How well does a sound component predict the sound?

Chinese characters are supposed to tell you how they are pronounced. Most of them carry a phonetic — a component borrowed for its sound, like 青 qīng inside 清 晴 請 情. Everyone learning traditional characters is told this. Nobody publishes how often it actually works.

So we measured it across the 3,000 most common traditional characters. Below is every phonetic family in the map with 3 or more members — 195 of them, covering 824 characters — ranked by how much of the head’s sound survives. The method is at the bottom, including what this measurement cannot tell you.

First, the honest problem

A ranking like this can cheat without meaning to. Our pipeline only draws a phonetic link where the readings already resemble each other — so scoring our own links and announcing that they agree would prove nothing.

The way out is to score every component against every character that contains it, a set we did not choose, and see whether identifying the phonetic separates the two populations. Across the 356 components that appear inside four or more characters:

That gap is the finding. “Characters hint at their sound” is only useful advice once you can say which piece is doing the hinting; guess wrong and you are at 3%. Everything below assumes you already know which component is the phonetic — which is what the map is for.

What the whole set looks like

Pooling all 824 characters in these 195 families, here is how much of the head’s reading each one kept:

• 29%Same reading — tone and all — 清 qīng from 青 qīng
• 28%Same syllable — the tone moved — 請 qǐng from 青 qīng
• 32%Near miss — rhymes, neighbouring consonant — 精 jīng from 青 qīng
• 9%Rhyme only — the rime survived, the consonant did not — 燒 shāo from 堯 yáo
• 1%No help — the sound has drifted away entirely

Read together: about 57% of the time the phonetic hands you the syllable outright, and 89% of the time it hands you something you could say and be understood. Tone is the part it almost never carries — only 29% match tone and all — which is why the score below ignores tone. A sound component never promised tone.

The families, ranked

Score is the share of the family whose reading you can predict from the head — same syllable, or the same rime with a neighbouring consonant. The bar breaks that down tier by tier, so a family that scores well on near misses looks different from one that scores well on exact matches. Names in color have a family page.

How this is computed

• The set. The 3,000 most common traditional characters, Taiwan readings (MOE via CNS11643, ranked by Taiwan word frequency). Families with fewer than 3 members are excluded — a family of two scores 0%, 50% or 100% and means nothing. Families with 4 or more also have a page of their own.
• The comparison. Readings are compared as zhuyin, which is already split into initial, final and tone. Two readings agree at the near tier when the rime matches and the initials fall in the same or a historically alternating articulatory group — 長 cháng → 張 zhāng is a real prediction, not a miss, because ㄔ and ㄓ are both retroflex.
• Primary readings only. A 多音字 with a second reading that happens to match the head would flatter its family, and a learner meeting a character cold gets the primary reading anyway.
• Tone is excluded from the score and published beside it. Requiring tone would score the writing system for a job it never claimed to do.

What this does not tell you

• It is bounded by the 3,000. A family that continues past the common set is scored only on the members inside it.
• Our identification favours precision over recall. Where a sound has drifted far enough, we would rather draw no link than a wrong one. Checked against the Dong Chinese wiki’s human-verified characters, our identification runs 94.5% precise at 56.8% recall — so real phonetic relationships with badly diverged readings are missing here rather than scored low. The ranking is honest about the links it has; it is not a complete census of Chinese phonetic series.
• It is synchronic. This measures modern Taiwan Mandarin. A component that predicted perfectly in Middle Chinese and lost the race to sound change scores badly here, which is correct for a learner and wrong for a historian.
• It is a computation, not an authority. Identification comes from the Dong Chinese character wiki, our own reading-comparison inference, and hand review; the sources are named per character at /credits. Corrections are welcome and get fixed in the data.