Table 1 Critical analysis of the reviewed models.
References | Models | Application | Advantages | Disadvantages |
|---|---|---|---|---|
Uthirapathy and Sandanam5 | LDA for topic modelling, BERT for sentiment classification | Sentiment classification in text datasets | BERT effectively captures contextual information; generalizes well across NLP tasks | Computationally expensive; requires large datasets for training |
Sharaf Al-deen et al.6 | DNN-MHAT (DNN with MHA, BiLSTM-CNN) | Short tweets and long reviews analysis | Multi-head attention improves long-term dependency detection | Increased model complexity; high computational cost |
Basiri et al.7 | ABCDM (Attention-based CNN-RNN with BiGRU and BiLSTM) | Sentiment analysis using GloVe embedding | Captures both local and contextual features effectively | Requires fine-tuning of multiple components |
Yasin and Arslankaya8 | ML-based (K-NN, NB, SVM) | Turkish tweets on climate change | Simple and interpretable ML models | Lower accuracy (max 74.63%); lacks deep learning-based improvements |
Yin et al.4 | NLP methodologies for COVID-19 vaccine sentiment extraction | Social media-based sentiment analysis | Helps in policy-making with sentiment trends | Limited to COVID-19 data; may not generalize to other topics |
El Barachi et al.10 | BiLSTM-based sentiment classification | Climate change and global warming sentiment analysis | High accuracy (87.01%—89.80%) | Requires large labelled datasets |
Rodrigues et al.11 | ML (NB) and DL (LSTM) for SA and spam detection | Twitter spam and sentiment classification | LSTM outperforms ML models in accuracy | Requires more resources for training |
Sivakumar and Srinivasulu12 | LSTM and fuzzy logic | Mobile phone reviews | Enhances SA with fuzzy logic | Increased complexity; requires fine-tuned aspect identification |
Naramula and Kalaivania13 | ML models (SVM, RF, K-NN) for sentiment extraction | Mobile phone reviews from Twitter | AI techniques improve accuracy | Limited scope to phone reviews |
Gandhi et al.26 | CNN and LSTM for SA | Movie reviews | High accuracy (87.74–88.02%) | It can be improved with advanced architectures |
Huang et al.15 | BERT-based linguistic model | Turkish verbal data preprocessing for ML | Streamlines text preprocessing | High computational requirements |
Jain et al.16 | BERT + DCNN for SA | Sentiment classification | DCNN improves classification performance | Model complexity increases training time |
Özçift et al.17 | BERT + CNN for sentiment classification | Large-scale movie reviews | Combines contextual and local features | Computationally intensive |
Pang et al.18 | Aspect-based SA with BERT | Extracting aspect-level information | Effective aspect-based classification | Requires labelled aspect-level data |
Manju and Gupta19 | BERT-based guided LDA | Restaurant reviews aspect-based SA | Improves aspect-based topic modelling | Can be resource-intensive |
Yousaf et al.20 | ML (Voting LR-SGD) with TF-IDF | Emotion recognition in Twitter data | Achieved 79% accuracy, 81% F1-score | Performance may degrade for multi-class classification |
Bansal et al.21 | MTL within DL framework; clustering-based summarization integrated with generative modelling | Sentiment classification and informative tweet selection during disaster-related events | Joint sentiment and informativeness detection. Improved clustering via generative modelling. Outperformed SOTA methods in disaster domains | Framework complexity increases due to MTL. The model does not generalize well beyond disaster events. Lacks real-time classification insights |
Dangi et al.22 | SATD model using various ML classifiers on lockdown vs. post-lockdown datasets | COVID-19 sentiment analysis on Twitter | Applied to real COVID-era Twitter data. Multi-model comparison offers performance robustness. High accuracy on recall and F1-score metrics | Limited to pandemic-era sentiment context. Lack of deep learning utilization for nuanced feature learning. Possible redundancy among models |
Dangi et al.,23 | COA-based Temporal Weight AdaBoost SVM | Drift-resistant tweet classification using social media streams | Tackles concept drift effectively. Handles class imbalance with chaotic optimization. Enhances SVM classification through boosting and weighting | Computationally intensive due to hybrid optimization. Optimization convergences vary. Designed mainly for short-term tweet streams |
Dangi et al.,24 | Comparative ML analysis: LSVM, LR, RF, SGD on COVID-related tweet classification | COVID-19 vs non-COVID-19 emotion classification | Identified strengths of different models: SGD: best overall performance. LR: best in ROC_AUC, recall. RF: best in precision and specificity. LSVM: best accuracy | No ensemble or hybrid DL methods used. Performance is dependent on balanced datasets. Does not account for temporal dynamics or linguistic shifts |
Dangi et al.34 | Deep learning with NLP | Sentiment analysis | -High performance in analyzing implicit sentiments -Handles high-dimensional data -Learns complex patterns | -High computational cost -Large space complexity -Sensitive to hyperparameters -Long training time |
