Simple code notes for explaining NLP building blocks

• Subword Segmentation Techniques
  • Let's compare various tokenizers, i.e., nltk, BPE, SentencePiece, and Bert tokenizer.
• Beam Decoding
  • Beam decoding is essential for seq2seq tasks. But it's notoriously complicated to implement. Here's a relatively easy one, batchfying candidates.
• How to get the last hidden vector of rnns properly
  • We'll see how to get the last hidden states of Rnns in Tensorflow and PyTorch.
• Tensorflow seq2seq template based on the g2p task
  • We'll write a simple template for seq2seq using Tensorflow. For demonstration, we attack the g2p task. G2p is a task of converting graphemes (spelling) to phonemes (pronunciation). It's a very good source for this purpose as it's simple enough for you to up and run.
• PyTorch seq2seq template based on the g2p task
  • We'll write a simple template for seq2seq using PyTorch. For demonstration, we attack the g2p task. G2p is a task of converting graphemes (spelling) to phonemes (pronunciation). It's a very good source for this purpose as it's simple enough for you to up and run.
• [Attention mechanism](Work in progress)
• POS-tagging with BERT Fine-tuning
  • BERT is known to be good at Sequence tagging tasks like Named Entity Recognition. Let's see if it's true for POS-tagging.
• Dropout in a minute
  • Dropout is arguably the most popular regularization technique in deep learning. Let's check again how it work.
• Ngram LM vs. rnnlm(WIP)
• Data Augmentation for Quora Question Pairs
  • Let's see if it's effective to augment training data in the task of quora question pairs.