Improving Accuracy in Predicting Comment Toxicity in Threads Applications: a Comparison of Bidirectional Encoder Representations From Transformers and Gated Recurrent Units

The present study will attempt the use of machine learning in the evaluation of toxicity within threaded comments. We have availed ourselves of two very strong algorithms, BERT and GRU; each is known to perform well in solving such complexity in regard to harmful comments. We have divided the results into two groups of ten outcome values each for the prediction of the level of toxicity. For an extended analysis, we ran 50 iterations for each group, summing up to a total of 100 iterations. In the analysis done through SPSS, we kept the 95% Confidence Interval CI, alpha level at <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$\mathbf{0. 0 5}$</tex>, and statistical power G*power at <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$\mathbf{8 0 \%}$</tex>. This is to ensure that all our predictions over the levels of toxicity within threaded comments will be as accurate and reliable as possible. Algorithms, which were used for content moderation, showed very impressive performance in predicting the levels of toxicity, standing at <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$\mathbf{9 3. 0 3 \%}$</tex> accuracy. While comparing BERT with GRU, the identification comes that BERT has outperformed GRU with an accuracy rate of 93.03% whereas GRU has shown an accuracy rate of 84.66 %, which clearly indicates that BERT is the superior classifier to improve content moderation in online discussions.