Methodology

Text you enter is sent from your browser to this site's server, which forwards it to a general-purpose machine translation service and returns the result. Nothing is translated on your device, and we do not train or fine-tune any model of our own.

The service we call is the same one behind most free translation widgets on the web. We are not claiming a proprietary engine — the difference here is the checked phrase data alongside it, not the machine output.

Identical requests are cached server-side for 24 hours so repeated lookups do not hit the upstream service unnecessarily, and there is a per-visitor rate limit for the same reason.

There is very little direct Ilocano–Tagalog training data in existence. In practice the engine translates Ilocano into English internally, then English into Tagalog. That intermediate step is the single largest source of error on this site.

It explains failures that otherwise look bizarre. Ilocano kannawanmeans “right” as in the direction; routed through the English word “right”, it comes back as Tagalog karapatan, a legal entitlement. Nothing in the Ilocano suggests a legal right — the English pivot introduced it.

The 428 entries were assembled from published references, not generated by the translator on this site. Where two sources agreed, we used the agreed form. Where they differed on spelling — Ilocano orthography is not fully standardised — we noted the variant in the entry.

Tagalog equivalents were written to match the sense of the Ilocano rather than to match it word for word, which is why some entries have a note explaining the difference.

Entries we could not tie to a citation are marked pending review and shown with that label on the page. They are not presented as verified.

• Ilocano phrasebook — Wikivoyage, CC BY-SA 4.0. Accessed 2026-07-26.
• 52 Useful Ilocano Words & Phrases to Learn for Travelers — Chill and Travel. Accessed 2026-07-26.

We took every translatable entry in the Ilocano phrasebook — 411of them, excluding those containing fill-in-the-blank placeholders — and sent each through this site's own translator, Ilocano into Tagalog. Each result was compared against the human-checked Tagalog.

Results were graded into four buckets:

• Exact — identical once punctuation and capitalisation are ignored.
• Close — 60% or more of the words overlap. A different but acceptable wording lands here.
• Partial — some overlap, but the meaning is incomplete or altered.
• Divergent — no meaningful overlap.

The headline figure of 58.9% is exact plus close. The measurement is reproducible: the script that produces it lives in the codebase and writes its raw results to a file, so the numbers on the accuracy page are generated rather than typed in by hand.

Two honest caveats. The grading is automatic, so a correct translation phrased differently from ours can be marked down — the true figure is probably a little better than 58.9%. And the test set is our own phrasebook: short, everyday phrases. Longer sentences with more context often fare better, because context is what rescues the engine from the ambiguity problem described above.

The same scoring runs across every pair the site publishes: 72 directions over 22,788 individual translations. One scoring function serves both measurements — two implementations would quietly produce two different accuracy scales, and the whole point of publishing a number is that it means something.

The reference translations come from a second dataset: 403 concepts held in parallel across the eight Philippine languages and English, so a single reviewed row produces every direction. Each pair is measured on the rows where both its languages have an entry, which is why the phrase counts differ between pairs — from 350 to 403 for the compiled set, against 411 for the cited Ilocano one.

Where a form could not be confirmed the column was left empty rather than filled with a plausible guess. That is what the uneven coverage below is:

The caveat that matters most: these seven languages are scored against compiled reference translations that no native speaker has yet reviewed. A language we have transcribed less confidently will score worse without the engine being any worse at it, and that risk is largest exactly where the score is lowest. The Ilocano figures rest on cited, checked data; the rest are a first measurement, not a settled one.

Re-running is incremental. A stored result is reused only when the phrase, its reference translation and its identifier are all unchanged, so correcting an entry re-measures that row and leaves the rest alone. Adding phrase data therefore costs only the new rows — and cannot silently leave a published figure describing a phrasebook that no longer exists.

Both datasets are published in full on the open-data page as CSV and JSON under CC BY-SA 4.0, so none of this has to be taken on trust.

Every phrase row has a “Suggest a correction” link. Submissions are recorded and checked against sources before anything changes; nothing is published automatically. See the editorial policy for how they are handled.