layout: default title: Normalization nav_order: 3 parent: Transforms


{: .no_toc }


{: .no_toc .text-delta }

  1. TOC {:toc}


Normalization is used to convert text to a unique, equivalent form. Software can normalize equivalent strings to one particular sequence, such as normalizing composite character sequences into pre-composed characters.

Normalization allows for easier sorting and searching of text. The ICU normalization APIs support the standard normalization forms which are described in great detail in Unicode Technical Report #15 (Unicode Normalization Forms) and the Normalization, Sorting and Searching sections of chapter 5 of the Unicode Standard. ICU also supports related, additional operations. Some of them are described in Unicode Technical Note #5 (Canonical Equivalence in Applications).


ICU 4.4 adds the Normalizer2 API (in Java, C++ and C), replacing almost all of the old Normalizer API. There is a design doc with many details. All of the replaced old API is now implemented as a thin wrapper around the new API.

Here is a summary of the differences:

  • Custom data: The new API uses non-static functions. A Normalizer2 instance can be created from standard Unicode normalization data, or from a custom (application-specific) data file with custom data processed by the new gennorm2 tool.
    • Examples for possible custom data include UTS #46 IDNA mappings, MacOS X file system normalization, and a combination of NFKC with case folding (see the Unicode FC_NFKC_Closure property).
    • By using a single data file and a single processing step for combinations like NFKC + case folding, the performance for such operations is improved.
  • NFKC_Casefold: ICU 4.4 supports the combination of NFKC, case folding and removing ignorable characters which was introduced with Unicode 5.2.
  • The old data file (used in Java, was hardcoded in the common library in C/C++) has been replaced with two new files, nfc.nrm and nfkc.nrm. If only canonical or only compatibility mappings are needed, then the other data file can be removed. There is also a new nfkc_cf.nrm file for NFKC_Casefold.
  • FCD: The old API supports FCD processing only for NFC/NFD data. Normalizer2 supports it for any data file, including NFKC/NFKD and custom data.
  • FCC: Normalizer2 optionally supports contiguous composition which is almost the same as NFC/NFKC except that the normalized form also passes the FCD test. This is also supported for any standard or custom data file.
  • Quick check: There is a new spanQuickCheckYes() function for an optimized combination of quick check and normalization.
  • Filtered: The new FilteredNormalizer2 class combines a Normalizer2 instance with a UnicodeSet to limit normalization to certain characters. For example, The old API's UNICODE_3_2 option is implemented via a FilteredNormalizer2 using a UnicodeSet with the pattern [:age=3.2:]. (In other words, Unicode 3.2 normalization now requires the data.)
  • Ease of use: In general, the switch to a factory method, otherwise non-static functions, and multiple data files, simplifies all of the function signatures.
  • Iteration: Support for iterative normalization is now provided by functions that test properties of code points, rather than requiring a particular type of ICU character iterator. The old implementation anyway simply fetched the code points and used equivalent code point test functions. The new API also provides a wider variety of such test functions.
  • String interfaces: In Java, input parameters are now CharSequence references, and output is to StringBuilder or Appendable.

The new API does not replace a few pieces of the old API:

  • The string comparison functions are still provided only on the old API, although reimplemented using the new code. They use multiple Normalizer2 instances (FCD and NFD) and are therefore a poor fit for the new Normalizer2 class. If necessary, a modernized replacement taking multiple Normalizer2 instances as parameters is possible, but not planned.
  • The old QuickCheck return values are used by the new API as well.

Data File Syntax

The gennorm2 tool accepts one or more .txt files and generates a .nrm binary data file for Normalizer2.getInstance(). For gennorm2 command line options, invoke gennorm2 --help.

gennorm2 starts with no data. If you want to include standard Unicode Normalization data, use the files in {ICU4C}/source/data/unidata/norm2/ . You can modify one of them, or provide it together with one or more additional files that add or remove mappings.

Hangul/Jamo data (mappings and ccc=0) are predefined and cannot be modified.

Mappings in one text file can override mappings in previous files of the same gennorm2 invocation.

Comments start with #. White space between tokens is ignored. Characters are written as hexadecimal code points. Combining class values are written as decimal numbers.

In each file, each character can have at most one mapping and at most one ccc (canonical combining class) value. A ccc value must not be 0. (ccc=0 is the default.)

Each line defines data for either a single code point (00E1) or a range of code points (0300..0314).

