Lexical data augmentation for sentiment analysis

Machine learning methods, especially deep models, have achieved impressive performance in various natural language processing tasks including sentiment analysis. However, models are more demanding for training data. Data augmentation techniques widely used to generate new instances based on modifications existing data or relying external knowledge bases address annotated scarcity, which hinders the full potential of machine techniques. This paper presents our work using part-of-speech (POS) focused lexical substitution (PLSDA) enhance algorithms We exploit POS information identify words be replaced and investigate different strategies find semantically related substitutions when generating instances. The choice tags as well a variety such semantic-based methods sampling discussed detail. Performance evaluation focuses comparison between PLSDA two previous substitution-based one is thesaurus-based, other lexicon manipulation based. Our approach tested five English analysis benchmarks: SST-2, MR, IMDB, Twitter, AirRecord. Hyperparameters candidate similarity threshold number newly generated optimized. Results show that six classifiers (SVM, LSTM, BiLSTM-AT, bidirectional encoder representations from transformers [BERT], XLNet, RoBERTa) trained with achieve accuracy improvement than 0.6% comparing averaged benchmarks. Introducing constraint well-designed can improve reliability methods. Consequently, significantly improves algorithms.