of product bundles of a given type, the product with the best sentiment over a concept has a sentiment value of 1.…”