Skip to content

LeoLaugier/conditional-auto-encoder-text-to-text-transfer-transformer

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 
 
 
 
 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

CAET5: Mitigating toxicity in online conversations using self-supervised transformers

CAET5 serves as code for fine-tuning pre-trained text-to-text transformers from Exploring the Limits of Transfer Learning with a Unified Text-to-Text Transformer on self-supervised attribute transfer tasks.

The code overrides objects from the T5 and the Mesh TensorFlow Transformer packages.

Table of Contents

Library

caet5

caet5 reproduces the structure of the T5 package.

caet5.data redefines Task objects. Please see the t5.data documentation for more details about the t5.data package.

We added:

  • functions to pre-process unpaired datasets made of several text files containing attribute-exclusive examples, with one example per line.
  • a text preprocessor function adapted to self-supervised attribute transfer.

We adapted functions from t5 that were not initially adapted to self-supervised attribute transfer.

caet5.evaluation adds metrics computed with (pre-trained / fine-tuned) parametric models (in particular transformers) and used to evaluate unsupervised attribute transfer:

  • sentence similarity (SIM)
  • attribute transfer accuracy (ACC)
  • perplexity (PPL)

caet5.models adapts the t5.models shims to unsupervised training, evaluation and inference methods for attribute transfer.

mesh_tensorflow_caet5

mesh_tensorflow_caet5 overrides objects of the Mesh TensorFlow Transformer package, to fit CAET5's training and evaluation approach.

Usage

The easiest way to try out CAET5 is with a free TPU on Colab.

Below we provide examples for how to fine-tune, evaluate and infer from a model from the model API. You can use these instructions to reproduce our results, fine-tune one of T5's released checkpoints with your own data and/or hyperparameters.

Dataset Preparation

You may either use a new or pre-existing Task_ll, or you may load examples from "raw" text files, each containing single attribute examples.

Using a Task_ll

Depending on your data source (see t5.data documentation), you will need to prepare your data appropriately.

Just make sure any file(s), either raw files or pre-processed file(s) loaded by your dataset_fn are accessible to the TPU (i.e., are in a GCS bucket).

Unsupervised Metric Preparation

In order to compute attribute transfer accuracy and perplexity, you need to store pre-trained parametric models. CAET5 currently supports BERT classification models fine-tuned on attribute classification and GPT2 language models, by default stored in gs://yourbucket/[metric]_binaries/[architecture]_[metric]_[mixture_or_task_name].pt where [metric] is "acc" or "ppl", and [architecture] is "bert" or "gpt2".

Installation

To install the CAET5 package, clone the github repo and run:

pip install .

Setting up TPUs on GCP

For details about setting up TPUs on GCP, please see the t5 documentation.

In order to run training or eval on Cloud TPUs, you must set up the following variables based on your project, zone and GCS bucket appropriately.

export PROJECT=your_project_name
export ZONE=your_project_zone
export BUCKET=yourbucket
export TPU_NAME=t5-tpu
export BASE_DIR=gs://yourbucket/
export MODELS_DIR_NAME=your_models_dir_name
export DATA_DIR_NAME=your_data_dir
export DATA_RAW_DIR_NAME=your_data_raw_dir_name

Fine-tuning

In order to fine-tune one of T5's pre-trained models, on an attribute transfer task called [mixture_or_task_name], please run:

caet5  \
  --tpu="${TPU_NAME}" \
  --gcp_project="${PROJECT}" \
  --tpu_zone="${ZONE}" \
  --bucket="${BUCKET}" \
  --base_dir="${BASE_DIR}" \
  --model_dir_name="${MODELS_DIR_NAME}" \
  --model_size=your_model_size \
  --data_dir_name="${DATA_DIR_NAME}" \
  --data_raw_dir_name="${DATA_RAW_DIR_NAME}" \
  --module_import=caet5.data.tasks \
  --use_model_api=True \
  --mode="finetune" \
  --train_steps=100000 \
  --mixture_or_task=[mixture_or_task_name] \
  --base_pretrained_model_dir="gs://t5-data/pretrained_models/" \
  --gin_file="dataset.gin" \
  --gin_file="objectives/denoise.gin" \
  --gin_file="models/cae_bi.gin" \
  --gin_file="train.gin" \
  --gin_file="sequence_lengths/[mixture_or_task_name]" \
  --gin_param="utils.tpu_mesh_shape.tpu_topology = '2x2'"

