Toxic-Spans-Detection

Repository for our code and experiments on SemEval-2021 Task-5 Toxic Spans Detection. We are still updating this repository and trying to make the code more efficient. We will be posting the updates here whenever possible. We would love to know about any issues found on this repository. Please create an issue for any queries, or you contact us at chhablani.gunjan@gmail.com.

Pre-print: http://arxiv.org/abs/2102.12254

Updates

[25 Feb 2021]: Repository is made public.

Usage

Setting Up

Install src using the following command:

python setup.py install

Install the requirements:

pip install -r requirements.txt

Baseline Models

RNNSL

python run_baseline_model.py --configs ./configs/rnnsl/default.yaml

Running this command will save the predictions for train, trial, and test, and corresponding F₁ scores.

SpaCy

cd baselines
python spacy_tagging.py

Running this command will save the predictions for train, trial, and test, and corresponding F₁ scores.

Training

For BERT-based models, the configurations are present in the configs directory. You can choose a config of your liking. An example for bert_token:

python train.py --train ./configs/bert_token/train.yaml --data ./configs/bert_token/dataset.yaml

This will save the logs, checkpoints, and the final model at the path specified in train.yaml.

Evaluating

Evaluation is done based on a checkpoint mentioned in eval.yaml configurations. Please ensure the correctness of the checkpoint path before continuing.

python eval.py --eval ./configs/bert_token/eval.yaml

Running this command will save the predictions for train, trial, and test, and corresponding F₁ scores.

Integrated Gradients

You can also apply Integrated Gradients to an example of your choosing. Change the sample_index in the configuration corresponding to the example. By default, the test dataset is used, but you can also choose to use this on other datasets. Please ensure correctness of the checkpoint path before continuing. An example command to run for roberta_token is:

python run_integrated_gradients.py --config ./configs/integrated_gradients/roberta_token.yaml

Running this command will save the word_importances, token_importances binary files, with format (words, importances) and (tokens, importances) respectively. A visualization will also be save in a file called viz.html.

Combining Predictions

We combine checkpoint predictions using intersection or union. An example for union of top-3 roberta checkpoints:

python /src/utils/combine_preds.py ./configs/combine_predictions/union_roberta_token_best_3_ckpts.yaml

Evaluating Prediction Files

You will need ground truths, predictions to be able to run this evaluation. We calculate F₁ scores on the given file using the following command:

python calculate_f1_scores.py --ground <ground> --preds <preds> --out <out>

where <ground> is path to the ground truths, <preds> is the prediction file, and <out> is the path where the scores are to be saved. Running this command will save the F₁ scores to the path.

