Much of the Text Mining needed in real-life boils down to Text Classification: be it prioritising e-mails received by Customer Care, categorising Tweets aired towards an Organisation, measuring impact of Promotions in Social Media, and (Aspect based) Sentiment Analysis of Reviews. These techniques can not only help gauge the customer’s feedback, but also can help in providing users a better experience.
Traditional solutions focused on heavy domain-specific Feature Engineering, and thats exactly where Deep Learning sounds promising!
We will depict our foray into Deep Learning with these classes of Applications in mind. Specifically, we will describe how we tamed Deep Convolutional Neural Network, most commonly applied to Computer Vision, to help classify (short) texts, attaining near-state-of-the-art results on several SemEval tasks consistently, and a few tasks of importance to Flipkart.
In this talk, we plan to cover the following:
Basics of Deep Learning as applied to NLP: Word Embeddings and its compositions a la Recursive Neural Networks, Convolutional Neural Networks, and Recurrent Neural Networks.
New Experimental results on an array of SemEval / Flipkart’s internal tasks: e.g. Tweet Classification and Sentiment Analysis. (As an example we achieved 95% accuracy in binary sentiment classification task on our datasets – up from 85% by statistical models)
Share some of the learnings we have had while deploying these in Flipkart!
Here is a mindmap explaining the flow of content and key takeawys for the audience: https://atlas.mindmup.com/2015/06/4cbcef50fa6901327cdf06dfaff79cf0/deep_learning_for_natural_language_proce/index.html
We have decided to open source the code for this talk as a toolkit. https://github.com/flipkart-incubator/optimus Feel free to use it to train your own classifiers, and contribute!