With 1.6 million subscribers and over a hundred fifty million content views, Chegg is a centralized hub where students come to get help with writing, science, math, and other educational needs.The content generated at Chegg is very unique. It is a combination of academic materials and language used by students along with images which could be handwritten. This data is unstructured and the only way to retrieve information from it is to do detailed NLP modeling for specific problems in search, recommendation systems, content tagging, finding relations between content, normalizing, personalized targeting, fraud detection etc. Deep Learning provides an efficient way to build high performance models without the necessity of feature engineering. However typically deep learning requires a huge amount of training data and is computationally expensive.

Transfer learning provides a path in between, it uses features from a related predictive modeling problems. Pre-trained word vectors or sentence vectors do not represent content at Chegg very well. Hence, we develop embeddings for characters, words and sentences that are optimized for building language models, question answering and text summarization using high performing GPUs. These embeddings are then made available for getting analytical insights and building models with machine learning techniques such as logistics regression to wide range of teams (consumer insights, analytics and ML model building). The advantage of this system is that previously unstructured content is associated with structured information developed using high performing GPU’s. In this talk I will give details of the architecture used to build the embeddings and the different problems that are solved using these embeddings.