Detection of Duplicate Content on Twitter
Ujwal Gadiraju · Research Repository (Delft University of Technology) · 2012
Depending on the needs of a user, searching on Twitter streams has proven to be very useful for a wide range of applications. Twitter however, has large volumes of information available in the form of tweets, that are shared by mil- lions of users. We have observed that many of these tweets contain mutually redundant information and due to this, there is a consequent impact on the search and retrieval process on Twitter streams. Redundancy in the information pre- sented to a user, hinders with the objective of a search engine to limit a user’s effort, by presenting content that has already been seen. Detecting redundant information in tweets is all the more challenging, due to the colloquial nature of the language that is used. In addition to this, the presence of tweet-specific constructs like hash-tags, shortened URLs and @-mentions further complicates the task of detecting mutually redundant tweets or duplicate tweets. Although it is relatively easy to automatically detect duplicate tweets which are syntactically identical (i.e., the text in the tweets is identical), it is challenging to detect tweets that are semantically equivalent (i.e., tweets have the same underlying meaning) but syntactically different. By analyzing a large Twitter benchmark dataset (provided as a part of the TREC microblog search challenge) used for the micro-blog search task in 2011, we ob- served that there is a varying extent to which a pair of tweets can be duplicates. We developed a scale that can be used to measure the extent of semantic over- lapping in a pair of tweets. Through our analysis, we identified the key aspects around which strategies for the detection of duplicate tweets can be designed, namely, terminological, semantic and linguistic aspects. We designed and imple- mented strategies for the detection of duplicate tweets from that perspective. Our main contributions are four-fold. Firstly, we developed a scale that can be used to measure the degree of a pair of duplicate tweets. Secondly, we analyzed the top-k search results on Twitter and observed the extent to which duplicates appear in top-k retrieval. Thirdly, we designed and implemented the strategies for the detection of duplicates. Our best strategy combination results in a commend- able F-measure of just over 80% (0.82). Finally, we develop a prediction model that can predict the duplicity order of a tweet-pair with an average accuracy of over 80%.