LLM-Driven Semantic Understanding and Automatic Categorization of Public Issue Reports in Smart Cities

LLM-Driven Semantic Understanding and Automatic Categorization of Public Issue Reports in Smart Cities

Hitesh Bhor, Bhushan Kumavat, Tilak Jain, Vedant Shirsath, Prof. P. G. Fegade, Prof. P. P. Patil, Prof. A. L. Rane

Master of Computer Applications

K.K. Wagh Institute of Engineering Education and Research Nashik, India

AbstractThe rapid digitalization of public services has led to a sharp increase in citizen-submitted issue reports, which are typically unstructured, multilingual, and multimodal in nature. Manual categorization of such reports is neither scalable nor consistent, resulting in delayed grievance resolution and poor resource utilization. This paper presents Jagruk, an LLM-driven pipeline for the semantic understanding, automatic classica- tion, and priority-based routing of civic complaint reports in smart city environments. The proposed system accepts unstruc- tured text along with optional image and geolocation inputs, and processes them through a multimodal natural language processing workow using instruction-tuned Large Language Models (LLMs). Each complaint is classied into a predened municipal category, assigned an urgency priority, and routed to the appropriate department, all within a single inference step. The system is evaluated against three baseline modelsTF-IDF with SVM, Multinomial Naive Bayes, and ne-tuned BERT on a labeled civic complaint dataset. The proposed LLM-based system achieves an overall accuracy of 94.2%, a macro F1- score of 92.9%, and a Mean Absolute Error (MAE) of 0.54 for priority prediction, outperforming all baselines. Deployment considerations including inference latency, data privacy, and fairness are also discussed, demonstrating the suitability of the system for real-world e-governance applications.

Index TermsLarge Language Models, semantic analysis, smart city, e-governance, complaint classication, natural lan- guage processing, grievance management, multimodal processing

1. Introduction

   The rapid digitalization of public services has resulted in a substantial increase in the volume of citizen-submitted issue reports that require timely processing and categorization. Man- ual methods of complaint management are not scalable and are prone to delays, inconsistencies, and inefcient resource allocation [1].

   In this paper, we propose Jagruk, an LLM-driven framework for the semantic understanding and automatic categorization of public issue reports in smart cities. The system harnesses Large Language Models to parse unstructured and multimodal inputs, and accurately classies, prioritizes, and routes complaints to the relevant municipal departments [2][4].

   Unlike conventional keyword-based or rule-based systems, the proposed framework captures contextual meaning, thereby accommodating informal language, code-mixed text, and mul- tilingual citizen reports [5]. This signicantly enhances the efciency, transparency, and responsiveness of e-governance platforms.

   1. Problem Statement

      Citizen complaints predominantly consist of unstructured, noisy text containing informal vocabulary and code-mixed language, with optional multimedia inputs such as images and location data. Manual processing of such heterogeneous inputs is error-prone, inconsistent, and fails to scale with growing complaint volumes. The primary challenge addressed in this work is the design of an automated system capable of accurately understanding, classifying, and prioritizing civic complaints while preserving fairness and protecting user pri- vacy.

   2. Key Contributions

      The principal contributions of this work are as follows:

      • A multimodal LLM-based pipeline that jointly processes text descriptions, images, and geolocation data for com- plaint understanding.

      • A single-inference classication and routing mechanism that assigns complaint category, urgency priority, and responsible department simultaneously.

      • A priority detection mechanism driven by sentiment and urgency cues, coupled with a community voting system to surface widely reported issues.

      • A real-time, end-to-end civic complaint management plat- form (Jagruk) deployed with a full-stack web architecture and integrated with municipal workow services.

      • Empirical evaluation demonstrating signicant perfor- mance improvements over TF-IDF+SVM, Naive Bayes, and ne-tuned BERT baselines.

   3. Importance in Smart Cities and E-Governance

      Automated complaint classication reduces response times, increases operational efciency, and supports data-driven decision-making in municipal governance. By enabling scal- able management of high-volume complaint streams, the sys- tem promotes transparency between government and citizens and improves the quality of public service delivery.

