An End-to-End System for Culturally-Attuned Driving Feedback using a Dual-Component NLG Engine

Iniakpokeikiye Peter Thompson; Reiter Ehud; Yi Dewei

REVIEW 4 major objections 6 minor 18 references

An End-to-End System for Culturally-Attuned Driving Feedback using a Dual-Component NLG Engine

T0 review · 4 major / 6 minor · reviewed 2026-08-05 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read This paper claims that a dual-component natural language generation engine, pairing legally grounded retrieval-based safety tips with a two-step reflective report pipeline, is a viable way to deliver culturally attuned driving feedback in a

desk verdict A clearly-written applied system paper with a plausible architecture, but its central factual-grounding claim is asserted, not demonstrated; the pilot shows the pipeline runs, not that the text is safe. read the letter →

arxiv 2509.04478 v1 pith:GT6YJ7RA submitted 2025-08-30 cs.CL

Iniakpokeikiye Peter Thompson , Yi Dewei , Reiter Ehud This is my paper · ORCID

classification cs.CL

keywords naturallanguagegenerationdata-to-textculturallyattunedfeedbackdrivingsafetyretrieval-augmentedtheoryofplannedbehaviourmobilesensingNigeria

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

The paper tries to show that an end-to-end mobile system can deliver safe-driving feedback that is both culturally attuned and factually reliable in a low-resource setting like Nigeria. Its main contribution is a dual-component NLG engine: a naive retrieval-augmented pipeline that grounds each safety tip in a specific section of the Nigerian Highway Code, and a reflective pipeline that uses the Theory of Planned Behaviour to generate persuasive weekly reports. A pilot with 90 drivers across four cities collected over two million data points and detected 154,230 unsafe events, supporting the claim that the architecture is viable. The paper is positioned as a blueprint for applying data-to-text and AI to social good in infrastructure-strained environments.

What carries the argument

The dual-component NLG engine. The Tips component is a Naive Retrieval-Augmented Generation pipeline that maps each detected unsafe behaviour by keyword to a pre-defined section of a local JSON encoding of the Nigerian Highway Code, guaranteeing factual grounding; the Reports component uses a two-step reflection process in which gpt-4-turbo drafts a Theory-of-Planned-Behaviour-grounded report and gpt-4o-mini refines it for factual consistency and tone. Supporting machinery includes a rule-based unsafe-event detector with context-dependent thresholds, a decision-tree classifier for alcohol-influenced trips, and a local-first architecture for offline operation.

What would settle it

Ask a licensed Nigerian driving-law expert to audit a sample of the generated tips against the full Highway Code; if a single tip contradicts or embellishes the retrieved regulation, the claimed legal-grounding guarantee fails.

Watch

Extended reading notes

Core claim

The paper's central claim is that its dual-component NLG engine—one arm using keyword-based retrieval from the Nigerian Highway Code to generate legally grounded safety tips, the other using a two-step reflective LLM pipeline guided by the Theory of Planned Behaviour to craft persuasive weekly reports—provides a viable framework for data-to-text systems aimed at road safety in low-resource settings. The pilot deployment with 90 drivers, which logged over 154,000 unsafe events, is offered as evidence that the system works end-to-end despite intermittent connectivity and noisy sensor data.

Load-bearing premise

The assumption that the keyword-based retrieval from the Nigerian Highway Code JSON plus the two-step LLM refinement guarantees that every generated tip is legally accurate; no evaluation of the generated text is reported.

Editorial extensions

If this is right

If correct, the system provides a working template for deploying data-to-text feedback systems in regions with poor connectivity and noisy sensor data.
The 154,230 unsafe events detected in the pilot indicate the rule-based detector can capture prevalent risky behaviours in Nigerian traffic at scale.
The planned pre-post evaluation will show whether the feedback reduces unsafe driving; positive results would make the framework transferable to other low-resource road-safety contexts.
The two-step LLM pipeline offers a cost-conscious pattern for generating safety-critical text: a strong model drafts, a cheaper model refines.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

The same keyword-retrieval-plus-LLM grounding pattern could be applied outside driving, for example to generate medical or financial advice where the source text is a fixed regulatory document.
The pilot reports no evaluation of the generated outputs themselves; auditing the tips against the Highway Code and the reports for factual and tonal consistency would settle the legal-grounding guarantee.
The two-step reflective design (expensive drafter, cheap reviser) could generalize to other content domains where factual fidelity matters more than generation diversity.

Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Desk Editor's Note

read the letter

Colleague,

Read the Thompson et al. driving-feedback system. The honest headline: this is a solid engineering paper with a plausible architecture, and it does not overclaim behaviour change. But the paper's own "legally-grounded" guarantee is the soft spot, and it is load-bearing.

What is actually new: the specific combination of a keyword-based retrieval pipeline for legal tips and a two-stage LLM report generator, deployed end-to-end on Android in Nigeria. Each component is standard (RAG with keyword retrieval, prompt chaining), and the authors cite SaferDrive as prior NLG driver feedback. The novelty is the configuration and the low-resource context, not a new method. That is fine for a systems paper, and they are honest that behaviour-change results are pending.

What the paper does well: the local-first architecture is thoughtfully engineered for intermittent connectivity and noisy sensor data. The alcohol-detection model is a real decision tree trained on a Nigerian dataset, described with features, and reported recall is high (though recall alone is weak). The paper is clearly written, the related work is adequate, and the limitations section acknowledges the main gaps.

Soft spots, in proportion:

- The 100% factual grounding claim in III.C.1 is not supported. Keyword retrieval selects the correct legal section assuming the behaviour-to-section mapping is right; it does not verify that the LLM-generated tip stays entailed by that section. No output evaluation, no constrained decoding, no post-hoc check. In a safety application, a paraphrased obligation can be legally wrong. The stress-test note is correct; this is the central weakness. - The two-step reflection is similarly asserted to reduce factual inconsistency, but no faithfulness evaluation is presented. Two examples are not evidence. - The alcohol model's 100% recall is reported without precision or cross-validation, and the rule-based event detection thresholds are not specified. This makes independent replication harder. - The pilot numbers (154k events, percentages) are descriptive counts with no baselines or error bars; the authors explicitly say the experiment is ongoing, so I would not count this as a flaw so much as a status report.

Overall: the core argument—that this architecture is a plausible reusable framework for low-resource road safety—holds up as a system claim. What does not hold up is the safety-critical factual-grounding guarantee. The authors are not pretending otherwise; they flag future evaluation. I would send this to a serious referee. It is a legitimate applied NLG system paper, and the missing output evaluation is precisely what a reviewer should demand. I would not desk-reject it.

Who is it for: people working on applied data-to-text, persuasive technology for road safety, or mobile AI in low-resource settings. I would cite it as related work for the architecture, not as evidence of effectiveness.

Recommendation: send to peer review, with the expectation that reviewers require either a verification mechanism for the grounding claim or a softened claim plus an evaluation plan with actual faithfulness metrics.

Referee Report

4 major / 6 minor

Summary. The paper presents an end-to-end Android system that detects unsafe driving events from smartphone sensors in Nigeria and generates two forms of culturally-attuned feedback: (1) legally-grounded safety tips derived from the Nigerian Highway Code via a keyword-based retrieval-augmented pipeline (the NRAG 'Tips' component), and (2) persuasive weekly behavioural reports generated through a two-step gpt-4-turbo to gpt-4o-mini refinement process grounded in the Theory of Planned Behaviour. The system also includes a decision-tree alcohol-detection model. The authors report a pilot deployment with 90 drivers, over 2,000,000 data points, and 154,230 detected unsafe events, and claim that the pilot demonstrates the viability of the approach. A future mixed-methods evaluation of behaviour change is planned but not yet conducted.