Directory Structure

.
├── baselines
│   ├── __init__.py
│   ├── models.py
│   ├── spacy_tagging.py
├── calculate_f1_scores.py # To calculate F1 scores given preds and grounds
├── configs
│   ├── bert_base_spans
│   │   ├── dataset.yaml
│   │   ├── eval.yaml
│   │   └── train.yaml
│   ├── bert_base_token
│   │   ├── dataset.yaml
│   │   ├── eval.yaml
│   │   └── train.yaml
│   ├── bert_crf_token
│   │   ├── dataset.yaml
│   │   ├── eval.yaml
│   │   └── train.yaml
│   ├── bert_multi_spans
│   │   ├── dataset.yaml
│   │   ├── eval.yaml
│   │   └── train.yaml
│   ├── bert_spans
│   │   ├── dataset.yaml
│   │   ├── eval.yaml
│   │   └── train.yaml
│   ├── bert_token
│   │   ├── dataset.yaml
│   │   ├── eval.yaml
│   │   └── train.yaml
│   ├── bert_token_spans
│   │   ├── dataset.yaml
│   │   ├── eval.yaml
│   │   └── train.yaml
│   ├── combine_predictions
│   │   ├── intersection_roberta_token_best_3_ckpts.yaml
│   │   ├── intersection_roberta_token_union_spanbert_spans_best_3_ckpts.yaml
│   │   ├── intersection_spanbert_spans_best_3_ckpts.yaml
│   │   ├── intersection_spanbert_spans_union_roberta_token_best_3_ckpts.yaml
│   │   ├── union_roberta_token_best_3_ckpts.yaml
│   │   ├── union_roberta_token_union_spanbert_spans_best_3_ckpts.yaml
│   │   ├── union_spanbert_spans_best_3_ckpts.yaml
│   │   └── union_spanbert_spans_union_roberta_token_best_3_ckpts.yaml
│   ├── integrated_gradients
│   │   ├── roberta_token.yaml
│   │   └── spanbert_spans.yaml
│   ├── rnnsl
│   │   └── default.yaml
│   ├── rnnsl_tsd_train_trial
│   │   └── default.yaml
│   ├── roberta_base_spans
│   │   ├── dataset.yaml
│   │   ├── eval.yaml
│   │   └── train.yaml
│   ├── roberta_base_token
│   │   ├── dataset.yaml
│   │   ├── eval.yaml
│   │   └── train.yaml
│   ├── roberta_crf_token
│   │   ├── dataset.yaml
│   │   ├── eval.yaml
│   │   └── train.yaml
│   ├── roberta_multi_spans
│   │   ├── dataset.yaml
│   │   ├── eval.yaml
│   │   └── train.yaml
│   ├── roberta_spans
│   │   ├── dataset.yaml
│   │   ├── eval.yaml
│   │   └── train.yaml
│   ├── roberta_token
│   │   ├── dataset.yaml
│   │   ├── eval.yaml
│   │   ├── roberta_token_best_3_ckpts
│   │   │   ├── eval_1500.yaml
│   │   │   ├── eval_4000.yaml
│   │   │   └── eval_5000.yaml
│   │   └── train.yaml
│   ├── roberta_token_spans
│   │   ├── dataset.yaml
│   │   ├── eval.yaml
│   │   └── train.yaml
│   ├── roberta_token_tsd_train_trial
│   │   ├── dataset.yaml
│   │   ├── eval.yaml
│   │   └── train.yaml
│   ├── spanbert_crf_token
│   │   ├── dataset.yaml
│   │   ├── eval.yaml
│   │   └── train.yaml
│   ├── spanbert_multi_spans
│   │   ├── dataset.yaml
│   │   ├── eval.yaml
│   │   └── train.yaml
│   ├── spanbert_spans
│   │   ├── dataset.yaml
│   │   ├── eval.yaml
│   │   ├── spanbert_spans_best_3_ckpts
│   │   │   ├── eval_2000.yaml
│   │   │   ├── eval_3500.yaml
│   │   │   └── eval_5000.yaml
│   │   └── train.yaml
│   ├── spanbert_spans_tsd_train_trial
│   │   ├── dataset.yaml
│   │   ├── eval.yaml
│   │   └── train.yaml
│   ├── spanbert_token
│   │   ├── dataset.yaml
│   │   ├── eval.yaml
│   │   └── train.yaml
│   ├── spanbert_token_spans
│   │   ├── dataset.yaml
│   │   ├── eval.yaml
│   │   ├── spanbert_token_spans_spans
│   │   │   └── eval.yaml
│   │   ├── spanbert_token_spans_token
│   │   │   └── eval.yaml
│   │   └── train.yaml
│   ├── toxicbert_spans
│   │   ├── dataset.yaml
│   │   ├── eval.yaml
│   │   └── train.yaml
│   ├── toxicbert_token
│   │   ├── dataset.yaml
│   │   ├── eval.yaml
│   │   └── train.yaml
│   ├── toxicroberta_spans
│   │   ├── dataset.yaml
│   │   ├── eval.yaml
│   │   └── train.yaml
│   └── toxicroberta_token
│       ├── dataset.yaml
│       ├── eval.yaml
│       └── train.yaml
├── data
│   ├── clean_versions (incorrect)
│   │   ├── clean_train.csv
│   │   ├── clean_train_trial.csv
│   │   ├── clean_trial.csv
│   │   └── modified_train.csv
│   ├── tsd_test.csv
│   ├── tsd_test_spans.csv
│   ├── tsd_train.csv
│   ├── tsd_train_trial.csv
│   └── tsd_trial.csv
├── eval.py
├── evaluation
│   ├── fix_spans.py
│   ├── fix_spans_test.py
│   ├── __init__.py
│   ├── metrics.py
│   ├── semeval2021.py
│   └── semeval2021_test.py
├── __init__.py
├── integrated_gradients.py
├── LICENSE
├── notebooks
│   ├── Exploratory Data Analysis & Preprocessing.ipynb
│   └── Span Length, Contiguous Spans stats.ipynb
├── print_predictions.py
├── README.md
├── requirements.txt
├── results
├── run_baseline_model.py
├── setup.py
├── src
│   ├── datasets
│   │   ├── __init__.py
│   │   ├── toxic_spans_crf_tokens.py
│   │   ├── toxic_spans_multi_spans.py
│   │   ├── toxic_spans_spans.py
│   │   ├── toxic_spans_tokens.py
│   │   └── toxic_spans_tokens_spans.py
│   ├── models
│   │   ├── auto_models.py
│   │   ├── bert_crf_token.py
│   │   ├── bert_multi_spans.py
│   │   ├── bert_token_spans.py
│   │   ├── __init__.py
│   │   ├── roberta_crf_token.py
│   │   ├── roberta_multi_spans.py
│   │   ├── roberta_token_spans.py
│   │   └── two_layer_nn.py
│   ├── modules
│   │   ├── activations.py
│   │   ├── embeddings.py
│   │   ├── __init__.py
│   │   ├── losses.py
│   │   ├── metrics.py
│   │   ├── optimizers.py
│   │   ├── preprocessors.py
│   │   ├── schedulers.py
│   │   └── tokenizers.py
│   ├── trainers
│   │   ├── base_trainer.py
│   │   ├── __init__.py
│   └── utils
│       ├── combine_preds.py
│       ├── configuration.py
│       ├── __init__.py
│       ├── logger.py
│       ├── mapper.py
│       ├── misc.py
│       ├── postprocess_predictions.py
│       └── viz.py
├── ToxicSpans_SemEval21.ipynb
└── train.py