A two-way mapping must map to a sequence of exactly two characters. Multi-code point ranges cannot have two-way mappings.

A one-way mapping can map to zero, one, two or more characters. Mapping to zero characters removes the original character in normalization.

The generator tool will apply each mapping recursively to each other. Groups of mappings that are forbidden by the Unicode Normalization algorithms are reported as errors. For example, if a character has a two-way mapping, then neither of its mapping characters can have a one-way mapping.

* Unicode 6.1         # Optional Unicode version (since ICU 49; default: uchar.h U_UNICODE_VERSION)
00E1=0061 0301        # Two-way mapping
00AA>0061             # One-way mapping
0300..0314:230        # ccc for a code point range
0315:232              # ccc for a single code point
0132..0133>0069 006A  # Range, each code point mapping to "ij"
E0000..E0FFF>         # Range, each code point mapping to the empty string

It is possible to override mappings from previous source files, including removing a mapping:


Data Generation Tool

Normally, data from one or more input files is combined as described above, processed, and a binary data file is written for use by the ICU library (same file for C++ and Java). The binary data file format changes occasionally in order to support additional functionality.

    bin/gennorm2 -v -o $ICU4C_DATA_IN/nfkc_cf.nrm -s $ICU4C_UNIDATA/norm2 nfc.txt nfkc.txt nfkc_cf.txt

For the complete set of options, invoke gennorm2 --help.

Instead of the binary data file, the processed data can be written into a C file. This is closely tied to the needs of the ICU library. The format may change from one ICU version to the next.

    bin/gennorm2 -v -o $ICU_SRC/icu4c/source/common/norm2_nfc_data.h -s $ICU4C_UNIDATA/norm2 nfc.txt **--csource**

With the --combined option, gennorm2 writes the combined data of the input files. The following example writes the combined NFKC_Casefold data. (New in ICU 60.)

    bin/gennorm2 -o /tmp/nfkc_cf.txt -s $ICU4C_UNIDATA/norm2 nfc.txt nfkc.txt nfkc_cf.txt **--combined**

With the “minus” operator, gennorm2 writes the diffs of the combined data from two sets of input files. (New in ICU 60.)

For example, the nfkc_cf.txt file in ICU contains the Unicode NFKC_CF mappings, extracted from the UCD file DerivedNormalizationProps.txt. It is not minimal. The following command line generates the minimal differences of NFKC_Casefold compared with NFKC.

    bin/gennorm2 -o /tmp/nfkc_cf-minus-nfkc.txt -s $ICU4C_UNIDATA/norm2 nfc.txt nfkc.txt nfkc_cf.txt **minus** nfc.txt nfkc.txt


class NormSample {
    // ICU service objects should be cached and reused, as usual.
    NormSample(UErrorCode &errorCode)
        : nfkc(*Normalizer2::getNFKCInstance(errorCode),
            fcd(*Normalizer2::getInstance(NULL, "nfc", UNORM2_FCD, errorCode) {}

    // Normalize a string.
    UnicodeString toNFKC(const UnicodeString &s, UErrorCode &errorCode) {
        return nfkc.normalize(s, errorCode);

    // Ensure FCD before processing (like in sort key generation).
    // In practice, almost all strings pass the FCD test, so it might make sense to
    // test for it and only normalize when necessary, rather than always normalizing.
    void processText(const UnicodeString &s, UErrorCode &errorCode) {
        UnicodeString fcdString;
        const UnicodeString *ps;  // points to either s or fcdString
        int32_t spanQCYes=fcd.spanQuickCheckYes(s, errorCode);
        if(U_FAILURE(errorCode)) {
            return;  // report error
        if(spanQCYes==s.length()) {
            ps=&s;  // s is already in FCD
        } else {
            // unnormalized suffix as a read-only alias (does not copy characters)
            UnicodeString unnormalized=s.tempSubString(spanQCYes);
            // set the fcdString to the FCD prefix as a read-only alias
            fcdString.setTo(FALSE, s.getBuffer(), spanQCYes);
            // automatic copy-on-write, and append the FCD'ed suffix
            fcd.normalizeSecondAndAppend(fcdString, unnormalized, errorCode);
            if(U_FAILURE(errorCode)) {
                return;  // report error
        // ... now process the string *ps which is in FCD ...
    const Normalizer2 &nfkc;
    const Normalizer2 &fcd;