Skip to content

Radically/radically

Repository files navigation

Radically

You may also like

Lookup Preprocessing Algorithm

A basic implementation of a lookup system from constituent radical to character is simple - depth first search starting from the root node, adding every encountered character in every IDS string into a hash table / dictionary / similar data structure. Indeed, this is done in this project.

車 -> 蓮,櫣,輿,轉 ...
车 -> 莲,舆,转 ...

The complication arises if we want to include radical frequency in our search - e.g. "find all characters with at least 3 occurrences of 人" should return 傘, 众, and 齒, but not 从, 纵, 齿, etc.

Decomposition of 蓮

  • There may be more than one IDS string, and thus decomposition per character.
  • The total number of characters in each head to leaf node path varies, as does the number of unique characters. We may (and I have!) run into a character with a very elaborate decomposition if we take a particular branch, and a decomposition that terminates early on another1.

The approach taken is to generate a list of the powerset2 of radicals for every possible head to leaf node path, EXCLUDING THE ROOT ITSELF (important!), e.g. the expected output would be

[
  { '艹': 1, '連': 1, '辶': 1, '車': 1, '十': 1, '丨': 1 },
  { '艹': 1, '連': 1, '辶': 1, '車': 1, '十': 2 },
  { '辶': 1, '莗': 1, '艹': 1, '車': 1, '十': 1, '丨': 1 },
  { '辶': 1, '莗': 1, '艹': 1, '車': 1, '十': 2 }
]

where there are 4 sets, or rather hash tabless, depicted by the 4 different colors above. The one in green corresponds to the 1st item in the list.

Define a recursive function, rec(char), where char is a single character and the output is as described above.

Taking the left branch from the root, 艹 can be decomposed into either 十十 or 十丨, hence rec(艹) -> [ { '十': 1, '丨': 1 }, { '十': 2 } ], and similarly, rec(連) -> [ { '辶': 1, '車': 1 } ].

The key observation is that for each character in an IDS string, we must pick 1 set from its list of sets before merging 1 set per character into a bigger set, to avoid double-counting radicals. Ergo, we must find the cartesian product between the powersets of each constituent character, then merge each resulting n-tuple.

Hence, the powersets of the IDS string 十連, although not explicitly defined here, are [ { '十': 1, '丨': 1, '辶': 1, '車': 1 }, { '十': 2, '辶': 1, '車': 1 } ].

Define another function mergeTwoFreqs(powersetA, powersetB) which merges two powersets together, adding the frequencies together if they are in both sets. Merging 3 or more sets is done by mergeTwoFreqs(powersetA, mergeTwoFreqs(powersetB, powersetC)...

Note how the first item of rec(艹), and then the second item, is merged with the only item of rec(連) in turn. In other words, for a given IDS string ABC, where rec(A) -> [ A0, A1, A2 ], rec(B) -> [ B0, B1, B2 ], and rec(C) -> [ C0, C1, C2 ], i.e. this is a permutation generation problem where we need to generate [ mergeFreqs((A0, B0, C0)), mergeFreqs((A0, B0, C1)), ... mergeFreqs((A2, B2, C2)) ].

Define another function permGen(n-tuple[]) which takes the length of each n-tuple and outputs int[] describing each permutation by their indices, e.g. [ [0,0,0], [0,0,1], ... [2,2,2] ] (keep on reading!)

Now, the only thing left to do is to merge the powersets of the IDS string, with the frequencies of the IDS string itself.

Going back to our example where the two powersets of 艹蓮 are [ { '十': 1, '丨': 1, '辶': 1, '車': 1 }, { '十': 2, '辶': 1, '車': 1 } ], we need to merge each of the powersets with { '艹': 1, '連': 1 } using mergeTwoFreqs to get two possible powersets for 蓮, namely,

[
  { '艹': 1, '連': 1, '辶': 1, '車': 1, '十': 1, '丨': 1 },
  { '艹': 1, '連': 1, '辶': 1, '車': 1, '十': 2 },
]

We do the same for the right subtree, to get the blue and green powersets.

[
  { '辶': 1, '莗': 1, '艹': 1, '車': 1, '十': 1, '丨': 1 },
  { '辶': 1, '莗': 1, '艹': 1, '車': 1, '十': 2 }
]

Finally, we append the left and right subtree results immediately above into a list, and flatten it to obtain our expected result far above.

The termination condition, i.e. a leaf node is considered to be a leaf node when the IDS string(s) contain only itself, i.e. there are no constituent characters, and thus we return an empty list of powersets.

