Exploring Data Augmentation Methods through Attribution

The following research explores data attribution methods to identify training subsets from a larger training set with varying contributions to a model's performance. Two approaches have been tested here - (1) Captum's Layer Integrated Gradients for feature attribution in text classification, and (2) a self-implemented perplexity-based text quality filtering method inspired by a recent Meta research paper for translation tasks. Additionally, we examine the effects of applying data augmentation to the top 10% and bottom 10% of training samples in both tasks using established techniques, including Easy Data Augmentation (EDA), An Easier Data Augmentation (AEDA), backtranslation (English ↔ French), and paraphrasing with the instruction-following variant of the Llama 3.1 (8B) model.

Project Structure

|- assets
|- attribution
  |-- lig.py
  |-- perplexity.py
|- augmentation
  |-- An Easier Data Augmentation
  |-- Backtranslation
  |-- Easy Data Augmentation
  |-- Paraphrasing
|- datasets
|- plots
|- train
  |-- bert.py
  |-- mT5.py

Datasets

The datasets for both tasks are available as .csv files in this repository.

• Task 1 - Classification

  • SST2 - binary classification dataset
  • SST5 - multi-label classification dataset
  • IMDB - binary classification dataset
  • AGNews10K - multi-label classification dataset

• Task 2 - Language Translation

  • English to Afrikaans
  • English to Welsh
  • English to Czech
  • English to Spanish
  • English to Romanian

Methodology

• Task 1 - Classification

• Task 2 - Language Translation