TagLM

nlp
language_model
review
Author

Shubham Gupta

Published

April 23, 2020

Introduction

  • This is the TagLM paper mentioned in Lecture 13 in the CS224N course title Semi-supervised sequence tagging with bidirectional language models

  • This paper demonstrates how we can use context embeddings from BiLSTM models and use it for sequence labelling tasks

Paper Introduction

  • Typically, RNN are used only on labelled data to learn the context embeddings of words.

  • Semi supervised approach used in this paper.

    • Train LM on large unlabelled corpus

    • Compute embedding at each position in the sequence(LM embedding)

    • Use embedding in supervised sequence tagging model

  • Using both forward and backward embeddings gives better performance than using forward only LM.

TagLM

Overview

  • Extract word and LM embeddings for every token

  • Prepare embedding by concatinating both embeddings

  • se them in supervised sequence tagging model

Baseline

  • Baseline model is hierachical neural tagging model

  • Obtain word and character embeddings. Concatenate them to form final embedding.

    \[ x_k = [c_k;w_k] \]

  • Char embedding: Can be obtained via CNN or RNN

  • Word embedding: Use pretrained embeddings

  • Use multiple bidirectional RNN to learn context embedding

  • For every token, concatenate forward and backward hidden states at each layer

  • 2nd layer will use the above output and predict next hidden state

  • Use GRU or LSTM depending on task

  • Use output of final layer to predict score for each possible tag using dense layer

  • Better to predict tags for full sequence than for a single token

  • THEREFORE, add another layer with parameters for each label bigram.

  • Compute sentence conditional random field loss(CRF) using forward-backward algorithm.

  • Use Viterbi algorithm to find most likely sequence

    Overview of architecture

  • LM embedding will be created by concatenating forward and backward embeddings. No parameter sharing between these two embeddings.

Experiments

  • Evaulation done on CoNLL 2003 NER and CoNLL 2000 chunking

  • Lot of detail around model architecture and training methods. Skipping this for now.

Analysis

  • Second RNN captures interactions between task specific context

  • Backward LM addition has significant performance gains

  • Model size makes a difference. Using bigger CNN model lead to  0.3 percent improvement

  • Also tried training the model JUST ON THE CoNLL data. Reduced model size

  • Including embeddings from this model decreased performance compared to baseline system.

  • Replacing task specific RNN with using LM embeddings with a dense layer and CRF reduced performance

  • Improvement shown by transferring knowledge from other tasks almost disappers when the initial model is trained on a large dataset.

Summary

  • First method to incorporate contextual word embeddings using bidirectional networks.
  • Significantly better than other methods at the time that use other forms of transfer learning or joint learning on NER and and Chunking tasks.
  • It works well with unlabelled data from a different domain as well.