Term weighting is an important step which has direct impact on the result in classical text classification. However, the behavior of the term weighting method may vary depending on different preprocessing techniques in sentiment analysis which considered as a text classification task. In this study, term weighted methods which are newly proposed for various research areas such as information retrieval, text classification and document filtering, performed to investigate effect on results for Twitter sentiment analysis. In feature extraction phase, two different models are used including Bag of Words (BoW) and character level N-gram. The experiments conducted on data sets consist of Turkish and English Twitter feeds. Sentiment classification of Twitter feeds performed using topic model generated with Latent Dirichlet Allocation (LDA) method. The Support Vector Machine (SVM) algorithm is employed in the classification stage. According to the experimental results, the most effective term weighting method that can be used in Twitter sentiment analysis studies is suggested.