Rumour Stance Classification using Discriminative Modeling of Text in Hawkes Process

Social networking sites like Twitter have become a huge source for sharing news. A lot of these news come from unreliable sources and lead to rumours. Analyzing various aspects of the spread of rumours has been an interesting challenge for data science researchers. We try to tackle the problem of classifying the stance of tweets following a rumour into one of supporting, denying, questioning or commenting. We give a Hawkes process (HP) based treatment to this problem. Hawkes process is a self-exciting point process which can analyze sequence of events where previous events can influence the upcoming events. Thus, Hawkes process is well suited for this problem to give importance to temporal aspects along with textual aspects. The previous work involving HP for rumour stance classification uses generative modeling of text which was separately included in the likelihood cost function. We propose a discriminative approach to text modeling using Hawkes process. The text features are now considered a part of the intensity function of Hawkes Process for the first time. We also provide novel models which make use of both time and text based kernels in Hawkes process. An important contribution is allowing use of neural networks as replacement for kernels to learn non-linear relationships between historical events. In another proposed model, we remove non-negative constraints on various parameters as well as the intensity function to learn negative influence of di↵erent classes. We show how approximation tricks can be used to solve the complex integrals in the likelihood function. Analysis and comparison of performances is shown between various models and it can been that in many cases discriminative modeling using Hawkes process performs better than generative modeling and also gives comparable performance to deep learning models. We initiate research in the direction of using various features as a part of HP intensity function