Sentiment analysis is a well known Natural Language Processing task that deals with finding out sentiments of opinion holders in text. It has been dealt by researchers at various levels of granularity - document level, sentence level and phrasal level. Our research focuses on phrasal level analysis. We propose using the imagery score in Dictionary of Affect by Dr. Whissel as an important score to differentiate between neutral and polar sentiments. This research gave us an intuition that there exists a correlation between neutrality and subjectivity. To prove the existence and strength of this correlation would be the focus of our future research. We published this work in EACL-2009 (see Publications).

We scan handwritten text of Males and Females and extract features that are used for gender classification. We focus on extracting document level features that are independent of the culture a person belongs to. On a small self acquired data set, our features seem to perform very well. We are collecting more data to make this work publishable. In future, we want to try our model on different languages (currently it’s on English.) We also want to extend this to cheating detection in assignments and examinations.  

Online Social Networks are becoming an essential part of our daily lives. People make friends, declare their personal attributes like their age, languages they speak, their hobbies, have online discussions on hot topics etc. It turns out, there is a lot of information out there about millions of people. Can we make use of this information to infer future friendships or hobbies of people? Are these social networks a real imitation of real life to begin with? These are questions that we try to answer in this research. We compared performance of many state of the art graph transduction algorithms like GFHF, LapSVM and GTAM with variations in weight matrices. We have obtained convincing results for a small self acquired data set. We published this work in SCA-2009 (see Publications).

Friendships and other social relations are conveyed through meta-data such as self-declared friendship links and emails. But a more accurate social network is present in the content of conversations (or emails). People express: who they meet, talk to, think about etc. using language. Our effort in this project is to automatically detect and classify social events that lead to stronger social relations such as friendships. We define social events as events between people where at least one person is aware of the other and of the event. In case both parties are mutually aware, we call it an Interaction event. In case exactly one party is aware of the other we call it Cognition event. We annotated part of the ACE-2005 corpus leveraging on the already existing entity annotations. We published a comprehensive list of social events at ACL-LAWIV, 2010 (see Publications). In a later publication at EMNLP 2010, we show results for our first attempt detect and classify these social events automatically.

What is Watson?

Watson playing Jeopardy!

1. 1.Patent: Multi-dimensional Feature Merger for Question Answering (disclosure submitted May 2012)
   

2. 2.Patent: Automatically Mining Inference Rules from Large Heterogenous Networks (disclosure submitted August 2012)
   

3. 3.Paper at CIKM 2012: Labeling by Landscaping: Classifying Tokens In Context By Pruning and Decorating Trees. See Publications.
   

4. 4.Paper at COLING 2012: Multi-dimensional Feature Merger for Question Answering. See Publications.

CV_Oct_2010.pdf - Updated Oct 17th 2010