The project "Protest Event in Local News" automates information extraction for protest event analysis from newspaper articles.

This repository provides trained models and prediction scripts.

The model files pytorch_model.bin and optimizer.pt are not staged to this repository. They can be shared via aws S3 buckets. To upload a file use this command:

aws s3 cp models/task-A/pytorch_model.bin s3://pea-fgz-models/task-A/pytorch_model.bin

Caution: uploaded files to this bucket will be public!

Task A: Relevance Classification

A binary classification task whether ot not an articles conatains a relevant protest event description.

Model: models/task-A/ Training data state: Bremen, Dresden, Leipzig, Stuttgart (2022-04-26) Performance: ca. F1 = 91 % in-sample dev, mcc = 82 % in-sample dev

Model Download

cd models/task-A
wget https://pea-fgz-models.s3.eu-central-1.amazonaws.com/task-A/pytorch_model.bin

Format Input Data

As input, a CSV file with ',' as separator, double quote escape strategy and column header is expected. For prediction, the column named text is used as input for prediction. The output file is a copy of the input CSV with three additional columns: pred_label, pred_score and pred_text. The latter contains the text after preprocessing that served as input to the classification model.

Run Prediction

For prediction, return to the project root directory and run (on the DSC server add CUDA_VISIBLE_DEVICES=0 srun before the command):

python task-A_prediction.py --input_csv ~/unlabeled_data.csv --output_csv ~/predicted_data.csv

Task B and C: Section identification and protest variable spans

A sequence tagging task to predict sentences containing relevant information for protest events (Task B, tagset ["PROTEST", "NONE"] and protest variables (Task C, tagset [CLAIM, FORM, TRAEGER, ZAHL, DATUM, NONE]).

Both tasks are tagged separately. Thus, tokens not tagged as PROTEST still may, for example, be tagged as ZAHL.

Model: Flair MultiTask models/task-B-C/best-model.pt Training data state: 607 gold documents from manual annotation (2022-04-01) Performance:

• ca. F1 = 71 % in-sample test for PROTEST tags (positive class), 80 % macro-F1
• ca. F1 = 62,5 % in-sample test for [CLAIM, FORM, TRAEGER, ZAHL, DATUM, NONE] (macro-avg)

Model Download

cd models/task-B-C
wget https://pea-fgz-models.s3.eu-central-1.amazonaws.com/task-B-C/best-model.pt

Format Input Data

As input, a CSV file with ',' as separator, double quote escape strategy and column header is expected. For prediction, the column doc_id is used to keep alignment of tokens to documents, and the column text is used as input for prediction. It makes sense, to only input documents classified as relevant in Task A.

The output file is a CoNLL-like format with tokens in rows and their predictions in colums.

Run Prediction

For prediction, return to the project root directory and run (on the DSC server add CUDA_VISIBLE_DEVICES=0 srun before the command):

python task-B-C_prediction.py --input_csv ~/unlabeled_data.csv --output_csv ~/predicted_data.csv