Measuring the Interestingness of News Articles
Raymond K. Pon (University of California - Los Angeles, USA), Alfonso F. Cardenas (University of California - Los Angeles, USA) and David J. Buttler (Lawrence Livermore National Laboratory, USA)
Copyright: © 2009
An explosive growth of online news has taken place. Users are inundated with thousands of news articles, only some of which are interesting. A system to filter out uninteresting articles would aid users that need to read and analyze many articles daily, such as financial analysts and government officials. The most obvious approach for reducing the amount of information overload is to learn keywords of interest for a user (Carreira et al., 2004). Although filtering articles based on keywords removes many irrelevant articles, there are still many uninteresting articles that are highly relevant to keyword searches. A relevant article may not be interesting for various reasons, such as the article’s age or if it discusses an event that the user has already read about in other articles. Although it has been shown that collaborative filtering can aid in personalized recommendation systems (Wang et al., 2006), a large number of users is needed. In a limited user environment, such as a small group of analysts monitoring news events, collaborative filtering would be ineffective. The definition of what makes an article interesting – or its “interestingness” – varies from user to user and is continually evolving, calling for adaptable user personalization. Furthermore, due to the nature of news, most articles are uninteresting since many are similar or report events outside the scope of an individual’s concerns. There has been much work in news recommendation systems, but none have yet addressed the question of what makes an article interesting.
Working in a limited user environment, the only available information is the article’s content and its metadata, disallowing the use of collaborative filtering for article recommendation. Some systems perform clustering or classification based on the article’s content, computing such values as TF-IDF weights for tokens (Radev et al., 2003). Corso (2005) ranks articles and new sources based on several properties, such as mutual reinforcement and freshness, in an online method. However, Corso does not address the problem of personalized news filtering, but rather the identification of interesting articles for the general public. Macskassy and Provost (2001) measure the interestingness of an article as the correlation between the article’s content and real-life events that occur after the article’s publication. Using these indicators, they can predict future interesting articles. Unfortunately, these indicators are often domain specific and are difficult to collect for the online processing of articles.
The online recommendation of articles is closely related to the adaptive filtering task in TREC (Text Retrieval Conference), which is the online identification of articles that are most relevant to a set of topics. The task is different from identifying interesting articles for a user because an article that is relevant to a topic may not necessarily be interesting. However, relevancy to a set of topics of interest is often correlated to interestingness. The report by Robertson and Soboroff (2002) summarizes the results of the last run of the TREC filtering task. Methods explored in TREC11 include a Rocchio variant, a second-order perceptron, a SVM, a Winnow classifier, language modelling, probabilistic models of terms and relevancy, and the Okapi Basic Search System.