University of Strathclyde at Headline Ranking TREC BLOG 2010

The University of Strathclyde participated in TREC BLOG Headline Ranking task only. Our general theme was to explore how the lexical changes in the BLOG corpus can reflect the importance of the articles appearing in the news corpus. Three (3) runs were submitted. For automated run "strath1", our algorithm identified the word unigrams, the frequencies of mentioning of which in the blog corpus increased substantially on the day of the query. Up to 100 such words were used as a query to return and rank the headlines using the Terrier platform and its PL2 model. Automated run “strath3” was similar to “strath1” except the weights were estimated based on the amount of the increase in the frequency of use and applied to the query words. “Strath2” was a manual run. Event descriptions were taken from the “Current Event” articles of Wikipedia Portal on the day of each topic (e.g. 22 January 2008 for the first topic). The description of each event was sent to Bing search engine. The words that occurred more frequently within the snippets than on the Web in average, were used to query the headlines corpus. Our participation was in close collaboration with the University of Glasgow group, which provided 1) the index to the news corpus 2) the daily statistics on the lexicons of the blog corpus and 3) the classification of the headlines into the required set of categories. HEADLINE RANKING TASK The top stories identification task (headline ranking) was first run as a pilot task in TREC 2009 to address the news dimension of the blogosphere (Macdonald et al., 2009). The task was exploring whether the blogosphere can be used to identify the most important news stories for a given day. In response to a date "query", systems were expected to provide a ranking of 100 headlines (stories) that they think were important on the specified day. For this year, Thomson-Reuters provided the TRC2 newswire corpus covering the same time-span as the corpus used in 2009. Not only the newer TRC2 corpus was much larger, it also included the full content of each story. The corpus was distributed by NIST for the TREC participants free of charge. Different from 2009, this task was treated as an online event detection: only the data created prior to the query date (e.g., an ontology or a Wikipedia article) was allowed to be used by an automatic run. Two main approaches were used during 2009 TREC to identify top stories (Macdonald et al., 2009; Balog et al., 2006): (i) News to Blogs, where mentioning of the headline in the blogs was typically counted as a vote for its importance, and (ii) Blogs to News, which generally proceed by the following steps: 1. observe the blog posts from the given date; 2. detect what differentiates these posts from the previous posts; 3. identify the emerging topics; 4. rank the headlines by their similarity to the emerging topics. The overall observation was that “Blogs to News” approaches worked better. This motivated our involvement and specific choices of techniques to explore as we elaborate in the next section. “Related Work” section follows our “Discussion Of Results,” followed, in turn, by “Conclusions, Limitations And Future Directions.”