README.md

ELMo Based Classifier

ELMo Introduction

ELMo, or Embeddings from Language Model, is a bidirectional language model based on RNN.

Bidirectional LSTM units are pre-trained with unlabeled data to generate ELMo representations. Then we can easily fine-tune the model to help us to classify labeled texts.

In this repository, we use the pre-trained ELMo with Tensorflow Hub.

MLP(Multi-Layer Perceptron) is used in our classification task. Accuracy might be higher if we used DNN or CNN rather than MLP.

Dataset

The classification task is subtask B of SemEval-2019 Task 8: Fact Checking in Community Question Answering Forums. The data for this task is provided in another repository: https://github.com/tsvm/factcheck-cqa.

Requirements

You should install python version 3.5 or later, Tensorflow version 1.x and Tensorflow Hub.

pip install tensorflow tensorflow-hub

Run

optional arguments:
  -h, --help            show this help message and exit
  --num_epochs NUM_EPOCHS
                        Number of epochs.
  --batch_size BATCH_SIZE
                        Batch size.
  --learning_rate LEARNING_RATE
                        Learning rate.
  --lamb LAMB           Regularization coefficient for weights.
  --output_path OUTPUT_PATH
                        Save path.
  --dataset_path DATASET_PATH
                        Data path.
  --labels LABELS       Labels of texts.
  --use_gpu USE_GPU     whether to use gpu.
  --do_train DO_TRAIN   whether to train the model.
  --do_eval DO_EVAL     whether to do the evaluation.
  --do_test DO_TEST     whether to predict test data.

Train and evaluate:

python run_classifier.py

For Windows users:

./run.bat