[Pycon] [new paper] "Arnav Arora" - NLP researchers on language modelling : What do they know? Do they know things? Let's find out!

[info a pycon.it]

Title: NLP researchers on language modelling : What do they know? Do they know things? Let's find out!
Duration: 45 (includes Q&A)
Q&A Session: 15
Language: it
Type: Talk

Abstract: Language is a complicated thing for computers to understand. In this talk, I aim to discuss why this is the case, how we break it down to different tasks to achieve the goal and finally, how far we've come in actually making some progress. The talk will be focused on the role deep learning has played in bringing about this development and what techniques within the field have turned out to be a big contributing factor like neural word embeddings, seq2seq, attention and more recently, unsupervised pretraining. The latter half of the talk will focus on language modelling as a prospective Imagenet equivalent task for NLP and the benefits of doing so along with the demo.

To keep the talk light, I'll show some practical demonstration of how easy it is to actually go ahead and use these pretrained models in daily life and/or work. These can be used for text classification tasks as well as text generation with few changes to the code. I'll be using jupyter notebooks/Google colab for the code and will use Pytorch, Allennlp, fastai etc for faster computation. A primer of these is not required but an overall idea of how neural network models function would be helpful in concentrating on the NLP parts of the talk. 

The talk will leave the audience with beneficial information as to what is possible using text (of which, everyone possesses a very large amount), what the state of the art in the research community is and how it can be reproduced in a few lines of code. I'll be covering advanced topics without the nitty gritty details and a good overview of the field to make it so that every level of audience can take something away from the talk. 

Tags: [u'Languages', u'deep learning', u'natural-language-processing', u'Pytorch', u'nlp']