Section: Application Domains
Crowd-sourced Information Filtering and Summarization
With the explosion of the People-centric Web, there is a proliferation of crowd-sourced content either under the form of qualitative reviews (mainly textual) and quantitative ratings (as 5 star ratings) regarding diverse products or services or under the form of various "real-time" feedback events (e.g., re-tweets, replies, likes, clicks, etc.) on published web content (ranging from traditional news, TV series, and movies to specialized blogs and posts shared over social networks). Such content captures the wisdom of the crowd and is valuable information source for building collaborative filtering systems and text summarization tools coping with information overload. For example, they can assist users to pick the most interesting web pages (e.g. Delicious) or to choose which movie to watch next (e.g. Netflix).
Implicit Feedback in Communities of a Place. We are initially interested in addressing one of the main limitation of collaborative filtering systems namely, the strong user engagement required to provide the necessary input (e.g., regarding their friends, tags or sites of preference) which is usual platform specific (i.e., for a particular social network, tagging, or bookmark system). The lack of user engagement translates into cold start and data sparsity. To cope with this limitation, we are developing a system called WeBrowse that passively observes network traffic to extract user clicks (i.e., the URLs users visit) for group of people who live, study, or work in the same place. Examples of such communities of a place are: (i) the students of a campus, (ii) the people living in a neighbourhood or (iii) researchers working in the same site. WeBrowse then promotes the hottest and most popular content to the community members sharing common interests.
Personalized Review Summarization. Finally, we are interested in helping people to take informed decisions regarding their shopping or entertainment activities. The automated summarization of a review corpus (for example, movie reviews from Rotten Tomatoes or IMDB; or restaurant reviews from Yelp) aims to assist people to form an opinion regarding a product/service of interest, by producing a coherent summary that is helpful and can be easily assimilated by humans. We are working on review summarisation methods that combine both objective (i.e., related to the review corpus) and subjective (i.e., related to the end-user interests) interestingness criteria of the produced reviews. In this respect we are exploiting domain models (e.g., Oscar's merit categories for movies) to elicit user preferences and mine the aspects of products/services actually commented in the textual sentences of reviews. For example, different summaries should be produced when a user is more interested in the actors performance rather than the movie story. We are particularly interested in extracting automatically the signatures of aspects (based on a set of seed terms) and rank review sentences on their importance and relevance w.r.t. the aspects they comment. Last but not least we are optimizing the automatically constructed summary w.r.t. to a number of criteria such as the number of the length of included sentences from the original reviews, the polarity of sentiments in the described aspects, etc.