Significance. If the central claims were fully supported, the paper would provide a useful framework for applying data-to-text technology to road safety in a low-resource setting. The system engineering is thoughtful: local-first storage for intermittent connectivity, on-device model conversion to ONNX, a rule-based event detector, and a two-component NLG design separating legal tips from persuasive reports are all sensible design choices. The paper is also honest about several limitations (limited city coverage, ongoing evaluation). However, the current evidence does not substantiate the stronger claims made in the abstract and Section III. In particular, the '100% factual grounding' guarantee for the Tips component is asserted without any output-level verification, and the pilot results are descriptive counts without baselines, ground truth, or statistical analysis. The claimed novelty of combining retrieval-augmented legal grounding with TPB-based reflective reports is plausible, but the evaluation is too preliminary to establish effectiveness. The paper offers no reproducibility artifacts (no code, data, or prompts), and the alcohol-detection claim rests on a self-citation [18] whose availab

major comments (4)

[§III.C.1] The claim that the keyword-based retrieval 'guarantees 100% factual grounding and explainability' conflates source selection with source-grounded generation. Retrieval ensures that the correct Highway Code section is selected (assuming the behaviour-to-section mapping is correct), but the LLM is then instructed to generate a tip from that context. Nothing in the paper verifies that the generated tip is entailed by the retrieved legal text, that it omits no conditions, or that it introduces no prohibitions. No constrained decoding, post-hoc verification, or output evaluation is reported. Because legal accuracy is safety-critical in a driving intervention, this is a load-bearing gap. The authors should either restrict the claim to 'source selection is deterministic' or provide an evaluation of tip-level fidelity against the retrieved regulations.
[§III.B.2 / §VI] The alcohol-detection model is reported to achieve '100% recall on the test set', but no precision, specificity, F1, confusion matrix, dataset size, or cross-validation protocol is given. A model that labels every trip as alcohol-influenced would achieve 100% recall in an imbalanced test set, so this statistic alone is not meaningful. The supporting details are deferred to self-citation [18], which is not available in this manuscript. Since the model is presented as addressing 'a key local safety issue', the authors should report full classification metrics and a clear evaluation protocol in this paper, or clearly mark the result as preliminary and cite the external source only after confirming its accessibility.
[§III.B.1 / §IV.A] The rule-based event detection engine is described only qualitatively ('dynamic thresholds', 'sensitivity... increased at higher speeds'), with no concrete thresholds, sensor sampling rates, or event definitions. The pilot results in Table I are raw counts without any comparison baseline, ground-truth labelling, or per-driver statistics. The conclusion that 'this initial analysis confirms that the system is effective at capturing the specific types of risky behaviours prevalent in this driving context' is therefore unsupported: the system may be capturing many events, but 'effective' requires validation against observed driving behaviour or a reference dataset. The paper should specify the detection parameters and provide at least a small validation study or clearly restrict the claim to system operational viability.
[§III.C.2 / §IV.B] The two-step refinement process is claimed to 'mitigate known failure modes of LLMs, namely factual inconsistency', but no faithfulness or consistency evaluation of the generated reports is presented. Figure 5 shows two examples, which is illustrative only. Without an evaluation of whether the final reports accurately reflect the driver's statistics, the claim that the hybrid approach 'ensuring the final report is factually consistent' is unsupported. A small human or automated faithfulness evaluation (e.g., comparing report statistics against trip data) would be needed to substantiate this design rationale.

minor comments (6)

[Abstract / §IV] The abstract states that the pilot 'demonstrates the viability of our approach'. This overstates the evidence: the pilot shows that the pipeline runs and collects data, but not that the feedback is effective or legally safe. Consider softening to 'demonstrates operational feasibility'.
[§V.B] Typographical issue: 'Repeated Measures ANOV A' should read 'Repeated Measures ANOVA'.
[References [17]] In reference [17], 'Journal of KONES' contains a stray 'V ol.' in the volume field: 'vol. V ol. 19, No. 4'.
[§III.B.2] The sentence 'This underscores a key takeaway... a hybrid approach that fuses human expertise with data-driven techniques' (also repeated in Section VI) is vague and not directly supported by the results presented in this paper, since the feature-engineering connection to the alcohol model is not detailed here.
[§IV.A] Table I gives percentages but no absolute counts or per-driver aggregates. Adding the raw counts and the number of drivers contributing to each event type would improve interpretability.
[§III.C] The paper does not specify the prompt templates used for either the Tips or Reports component. Since the prompts are central to the claimed behaviour of the NLG engine, including them in an appendix would improve reproducibility.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the paper is a systems/architecture description with no derivational chain that reduces to its inputs.

