Abstract
Transformer-based language models are widely adopted for sentiment analysis due to their strong contextual representation capabilities; however, their high computational complexity can limit practical deployment in resource-constrained environments. This work presents a systematic comparative study of static and contextual word representations for sentiment classification across architecturally distinct model families under standardized preprocessing and evaluation conditions. We evaluate Long Short-Term Memory (LSTM) networks with Word2Vec, GloVe, and FastText embeddings, and compare them with fine-tuned BERT-base and DistilBERT transformer baselines using the same preprocessing, training, and evaluation protocols across five publicly available datasets spanning formal reviews and noisy social media text. In addition to overall accuracy, we report macro- and weighted-F1 scores across three random seeds to estimate experimental variance, and apply McNemar’s test (with Bonferroni correction for multiple comparisons) to assess pairwise model differences on each dataset’s test set. Contingency tables required for each McNemar comparison are provided in the supplementary material. Detailed error analyses using confusion matrices and out-of-vocabulary (OOV) statistics examine the impact of lexical coverage on model performance. Experimental results indicate that FastText consistently outperforms other static embeddings, particularly on datasets with high lexical variability, while offering substantially lower computational overhead compared to the transformer baselines. The primary contribution of this study is a diagnostic multi-dataset evaluation framework that quantifies the relationship between lexical coverage and subword modeling benefit, and characterises the efficiency–accuracy trade-off across embedding paradigms. All architectural and training differences between model families are explicitly documented and their confounding effects discussed rather than assumed away. These findings provide practical guidance for selecting embedding strategies by jointly considering accuracy and deployment efficiency.

This publication has 26 references indexed in Scilit: