Resolving Morphological Sentiment Ambiguity in Kazakh Social Media Text

Sentiment analysis in low-resource languages suffers from scarce labeled data, and sometimes agglutinative languages add a second difficulty. Suffixes can reverse the polarity of a root, so one stem yields word forms of opposite sentiment, and each form appears only rarely in a small corpus. This study examines this morphological sentiment ambiguity in Kazakh and tests four hypotheses about the role of explicit lexical knowledge. The experimental instrument was SezimEmoLex, a Kazakh lexicon that adapts NRC EmoLex seeds to Kazakh morphology and labels word forms with positive, negative, or neutral polarity and eight emotion categories. Its features were added to a support vector machine (SVM), a long short-term memory network (LSTM), and a Kazakh BERT model (KazBERT), all evaluated on 5,000 annotated social media comments. The lexical features raised the accuracy of all three models and narrowed the gap between the SVM and KazBERT from 13.9 to 9.9 points. Most residual errors involved the neutral class, where negation, contrast, sarcasm, and language mixing override word-level evidence. Wordform-level lexical knowledge thus substitutes for part of what pretraining provides and makes lightweight models practical for agglutinative low-resource languages.