Table 1 Comparative summary of recent SCI/SCIE works on complaint analytics, multimodal reasoning, sentiment modelling, and toxicity detection (2022–2025).
Study (year) | Method/model | Modality | Novelty | Dataset/domain | Key limitations |
|---|---|---|---|---|---|
Vairetti et al. (2023)7 | Deep learning and MCDM prioritisation | Text | Hybrid ranking using operational KPIs | Service complaints (industry) | No speech support; no toxicity modelling; no zero-shot routing |
Schupp et al. (2025)1 | Topic mining and anomaly detection | Text | Proactive detection of recurring issues | Government helpdesk logs | Backend analytics only; no per-complaint decision logic |
Kim & Park (2024)8 | Materiality-based classifier | Text | Material vs. immaterial complaint modelling | Large-scale review corpus | No multimodal integration; no behavioural analysis |
Liang & Wang (2024)9 | Hierarchical complaint classifier | Text | Label-aware multi-branch deep network | Clinical chief-complaint data | Single-task; no sentiment or toxicity inference |
Marques et al. (2023)10 | Federated NLP classifier | Text | Privacy-preserving distributed training | Banking complaint logs | No multimodal reasoning; no urgency estimation |
Singh et al. (2022)11 | Multimodal complaint detector | Text and Image | Emotion- and sentiment-aware fusion | CESAMARD dataset | Not civic-oriented; no speech handling or routing |
Singh et al. (2023)12 | Bi-transformer multimodal ABSA | Text and Image | Aspect-level complaint and cause detection | Public review datasets | No urgency modelling; no toxicity analysis |
Yin et al. (2025)13 | PWCR speech-based complaint detector | Speech | Paralinguistic and temporal modelling | Call-centre audio | No routing; no urgency scoring; no abuse logic |
Devanathan et al. (2023)14 | Federated multimodal meta-learning | Text and Image | Cross-client multimodal generalisation | Distributed platforms | Complaint detection only; no toxicity or urgency components |
Koto et al. (2024)18 | Zero-shot sentiment via multilingual lexicon | Text | Zero-shot affect modelling across 34 languages | Multilingual corpora | No routing integration; no abuse detection |
Bansal et al. (2024)23 | Domain-adapted abuse classifier | Text | Regularised transformer for robustness | Six toxicity datasets | No behavioural history; no multimodal reasoning |
Lee et al. (2025)24 | Fairness-aware abuse detection | Text | Adversarial bias mitigation framework | Benchmark abusive-language corpora | No urgency component; not civic-specific |
Nguyen et al. (2025)25 | Euphemistic toxicity detector | Text | Contrastive euphemism modelling | Social media datasets | Standalone moderation; no routing or escalation integration |
