inputx-scoring

Probability-native candidate scoring primitive for IME engines.

score(W | i) = log_prior(W) + log_likelihood(i | W)

This crate ships only the schema — the Candidate struct, the MatchType enum, the Source tag, and the additive sort key. It has no opinion on how producers derive log_prior or log_likelihood; that lives in the consuming engine (Inputx wubi / pinyin / nihongo cement crates, or any third-party IME that wants the same Bayesian shape).

Why log-space, why fixed-point

P(W|i) ∝ P(i|W) · P(W) is the standard Bayesian decomposition for candidate ranking. In log space, multiplication becomes addition — so each candidate's score is the sum of two independent factors, and producers (dict lookups, n-gram scorers, fuzzy matchers) can contribute to either term without coordinating a single global formula.

i32 Q4 fixed-point (Q4 = 16) gives ~0.0625 log-unit resolution over ±67 million log units — far more headroom than any realistic corpus needs. Integer arithmetic also keeps the path deterministic across platforms (no f64 rounding drift between mac / iOS / Linux).

Schema

pub const Q4: i32 = 16;

pub enum Source { Wubi = 0, Pinyin = 1, Japanese = 2 }

pub enum MatchType {
    Exact,
    Prefix(u16),                       // proximity_milli
    Fuzzy(u16),                        // edit_cost_milli
    Composed { bigram_links: u8 },     // chain_len - 1
}

pub struct Candidate<'a> {
    pub word: &'a str,
    pub log_prior: i32,
    pub log_likelihood: i32,
    pub match_type: MatchType,
    pub source: Source,
}

pub fn score(c: &Candidate<'_>) -> i32 { c.log_prior + c.log_likelihood }

Usage

use inputx_scoring::{Candidate, MatchType, Source, Q4, score};

let c = Candidate {
    word: "继续",
    log_prior: Q4 * 10,
    log_likelihood: Q4 * 5,
    match_type: MatchType::Exact,
    source: Source::Pinyin,
};
assert_eq!(score(&c), Q4 * 15);

no_std

Disable the std feature for #![no_std] builds (the schema itself is core only; the std feature gates optional formatting helpers).

inputx-scoring = { version = "1.4", default-features = false }

License

Dual-licensed under MIT or Apache-2.0.