Semeval 2020, Task 11

Overview

This repository provides code for the SemEval-2020 Task 11 competition (Detection of Propaganda Techniques in News Articles).

The competition webpage: https://propaganda.qcri.org/semeval2020-task11/

The description of the architecture of models can be found in our paper Aschern at SemEval-2020 Task 11: It Takes Three to Tango: RoBERTa, CRF, and Transfer Learning.

Requirements

pip install -r ./requirements.txt

Project structure

configs: yaml configs for the system
datasets: contains the task datasets, which can be downloaded from the team competition webpage
results: the folder for submissions
span_identification: code for the task SI
- ner: pytorch-transformers RoBERTa model with CRF (end-to-end)
- dataset: the scripts for loading and preprocessing source dataset
- submission: the scripts for obtaining and evaluating results
technique_classification: code for the task TC (the folder has the same structure as span_identification)
tools: tools provided by the competition organizers; contain useful functions for reading datasets and evaluating submissions
visualization_example: example of visualization of results for both tasks

Running the models

All commands are run from the root directory of the repository.

Span Identification

Configure configs/si_config.yml file, if it is needed. data_dir is the path to the cache of original train/eval sub-datasets and their BIO versions. In addition to using the config, it is also possible to specify arguments through the command line.
Split the dataset for local evaluation (if --overwrite_cache, previous files will be replaced). It will produce files with the BIO-format tagging for spans (B-PROP, I-PROP, O) in your --data_dir.
```
python -m span_identification --config configs/si_config.yml --split_dataset --overwrite_cache
```
Train and eval model (the model parameters are specified in the config, you need to change the paths). The use of CRF is regulated by the flag --use_crf. For the first run you can use --model_name_or_path roberta-large.
```
python -m span_identification --config configs/si_config.yml --do_train --do_eval
```
Apply the trained model to the test_file (in BIO-format) specified in the config. It will be created based on the test_data_folder folder in case of missing or if the flag --overwrite_cache is specified.
```
python -m span_identification --config configs/si_config.yml --do_predict
```
Create the submission file output_file in the result folder. It will obtain spans from the result files with the token labeling specified in predicted_labels_files. At the aggregation stage, the span prediction results are simply joined.
```
python -m span_identification --config configs/si_config.yml --create_submission_file
```
In case you have the correct markup in the test_file or gold --gold_annot_file (source competition format), you can run the evaluation competition script.
```
python -m span_identification --config configs/si_config.yml --do_eval_spans
```
Use visualization_example/visualization.ipynb if you want to visualize labels.

Technique Classification

Here you need almost the same commands and settings as in the SI task.

Configure configs/tc_config.yml file, if it is needed.

Split the dataset for local evaluation.

python -m technique_classification --config configs/tc_config.yml --split_dataset --overwrite_cache

Train and eval model. We used two setups with and without flags --join_embeddings --use_length (to get our RoBERTa-Joined). For the first run you can use --model_name_or_path roberta-large.

python -m technique_classification --config configs/tc_config.yml --do_train --do_eval

or distributed

CUDA_VISIBLE_DEVICES=0,1,2,3 python -m torch.distributed.launch --nproc_per_node 4 technique_classification --config configs/tc_config.yml --do_train --do_eval

Apply the trained model to the test_file specified in the config. It will be created based on the test_data_folder folder and test_template_labels_path file in case of missing or if the flag --overwrite_cache is specified.
```
python -m technique_classification --config configs/tc_config.yml --do_predict --join_embeddings --use_length
```
Create the submission file output_file. It will combine predictions from the list predicted_logits_files with coefficients specified in --weights (optional) and apply some post-processing.
```
python -m technique_classification --config configs/tc_config.yml --create_submission_file
```
In case you have the correct markup in the test_file or gold --test_labels_path (source competition format), you can check your accuracy (micro f1-score) and f1-score per classes.
```
python -m technique_classification --config configs/tc_config.yml  --eval_submission
```
Use visualization_example/visualization.ipynb if you want to visualize labels.

Our pretrained RoBERTa-CRF (SI task) and RoBERTa-Joined (TC task) models are available in Google Drive.

Citation

If you find this repository helpful, feel free to cite our publication Aschern at SemEval-2020 Task 11: It Takes Three to Tango: RoBERTa, CRF, and Transfer Learning:

@inproceedings{chernyavskiy-etal-2020-aschern,
    title = "Aschern at {S}em{E}val-2020 Task 11: It Takes Three to Tango: {R}o{BERT}a, {CRF}, and Transfer Learning",
    author = "Chernyavskiy, Anton  and
      Ilvovsky, Dmitry  and
      Nakov, Preslav",
    booktitle = "Proceedings of the Fourteenth Workshop on Semantic Evaluation",
    month = dec,
    year = "2020",
    address = "Barcelona (online)",
    publisher = "International Committee for Computational Linguistics",
    url = "https://www.aclweb.org/anthology/2020.semeval-1.191",
    pages = "1462--1468"
}

SimPPL/semeval2020_task11