Date of Award


Degree Type


Degree Name

Doctor of Philosophy (PhD)


Computer Science

First Advisor

Rajshekhar Sunderraman

Second Advisor

Anu Bourgeois

Third Advisor

Jonathan Ji

Fourth Advisor

Dror Walter


Social media has become an integral component of the daily life. There are millions of various types of content being released into social networks daily. This allows for an interesting view into a users' view on everyday life. Exploring the opinions of users in social media networks has always been an interesting subject for the Natural Language Processing researchers. Knowing the social opinions of a mass will allow anyone to make informed policy or marketing related decisions. This is exactly why it is desirable to find comprehensive social opinions. The nature of social media is complex and therefore obtaining the social opinion becomes a challenging task. Because of how diverse and complex social media networks are, they typically resonate with the actual social connections but in a digital platform. Similar to how users make friends and companions in the real world, the digital platforms enable users to mimic similar social connections. This work mainly looks at how to obtain a comprehensive social opinion out of social media network. Typical social opinion quantifiers will look at text contributions made by users to find the opinions. Currently, it is challenging because the majority of users on social media will be consuming content rather than expressing their opinions out into the world. This makes natural language processing based methods impractical due to not having linguistic features. In our work we look to improve a method named stance inference which can utilize multi-domain features to extract the social opinion. We also introduce a method which can expose users opinions even though they do not have on-topical content. We also note how by introducing weak supervision to an unsupervised task of stance inference we can improve the performance. The weak supervision we bring into the pipeline is through hashtags. We show how hashtags are contextual indicators added by humans which will be much likelier to be related than a topic model. Lastly we introduce disentanglement methods for chronological social media networks which allows one to utilize the methods we introduce above to be applied in these type of platforms.


File Upload Confirmation