Due to the scarcity of linguistic resources and annotated datasets, sentiment analysis in low-resource languages like Gujarati poses particular difficulties. The effectiveness of lexicon-based techniques in conjunction with conventional machine learning models for sentiment categorization of Gujarati news headlines is investigated in this work. This work drastically reduces the amount of manual labor needed to categorize 21,616 news headlines by introducing a lexicon-based method for automating sentiment annotation. The approach successfully captures subtle manifestations of sentiment by utilizing an enhanced lexicon that includes both synonyms and antonyms of sentiment-bearing terms. By using two custom dictionaries, Synonyms-dict and Opposites-dict, the suggested annotation method assigns a sentiment score to each word in the headlines, allowing sentiments to be identified even when there are no direct matches. Features like Count Vectorization, TF-IDF, and N-Grams were used to train machine learning models including Random Forest, Support Vector Machines, and Logistic Regression using this annotated dataset. With a high accuracy of 72.75%, the results demonstrate that Random Forest in conjunction with N-Grams performed better than other combinations. This work underscores the importance of synonym-antonym dictionaries in enhancing annotation accuracy and shows the promise of automated lexicon-based annotation for extensive sentiment analysis in low-resource languages. To improve performance even further, future research will include transformer-based techniques and neural embeddings.