   4. Challenges in Manual Categorization

   Existing manual systems are slow, inconsistent, and inca- pable of handling the linguistic diversity of citizen-generated content. They do not scale with growing complaint databases, and their inability to interpret contextual or multimodal data results in frequent misrouting and diminished service efcacy.

2. Literature Review

   1. Existing Complaint Management Systems

      Classical complaint management systems in e-governance rely on manual sorting or rule-based routing. These systems enforce rigid workows that lack exibility and become ineffective as complaint volume and diversity increase [6], [7]. Earlier systems were unable to automatically respond to unstructured complaints, leading to persistent delays and unresolved grievances. More recent attempts at automation still struggle with contextual understanding, particularly in multilingual settings [5]. Unlike these systems, the proposed approach requires no predened rules and handles informal, code-mixed text through LLM-based semantic inference.

   2. NLP-Based Approaches

      Natural language processing has been widely applied to complaint classication. Early approaches employed TF-IDF with N-gram features and machine learning classiers such as Naive Bayes and SVM. These methods require extensive manual feature engineering and perform poorly on informal, noisy, or semantically ambiguous text. The introduction of deep learning models, particularly LSTMs with attention mechanisms, improved classication accuracy. Transformer- based models such as BERT have further advanced semantic classication capabilities [1], [3], establishing new perfor- mance benchmarks across multiple complaint classication tasks. However, BERT-based systems require substantial la- beled data for ne-tuning and do not generalize well to out- of-distribution complaint categories without retraining [9].

   3. Machine Learning versus Large Language Models

      Standard machine learning models are computationally ef- cient but exhibit limited semantic understanding and con- textual adaptability. Large Language Models, by contrast, support zero-shot and few-shot classication [4], enabling robust handling of diverse and evolving complaint categories with minimal supervision. Recent work has demonstrated that instruction-tuned LLMs have been shown to outperform in several short-text classication scenarios, particularly when training data is scarce or categories are loosely dened [11]. Howver, LLMs introduce challenges related to computational

      cost and inference latency. Recent research indicates that hybrid approaches combining LLM reasoning with lightweight traditional models can achieve a favorable balance between accuracy and efciency [8], [9], [13][15]. Furthermore, mul- timodal extensions of transformer architectures have shown that incorporating visual context alongside text consistently improves classication performance in ambiguous scenarios [12].

   4. Comparative Summary

   Table I summarizes the key characteristics of representative approaches in the literature. The proposed system advances beyond prior work by combining multimodal inputs, zero-shot LLM classication, and integrated priority estimation within a single deployable platform.

   TABLE I

   Comparison of Existing Complaint Classification Approaches

   Approach	Multimodal	Multilingual	Priority	Accuracy
   TF-IDF + SVM [7]	No	No	No	76%
   BERT Fine-tuned [1]	No	Limited	No	86%
   RailNeural [6]	No	No	No	83%
   Zero-shot LLM [4]	No	Yes	No	88%
   Proposed (Jagruk)	Yes	Yes	Yes	94.2%

3. Proposed System

   Jagruk is a web-based civic complaint management plat- form designed for smart city environments, where the high volume, diversity, and unstructured nature of citizen reports renders manual sorting impractical. Unlike traditional rule- based approaches, Jagruk employs an instruction-tuned Large Language Model for semantic understanding and autonomous decision-making.

   Every submitted complaint is processed by the LLM, which generates structured metadata encompassing the issue category, estimated urgency, responsible department, and a concise rationale. This metadata drives automated routing and prioritization. The platform is built on a full-stack architecture comprising React for the frontend, Node.js/Express for back- end orchestration, MongoDB for persistent storage, and several external microservices for media handling, authentication, geolocation, and LLM access.

   A. System Architecture

   Jagruk employs a modular, multi-tier architecture organized into four primary layers, as illustrated in Fig. 1.

   Presentation Layer: A React single-page application pro- vides role-based interfaces for citizens and administrators. Citizens submit complaints with optional image and location data, while administrators access dashboards for complaint tracking, queue management, and analytics.

   Service Layer: Built with Node.js and Express, this layer exposes RESTful APIs connecting the frontend to backend services. It handles input validation, orchestration logic, and integration with external services.

   Intelligence Layer: An instruction-tuned Large Language Model, accessed via the OpenRouter API, classies structured

   prompts into JSON-formatted outputs for automated catego- rization and prioritization.

