SemEval2020

OffensEval - Task 12

System developed for the SemEval-2020 competition on MultilingualOffensive Language Identification in Social Media where we achieved 2nd position in Turkish sub-task and 6th position in Greek Subtask. The goal of task was the identifying the Offensive Language in Social Media.

Macro Averaged F1-Scores and Accuracies on validation and test sets

Model	Language	Embedding	UnderSampling	Val F1	Val Acc	Test F1	Test Acc
CNN	Tur	Twitter W2V	Without	0.739	0.857	0.747	0.861
CNN	Tur	Twitter W2V	With	0.733	0.847	0.738	0.857
BiLSTM	Tur	Twitter W2V	Without	0.733	0.867	0.747	0.864
BiLSTM	Tur	Twitter W2V	With	0.755	0.857	0.748	0.865
CNN-LSTM	Tur	Twitter W2V	Without	0.742	0.855	0.766	0.863
CNN-LSTM	Tur	Twitter W2V	With	0.751	0.867	0.773	0.865
CNN-LSTM	Tur	Public FastText	Without	0.722	0.862	0.710	0.856
CNN-LSTM	Tur	Public FastText	With	0.720	0.851	0.726	0.852
CNN-LSTM	Tur	Public W2V	Without	0.711	0.856	0.717	0.856
CNN-LSTM	Tur	Public W2V	With	0.731	0.85	0.739	0.857
CNN-LSTM	Tur	Twitter FastText	Without	0.743	0.851	0.756	0.853
CNN-LSTM	Tur	Twitter FastText	With	0.767	0.874	0.753	0.864
BiLSTM-Att	Tur	Public W2V	Without	0.679	0.847	0.681	0.844
BiLSTM-Att	Tur	Public W2V	With	0.707	0.827	0.721	0.844
BiLSTM-Att	Tur	Public FastText	Without	0.684	0.849	0.698	0.850
BiLSTM-Att	Tur	Public FastText	With	0.698	0.839	0.726	0.846
BiLSTM-Att	Tur	Twitter FastText	Without	0.735	0.852	0.721	0.867
BiLSTM-Att	Tur	Twitter FastText	With	0.738	0.866	0.747	0.861
BiLSTM-Att	Tur	Twitter W2V	Without	0.763	0.859	0.781	0.871
BiLSTM-Att	Tur	Twitter W2V	With	0.748	0.870	0.760	0.868
BERTurk	Tur	-	Without	0.814	0.888	0.806	0.877
BERTurk	Tur	-	With	0.789	0.866	0.808	0.873
GreekBERT	Greek	-	Without			0.832
Ensemble	Tur	-	Without	0.822	0.896	0.813	0.887
Ensemble	Tur	-	With	0.809	0.881	0.816	0.883

Submitted models are written in bold letters. All models are written in Google Collaboratory environment to take advantage of free GPU.