wordmap/dist/algorithms/NgramFrequency.js

Multi-Lingual Word Alignment Prediction

"use strict";
Object.defineProperty(exports, "__esModule", { value: true });
const GlobalAlgorithm_1 = require("../GlobalAlgorithm");
/**
 * This algorithm calculates the frequency of n-gram occurrences.
 */
class NgramFrequency extends GlobalAlgorithm_1.default {
    constructor() {
        super(...arguments);
        this.name = "n-gram frequency";
    }
    /**
     * Performs a numerical addition with the value of a key in a number object.
     * TODO: move this into it's own class?
     *
     * @param {NumberObject} object
     * @param {string} key
     * @param {number} value
     */
    static addObjectNumber(object, key, value) {
        if (!(key in object)) {
            object[key] = 0;
        }
        object[key] += value;
    }
    /**
     * Performs a numerical division.
     * Division by zero will result in 0.
     * TODO: move this into a math utility?
     *
     * @param {number} dividend
     * @param {number} divisor
     * @return {number}
     */
    static divideSafe(dividend, divisor) {
        if (divisor === 0) {
            return 0;
        }
        else {
            return dividend / divisor;
        }
    }
    /**
     * Load data into the predictions
     * @param  predictions [description]
     * @param  cIndex      [description]
     * @param  saIndex     [description]
     * @return             [description]
     */
    execute(predictions, cIndex, saIndex) {
        const alignmentFrequencyCorpusSums = {};
        const alignmentFrequencyAlignmentMemorySums = {};
        for (const p of predictions) {
            // alignment permutation frequency within the corpus/alignment memory
            const alignmentFrequencyCorpus = cIndex.permutations.alignmentFrequency.read(p.alignment);
            const alignmentFrequencyAlignmentMemory = saIndex.alignmentFrequency.read(p.alignment);
            // n-gram permutation frequency within the corpus/alignment memory
            // looked up by n-gram
            // TODO: rename to something like this.
            // const sourceNgramFrequencyInCorpusPermutations
            const ngramFrequencyCorpusSource = cIndex.permutations.sourceNgramFrequency.read(p.source);
            const ngramFrequencyAlignmentMemorySource = saIndex.sourceNgramFrequency.read(p.source);
            const ngramFrequencyCorpusTarget = cIndex.permutations.targetNgramFrequency.read(p.target);
            const ngramFrequencyAlignmentMemoryTarget = saIndex.targetNgramFrequency.read(p.target);
            // permutation frequency ratio
            const sourceCorpusPermutationsFrequencyRatio = NgramFrequency.divideSafe(alignmentFrequencyCorpus, ngramFrequencyCorpusSource);
            const targetCorpusPermutationsFrequencyRatio = NgramFrequency.divideSafe(alignmentFrequencyCorpus, ngramFrequencyCorpusTarget);
            const sourceAlignmentMemoryFrequencyRatio = NgramFrequency.divideSafe(alignmentFrequencyAlignmentMemory, ngramFrequencyAlignmentMemorySource);
            const targetAlignmentMemoryFrequencyRatio = NgramFrequency.divideSafe(alignmentFrequencyAlignmentMemory, ngramFrequencyAlignmentMemoryTarget);
            // store scores
            p.setScores({
                sourceCorpusPermutationsFrequencyRatio,
                targetCorpusPermutationsFrequencyRatio,
                sourceAlignmentMemoryFrequencyRatio,
                targetAlignmentMemoryFrequencyRatio
            });
            // sum alignment frequencies
            NgramFrequency.addObjectNumber(alignmentFrequencyCorpusSums, p.key, alignmentFrequencyCorpus);
            NgramFrequency.addObjectNumber(alignmentFrequencyAlignmentMemorySums, p.key, alignmentFrequencyAlignmentMemory);
        }
        // calculate filtered frequency ratios
        for (const p of predictions) {
            const alignmentFrequencyCorpus = cIndex.permutations.alignmentFrequency.read(p.alignment);
            const alignmentFrequencyAlignmentMemory = saIndex.alignmentFrequency.read(p.alignment);
            // TODO: instead of generating filters of alignmentFrequencyCorpus etc
            // we want to generate filtered ngramFrequencyCorpusSource and ngramFrequencyCorpusTarget
            // see notes in ngram_frequency line 160.
            // alignment frequency in the filtered corpus and alignment memory
            const alignmentFrequencyCorpusFiltered = alignmentFrequencyCorpusSums[p.key];
            const alignmentFrequencyAlignmentMemoryFiltered = alignmentFrequencyAlignmentMemorySums[p.key];
            // source and target frequency ratio for the corpus and alignment memory
            const frequencyRatioCorpusFiltered = NgramFrequency.divideSafe(alignmentFrequencyCorpus, alignmentFrequencyCorpusFiltered);
            const frequencyRatioAlignmentMemoryFiltered = NgramFrequency.divideSafe(alignmentFrequencyAlignmentMemory, alignmentFrequencyAlignmentMemoryFiltered);
            // store scores
            p.setScores({
                // alignmentFrequencyCorpusFiltered,
                // alignmentFrequencyAlignmentMemoryFiltered,
                // TODO: we aren't using these at the moment
                frequencyRatioCorpusFiltered,
                frequencyRatioAlignmentMemoryFiltered
            });
        }
        return predictions;
    }
}
exports.default = NgramFrequency;