   Persistence Layer: MongoDB stores enriched complaint records together with full audit trails, enabling traceability and compliance monitoring.

   External microservices include Appwrite for authentication and role management, Cloudinary for image storage and delivery, OpenCage for reverse geocoding, and OpenRouter for LLM access.

   Fig. 1. System architecture of the Jagruk LLM-driven civic issue platform.

   C. LLM-Based Processing

   Upon receiving a validated input payload, the backend constructs a structured prompt comprising two components.

   The system prompt species the classication task, de- nes the permitted complaint categories and urgency levels, and mandates a strict JSON output schema with the elds: category, priority, rationale, and department.

   The user prompt contains the dynamically populated com-

   plaint description, the image URL (if available), and the resolved geographic address. This composite prompt is sub- mitted to the instruction-tuned LLM via the OpenRouter API. Within a single inference step, the LLM simultaneously performs complaint classication, urgency estimation based on contextual signals (e.g., safety-related language), and de- partment assignment. The response is validated against the required schema. If the output is malformed or contains unsupported labels, the system retries with stricter output constraints. Persistent failures result in the complaint being

   agged as unclassied and queued for manual review.

   D. Output Generation

   Processed complaints are surfaced through role-specic interfaces. Citizens can track their submissions via a dashboard showing the assigned category, current status (Pending, In Progress, or Resolved), and an estimated resolution time based on predened SLA thresholds. A representative view of the citizen interface is shown in Fig. 2.

   Administrators access a unied dashboard that supports complaint sorting by urgency priority (Fig. 3). A map- based interface visualizes the geographic distribution of complaints (Fig. 4). An analytics module displays com- plaint volume trends, category distributions, and geographic hotspots to support operational decision-making. All system actionsincluding status transitions and resolution events are recorded in a timestamped audit log with user identiers and state change records.

   Fig. 2. Citizen complaint submission interface in Jagruk.

   B. Input Modalities

   The system accepts three categories of input: a natural lan- guage complaint description (mandatory), an optional image, and optional geolocation data. The text eld imposes no length or language restriction, accommodating reports ranging from brief phrases to detailed descriptions, including multilingual and code-mixed content.

   When an image is present, it is uploaded to Cloudinary and its CDN URL is incorporated into the processing pipeline as additional visual context. When location access is granted by the user, the browsers Geolocation API provides latitude and longitude coordinates, which are converted to a human- readable address using the OpenCage reverse geocoding ser- vice. All three inputs are combined into a unied context payload by the backend before being forwarded for LLM processing.

   Fig. 3. Citizen dashboard showing complaint status and SLA tracking.

   E. System Workow

   The end-to-end processing pipeline consists of ten sequen- tial stages, as illustrated in Fig. 5:

   Step 1 User Authentication: The citizen logs in via Appwrite-managed authentication. Step 2 Complaint Sub- mission: The complaint form is completed with a text descrip- tion and optional image and location inputs. Step 3 Image

   Fig. 4. Community and map-based complaint visualization dashboard.

   Upload: If present, the image is uploaded to Cloudinary and a CDN URL is returned. Step 4 Geolocation Resolution: Lat- itude/longitude coordinates are converted to a human-readable address via OpenCage. Step 5 Prompt Construction: The backend assembles a structured prompt from the complaint text, image URL, and resolved address. Step 6 LLM Inference: The prompt is dispatched to the instruction-tuned LLM via the OpenRouter API. Step 7 Response Validation: The LLM output is parsed and validated against the required JSON schema; malformed responses trigger a retry. Step 8 Data Persistence: The validated, enriched complaint record is stored in MongoDB. Step 9 Routing and Prioritization: A rule-based post-processing layer maps the predicted category to the responsible department and applies SLA-based priority thresholds. Step 10 Resolutin Logging: Administrators update complaint status and record resolution actions, com- pleting the complaint lifecycle.

4. Methodology

   The development of the Jagruk system was guided by ve principal design decisions: data collection, preprocess- ing, model selection, prompt engineering, and categorization logic. Together, these elements ensure that the architecture is accurate, scalable, and robust for real-world smart city deployments.

