Hierarchical Opinion Mining Based On Social Network And Content

The advent of social media platforms such as Tweet has provided users with unprecedented methods to show their preferences.A large portion of users manifest themselves in chronological behaviors,scilicet in following,commenting on or forwarding others' texts.This poses several new challenges to existing approaches in the field of opinion mining.Firstly,users' opinions exhibit a hierarchical structure.Secondly,users are not necessarily influenced by all the surrounding contexts,they tend to ignore those microblogs containing topics falling outside of their interests.Thirdly,users gather together through social interactions instead of explicit texts.Fourthly and lastly,users may update their opinions over time.These characteristics inspire us to explore a unified architecture to take into account miscellaneous data and user modes for classifying user-level topic-dependent stances.Since opinions expressed in text exhibit a hierarchical tree structure,we develop the Hierarchical Opinion Phrase(HOP)model to cope with multi-grained opinions and N-gram expressions.As a distinction from other hierarchical topic models,the number of stances is fixed and placed at level-2,making it possible to leverage prior information such as hashtags or sentiment lexicons.Phrases featured in opinions are generated by hierarchical Pitman-Yor processes.Experimental results demonstrate the effectiveness of our model in comparison with existing approaches for hierarchical topic discovery and text-level stance classification when incorporated with document or lexicon priors.Considering that a user's opinion can either reside in his text or originate from his social relationships,we propose the Neural Opinion Dynamics(NOD)model.NOD is built on Recurrent Neural Networks(RNNs),which hopefully depicts the user's chronological behaviors.An attention mechanism which allocates higher signals to the user's concerning materials is introduced to capture the three key factors: the user's past opinions,the user's neighborhood opinions and contextual information about topics.We perform experiments in an online learning setup that data stream is split into epochs temporally and the model is updated at each epoch sequentially to allow for the prediction of user-level topic-specific stances in the following epoch,which has demonstrated the effectiveness of the model.