Recipe for data augmentation based on conditional text generation

The repository contains the recipe to run data augmentation based on conditional text ge neration in order to automatically generate labeled data using a training set, following the approach described by Not Enough Data? Deep Learning to the Rescue! and Natural Language Generation for Effective Knowledge Distillation.

First, we fine-tune a causal language model like GPT-2 on the training set. Each sample contains both the label and the sentence. We mark the end of the sample by a special token EOS.

Second, we synthesize labeled data. Given a class label, we use the fine-tuned language model to predict the sentence until EOS.

We show here an example how to augment GLUE SST-2 training set.

Installation

The code runs with Python3. It is based on transformers (3.6.1) and datasets (1.6.2) libraries from HuggingFace. The script run_clm.py from transformers examples is used for fine-tuning GPT2.

Prepare the data

This following code loads and prepares the data by adding EOS token at the end of each sample.

python prepare_data.py

Fine-tuning Causal Language Model

The following code fine-tunes GPT-2 (gpt2-medium) on SST-2. The loss here is that of causal language modeling.

DATASET=SST-2
RAIN_FILE=$DATASET/train.txt
TEST_FILE=$DATASET/validation.txt
MODEL=gpt2-medium
MODEL_DIR=model/$MODEL-$DATASET

python3 transformers/examples/pytorch/language-modeling/run_clm.py \
    --model_name_or_path $MODEL \
    --train_file $TRAIN_FILE \
    --validation_file $TEST_FILE \
    --do_train \
    --do_eval \
    --output_dir $MODEL_DIR \
    --overwrite_output_dir

Data augmentation

The following code generates samples, given the label.

DATASET=SST-2
NUM_RETURN_SEQUENCES=800000
MODEL=gpt2-medium
MODEL_DIR=model/$MODEL-$DATASET
OUTPUT=augmented-glue-sst2/train.tsv

python data_augment.py \
 --input $DATASET/train.txt \
 --model_type gpt2 \
 --model_name_or_path $MODEL_DIR \
 --output $OUTPUT \
 --num_return_sequences $NUM_RETURN_SEQUENCES

Citation

Not Enough Data? Deep Learning to the Rescue!, Ateret Anaby-Tavor, Boaz Carmeli, Esther Goldbraich, Amir Kantor, George Kour, Segev Shlomov, Naama Tepper, Naama Zwerdling, AAAI 2020.

Natural Language Generation for Effective Knowledge Distillation, Raphael Tang, Yao Lu, Jimmy Lin, DeepLo 2019.