   1. Data Collection

      The evaluation dataset was assembled from two sources. The primary source was a controlled pilot deployment con- ducted in collaboration with a municipal body over a twelve- week period, during which citizens submitted complaints through the platform. Each record was independently labeled by three municipal domain experts to establish consensus- quality ground truth annotations for complaint category and urgency level.

      To address class imbalance and limited sample counts in underrepresented categories, a supplementary dataset was compiled from publicly available civic complaint portals. All records were anonymized using rule-based PII removal (elimi- nating names, phone numbers, and addresses), and underrepre- sented categories were augmented via stratied oversampling. The nal labeled dataset comprised approximately 4,800 com- plaint records across six municipal service categories.

      Fig. 5. Ten-stage end-to-end processing pipeline in Jagruk.

   2. Preprocessing

      A lightweight preprocessing pipeline was applied to normal- ize user-generated text. Unicode characters were normalized to NFC form to ensure consistent encoding across devices. Re- dundant whitespace was removed to improve tokenization con- sistency. Minimum input length was enforced at ten characters, with shorter inputs agged for user clarication. Excessively long inputs were segmented at sentence boundaries to conform to model context limits. Crucially, no stemming, stop-word removal, or spelling correction was applied, as instruction- tuned LLMs are inherently robust to noisy and non-standard text inputs.

   3. Model Selection

      Multiple commercially available instruction-tuned LLMs, accessed via the OpenRouter API, were evaluated on a strati- ed hold-out set comprising 20% of the dataset (approximately 960 samples). Models were assessed across three criteria: macro F1-score, P95 inference latency, and cost per 1,000 requests. The selected model achieved a macro F1-score exceeding 90%, maintained a P95 latency below three seconds under simulated concurrent load, and operated within accept- able cost bounds. OpenRouter was retained as the integration layer to support model-agnostic switching and failover logic.

   4. Prompt Engineering

      Prompt design follows a two-component structure. The static system prompt denes the classication task, enumerates permitted categories with brief descriptions, species urgency levels (Low, Medium, High, Critical), and mandates strict JSON output formatting. The dynamic user prompt is popu- lated at runtime with the complaint description, optional image URL, and resolved geographic address.

      A two-shot learning approach was incorporated into the sys- tem prompt, providing two labeled examples per semantically similar category pair. This improved classication performance on categories with overlapping lexical characteristics, with minimal latency overhead. Responses that are malformed (non- JSON, unsupported labels, or missing required elds) trigger an automatic retry with stricter output constraints. Persistent failures are escalated to a human review queue.

   5. Categorization Logic

   Following LLM output validation, a rule-based post- processing layer maps each predicted category to the cor- responding municipal department using a predened cong- uration table. The system supports multi-department routing in cases where a complaint spans multiple service domains. Priority levels govern both SLA deadlines and escalation mechanisms: Critical and High priority complaints trigger immediate notications to relevant authorities. A commu- nity voting mechanism additionally allows citizens to upvote widely reported issues, dynamically elevating their priority to ensure prompt attention to systemic problems.

5. Results and Discussion

   The proposed system was evaluated against three base- line models: SVM with TF-IDF features, Multinomial Naive Bayes, and a ne-tuned BERT model. Performance was as- sessed using accuracy, macro precision, macro recall, macro F1-score, Recall@K for routing effectiveness, and Mean Ab- solute Error (MAE) for priority prediction.

   1. Classication Performance

      The Jagruk LLM-based system achieves an overall accuracy of 94.2%, a macro precision of 93.1%, a macro recall of 92.7%, and a macro F1-score of 92.9%, surpassing the ne- tuned BERT baseline across all metrics by a substantial margin. Error analysis indicates that the majority of misclassi- cations occur on very short inputs (fewer than 15 characters) or highly code-mixed descriptions that blend two or more languages within a single sentence. Despite these edge cases, the system demonstrates consistent robustness across diverse linguistic patterns, underscoring the effectiveness of LLM- based semantic understanding in real-world civic applications. The superior performance of the proposed system over BERT and classical ML baselines can be attributed to three factors: (i) the instruction-tuned LLMs broad world knowl- edge enables interpretation of informal and ambiguous com- plaints without task-specic ne-tuning on large labeled cor- pora; (ii) two-shot prompting provides sufcient in-context guidance for discriminating between semantically adjacent categories; and (iii) the integrated multimodal context (im- age URL and geolocation) supplies disambiguation signals

      unavailable to text-only baselines.

