<![CDATA[Text mining is an important approach in analyzing text data, particularly for detecting negative sentiments such as cyberbullying on social media. Twitter (X), as an open platform, often serves as a space for the proliferation of hate speech and abusive behavior recorded in text form. This study aims to improve the performance of sentiment classification models on Twitter (X) data by combining the Improved Word Vector (IWV) and Bidirectional Encoder Representations from Transformers (BERT) methods, evaluated using precision, recall, and F1-score metrics. The dataset used consists of 9,874 Indonesian-language tweets labeled into three categories: Hate Speech (HS), Abusive, and Neutral. This data is sourced from previous research and is the result of re-annotation of the original dataset of 13,169 tweets. IWV is formed from a combination of Word2Vec, GloVe, POS tagging, and emotion lexicon features designed to enrich word representation semantically. The preprocessing process is carried out through several important stages, namely tokenization, filtering, stemming/lemmatization, and normalization. The IWV extraction results were then combined with BERT embedding through concatenation to produce high-dimensional vector representations. Evaluation was performed using precision, recall, and F1-score metrics. The test results showed that the combined IWV+BERT model was able to produce better performance than BERT alone. The use of data that has been balanced through balancing techniques also contributed to the improvement in accuracy, with the highest accuracy value reaching 91%. This finding indicates that the integration of word representation features from IWV and sentence context from BERT can improve the effectiveness of text mining in sentiment analysis related to cyberbullying on social media]]>