Eval

In order to evaluate a model in the CAET5 framework, you need to specify the model directory and which checkpoint step(s) to evaluate. So, to evaluate on the [mixture_or_task_name] task on all checkpoints, use the following command:

caet5 --tpu="${TPU_ADDRESS}" \
      --gcp_project="${PROJECT}" \
      --tpu_zone="${ZONE}" \
      --bucket="${BUCKET}" \
      --base_dir="${BASE_DIR}" \
      --model_dir_name="${MODELS_DIR_NAME}" \
      --model_size="${MODEL_SIZE}" \  
      --data_dir_name="${DATA_DIR_NAME}" \
      --module_import=caet5.data.tasks \
      --use_model_api=True \
      --mode="eval" \      
      --mixture_or_task=[mixture_or_task_name] \
      --base_pretrained_model_dir="gs://t5-data/pretrained_models/" \
      --checkpoint_mode="all" \
      --gin_file="dataset.gin" \
      --gin_file="models/cae_bi.gin" \
      --gin_file="sequence_lengths/[mixture_or_task_name]" \
      --gin_param="utils.tpu_mesh_shape.tpu_topology = '2x2'"

To evaluate a specific checkpoint, simply set the eval_checkpoint_step parameter to appropriate checkpoint.

--gin_param="eval_checkpoint_step = 100000"

Decode

In order to produce predictions from a model in the CAET5 framework, you need to use the infer.gin file, specify the model directory and which checkpoint step(s) to use for decoding. Assuming you have a text file of input sequences and destination attribute stored at /path/to/intputs.txt, an example command would be:

caet5 --tpu="${TPU_ADDRESS}" \
      --gcp_project="${PROJECT}" \
      --tpu_zone="${ZONE}" \
      --bucket="${BUCKET}" \
      --base_dir="${BASE_DIR}" \
      --model_dir_name="${MODELS_DIR_NAME}" \
      --model_size="${MODEL_SIZE}" \  
      --data_dir_name="${DATA_DIR_NAME}" \
      --module_import=caet5.data.tasks \
      --use_model_api=True \
      --mode="predict" \      
      --mixture_or_task=[mixture_or_task_name] \
      --base_pretrained_model_dir="gs://t5-data/pretrained_models/" \
      --checkpoint_mode="latest" \
      --input_file='/path/to/inputs.txt' \
      --output_file='/tmp/outputs.txt' \
      --predict_batch_size=your_predict_batch_size \
      --gin_file="dataset.gin" \
      --gin_file="models/cae_bi.gin" \
      --gin_file="infer.gin" \
      --gin_file="sequence_lengths/[mixture_or_task_name]" \
      --gin_param="utils.tpu_mesh_shape.tpu_topology = '2x2'"

How to Cite

If you extend or use this work, please cite the paper where it was introduced:

@inproceedings{laugier-etal-2021-civil,
    title = "Civil Rephrases Of Toxic Texts With Self-Supervised Transformers",
    author = "Laugier, L{\'e}o  and
      Pavlopoulos, John  and
      Sorensen, Jeffrey  and
      Dixon, Lucas",
    booktitle = "Proceedings of the 16th Conference of the European Chapter of the Association for Computational Linguistics: Main Volume",
    month = apr,
    year = "2021",
    address = "Online",
    publisher = "Association for Computational Linguistics",
    url = "https://aclanthology.org/2021.eacl-main.124",
    pages = "1442--1461",
    abstract = "Platforms that support online commentary, from social networks to news sites, are increasingly leveraging machine learning to assist their moderation efforts. But this process does not typically provide feedback to the author that would help them contribute according to the community guidelines. This is prohibitively time-consuming for human moderators to do, and computational approaches are still nascent. This work focuses on models that can help suggest rephrasings of toxic comments in a more civil manner. Inspired by recent progress in unpaired sequence-to-sequence tasks, a self-supervised learning model is introduced, called CAE-T5. CAE-T5 employs a pre-trained text-to-text transformer, which is fine tuned with a denoising and cyclic auto-encoder loss. Experimenting with the largest toxicity detection dataset to date (Civil Comments) our model generates sentences that are more fluent and better at preserving the initial content compared to earlier text style transfer systems which we compare with using several scoring systems and human evaluation.",
}

Releases

No releases published

Packages

No packages published