The proliferation of digital public service platforms and the expansion of e-government initiatives have significantly increased the volume and diversity of citizen-generated feedback. This trend emphasizes the need for classification systems that are not only tailored to specific administrative domains but also robust to the linguistic, contextual, and structural variability inherent in user-submitted content. This study investigates the comparative effectiveness of large language models (LLMs) alongside instruction-tuned embedding models in the task of categorizing public transportation complaints. LLMs were tested using a few-shot inference, where classification is guided by a small set of in-context examples. Embedding models were assessed under three paradigms: label-only zero-shot classification, instruction-based classification, and supervised fine-tuning. Results indicate that fine-tuned embeddings can achieve or exceed the accuracy of LLMs, reaching up to 90 percent, while offering significant reductions in inference latency and computational overhead. E5 embeddings showed consistent generalization across unseen categories and input shifts, whereas BGE-M3 demonstrated measurable gains when adapted to task-specific distributions. Instruction-based classification produced lower accuracy for both models, highlighting the limitations of prompt conditioning in isolation. These findings position multilingual embedding models as a viable alternative to LLMs for classification at scale in data-intensive public sector environments.

Public transport agencies must triage thousands of multilingual complaints every day, yet the cost of training and serving fine-grained sentiment analysis models limits real-time deployment. The proposed “one encoder, any facet” framework therefore offers a reproducible, resource-efficient alternative to heavy fine-tuning for domain-specific sentiment analysis or opinion mining tasks on digital service data. To the best of our knowledge, we are the first to test this paradigm on operational multilingual complaints, where public transport agencies must prioritize thousands of Russian- and Kazakh-language messages each day. A human-labelled corpus of 2400 complaints is embedded with five open-source universal models. Obtained embeddings are matched to semantic “anchor” queries that describe three distinct facets: service aspect (eight classes), implicit frustration, and explicit customer request. In the strict zero-shot setting, the best encoder reaches 77% accuracy for aspect detection, 74% for frustration, and 80% for request; taken together, these signals reproduce human four-level priority in 60% of cases. Attaching a single-layer logistic probe on top of the frozen embeddings boosts performance to 89% for aspect, 83–87% for the binary facets, and 72% for end-to-end triage. Compared with recent fine-tuned sentiment analysis systems, our pipeline cuts memory demands by two orders of magnitude and eliminates task-specific training yet narrows the accuracy gap to under five percentage points. These findings indicate that a single frozen encoder, guided by handcrafted anchors and an ultra-light head, can deliver near-human triage quality across multiple pragmatic dimensions, opening the door to low-cost, language-agnostic monitoring of digital-service feedback.

Accurate bus travel time prediction is a critical requirement for managing urban mobility and enabling reliable public transit operations. While existing research often relies on the fusion of multi-source, heterogeneous data, such approaches are less feasible in cities with limited access to external information. To address this gap, this study proposes a dataefficient approach to travel time prediction using only commonly available historical transit data. Specifically, a Hybrid Bidirectional Long Short-Term Memory (BiLSTM) model is applied to predict bus travel times without requiring external contextual inputs. The model is evaluated on real-world bus routes in Kandy and Astana, where it consistently outperforms attention-based ConvLSTM and stacked ensemble baselines. On Route 1 in Kandy, the proposed model achieves a Mean Absolute Error (MAE) of 5.44 seconds and a Mean Absolute Percentage Error (MAPE) of $\mathbf{2. 4 1 \%}$, markedly better than competing methods. Across three routes in Astana, the model reports MAE values between 42.16 and 50.45 seconds and MAPE between 34.23 % and 39.84 %, improving upon baselines by up to 20 %. Cross-validation further confirms the robustness of the approach, with consistently lower errors across all folds. The results show that accurate bus travel time prediction is achievable, even in a variety of data-limited environments, offering a practical and versatile solution for transport agencies.

This study investigates the development and performance evaluation of transformer-based models for the automatic classification of public transportation passenger reviews, aiming to enhance feedback processing while optimizing issue resolution. Efficient handling of passenger feedback is crucial for improving public transportation services, as unresolved complaints or operational inefficiencies can decrease passenger satisfaction and create logistical challenges. Traditional text classification approaches, such as keyword-based methods or classical Machine Learning (ML) algorithms, struggle with multilingual and heterogeneous textual data, particularly in low-resource languages. This study addresses this gap by systematically comparing transformer-based architectures for review classification in Russian and Kazakh, demonstrating their effectiveness in real-world applications. A key contribution of this research lies in evaluating both language-specific and multilingual transformers on passenger-generated feedback, offering insights into their generalization capabilities. Unlike previous studies, which predominantly focus on English-language datasets, this work introduces a newly created, manually labeled dataset covering diverse real-world scenarios in Russian and Kazakh, enabling an objective comparative analysis. Three transformer models DeepPavlov/rubert-base-cased, XLM-RoBERTa-base, and XLM-RoBERTa-large were trained and tested to assess their ability to process complex multilingual input. Experimental results indicate that XLM-RoBERTa-large achieves the highest classification accuracy (90%), particularly for code-mixed and multilingual reviews, whereas rubert-base-cased performs consistently well for Russian-language feedback (87.667%), reinforcing its suitability for monolingual classification tasks. XLM-RoBERTa-base exhibits a balanced trade-off between accuracy and robustness, making it a viable option for heterogeneous review processing (89.5%). Despite their effectiveness, transformer-based models still encounter challenges related to data balancing and the handling of underrepresented classes, particularly in scenarios with uneven language distributions or domain-specific terminology. These findings confirm that transformer models significantly enhance the automation of passenger feedback classification, providing a scalable solution for public transportation providers.