I am working with Prof. Pushpak Bhattacharyya of CSE Department, IIT Bombay, in the area of Natural Language Processing.

The title of my thesis was: Towards Evaluating Lexico-Semantic Networks such as Wordnets.

The basic idea is as follows:
Lexico-semantic networks is a fancy word for NLP resources that essentially are dictionaries, thesauri, ontologies (sometimes) and more on steroids. Lexico because they provide information at the level of words as a dictionary would, Semantic because they provide knowledge at the meaning/conceptual level using relationships between words such as synonymy, parent-child links etc, and Networks because they can be visualised as a graph of words, concepts, and meanings if needed. We can also call them Lexical Knowledge Bases because some of these also have encyclopaedic or ontological information.

Examples of these resources as Princeton (English) Wordnet, Wordnets in Hindi and Marathi being developed at CFILT, IIT Bombay, Hownet, ConceptNet, MindNet, FrameNet, VerbNet, and so on. These resources have proved to be very useful in a bunch of applications in the area of NLP, Information Retrieval, Text Mining, and so on. Hence, many research and commercial projects have begun to use such resources.

In this project, we are trying to answer the question: "What is a good lexico-semantic network?". This means asking what are the parameters to evaluate a wordnet and how to go about doing so. Our literature survey showed that little to no work has been done in this area, so it seemed like a good and useful area to research.

In my project, I began by exploring structural properties of Wordnets using the famous Small World theories. This yielded some interesting observations. I then moved on to something more concrete: trying to evaluate the quality of synsets (which form the nodes in a Wordnet graph and are made of synonyms that express a specific concept). This led me to investigate the sub-problem of synonymy verification.

In synonymy verification, given a synset of words, how does one verify whether the words are indeed synonyms of each other? I have come up with some statistical and rule-based techniques to achieve this.

Some reading on the subject:

1. Stage 1 Report and Stage 1 Presentation
   
2. Stage 2 Report and Stage 2 Presentation
   
3. A small talk I gave to the CS460-IT632 NLP class: 'An introduction to the worlds of Small Worlds'
   
4. My Seminar on Lexical Knowledge Bases: Report, Presentation