Tasks

Done

Add Directory Structure
Add Usage

Ongoing

Update README
- Add Approaches
- Add Results and Analysis

To-Do

Fix Linting Issues
Update Docs

Approaches

Results and Analysis

Data

Name	Value
Train Data Size	7939
Trial Data Size	690
Test Data Size	2000

	Train	Trial	Test
Spaces Marked as Toxic (sum)	13278.00	830	575.00
Spaces Marked as Toxic (mean)	1.67	1.20	0.29
Spaces Marked as Toxic (std)	7.72	4.48	3.19
Words Cut in Spans (sum)	263	26	8
Words Cut in Spans (mean)	0.03	0.04	0.00
Words Cut in Spans (std)	0.20	0.23	0.06
Spans start/end w space (sum)	20	1	1
Spans start/end w space (mean)	0.00	0.00	0.00
Spans start/end w space (std)	0.05	0.00	0.02

Based on BERT Tokenizer:

	Train	Trial	Test
#Tokens(mean)	47.5	46.1	43.12
#Tokens(std)	45.46	43.82	39.88
#Tokens(max)	335	234	291
#Tokens(min)	1	1	2
#Words(mean)	35.95	35.01	32.86
#Words(std)	34.97	34.42	31.01
#Words(max)	192	182	186
#Words(min)	1	1	1
#Chars(mean)	204.57	199.47	186.41
#Chars(std)	201.37	196.63	178.76
#Chars(max)	1000	998	1000
#Chars(min)	4	5	6

	Train	Trial	Test
Num Contiguous Spans(Mean)	1.3	1.31	0.93
Num Contiguous Spans(Std)	0.84	0.74	0.62
Num Contiguous Spans(Max)	25	6	7
Num Contiguous Spans(Min)	0	0	0
Len Contiguous Spans(Mean)	13.51	11.3	7.89
Len Contiguous Spans(Std)	38.57	20.76	17.86
Len Contiguous Spans(Max)	994	350	713
Len Contiguous Spans(Min)	1	1	3
Per Contiguous Spans(Mean)	0.14	0.14	0.09
Per Contiguous Spans(Std)	0.2	0.2	0.14
Per Contiguous Spans(Max)	1	1	1
Per Contiguous Spans(Min)	0	0	0

References

Citation

You can cite our work as:

@misc{chhablani2021nlrg,
      title={NLRG at SemEval-2021 Task 5: Toxic Spans Detection Leveraging BERT-based Token Classification and Span Prediction Techniques},
      author={Gunjan Chhablani and Yash Bhartia and Abheesht Sharma and Harshit Pandey and Shan Suthaharan},
      year={2021},
      eprint={2102.12254},
      archivePrefix={arXiv},
      primaryClass={cs.CL}
}

If you use any part of our code in your work, please use the following citation:

@misc{chhablani2021nlrggithub,
  author = {Gunjan Chhablani and Yash Bhartia and Abheesht Sharma and Harshit Pandey and Shan Suthaharan},
  title = {gchhablani/toxic-spans-detection},
  year = {2021},
  publisher = {GitHub},
  journal = {GitHub repository},
  howpublished = {\url{https://github.com/gchhablani/toxic-spans-detection}},
}

Name		Name	Last commit message	Last commit date
Latest commit History 228 Commits
baselines		baselines
configs		configs
data		data
evaluation		evaluation
notebooks		notebooks
results/test_predictions		results/test_predictions
src		src
.gitignore		.gitignore
LICENSE		LICENSE
README.md		README.md
ToxicSpans_SemEval21.ipynb		ToxicSpans_SemEval21.ipynb
__init__.py		__init__.py
calculate_f1_scores.py		calculate_f1_scores.py
eval.py		eval.py
example_markdown_table.md		example_markdown_table.md
example_table.txt		example_table.txt
generate_examples_table.py		generate_examples_table.py
integrated_gradients.py		integrated_gradients.py
neural_sequence_labeler.model.png		neural_sequence_labeler.model.png
print_predictions.py		print_predictions.py
requirements.txt		requirements.txt
run_baseline_model.py		run_baseline_model.py
setup.py		setup.py
train.py		train.py

License

gchhablani/toxic-spans-detection

Folders and files

Latest commit

History

Repository files navigation

Toxic-Spans-Detection

Updates

Usage

Setting Up

Baseline Models

Training

Evaluating

Integrated Gradients

Combining Predictions

Evaluating Prediction Files

Directory Structure

Tasks

Done

Ongoing

To-Do

Approaches

Results and Analysis

Data

References

Citation

About

Resources

License

Stars

Watchers

Forks

Languages