full rationale

The paper presents an end-to-end mobile NLG system and a pilot deployment. It contains no equations or formal derivation chain; the central claim is the design and viability of a dual-component NLG engine. The components are described as concrete pipelines (NRAG tips and reflective reports), and their outputs are not derived from fitted parameters. The '100% factual grounding' claim in Section III.C.1 is an architectural statement about keyword-based retrieval mapping behaviours to a predefined JSON section; it does not predict a quantity from a fitted value. The self-citation [18] summarizes an auxiliary alcohol-detection model developed in prior work, and the paper explicitly states its development is detailed elsewhere. This citation is not load-bearing for the NLG engine claim, and no uniqueness theorem or ansatz is imported. The pilot results are descriptive event counts, not a prediction forced by construction. Limitations are acknowledged but do not indicate circularity. Therefore no significant circularity is found.

Assumptions & free parameters 3 free parameters · 4 assumptions · 0 invented entities

The central claims rest on several domain assumptions (TPB validity, Highway Code completeness, sensor detection accuracy, LLM reliability) that are asserted but not independently validated. No new physical entities are introduced. The main free parameters are the hand-tuned event-detection thresholds and the fitted ML model parameters, neither of which is reported in this paper.

free parameters (3)

Rule-based event detection thresholds
The 'dynamic thresholds' for harsh acceleration, braking, swerving, and speeding (Section III.B.1) are hand-chosen, but values are not stated. All pilot event counts depend on these thresholds.
Alcohol detection Decision Tree parameters = not provided (see [18])
The 100% recall claim depends on model parameters fitted to a Nigerian dataset; no precision, cross-validation, or artifact is reported in this paper.
Trip debouncing timing parameters
Trip start/stop detection uses a 'debouncing mechanism' (Section III.A) with unspecified time constants that affect trip segmentation and hence event counts.

assumptions (4)

domain assumption Theory of Planned Behaviour (TPB) is a valid framework for persuasive feedback in this population.
The Reports pipeline is grounded in TPB (Sections II and III.C.2), but no evidence is provided that TPB-based messages persuade Nigerian drivers.
domain assumption The Nigerian Highway Code JSON is accurate, complete, and legally authoritative.
The NRAG Tips pipeline maps detected behaviours to pre-defined JSON sections (Section III.C.1); incorrect or incomplete content would generate legally wrong tips despite the '100% factual grounding' claim.
domain assumption Smartphone sensor data and the rule-based event detector correctly identify unsafe driving events.
Pilot event counts (Table I) are treated as measures of unsafe behaviour, but no ground-truth validation of thresholds or sensor reliability is reported (Section IV.A).
domain assumption LLM prompt-chaining can produce factually consistent and culturally appropriate text.
The two-step reflection process is assumed to mitigate LLM factual inconsistency (Section III.C.2), but no evaluation of output quality is presented.

how reviews work

0 comments

Cite this review

Pith. "Pith review of An End-to-End System for Culturally-Attuned Driving Feedback using a Dual-Component NLG Engine." pith.science (2026). https://pith.science/paper/GT6YJ7RA

@misc{pith2026250904478,
  author       = {Pith},
  title        = {Pith review of: An End-to-End System for Culturally-Attuned Driving Feedback using a Dual-Component NLG Engine},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/GT6YJ7RA}},
  note         = {Machine review of arXiv:2509.04478}
}

read the original abstract

This paper presents an end-to-end mobile system that delivers culturally-attuned safe driving feedback to drivers in Nigeria, a low-resource environment with significant infrastructural challenges. The core of the system is a novel dual-component Natural Language Generation (NLG) engine that provides both legally-grounded safety tips and persuasive, theory-driven behavioural reports. We describe the complete system architecture, including an automatic trip detection service, on-device behaviour analysis, and a sophisticated NLG pipeline that leverages a two-step reflection process to ensure high-quality feedback. The system also integrates a specialized machine learning model for detecting alcohol-influenced driving, a key local safety issue. The architecture is engineered for robustness against intermittent connectivity and noisy sensor data. A pilot deployment with 90 drivers demonstrates the viability of our approach, and initial results on detected unsafe behaviours are presented. This work provides a framework for applying data-to-text and AI systems to achieve social good.

Figures

Figures reproduced from arXiv: 2509.04478 by the authors.

