Good-Enough Data Augmentation

A simple rule-based data augmentation scheme aimed at encouraging generalization in sequence-to-sequence models.

Jacob Andreas, ACL 2020. https://arxiv.org/abs/1904.09545

Data:

• Semantic parsing dataset (Finnegan-Dolak et al.): https://github.com/jkkummerfeld/text2sql-data

• SCAN dataset (Lake and Baroni): https://github.com/brendenlake/SCAN

• Language modeling: Under data/lm. All data is from Wikipedia, except for the Na data, which is derived from "Cross-Lingual Word Embeddings for Low-Resource Language Modeling" (Adams et al. 2017). Note different train/test splits for Na.

• Human sequence-to-sequence learning (Lake, Linzen and Baroni): Fig 2 in https://arxiv.org/pdf/1901.04587.pdf

• COGS dataset (Kim and Linzen): https://github.com/najoungkim/COGS

To use on a new dataset:

1. Point torchdec at https://github.com/jacobandreas/torchdec.
2. Create a new data loader under data (look at data/colors.py for a minimal example).
3. Update get_dataset in train.py to use the new loader.
4. Run the experiment pipeline (look at exp/scan_jump/retrieval/run.sh for an example).

The wug_size and wug_count flags (defined in data/builder.py) determine the number and size of the fragments that will be extracted from each template. the template_sim flag determines whether the whole string or a fixed-size window will be used for evaluating template similarity; sim_window_size determines the window size. The number and diversity of generated templates can be further controlled using the variants and n_sample flags.