Once again, in pseudocode:

def mergeTwoFreqs(freqsA, freqsB) {
    res = { ...freqsB };
    for (let char of Object.keys(freqsA)) {
        if (!res[char]) res[char] = 0;
        res[char] += freqsA[char];
    }
    return res;
};

// given [ [ A0, A1, A2 ], [ B0, B1, B2 ], [ C0, C1, C2 ] ], return [ mergeFreqs((A0, B0, C0)), mergeFreqs((A0, B0, C1)), ... mergeFreqs((A2, B2, C2)) ]
// A0 .. C2 are *powersets*.
def permuteAndMerge(freqsArr) {
    // [ [0,0,0], [0,0,1], ... [2,2,2] ]
    permutations = permGen(freqsArr.map(getSize))

    res = []
    for permutation in permutations {
        mergedPowerset = {}
        // take one from each powerset
        elements = permutation.map((elem, i) => freqsArr[i][elem])
        for element in elements {
            mergedPowerset = mergeTwoFreqs(mergedPowerset, element)
        }
        res.append(mergedPowerset)
    }
    return res
}

def rec(char) {
    // each index on this array corresponds to a fork in the black arrow, i.e. a different decomposition in the picture above
    freqsAtThisNode = []

    // ['⿰木⿺辶莗', '⿰木蓮']
    idsStrings = getIDSStrings(char)

    // record the frequencies of each IDS string
    // itself in the array
    for i = 0; i < idsStrings.length; i++ {
        freqsAtThisNode.append({});
        for idsChar in idsStrings[i] {
            if idsChar !== char and isValidHanChar(char) {
                if (!freqsAtThisNode[i][idsChar]) {
                    freqsAtThisNode[i][idsChar] = 0;
                }
                freqsAtThisNode[i][idsChar] += 1;
            }
        }
    }

    res = []
    for i = 0; i < idsStrings.length; i++ {
        // eureka
        freqs = permuteAndMerge(Object.keys(freqsAtThisNode[i]).map((key) => rec(key)))
        // merge the powersets with the freqs of the corresponding IDS string
        freqs = freqs.map((freq) => mergeTwoFreqs(freq, freqsAtThisNode[i]))
        // flatten
        res = res.concat(freqs)
    }
    return res
}

N.B. if there are no valid characters in the IDS string, freqs in the 2nd for loop will be empty, and it will vanish during flattening.

[1] AFAIK, there is a lot of human subjectivity involved in the CHISE IDS dataset.

[2] "Powerset" here refers to a hash table, the keys of which are ALL the constituent radicals of a particular character as we traverse the tree, and the values of which are their frequencies. I refer to this concept as freqs in the code often.

Generated Datasets

These JSON datasets are generated by npm run etl using my preprocessed IDS data and output into the public/json folder for consumption by the frontend.

baseRadicals.json: string[], a list of radicals which cannot be further decomposed.

forwardMap.json: { [key: string]: string[]; }, a map of component to characters that it is present in.

variantsIslandsLookup.json: { islands: string[][]; chars: { [key: string]: number[] }; }, islands is a list of lists of related characters. chars is a map of individual character to the index in islands of the island containing itself.

variantsMap.json: { [key: string]: number[]; }, map of character to list of types that it is classified as (e.g. Shinjitai, Simplified Chinese)

readings.json: { [key: string]: { [key: string]: string; }; }, map of character to Unihan reading fields.

reverseMap.json (large! >150MB. Not actually used by the frontend directly.):

{
  [key: string]: {
    utf_code: string;
    ids_strings: {
      ids: string;
      locales: string;
    }[];
    charFreqs?: powerset[];
  };
}

ids_strings is a list of IDSs and the locales that this IDS decomposition corresponds to.

Refer above2 for what a "powerset" means here.

reverseMapCharFreqsOnly.json and reverseMapIDSOnly.json are optimized versions of reverseMap.json in order to fit within Apple's 50MB service worker cache limit for PWAs.

reverseMap.json is gzipped and served as reverseMap.json.gz in the deployed version.

Rules of thumb

The IDS sequences as provided by CHISE use Kangxi radicals and their actual CJK character counterparts interchangeably.

Decomposition which uses U+2E81 (Kangxi radical)

Decomposition which uses U+20086 (CJK character)

All Kangxi radicals are converted to their corresponding CJK characters as part of the ETL process.

All variants of a character, including transitive ones, should be retrivable from any of the characters involved. e.g. searching for 发 (SC) should return 發 (TC), 髮 (TC), 発 (JA), and searching for 鄕 (KR) should return 郷 (JA), 鄉 (TC), and 乡 (SC), along with other less-commonly used variants.

Licenses

SPDX-License-Identifier: GPL-2.0-or-later

  • The products of npm run etl (most of the files in the public/json folder) are licensed under the GNU General Public License v2.0 or later.

  • I have plans to dual-license Radically under the MIT license.

    • The main obstacles are the lack of a non-copyleft, small, 100% CJK Unicode coverage webfont, and most of the etl folder relying on said GPLv2 licensed data.
    • You are more than welcome to contact me via email, the chat links above, or opening an issue if you have a need to use (parts of) it, or the algorithm in non-copyleft settings.
    • If you wish to contribute to Radically, please keep this in consideration.