      Table II and Fig. 6 present the quantitative results across all evaluated models.

      TABLE II

      Classification performance of all evaluated models. Bold values indicate the best result in each column.

      Model	Acc. (%)	Prec. (%)	Rec. (%)	F1 (%)	MAE
      TF-IDF + SVM	76.3	73.7	72.4	73.0	1.42
      Na¨ve Bayes	71.8	69.2	68.1	68.6	1.67
      Fine-tuned BERT	86.4	83.7	81.4	82.5	0.98
      Proposed LLM	94.2	93.1	92.7	92.9	0.54

   2. Per-Category Analysis

      Table III presents macro F1-scores disaggregated by mu- nicipal service category. The highest F1-score is achieved for Green Spaces & Parks (94.8%), likely attributable to the relatively unambiguous lexical markers associated with this category (e.g., tree, garden, park). Public Safety & Nuisance records the lowest F1-score (91.5%), primarily due to lexical overlap with other categories such as Roads & Infrastructure (e.g., complaints about construction noise at night). Two-shot prompting partially mitigates this ambiguity, though complete disambiguation remains challenging without increasing prompt complexity. Notably, all categories achieve F1-scores above 91%, indicating stable and consistent model performance across the full range of complaint types.

      Fig. 7. (a) Recall@K routing accuracy at K=1, 3, and 5 for all models. (b) MAE for priority prediction across all models. Lower MAE indicates more accurate urgency estimation.

      Fig. 6. Comparison of acuracy, precision, recall, and F1-score across all evaluated models. The proposed LLM system consistently outperforms all baselines.

      TABLE III

      Per-category macro F1 scores. LLM results compared with

      BERT AND TF-IDF+SVM.

      Category	F1 (LLM)	F1 (BERT)	F1 (SVM)
      Roads & Infrastructure	94.1	91.8	68.5
      Sanitation & Waste Mgmt.	91.8	89.2	63.4
      Water Supply & Drainage	93.3	90.7	65.1
      Electricity & Lighting	92.0	90.1	66.8
      Public Safety & Nuisance	91.5	88.9	62.7
      Green Spaces & Parks	94.8	92.3	70.2

   3. Routing Effectiveness and Priority Estimation

      Routing performance, measured as Recall@K, is illustrated in Fig. 7. At K=1, the proposed system achieves 87.4% accuracy, outperforming BERT (79.2%) and TF-IDF+SVM (68.5%). At K=3, accuracy improves to 96.1%, and at K=5 it reaches 98.6%, demonstrating strong multi-department routing reliability even in complaint scenarios involving overlapping departmental responsibilities. These results indicate that the correct department is consistently identied within a short candidate list, which is particularly valuable in municipal contexts where multiple departments may share jurisdiction.

      Priority prediction is evaluated using Mean Absolute Error (MAE). The proposed system achieves an MAE of 0.54, substantially lower than BERT (0.98), TF-IDF+SVM (1.42), and Naive Bayes (1.67). The improvement is most pronounced for high-urgency complaints, where accurate priority ordering is critical for timely governmental response.

   4. Deployment Considerations

   Beyond classication accuracy, several practical deployment aspects were evaluated. Under a simulated concurrent load, the system maintained a P95 inference latency of 2.7 seconds, sat- isfying real-time processing requirements for civic complaint management.

   Data privacy was enforced through rule-based PII removal applied prior to LLM submission, ensuring that sensitive per- sonal information is not exposed to external model providers.

   A fairness analysis revealed no statistically signicant per- formance bias across geographic regions or language groups, though further validation on larger and more diverse datasets is recommended before large-scale municipal deployment.

   The system incorporates human-in-the-loop safeguards, in- cluding a community reporting mechanism and a manual verication queue for low-condence classications. These features reduce the risk of systematic misclassication and support accountable decision-making in production deploy- ments.

