Samraj M
- Research Program Mentor
MS at University of Illinois at Urbana Champaign (UIUC)
Expertise
Natural language processing, deep learning, (anti)social computing
Bio
My main academic passion lies in natural language processing with a social impact. For example, I'm really interested in topics like fake news detection or natural language generation for persuasive argumentation. I generally work with deep learning, traditional natural language techniques, and other data science methods (e.g. Bayesian analysis). If you are interested in automating something with text to make a social impact, I'd love to have a chat! As for my personal interests, I'm often at the gym either weightlifting or boxing. As of recently, I've picked up soccer and skateboarding. I also love to cook and bake (but I challenge myself to do it without a recipe). If I'm not doing any of these things, you'll probably find me gaming.Project ideas
Fake News Detection
How can we automatically determine whether a news article or post on social media provides accurate information? Anyone can post on the internet and with the amount of information available, it becomes difficult for a human to cross-check everything they read online. While creating a machine learning classifier to detect fake news is not a new task, it has many challenges. For example, should we use an existing knowledge base for fact-checking what is said online? How do we collect all this information, ensure it is general enough to cover a wide variety of cases, and store this information efficiently? Should we instead use features like the age of the authors account to detect if an article is likely fake? It can be easy to make mistakes with this approach and there is not a clear-cut way to tell whether the information is fake. There are many routes you can take with this project depending on your existing knowledge and your interests. You would likely have to create or take an existing fake news dataset and train a classifier on it. From there, we can make novel contributions by incorporating knowledge from sources like Wikipedia or using external confounds to further improve the performance of our classifier.
Persuasive Natural Language Generation
Automatic language generation techniques often generate bland text, so can we make these texts more appealing and persuasive? What does it mean for text to be persuasive and how to do we teach this to a machine? This system would be useful in automatic argument generation systems (e.g. Project Debater by IBM) or in generating text for advertising. In this project, you would likely create/find a dataset of persuasive text and learn what makes the text persuasive. You would then transfer this knowledge into a natural language generation model and evaluate the persuasiveness of the text that you generate.