Sentiment Analysis of TikTok Comments on Law Enforcement Misconduct Cases Using BiLSTM with RoBERTa Labeling
Abstract
This study examines the sentiment distribution of TikTok comments on four law enforcement misconduct cases, explores how the comments relate to officer accountability and expectations of government responsiveness, and evaluates sentiment classification with a bidirectional long short-term memory (BiLSTM) network. The study uses a quantitative approach with comments from four purposively selected TikTok posts. Web scraping collects 3,685 comments, which decrease to 3,446 after the removal of empty and duplicate texts. RoBERTa assigns negative, neutral, and positive labels, and the normalized texts with these labels train the BiLSTM. A stratified split reserves 20% of the data for testing, and random oversampling balances the classes in the training data. The labeling results show negative sentiment at 72.37%, followed by positive at 16.37% and neutral at 11.26%. On 690 test comments, the BiLSTM obtains an accuracy of 71.01% and a macro-F1 of 0.580. The accuracy is lower than the 72.32% of a majority-class baseline, but the macro-F1 is higher than the baseline of 0.280. The findings suggest that the comments express criticism, demands for sanctions, and sympathy toward victims, so positive sentiment does not always signal support for officers. Recognition of the neutral and positive classes remains limited, and the automatic labels lack human validation.
Keywords
Share and Cite
Article Metrics

This work is licensed under a Creative Commons Attribution-NonCommercial 4.0 International License.

