Deep Learning Conf 2016

A conference on deep learning.


Text made Understandable by Machines

Submitted by Ashish Kumar (@ashish122) on Monday, 30 May 2016

Section: Full talk Technical level: Intermediate


Understanding language is a trivial task for humans, but when it comes to mimic that task by machines it doesn’t remain that trivial. For humans, everything(image, text, speech etc.) is in terms for electrical impulses. In the same way for machines, everything is numbers either in the vector form (in the case of text or speech) or matrix form (in the case of images or videos). Deep learning has recently shown many promises for Natural Language Processing(NLP) applications. Traditionally in most NLP approaches documents or sentences are represented by a sparse bag-of-words representation.
A lot of work has been done, which goes beyond this by adopting a distributed representation of words by constructing a so-called “neural embedding” or vector space representation for each word(word2vec), sentence(thought vectors) or document(doc2vec).


1) Introduction and the importance of Word Embedding
2) Old methods used for Text representaion
3) Word2Vec and its pros and cons
4) Thought Vectors and its pros and cons
5) Doc2Vec and its pros and cons

Speaker bio

I’m a software engineer at Snapshopr. You can also go through my profile


  • Gaurav Agarwal 3 years ago

    not to take anything away from this excellent topic, wanted to highlight that Word2Vec and the extension Doc2Vec are not examples of deep neural nets but are 2-layer NNs.

  • Ashish Kumar (@ashish122) Proposer 3 years ago

    I agree with you @Gaurav. But talking about deep learning on text, the most important thing is to come up with word embeddings. If your word-embeddings are not good then you will not get good results and porbably your system will just go mad. Hence I hope you understand that these algorithms are as important as the language models that you construct later.

Login with Twitter or Google to leave a comment