Searching and Filtering Tweets: CSIRO at the TREC 2012 Microblog Track
Sarvnaz Karimi, Jie Yin, Paul Thomas · 2012
We report on the participation of the CSIRO1 team in the TREC 2012 Microblog Track. We participated with four au-tomatic runs for the adhoc search task and four automatic runs for the filtering task. In the adhoc search task, we ex-periment with different pre-processing and query expansion techniques. Our most important finding is highlighting the value of systematic pre-processing of tweets and its impact on improving the effectiveness of search. In the filtering task, we apply different feature extraction and classification techniques. We demonstrate the potential of using SVM classifiers for filtering tweets for a given topic. 1. MICROBLOG TRACK The Microblog Track was introduced in 2011 to TREC to encourage the information retrieval community to ex-plore techniques for searching and filtering information in microblogs, specifically Twitter. The search task is designed as a typical search scenario where users are interested to see recent relevant tweets for their time-stamped queries. The fil-tering task on the other hand has double-timestamped queries to specify a time span that started from the last time that the user queried for this topic, till when the latest tweet was tweeted. It assumes that the user has already seen the old tweets prior to the previous tweet-time. An example query for both tasks is shown in Figure 1. In this figure, and can be used interchangeably. Topic MB010 Query with one time-stamp: Egyptian protesters attack museum