Governments and social media providers put high effort to tackle massive negative contents in social media. Those contents are mostly containing religion, race, and inter-group issues, cyberbullying, and also body shamming, which usually appears together with offensive languages. It becomes difficult to overcome because of a large number of internet users in Indonesia. Hence, we need a system that can automatically detect the negative contents. This paper utilizes Neural Network (NN) models for not only classifying the words as (non)offensive words but also considering the structure of the sentence to get its context. There are two NN models analyzed in this paper: Artificial Neural Network (ANN) and Recurrent Neural Network (RNN). The computer simulation results show that the RNN has better performances than the ANN with the accuracy of training, validation, and testing 94%, 84%, and 84%, respectively. Pemerintah dan penyedia layanan media sosial di Indonesia berusaha keras untuk mengatasi maraknya konten negatif di media sosial. Konten negatif yang sering ditemui diantaranya isu suku, agama, ras, dan antargolongan (SARA), cyberbullying, serta body shamming, yang biasanya muncul disertai kalimat-kalimat umpatan. Hal tersebut menjadi sulit untuk diatasi karena jumlah pengguna internet di Indonesia yang sangat besar, sehingga perlu adanya sebuah sistem yang dapat mendeteksinya secara otomatis. Penelitian ini mengusulkan sistem dengan model Neural Network untuk deteksi konten negatif di media sosial dengan cara mempertimbangkan konteks kalimat atau frasa, tidak hanya kata-per-kata. Ada dua model NN yang dianalisis di penelitian ini, yaitu Artificial Neural Network (ANN) dan Recurrent Neural Network (RNN). Model RNN menunjukkan performa yang lebih baik dibandingkan dengan model ANN dengan akurasi training, validasi, dan test masing-masing adalah 94%, 84%, dan 84%.

No abstract