Skip to content

Latest commit

 

History

History
157 lines (108 loc) · 6.69 KB

README.md

File metadata and controls

157 lines (108 loc) · 6.69 KB
Raw

Towards identification of bias by source selection, commission and omission in a collection of news articles by identification of reused text fragments

This repository holds the source code of the master thesis title above. Click to Demo.

Screenshot:

alt text

Step 0: Prerequisites

Python v3.7.0+

Node.js v18.12.1+

npm v8.19.2+

Step 1: Dataset preparation

Dataset is collection of news articles of a specific topic and store as a row in a CSV file. Each row must have following fields:

Field Description Example
id Integer 1
title Article's headline Russia Demanded Spy Serving Life Sentence In Germany Swapped In Exchange For Paul Whelan
text Article's main content A report by CNN says that Russia demanded that Russian spy, Vadim Krasikov, be released along with Viktor Bout, a Russian arms dealer who was serving a 25-year sentence in the US, in exchange for Paul Whelan and Brittney Griner
outlet_name Outlet name (optional) CNN
datetime Published date with format dd/MM/yyyy hh:mm 9/12/2022 11:12
label Polarity assigned by outlet (LEFT, CENTER, RIGHT) CENTER

Step 2: Analysis of bias by COSS

Clone the project

$ git clone https://github.com/gipplab/news-story-identification

Install dependencies

$ pip install -r requirements.txt

Put your dataset in ./data folder, create more folders for your dataset if needed

Open the file ./consts.py, find the variable dataset and create a new dictionary object to specify your data files and directory path, a dataset can have multiple versions if needed. Example:

dataset = {
    'GROUNDNEWS': {
        'one_topic': {
            'FOLDER': './data/ground.news',
            'FILES': ['./0/data.csv']
        },
        'Full': {
            'FOLDER': './data/ground.news',
            'FILES': ['./0/data.csv', './1/data.csv', './2/data.csv', './3/data.csv', './4/data.csv', './5/data.csv']
        }
    }
}

Open the file ./by_coss.py, find the variables DATASET and DATASET_VERSION and update them with your dataset name and version of choice. Example:

DATASET = 'GROUNDNEWS'
DATASET_VERSION = 'Full'

Run following command

$ python ./by_coss.py

Results are generated in the folder ./data/<your_dataset_path>/by_coss.

Step 3: Visualization

Go to the folder vis.

$ cd vis

Install dependencies

$ ./vis> npm install

Copy your analysis results from the folder ./data/<your_dataset_path>/by_coss in Step 2 to ./public/

Open the file ./environments.js, find the variable dataFolder and update with your new path which you've just copied your results to (keep the keyword exports). Example

exports.dataFolder = `./public/data/ground.news/by_coss/`;

Run the following command

$ ./vis> node index.js

Open your favorite browser and visit http://localhost/ and start exploring results.

Other configurables

Change transformers models

Path to the paraphrase identification and polarity classification models can be edited by opening the file ./consts.py, find and edit the path for these two variables: paraphrase_identifier_modelname and polarity_classifier_path. The path can be a name to a huggingface model or in local folder.

Example:

# Hugging Face model
paraphrase_identifier_modelname = 'sentence-transformers/paraphrase-MiniLM-L3-v2'

# Local folder
polarity_classifier_path = './checkpoint-1130-paraphrase-albert-small-v2' #

Fine tune Polarity classfication model

The polarity classification model is fined-tune by either changing the pre-trained model in the variable polarity_classifier_path or edit training parameters, which is an object variable polarity_classification_configurations in the ./consts.py file. Example:

polarity_classification_configurations = {
    "modelname": paraphrase_identifier_modelname,
    "output_dir": "model",
    "learning_rate": 2e-5,
    "per_device_train_batch_size": 16,
    "per_device_eval_batch_size": 16,
    "num_train_epochs": 2,
    "weight_decay": 0.01,
    "evaluation_strategy": "epoch",
    "save_strategy": "epoch",
    "load_best_model_at_end": True,
    "push_to_hub": False,
}

Explanations of the parameters: https://huggingface.co/docs/transformers/v4.29.1/en/main_classes/trainer#transformers.TrainingArguments

Training dataset are structured exactly the same in Step 1.

To specify the training dataset location, open the file ./polarity_classification_automodel.py, find and update the variables DATASET and DATASET_VERSION similarly in Step 2.

Run following command to start fine-tuning

$ python polarity_classification_automodel.py

Resulting model is stored under the folder ./model or the location which is put in the variable polarity_classification_configurations['output_dir'] above.