**Figure 2.** Sequence diagram of component interactions. [PITH_FULL_IMAGE:figures/full_fig_p002_2.png] view at source ↗

**Figure 3.** High-level architecture of the dual-component NLG [PITH_FULL_IMAGE:figures/full_fig_p002_3.png] view at source ↗

Figures from the paper (2 more)

**Figure 4.** Figure 4: The two-step reflection process for generating persua [PITH_FULL_IMAGE:figures/full_fig_p003_4.png]

**Figure 5.** Figure 5: Examples of final system outputs as displayed in the mobile application. [PITH_FULL_IMAGE:figures/full_fig_p004_5.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

18 extracted references · 17 canonical work pages

[18]

Mobile Phone Sensor-based Nigerian Driving Dataset to Detect Alcohol-influenced Behaviours

I. P. Thompson, Y . Dewei, and R. Ehud, “Mobile Phone Sensor-based Nigerian Driving Dataset to Detect Alcohol-influenced Behaviours.”

work page
[1]

Global status report on road safety 2023,

World Health Organization, “Global status report on road safety 2023,” 2023

work page 2023
[2]

Frsc records fewer crashes, higher death toll in 2024,

F. Okpale, “Frsc records fewer crashes, higher death toll in 2024,” Businessday NG, 2025, accessed: June 7, 2025

work page 2024
[3]

Understanding risky driving behaviors among young novice drivers in nigeria: A latent class analysis coupled with association rule mining approach,

M. S. Labbo, L. Qu, C. Xu, W. Bai, E. A. Atumo, and X. Jiang, “Understanding risky driving behaviors among young novice drivers in nigeria: A latent class analysis coupled with association rule mining approach,” Accident Analysis & Prevention , vol. 200, p. 107557, 2024

work page 2024
[4]

Saferdrive: An nlg-based behaviour change support system for drivers,

D. Braun, E. Reiter, and A. Siddharthan, “Saferdrive: An nlg-based behaviour change support system for drivers,” Natural Language En- gineering, vol. 24, no. 4, pp. 551–588, 2018

work page 2018
[5]

Data-to-text: A constrained approach to diverse candidate text generation and ranking,

W. S. Chen, J. Boylan, and D. Meyer, “Data-to-text: A constrained approach to diverse candidate text generation and ranking,” in 2023 9th International Conference on Computer and Communications (ICCC) , 2023, pp. 1022–1029

work page 2023
[6]

The influence of road safety culture on driver behaviour: A study of nigerian drivers,

C. C. Uzondu, “The influence of road safety culture on driver behaviour: A study of nigerian drivers,” Ph.D. dissertation, University of Leeds, 2019

work page 2019
[7]

Nigeria: Freedom on the net 2021 country report,

Freedom House, “Nigeria: Freedom on the net 2021 country report,” Freedom House, Tech. Rep., 2021

work page 2021

Show all 18 references

[8]

Identification and analysis of risky driving behaviors at intersections based on vehicle trajectories,

Y . Zheng, W. Li, and K. Li, “Identification and analysis of risky driving behaviors at intersections based on vehicle trajectories,” in 2022 8th International Conference on Computer and Communications (ICCC) , 2022, pp. 1113–1118

2022
[9]

Smart- phone sensing for understanding driving behavior: Current practice and challenges,

E. Mantouka, E. Barmpounakis, E. Vlahogianni, and J. Golias, “Smart- phone sensing for understanding driving behavior: Current practice and challenges,” International Journal of Transportation Science and Technology, vol. 10, no. 3, pp. 266–282, 2021

2021
[10]

Culture of distracted driving among intra-city com- mercial bus drivers in ile-ife, south-western nigeria,

O. M. Olapoju, “Culture of distracted driving among intra-city com- mercial bus drivers in ile-ife, south-western nigeria,” Transportation Research Part F: Traffic Psychology and Behaviour , vol. 42, pp. 425– 432, 2016

2016
[11]

Survey of the State of the Art in Natural Language Generation: Core tasks, applications and evaluation,

A. Gatt and E. Krahmer, “Survey of the State of the Art in Natural Language Generation: Core tasks, applications and evaluation,” Jan. 2018, arXiv:1703.09902 [cs]. [Online]. Available: http://arxiv.org/abs/1703.09902

2018 arXiv
[12]

Why is constrained neural language generation particularly challenging?

