seq2seq-for-data-augmentation

Kit of functions to help modify a text dataset with a seq2seq model

Features

Append rows to data

Send the encoder and decoder model names from HuggingFace, the data as a DataFrame pd.DataFrame([[txt1, label1], [txt2, label2]], columns=['text', 'label']), model's maximum sequence length (default is 512), frequency of flipping a row (default is 0.5), random_state (given to train_test_split), and whether to append a row if it comes out identical to the original (default is False)

initial_df = pd.DataFrame([["Bueno", 0], ["La biblioteca", 1], ["La maestra es tonta", 0]],
  columns=['text', 'label']
)
append_sequenced(
    "monsoon-nlp/es-seq2seq-gender-encoder",
    "monsoon-nlp/es-seq2seq-gender-decoder",
    initial_df,
    seq_length=512,
    frequency=0.5,
    random_state=1,
    always_append=False
)

If randomly selected, the input ["La maestra es tonta", 1] will result in ["el maestro es tonto", 1] being appended to the returned DataFrame.

If always_append=True, the "bueno" and "la biblioteca" rows will be included, unmodified.

Replace rows in data

Send the encoder and decoder model names from HuggingFace, the data as a DataFrame pd.DataFrame([[txt1, label1], [txt2, label2]], columns=['text', 'label']), model's maximum sequence length (default is 512), frequency of flipping a row (default is 0.5), and random_state (given to train_test_split)

initial_df = pd.DataFrame([["Bueno", 0], ["La biblioteca", 1], ["La maestra es tonta", 0]],
  columns=['text', 'label']
)
replace_sequenced(
    "monsoon-nlp/es-seq2seq-gender-encoder",
    "monsoon-nlp/es-seq2seq-gender-decoder",
    initial_df,
    seq_length=512
    frequency=0.5,
    random_state=12
)

If randomly selected, the input ["La maestra es tonta", 1] will be replaced with ["el maestro es tonto", 1].

Applied

Working with SimpleTransformers to show data augmentation improves accuracy of classification and regression tasks:

https://colab.research.google.com/drive/194ITDA1AjxAx_4ZLjoRFQI1aWzsl7xU8?usp=sharing

Dependencies

pip install pandas scikit-learn transformers

License

Open source, MIT license