I work on the core Machine Learning/Natural Language Processing platform at Neuron. I am also responsible for managing and integrating the work done by the team over the NLP...more platform, services and products.
I was a GSOC Intern at Mifos during the 2014 Summer. During the period of 4 months I developed a Batch API over Mifos Platform which provides a highly efficient system which...more reduces network load by combining multiple HTTP request into a Single HTTP Request Array.
Source Code for the Batch API is available at: https://github.com/rishy/mifosx/tree/Batch-API Source Code for Community App changes: https://github.com/rishy/community-app/tree/MIFOSX-1425
Facebook's latest campaign in Britain comes after vocal criticism of the network's role in spreading false information during the 2016 American… · More presidential election, as well as during a series of elections in Europe this year.
Data-Tag is an evolved system to classify textual data and web pages using NLP techniques, rather than not so intelligent Keyword-based Tagging. It uses NLTK… · More to categorize data tokens into various "Word-Classes" and then using Open Data from Wikipedia applies Word-Sense Disambiguation algorithm to "smartly" tag the input data.
This project consists of two parts. First being the exploratory analysis of Big Data related to Github Repos and Users to find some of the recent trends in… · More Github repositories and demographics related to certain programming languages and users. Based on the we extend this project to create a recommender system which can help a user with possible repos he/she can contribute to.
Unsupervised and semi-supervised approaches to find interest
It used deep unsupervised and semi-supervised learning to predict the interests graphs of people so as to assist other companies to market their product to… · More precise audience. This results in high conversion rates.
Unsupervised Semantic Relatedness on variable length text
NLP/Deep Learning Reseacher · This model finds the contextual similarity between two documents and returns the similarity score in the range of 0-1. We… · More used a novel technique to train the word vectors in a particular way, so that these can be further convoluted into a richer document embedding, which can then be compared with the document embeddings of any other text.
This works particularly well for unstructured text like conversations(using slangs, idioms), tweets, etc., considering the fact that CNNs don't rely as… · More much on the "sequential" order of the text, compared to RNNs. Furthermore, in addition to word embeddings this model also considers character level embeddings for predicting the sentiment polarity of the text.
This is a generic model to find the context(s) of a document. Furthermore, it is highly flexible in the sense that by just providing the new classes in… · More "natural language", we were able to use the same model for a specific set of classes. The output is the probability distribution of the text among the defined classes.
Keyword Extraction using Joint Word Embeddings and Topic Mod
This model returns highly relevant Keywords in a document. Ranking of the Keyword is guided by the overall context of the text as well as considering the… · More combined word embeddings of multi-word nouns(eg: pink_floyd_the_band).