Predicting Interesting Things in Text

Proceedings of COLING 2014, the 25th International Conference on Computational Linguistics: Technical Papers, pages 1477–1488, Dublin, Ireland, August 23-29 2014. Predicting Interesting Things in Text Michael Gamon Microsoft Corp. One Microsoft Way Redmond, WA 98052 mgamon@microsoft.com Arjun Mukherjee Department of Computer Science University of Houston Houston, TX 77004 ar- jun4787@gmail.com Patrick Pantel Microsoft Corp. One Microsoft Way Redmond, WA 98052 ppantel@microsoft.com Abstract While reading a document, a user may encounter concepts, entities, and topics that she is interested in exploring more. We propose models of “interestingness”, which aim to predict the level of interest a user has in the various text spans in a document. We obtain naturally occurring interest signals by observing user browsing behavior in clicks from one page to another. We cast the problem of predicting interesting- ness as a discriminative learning problem over this data. We leverage features from two principal sources: textual context features and topic features that assess the semantics of the document transition. We learn our topic features without supervision via probabilistic inference over a graphical model that captures the latent joint topic space of the documents in the transition. We train and test our models on millions of real- world transitions between Wikipedia documents as observed from web browser session logs. On the task of predicting which spans are of most interest to users, we show significant improvement over various baselines and highlight the value of our latent semantic model. 1 Introduction Reading inevitably leads people to discover interesting concepts, entities, and topics. Predicting what interests a user while reading a document has important applications ranging from augmenting the doc- ument with supplementary information, to ad placement, to content recommendation. We define the task of predicting interesting things (ITs) as ranking text spans in an unstructured document according to whether a user would want to know more about them. This desire to learn more serves as our proxy for interestingness. There are many types of observable behavior that indicate user interest in a text span. The closest one to our problem definition is found in web browsing, where users click from one document to another via named anchors. The click process is generally governed by the user’s interest (modulo erroneous clicks). As such, the anchor name can be viewed as a text span of interest for that user. Furthermore, the frequency with which users, in aggregate, click on an anchor serves as a good proxy for the level of interest1. What is perceived as interesting is influenced by many factors. The semantics of the document and candidate IT are important. For example, we find that when users read an article about a movie, they are more likely to browse to an article about an actor or character than to another movie or the director. Also, user profile and geo-temporal information are relevant. For example, interests can differ depend- ing on the cultural and socio-economic background of a user as well as the time of the session (e.g., weekday versus weekend, daytime versus …

Content truncated to 500 words