Acro Helper: contextual help without leaving the page

Acro Helper is a browser extension I built to help people read technology articles when the terminology gets in the way. An unfamiliar acronym can be a small interruption for an experienced reader. For someone less comfortable with technology, it can mean opening another tab, searching for a definition, and trying to find their place again.

The code is open source: https://github.com/unseen1980/acro-helper. There is also a video demonstration.

Preprint paper: https://arxiv.org/abs/2603.17592

The interaction is deliberately small. Hover over a recognised technical acronym and a tooltip explains it in place. The article stays where it is, and the reader decides whether they need the explanation. This work began as my MSc Computing dissertation; the paper presents the system and the evaluation of that original project.

Underneath that simple interaction is a pipeline combining a curated technical dictionary with language-model-generated definitions. The dictionary handles known terms, while the language model can identify and explain additional acronyms using the surrounding text. A classification stage first checks whether the page is about technology, using Google Cloud’s NLP taxonomy and an OpenAI fallback. Without that check, everyday words can easily be mistaken for technical abbreviations.

The extension adds definitions to the page’s text without replacing existing links or changing the article’s structure. It also provides a separate dictionary search page. That gives readers another route to a definition when the unfamiliar term comes from somewhere other than the current web page.

I evaluated it with 25 participants recruited from a community centre, a further-education college, and staff whose work did not involve daily computer use. Participants read a technology article, tried the extension, completed questionnaires, and took part in interviews. The study took place in 2023, with participants reporting low to intermediate levels of digital literacy.

The responses were encouraging: 92% reported at least moderate improvement in understanding technical terms, and 96% agreed that it saved time compared with searching manually. All 25 said the tooltips did not distract them from reading. These are participants’ reports from a supervised study, rather than proof of a lasting improvement in digital literacy.

The timing results also show why the dictionary still matters. Across ten benchmark runs, dictionary-based tooltip injection averaged about 2.1 seconds, compared with 16.4 seconds for the language-model path. Participants’ manual searches for “CPU” averaged 17.2 seconds per search. The extension timings cover page processing and tooltip injection, so they should not be read as identical tasks measured under identical conditions.

There is a trade-off here. Around 70% of participants preferred the AI-generated explanations for their concision and relevance to the article, even though the dictionary path was considerably faster. People also asked for adjustable text size and a choice between brief and detailed definitions. Those details matter when the purpose of the tool is to make reading easier.

The sample is small, participation was self-selected, and the supervised setting limits what can be concluded about everyday use. Ambiguous terms remain a problem too: “POST” does not always mean “Power On Self Test”. Caching definitions and resolving ambiguity locally are possible next steps, rather than features I am claiming the evaluated system already has.

What I take from the project is the value of putting help at the point where it is needed. A useful explanation is easier to act on when getting it does not require abandoning the thing you were trying to understand.

Thanks for reading - Christos