COVID19-Fake-News-Detection-in-English link

class identification

• Fake (1): post or tweet provide information cosidered as fake.
• Real (0): post or tweet which provided verified information.

Preprocessing

"👑🦠": "coronavirus","📍":"update","🆕":"update","📌":"update","📢":"update","📣":"update","🔴":"warning",
"👍":"approved","🔰":"driving sign","🙏":"thank you","😷":"wear a mask","❗":"excitement","🌍":"europe africa",
"🌎":"america","🌏":"asia austrila","🌐":"globe","🏫":"school","📺":"television","👏":"hand washing",
"🤔":"thinking","🏠":"stay at home","🏡":"stay at home","🕌":"mosque","📈":"increasing","📉":"decreasing",
"🙌":"hand washing","👐":"hand washing","👦":"boy","❌":"not","🙄":"rolling eye","📏":"keep distance",
"👥":"social distance","💉":"vaccination","\U0001f9ea""vaccination","🧪":"vaccination" ,"💊":" pill",
"🗣️":"speak" ,"🚍":"transportation","🚌":"transportation","💺":"seat" ,"🐍":"snake","🏥":"hospital",
"🚑":"ambulance"  ,"🚨":"emergency","🏪":"convenience store","😞":"disappointed" ,"🎉":"party popper",
"📰":"newspaper","🤱":"breast feeding","👀":"astonished","🧤":"gloves","🚶🏿":"social distance","🚥":"traffic light",
"↗":"up right","⚕":"medicine and health care services","↔️":"keep distance","🦠":"virus","💧":"droplet",
"💦":"droplet","🧴":"sanitizer","🧼":"hand wash","🚫":"not","🚰":"hand wash","🤝":"not handshake",
"🧑‍🤝‍🧑":"not close contact","🧫":"test for covid-19","🧪":"test for covid-19","🌡️":"body temperature measured",
"🔬":"test for covid-19","📝":"contact tracing","🧑‍⚕️":"health worker","🐦":"hand wash","🛡":"shielding required",
"♿️":"accessibility requirements","🆘":"emergency help required"

Training and Test files prepare

• train_data.csv is preprocessed file from actual Constraint_English_Train.csv
• test_data.csv is preprocessed file from actual Constraint_English_Val.csv

Wordcloud for train and test data

Training Procedure

• Run python train.py file from src directory
• See the all train and validation history from logs/info.log
• See train and validation history visualization from train_loss_his/acc_curves.png and train_loss_his/loss_curves.png

Contribution

• We have done regular preprocessing to replace emojis with tokens , replace all of the mentions with Twitter handles, split all of the hashtags into separate words. [preprocessing]
• Extracted news headlines from URL which was used in user posts and concatenate that with user post if post and headline not syntactically similar.
• Used transformers-based variant of Bert and Roberta models alongside different architecture on top of Bert and Roberta models.
• Utilized Ensemble-based approach based on top three model.

To run all of the bert-base-uncased model

• Enter bert-base-uncased folder[I have shared a bert base pretrain model folder use that as a pretrain path]
• chmod +x run.sh for permission
• ./run.sh put that command from terminal to run all of the bert base uncased model

To run all of the bert-large-uncased model

• Enter bert-large-uncased folder[I have shared a bert large pretrain model folder use that as a pretrain path]
• chmod +x run.sh for permission
• ./run.sh put that command from terminal to run all of the bert large uncased model

To run all of the covid bert model

• Enter covid-twitter-bert folder
• chmod +x run.py for permission
• ./run.sh put that command from terminal to run all of the covid model

To run all of the roberta large model

• Enter roberta base folder
• chmod +x run.sh for permission
• ./run.sh put that command from terminal to run all of the roberta large model

our best model achieved the weighted F1-score of 0.978 on the test
set (26th place in the leaderboard) out of 166 submitted teams in total.