Text annotation is an essential element of the natural language processing approaches. The manual annotation process performed by humans has various drawbacks, such as subjectivity, slowness, fatigue, and possibly carelessness. In addition, annotators may annotate ambiguous data. Therefore, we have developed the concept of automated annotation to get the best annotations using several machine-learning approaches. The proposed approach is based on an ensemble algorithm of meta-learners and meta-vectorizer techniques. The approach employs a semi-supervised learning technique for automated annotation to detect hate speech. This involves leveraging various machine learning algorithms, including Support Vector Machine (SVM), Decision Tree (DT), K-Nearest Neighbors (KNN), and Naive Bayes (NB), in conjunction with Word2Vec and TF-IDF text extraction methods. The annotation process is performed using 13,169 Indonesian YouTube comments data. The proposed model used a Stemming approach using data from Sastrawi and new data of 2245 words. Semi-supervised learning uses 5%, 10%, and 20% of labeled data compared to performing labeling based on 80% of the datasets. In semi-supervised learning, the model learns from the labeled data, which provides explicit information, and the unlabeled data, which offers implicit insights. This hybrid approach enables the model to generalize and make informed predictions even when limited labeled data is available (based on self-learning). Ultimately, this enhances its ability to handle real-world scenarios with scarce annotated information. In addition, the proposed method uses a variety of thresholds for matching words labeled with hate speech ranging from 0.6, 0.7, 0.8, to 0.9. The experiments indicated that the DT-TF-IDF model has the best accuracy value of 97.1% with a scenario of 5%:80%:0.9. However, several other methods have accuracy above 90%, such as SVM (TF-IDF and Word2Vec) and KNN (Word2Vec), based on both text extraction methods in several test scenarios.

Information and news about Covid-19 received various responses from social media users, including Twitter users. Changes in netizen opinion from time to time are interesting to analyze, especially about the patterns of public sentiment and emotions contained in these opinions. Sentiment and emotional conditions can illustrate the public's response to the Covid-19 pandemic in Indonesia. This research has two objectives, first to reveal the types of public emotions that emerged during the Covid-19 pandemic in Indonesia. Second, reveal the topics or words that appear most frequently in each emotion class. There are seven types of emotions to be detected, namely anger, fear, disgust, sadness, surprise, joy, and trust. The dataset used is Indonesian-language tweets, which were downloaded from April to August 2020. The method used for the extraction of emotional features is the lexicon-based method using the EmoLex dictionary. The result obtained is a monthly graph of public emotional conditions related to the Covid-19 pandemic in the dataset.

A cross-domain sentiment analysis (CDSA) study in the Indonesian language and tree-based ensemble machine learning is quite interesting. CDSA is useful to support the labeling process of cross-domain sentiment and reduce any dependence on the experts; however, the mechanism in the opinion unstructured by stop word, language expressions, and Indonesian slang words is unidentified yet. This study aimed to obtain the best model of CDSA for the opinion in Indonesia language that commonly is full of stop words and slang words in the Indonesian dialect. This study was purposely to observe the benefits of the stop words cleaning and slang words conversion in CDSA in the Indonesian language form. It was also to find out which machine learning method is suitable for this model. This study started by crawling five datasets of the comments on YouTube from 5 different domains. The dataset was copied into two groups: the dataset group without any process of stop word cleaning and slang word conversion and the dataset group to stop word cleaning and slang word conversion. CDSA model was built for each dataset group and then tested using two types of tree-based ensemble machine learning, i.e., Random Forest (RF) and Extra Tree (ET) classifier, and tested using three types of non-ensemble machine learning, including Naïve Bayes (NB), SVM, and Decision Tree (DT) as the comparison. Then, It can be suggested that the accuracy of CDSA in Indonesia Language increased if it still removed the stop words and converted the slang words. The best classifier model was built using tree-based ensemble machine learning, particularly ET, as in this study, the ET model could achieve the highest accuracy by 91.19%. This model is expected to be the CDSA technique alternative in the Indonesian language.

Supervised sentiment analysis ideally uses a fully labeled data set for modeling. However, this ideal condition requires a struggle in the label annotation process. Semi-supervised learning (SSL) has emerged as a promising method to avoid time-consuming and expensive data labeling without reducing model performance. However, the research on SSL is still limited and its performance needs to be improved. Thus, this study aims to create a new SSL-Model for sentiment analysis. The Ensemble Classifier SSL model for sentiment classification is introduced. The research went through pre-processing, vectorization, and feature extraction using TF-IDF and n-grams. Support Vector Machine (SVM) or Random Forest for tokenization was used to separate unigram, bigram, and trigram in model generation. Then, the outputs of these models were combined using stacking ensemble approach. Accuracy and F1-score were used for the evaluation. IMDB datasets and US Airlines were used to test the new SSL models. The conclusion is that the sentiment annotation accuracy is highly dependent on the suitability of the dataset with the machine learning algorithm. In IMDB dataset, which consists of two classes, it is better to use SVM. In the US Airlines consisting of three classes, SVM is better at improving the model performance against the baseline, but RF is better at achieving the baseline performance even though it fails to maintain the model performance.