Two of my former PhD students, Jawwad Baig and Iniakpokeikiye Thompson, have worked on developing systems that give feedback to drivers on unsafe driving habits; evaluations showed that both of these systems were associated with reductions in unsafe driving. I have described this work in general in a previous blog. Jawwad and Thompson have now written a paper for INLG (Arxiv) which looks at how the different contexts in UK and Nigeria motivated changes in the type of feedback provided to users in these countries.

I think this is a very important and underexplored topic in NLG. Of course there is a lot of work on adapting systems to different languages, and also some work on localising visual features such as colour and images. But in an NLG context, we also need to localise the content of generated texts to fit country, culture, and context. Jawwad and Thompson’s work provides a really nice case study of this process.

Content depends on data

The first point is that content depends on available data. In a driving feedback context, a major difference between the UK and Nigeria is that reliable speed limit data for individual roads is available in UK (Jawwad used data from Damoov), but not in Nigeria. Hence the UK system can give feedback on exceeding the speed limit, but the Nigerian system cannot.

On the other hand, the Nigerian system uses a model to check driving after drinking, which was built from a dataset where Nigerian drivers self-report alcohol consumption; since such driving is often not stigmatised in Nigeria, we believe self-reports are at least somewhat accurate. The UK system does not do this, because we do not believe that self-reports from UK drivers would be accurate, since drunk driving is usually not socially acceptable in UK.

The key point is that an NLG system cannot generate content if it lacks appropriate input data, and what data is available is different in different countries. Jawwad and Thompson’s paper highlights that an “Evidence Audit” must be done to determine what data is available.

Content depends on stakeholder requirements

It is also of course true that content depends on what stakeholders (such as drivers and road safety organisations) want and need. For example, in the UK most drivers have basic knowledge of traffic rules and safe-driving principles, not least because they have to pass a test to get a driving licence. However in Nigeria, many people have very little knowledge about traffic rules and safe driving, perhaps in some cases because people have fraudulent licences. Hence the Nigerian system gives “Tips” which summarise relevant Nigerian laws (including fines); the UK system does sometimes give information about fines, but does not describe traffic laws.

Jawwad and Thompson’s paper describes other such examples, and highlights the need for explicit “Requirements elicitation” to understand stakeholder requirements

Content depends on what establishes credibility

A related but more subtle point is that users need to trust the app and believe that its advice is credible, and techniques for establishing credibility depend on culture. In particular, work with stakeholders suggested that the credibility of reports in Nigeria would be enhanced by citing specific laws and acknowledging local road realities; while in the UK work similar work suggested that personalising content to user circumstances increased trust. Of course it may be true that citing laws would also enhance the UK system, and personalisation would also enhance the Nigerian system, but the above reflects what stakeholders told Jawwad and Thompson.

Jawwad and Thompson’s paper says that “Message enrichment” is a key part of building an NLG system; this includes enriching core content about driving behaviour with information (such as citing laws) which enhances credibility and trust.

Country and context also matters in healthcare

The kind of adaptation that Jawwad and Thompson discuss is not limited to driving feedback! In a patient-information context, for example, we have seen problems where US-focused language models produce texts which is inappropriate for UK patients, for example recommending that patients talk to insurance providers, or phrasing texts in a way which perhaps are common in US, but are annoying to many UK readers.

One can imagine many other problems which could arise. For example, in rural Nigeria, telling patients to talk to their GP is not useful for patients who do not have easy access to a GP.

Final thoughts

Its not surprising that generated content needs to be adapted for different countries and contexts, but I have seen very little written about this in the NLG literature. I think Jawwad and Thompson’s paper is a great contribution in this space, because it is a concrete case study based on systems that work and are effective. I hope to see more papers on this topic!

Reference

Iniakpokeikiye Peter Thompson, Jawwad Baig, Ehud Reiter, Dewei Yi (2026). One Feedback System Does Not Fit All: Localising Data-to-Text Driver Coaching for the United Kingdom and Nigeria. To be presented at INLG-2026 (Arxiv)