6. Implementation

   The proposed system is implemented as a full-stack web application. The frontend is developed in React.js as a single- page application. The backend is built with Node.js and Express, providing RESTful API orchestration and complaint processing logic. Complaint records are persisted in Mon- goDB, with each document storing the enriched complaint metadata, processing audit trail, and status history.

   LLM inference is accessed through the OpenRouter API, enabling model-agnostic complaint classication, priority es- timation, and department routing. External service integrations include Cloudinary for media storage and delivery, OpenCage for reverse geocoding, and Appwrite for user authentication and role-based access control. The system is deployed on a cloud-hosted environment and is designed to support horizon- tal scaling to accommodate variable complaint volumes.

   1. System Workow Implementation

      The backend processing pipeline is structured as follows. User input is rst validated and normalized to ensure consis- tency. A structured prompt is assembled from the complaint description, optional image URL, and resolved geographic address, and submitted to the LLM. The LLM response, for- matted as a JSON object containing category, priority, department, and rationale, is validated against the required schema. The validated complaint is then persisted to MongoDB. A rule-based post-processing layer resolves the predicted category to a specic department and applies SLA-based priority thresholds. Administrators manage the complaint lifecycle through the admin dashboard, with all status transitions logged to the audit trail.

   2. Tools and Technologies

      • Frontend: React.js (Single Page Application)

      • Backend: Node.js + Express (RESTful API)

      • Database: MongoDB (document store with audit log- ging)

      • LLM Access: OpenRouter API (model-agnostic, with failover)

      • Cloud Services: Cloudinary (media), OpenCage (geolo- cation), Appwrite (auth)

      • Deployment: Cloud-hosted, horizontally scalable mi- croservice architecture

7. Advantages

   The proposed system offers several advantages over tradi- tional complaint management approaches:

   Semantic comprehension: The LLM-based pipeline ac- curately interprets high-variability inputs, including informal vocabulary, code-mixed text, and multilingual descriptions, without requiring explicit translation or preprocessing.

   End-to-end automation: Classication, prioritization, and routing are performed automatically within a single inference step, eliminating manual intervention and improving consis- tency.

   Scalability: The cloud-hosted architecture supports real- time processing of high-volume complaint streams without degradation in throughput or accuracy.

   Multilingual support: The system accommodates com- plaint descriptions in multiple languages natively, broadening accessibility for diverse citizen populations.

8. Limitations

   The following limitations have been identied in the current implementation:

   Classication inconsistency on ambiguous inputs: The model may produce inconsistent or incorrect classications for very short, highly ambiguous, or out-of-domain complaint descriptions that lack sufcient contextual signals.

   Computational cost: LLM inference incurs non-trivial API costs and computational overhead compared to lightweight classiers, which may constrain deployment at extreme scale. Inference latency under heavy load: Response times may degrade under peak concurrent request volumes, necessitating

   load balancing and request queuing strategies.

   Data dependency: System performance is sensitive to the quality and linguistic diversity of the training and evaluation datasets; poor-quality inputs may reduce classication accu- racy.

   Potential model bias: As with all pre-trained language models, the underlying LLM may reect biases present in its pre-training corpus, potentially affecting fairness across demographic or geographic groups [10].

9. Future Scope

   Several directions are identied for future enhancement of the Jagruk system:

   Hybrid model architecture: Combining LLM-based se- mantic understanding with lightweight ML classiers for a rst-pass lter could reduce API costs while preserving accu- racy for ambiguous cases.

   Domain-adaptive ne-tuning: Fine-tuning the LLM on city- or region-specic complaint corpora may improve clas- sication accuracy for domain-specic terminology and local governance structures.

   Live government API integration: Diect integration with municipal ERP and ticketing systems would enable automated complaint assignment and real-time status synchronization.

   Explainability: Incorporating model explanation mecha- nisms (e.g., rationale generation or attention visualization) would increase transparency and support audit-trail require- ments in e-governance contexts.

   Extended input modalities: Support for voice and video complaint submissions would broaden accessibility and cap- ture richer contextual information for improved classication.

10. Conclusion

This paper presented Jagruk, an LLM-driven civic complaint management system designed for smart city environments. The system processes heterogeneous citizen inputsincluding unstructured text, images, and geolocation datathrough a multimodal NLP pipeline that classies complaints, estimates urgency priority, and routes each issue to the appropriate municipal department within a single LLM inference step.

