Leveraging Language Representation Learning and Semantic Features for Improving Hashtag Recommendation on Microblogging Platforms

Abstract

Communication and information technologies have experienced significant growth over the past few years, with the emergence of social networks representing one of the most important developments. These networks have played a major role in this evolution, as users increasingly share information through knowledge bases, participate in virtual communities, and engage in microblogging platforms. Consequently, vast amounts of information are published online every day and made available to the public. This information is highly valuable and can contribute to the improvement of various applications when effectively exploited.

Microblogging platforms combine features of blogging and instant messaging, allowing users to share their ideas, opinions, moods, and experiences with others in real time. These platforms have become extremely popular, with Twitter serving as a prominent example. Users can create hashtags and include them in their tweets, subject to the platform’s character limit. Hashtags have become a widely used means of categorizing microblog posts on social media. By placing the hash symbol (#) before a word or phrase without spaces, users create a clickable link that directs other users to a feed of posts containing the same hashtag.

Twitter users include hashtags in their tweets for two main purposes: (a) to categorize their posts, making it easier for others to discover and follow discussions related to specific topics or themes, and (b) to follow and monitor trending topics. Consequently, hashtag recommendation has become an important task in various data mining applications. However, existing hashtag recommendation systems often rely primarily on syntactic features or conventional tag recommendation techniques, limiting their ability to capture the semantic content of microblog posts.

This thesis addresses the broader challenge of improving natural language understanding for textual data, particularly microblogs. To this end, we propose several effective approaches to text representation, ranging from simple statistical techniques to more sophisticated methods that leverage multiple semantic knowledge bases and neural network architectures.

The contributions of this thesis demonstrate the effectiveness of semantic knowledge bases as sources of background knowledge for natural language processing. Furthermore, they highlight the potential of word embedding techniques for text representation. We evaluate the performance of the proposed approaches using hashtag recommendation systems. Our experimental results show that these approaches outperform several state-of-the-art methods and achieve promising results.