Author: datageeko

Why some linguistics is necessary for NLP

Disclaimer: I’m no means an expert in linguistics and below is the opinion of my personal research. Feel free to correct me. After sitting in the NLP classes for the last 3 weekends, my classmates exclaimed that it felt like they went through 3 adult English classes. It didn’t help that the workshops are designed

Backprop, Autograd and Squeezing in larger batch using PyTorch

Backprogation is a beautiful play of derivatives which we have taken for granted. We often do a simple one-liner: to leverage the power of automatic differentiation in many deep learning framework without much thought. Today let’s look at a developed view of backpropagation as backward flow in real-valued circuits. Motivation Given some function ,we are

Crossing the language barrier with NLP

One of the biggest open problems in NLP is the unavailability of many non-English dataset. Dealing with low-resource/low-data setting can be quite frustrating when it seems impossible to transfer the same success we saw in various English NLP tasks. In fact, there are voices within the NLP community to advocate research and focus on low-resource

Model building and performance tips for PyTorch

Here are some key observations and lessons learned from building a brand new Seq-to-Seq model for sentence summarization and training it against a 1 million samples dataset. General 1) Always maintain codes in Git repo; it’s an efficient way to ensure that different training machines have the exact same codes. Avoid the temptation to just

Transfer learning and beyond

Transfer learning has proven to be useful in NLP in the recent years. As many called the “Imagenet moment” when the likes of large pretrained language models such as BERT, GPT, GPT2 have sprung out from the big research labs, they have been extended in various methods to achieve further state of the art results

Recent Advances in Abstractive Summarization Using Deep Learning Part 2

This post is a continuation to the previous post here. We continue to track the recent progress and trend in abstractive summarization in 2018. The earlier efforts in abstractive summarisation focuses on problems that are related to natural language generation rather than the summarization task itself. Some problems that were tackled: Unfactual information (copy mechanism)

Go Even Deeper with Char-Gram+CNN

This is a repost from my kernel at Kaggle, which has received several positive responses from the community that it’s helpful to them. This is one of my kernels that tackles the interesting Toxic Comment Classification Challenge at Kaggle, which aims to identify and classify toxic online comments. In this notebook, we are going to