Experimental evaluation on a labeled municipal complaint dataset demonstrates that the proposed system achieves a macro F1-score of 92.9% and a routing Recall@1 of 87.4%, substantially outperforming TF-IDF+SVM, Multinomial Naive Bayes, and ne-tuned BERT baselines. The system maintains a P95 inference latency of 2.7 seconds under concurrent load, conrming its suitability for real-time deployment. Privacy, fairness, and human-in-the-loop safeguards further support responsible deployment in production e-governance settings.

These results indicate that LLM-based semantic understand- ing can meaningfully advance the state of civic complaint management, offering a scalable, accurate, and equitable alter- native to conventional approaches. Future work will focus on hybrid architectures, domain-adaptive ne-tuning, and direct integration with municipal service platforms.

References

1. J. Devlin, M.-W. Chang, K. Lee, and K. Toutanova, BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding, in Proc. 2019 Conf. North American Chapter of the ACL: Human Language Technologies (NAACL-HLT), Minneapolis, MN, pp. 41714186, Jun. 2019.

2. T. B. Brown et al., Language Models are Few-Shot Learners, in

   Advances in Neural Information Processing Systems (NeurIPS), vol. 33,

   pp. 18771901, 2020.

3. A. Vaswani et al., Attention is All You Need, in Advances in Neural Information Processing Systems (NIPS), vol. 30, 2017.

4. S. Gretz et al., Zero-shot Topical Text Classication with LLMs an Experimental Study, in Findings of the Assoc. for Computational Linguistics: EMNLP 2023, Singapore, pp. 96479676, Dec. 2023.

5. M. Rawat and N. Kaushik, NLP Based Grievance Redressal System for Indian Railways, arXiv preprint arXiv:2111.08999, Nov. 2021.

6. M. Gupta, A. Singh, R. Jain, A. Saxena, and S. Ahmed, Multi-class Railway Complaints Categorization Using Neural Networks: RailNeu- ral, Journal of Rail Transport Planning & Management, vol. 20, p. 100265, Dec. 2021. DOI: 10.1016/j.jrtpm.2021.100265.

7. S. Bhosale, S. Patankar, K. Kadam, R. Dhere, and M. Desai, Sur- vey on Civil Complaints Management System by Using Machine Learning Techniques, International Journal of Advanced Research in Science, Communication and Technology (IJARSCT), Jun. 2021. DOI: 10.48175/IJARSCT-1449.

8. H. Touvron et al., Llama 2: Open Foundation and Fine-Tuned Chat Models, arXiv preprint arXiv:2307.09288, Jul. 2023.

9. W. X. Zhao et al., A Survey of Large Language Models, arXiv preprint arXiv:2303.18223, Apr. 2023.

10. R. Navigli, S. Conia, and B. Ross, Biases in Large Language Models: Origins, Inventory, and Discussion, ACM Journal of Data and Infor- mation Quality, vol. 15, no. 2, pp. 121, 2023.

11. Z. Wei et al., Zero-Shot Information Extraction via Chatting with ChatGPT, arXiv preprint arXiv:2302.10205, Feb. 2023.

12. J. Li et al., BLIP-2: Bootstrapping Language-Image Pre-Training with Frozen Image Encoders and Large Language Models, in Proc. Int. Conf. Machine Learning (ICML), Honolulu, HI, pp. 1973019742, Jul. 2023.

13. J. Huang and K. C.-C. Chang, Towards Reasoning in Large Language Models: A Survey, in Findings of the Assoc. for Computational Lin- guistics: ACL 2023, Toronto, Canada, pp. 10491065, Jul. 2023.

14. N. Muennighoff et al., MTEB: Massive Text Embedding Benchmark, in Proc. 17th Conf. European Chapter of the ACL (EACL), Dubrovnik, Croatia, pp. 20062029, May 2023.

15. S. Shen et al., HuggingGPT: Solving AI Tasks with ChatGPT and its Friends in Hugging Face, in Advances in Neural Information Processing Systems (NeurIPS), vol. 36, pp. 3815438180, 2023.