How accurate is Ilocano to Tagalog translation?

We ran all 411 translatable entries in our Ilocano↔Tagalog phrasebook through the same engine this site uses, and compared each result against the human-checked Tagalog. Categories are our own; the comparison ignores punctuation and casing.

Every other Ilocano translation site we looked at asserts that its results are accurate. None publishes a number. This is ours, and the method is described in full on the methodology page so you can judge whether it is a fair test.

Phrases tested

411

Exact match

212

Close

30

Wrong or partial

169

Translating out of Ilocano works better than translating into it. Across 411 phrases, Ilocano as the source scored 58.9% and 56.0%; Ilocano as the target scored 50.1% and 47.2% — a gap of 8.8 points. We have now measured this at three sample sizes and the direction of the effect held every time, though its size moved between roughly nine and fourteen points, so treat the exact number as approximate.

This is what a low-resource language looks like from the inside. Recognising Ilocano and mapping it onto a well-resourced language is the easier half of the problem. Producing correct Ilocano — choosing the right affix, the right linker, the right register — needs far more training data than exists, so the engine falls back on passing words through unchanged or borrowing a form from the source language.

If you need Ilocano in the output and it matters, have a speaker check it. The direction you are translating changes how much you can trust the result, and every tool we know of presents both directions as equally reliable.

We have since run the same measurement on every pair the site publishes — 72 directions across eight Philippine languages plus English, 22,788 individual translations, scored identically. Mean agreement is 54.5%.

The headline above — that translating out of Ilocano beats translating into it — held, though at 2.2 points rather than the nine measured here. What the Ilocano-only test could not show is that the effect reverses for several languages entirely. The widest gap belongs to Hiligaynon, at 13.4 points: it is the worst target language of the nine by a clear margin, across every source we measured it from.

The full ranking, the per-language breakdown and the limits of the comparison are on the Philippine-languages measurement.

This page used to tell you to translate whole sentences rather than isolated words, on the reasoning that context disambiguates. That was an argument, so we measured it. It does not hold.

Longer input did not translate more reliably. Going into Tagalog, single words and long phrases scored about the same; going into Ilocano, every multi-word bucket did worse than single words.

Two limits on this. The multi-word buckets are small — 39 and 68 entries — so a few results move them several points. And our automatic grading treats lengths differently: a long phrase translated acceptably but worded differently can fall below the word-overlap threshold and be marked down, while a single word is judged almost binary. We cannot cleanly separate those effects from a real one.

What we can say is narrower than what we used to say: there is no evidence here that feeding the engine more context makes it more reliable, so we no longer advise it as a fix.

Reliability is not evenly spread. Concrete vocabulary does well; anything involving people, relationships or politeness does badly.

The most damaging failure, because nothing signals it went wrong. When the engine does not recognise an Ilocano word it echoes the input. Ten of our test phrases came back identical to what went in despite having a perfectly ordinary Tagalog equivalent.

If the output looks like the input, it almost certainly failed. A further 17 words in our set were identical in both languages legitimately — 'lima', 'pito', 'itlog', 'kape' — so identical output is not proof of failure on its own.

The engine has little direct Ilocano↔Tagalog training data, so it pivots through English. Where the English word is ambiguous, the wrong sense gets carried into Tagalog.

We assumed a full sentence would rescue the right sense, and then measured it. It does not: longer input scored no better than single words going into Tagalog, and worse going into Ilocano. Check the output instead of trusting context to fix it.

Short greetings sometimes come back with filler the original did not contain.

Read the output for length. If it is noticeably longer than the source, something was invented.

Kinship and courtesy terms come back in a different register than the source, which matters in a language where address terms carry respect.

Have output checked by a speaker before using it in anything formal or public.

• Legal, medical or financial documents.
• Anything going into print or public signage.
• Names, addresses and place names — these are passed through or mangled unpredictably.
• Regional Ilocano vocabulary outside the Ilocos Norte/Sur standard.
• Anything where formality or respect matters, given the register drift above.

For any of those, use a professional translator. A free tool that is right about right about three times in five is a helpful draft, not a substitute for one.

• Do not assume a longer sentence will fix a bad result. We tested that and it is not true — see the section above.
• Compare the output against our checked phrasebook for anything common.
• Watch for output identical to your input, and for output noticeably longer than your input. Both are reliable signals that something went wrong.
• Use standard Ilocano rather than regional vocabulary where you can — the engine has effectively no coverage outside the Ilocos Norte and Sur varieties.