Design Docs‎ > ‎Compact Encodings‎ > ‎

Compact Encoding of Names

What? Person names, company names, place names, etc.

Problem: Encode name strings in a compact way.

Assumption: Match names case-insensitively. If this is the case, then we can normalize names using Unicode NFKC_Casefold or similar.

Assumption: We decode more often than encode, and don't need random access on the encoded byte sequence.

Observations: Names mostly consist of letters, spaces and occasional punctuation. After normalization, there are no uppercase letters. Names normally use letters from a single script. No control codes, few format control characters (probably no RLM/LRM), few symbols.

Ideas:

  • Optimize for high-runner cases, with an escape mode for all of Unicode.
  • For the encoding process: Try out different sub-encodings for a name string, choose the one that yields the smallest number of bytes.
  • We probably need to indicate inside the encoded byte sequence which sub-encoding we use. (As opposed to out-of-band information.)
  • With names normally using a single script, a single sub-encoding should suffice in most cases, without the need to switch mid-stream. (Unlike SCSU for example which is designed for switching between modes.)
  • Therefore, the first byte can indicate the sub-encoding used in the whole sequence. For the most common sub-encodings, the first byte can also contain the first letter of the name, or part of it (a lead byte).
  • Example sub-encodings:
    • Use 5 bits per character for Latin a-z, space, and some other characters commonly used in Latin names. Use 32 values in the first byte to indicate this sub-encoding and the first character.
      • With a five-bit encoding, make sure we can tell which quintet is the last. We need to be careful with a trailing 0 quintet. Set some of the bits following it, or fail the 5-bit sub-encoding if the last quintet is a 0 and there are no spare bits to set.
    • Use 5 bits per character for the 32 most common Cyrillic letters (out of the 33 main letters). Use 32 values in the first byte to indicate this sub-encoding and the first character.
    • Use 5 bits per character for the 32 most common Arabic letters. Use 32 values in the first byte to indicate this sub-encoding and the first character.
    • Use 8 bits per character for Hiragana & Katakana (U+3040..U+30FF). Throw in the 32 characters from 5-bit Latin and maybe U+3000..U+301F.
    • Maybe use 5 or 6 bits per character for various other small scripts, although the majority of names will be covered by sub-encodings listed here.
    • Maybe use 8 bits per character for Indic/Thai/Lao U+0900..U+0EFF where the first byte would be one of 6 values indicating the lead byte for the remaining truncated character codes. 2..4 byte values (e.g., fc..ff) could be reserved for cherry-picked characters like ZWNJ and ZWJ.
    • Maybe use 12 bits per character for U+0000..U+0FFF for most small scripts not covered by more efficient sub-encodings, and as an escape valve for Cyrillic and Arabic etc. that don't fit into their 5-bit sub-encodings. Maybe throw in a few other common characters like maybe ZWJ & ZWNJ, replacing control codes.
    • Alternative 12-bit-per-character encoding, considering that Latin text is better handled by the Unicode mode below:
      • U+0300..U+07FF -> 0..0x4ff  // 0x500 combining marks, Greek, Cyrillic, Arabic, Hebrew...
      • U+0900..U+10FF -> 0x500..0xcff  // 0x800 Indic scripts, Thai, Lao, Tibetan, Burmese, Georgian
      • U+1200..U+13FF -> 0xd00..0xeff  // 0x200 Ethiopic, Cherokee
      • U+1780..U+17FD -> 0xf00..0xf7d  // 0x7E Khmer
      • U+200C..U+200D -> 0xf7e..0xf7f  // 2 ZWNJ & ZWJ
      • U+2D80..U+2DDF -> 0xf80..0xfdf  // 0x60 Ethiopic Extended
      • space, a-z, some other characters -> 0xfe0..0xfff  // 0x20 same as 5-bit Latin
    • Use 2 bytes per character for Unihan (U+4E00..U+9FFF) and Hangul (U+AC00..U+D7FF). Use 126 values in the first byte to indicate this sub-encoding and the lead byte of the first character. Maybe use another 26 values in the first byte to also indicate this sub-encoding, but with a-z as the first character.
      • Throw in the 32 characters from 5-bit Latin as single bytes E0..FF, pushing rarely-used high-BMP characters to a 3-byte-per-character encoding, all while staying in this mostly-16-bit mode. We could also use some other single bytes (e.g., D8..DF) for a few other common characters like maybe ZWJ & ZWNJ, for 3-byte modes for rest-of-BMP and plane 2, and a 4-byte mode for other supplementary characters. (Or 17 byte values for rest-of-Unicode indicating bits 20..16.)
      • Note also that U+3400..U+4DFF and most of U+2000..U+2FFF occur rarely, so (some of) their lead bytes might be used for something else too.
      • These extra modes turn this mode into a reasonable all-Unicode mode.
Comments