C. Garbacea and Q. Mei, “Why is constrained neural language generation particularly challenging?” Apr. 2025, arXiv:2206.05395 [cs] version: 2. [Online]. Available: http://arxiv.org/abs/2206.05395

2025 arXiv
[13]

Persuasive technology for health and wellness: State of the art and emerging trends,

R. Orji and K. Moffatt, “Persuasive technology for health and wellness: State of the art and emerging trends,” in Journal of medical Internet research, vol. 20, no. 1. JMIR Publications Inc., Toronto, Canada, 2018, p. e10804

2018
[14]

The theory of planned behavior,

I. Ajzen, “The theory of planned behavior,” Organizational Behavior and Human Decision Processes , vol. 50, no. 2, pp. 179–211, 1991

1991
[15]

Efficacy of the Theory of Planned Behaviour: a meta-analytic review,

C. J. Armitage and M. Conner, “Efficacy of the Theory of Planned Behaviour: a meta-analytic review,” The British Journal of Social Psychology, vol. 40, no. Pt 4, pp. 471–499, Dec. 2001

2001
[16]

An integrated behavior change model for physical activity,

M. S. Hagger and N. L. D. Chatzisarantis, “An integrated behavior change model for physical activity,” Exercise and Sport Sciences Re- views, vol. 42, no. 2, pp. 62–69, Apr. 2014

2014
[17]

Measuring driver behaviour - indicators for traffic safety,

M. Niezgoda, T. Kami ´nski, and M. Kruszewski, “Measuring driver behaviour - indicators for traffic safety,” Journal of KONES , vol. V ol. 19, No. 4, pp. 503–511, 2012. [Online]. Avail- able: http://yadda.icm.edu.pl/baztech/element/bwmeta1.element.baztech- article-BUJ8-0021-0008

2012

Pith tools

Reviewed August 5, 2026 · model on record in the stance chip above.

[1] [18]

Mobile Phone Sensor-based Nigerian Driving Dataset to Detect Alcohol-influenced Behaviours

I. P. Thompson, Y . Dewei, and R. Ehud, “Mobile Phone Sensor-based Nigerian Driving Dataset to Detect Alcohol-influenced Behaviours.”

work page

[2] [1]

Global status report on road safety 2023,

World Health Organization, “Global status report on road safety 2023,” 2023

work page 2023

[3] [2]

Frsc records fewer crashes, higher death toll in 2024,

F. Okpale, “Frsc records fewer crashes, higher death toll in 2024,” Businessday NG, 2025, accessed: June 7, 2025

work page 2024

[4] [3]

Understanding risky driving behaviors among young novice drivers in nigeria: A latent class analysis coupled with association rule mining approach,

M. S. Labbo, L. Qu, C. Xu, W. Bai, E. A. Atumo, and X. Jiang, “Understanding risky driving behaviors among young novice drivers in nigeria: A latent class analysis coupled with association rule mining approach,” Accident Analysis & Prevention , vol. 200, p. 107557, 2024

work page 2024

[5] [4]

Saferdrive: An nlg-based behaviour change support system for drivers,

D. Braun, E. Reiter, and A. Siddharthan, “Saferdrive: An nlg-based behaviour change support system for drivers,” Natural Language En- gineering, vol. 24, no. 4, pp. 551–588, 2018

work page 2018

[6] [5]

Data-to-text: A constrained approach to diverse candidate text generation and ranking,

W. S. Chen, J. Boylan, and D. Meyer, “Data-to-text: A constrained approach to diverse candidate text generation and ranking,” in 2023 9th International Conference on Computer and Communications (ICCC) , 2023, pp. 1022–1029

work page 2023

[7] [6]

The influence of road safety culture on driver behaviour: A study of nigerian drivers,

C. C. Uzondu, “The influence of road safety culture on driver behaviour: A study of nigerian drivers,” Ph.D. dissertation, University of Leeds, 2019

work page 2019

[8] [7]

Nigeria: Freedom on the net 2021 country report,

Freedom House, “Nigeria: Freedom on the net 2021 country report,” Freedom House, Tech. Rep., 2021

work page 2021