fune/toolkit/components/translation/LanguageDetector.sys.mjs

/* This Source Code Form is subject to the terms of the Mozilla Public
 * License, v. 2.0. If a copy of the MPL was not distributed with this file,
 * You can obtain one at http://mozilla.org/MPL/2.0/. */

// workerManager is exported for tests.
import { clearTimeout, setTimeout } from "resource://gre/modules/Timer.sys.mjs";

const WORKER_URL = "resource://gre/modules/translation/cld-worker.js";

/**
 * The length of the substring to pull from the document's text for language
 * identification.
 *
 * This value should ideally be one that is large enough to yield a confident
 * identification result without being too large or expensive to extract.
 *
 * At this time, this value is not driven by statistical data or analysis.
 *
 * For the moment, while we investigate which language identification library
 * we would like to use, keep this logic in sync with language-id-engine.sys.mjs
 */
const DOC_TEXT_TO_IDENTIFY_LENGTH = 1024;

export var workerManager = {
  // Since Emscripten can handle heap growth, but not heap shrinkage, we
  // need to refresh the worker after we've processed a particularly large
  // string in order to prevent unnecessary resident memory growth.
  //
  // These values define the cut-off string length and the idle timeout
  // (in milliseconds) before destroying a worker. Once a string of the
  // maximum size has been processed, the worker is marked for
  // destruction, and is terminated as soon as it has been idle for the
  // given timeout.
  //
  // 1.5MB. This is the approximate string length that forces heap growth
  // for a 2MB heap.
  LARGE_STRING: 1.5 * 1024 * 1024,
  IDLE_TIMEOUT: 10 * 1000,

  detectionQueue: [],

  detectLanguage(aParams) {
    return this.workerReady
      .then(worker => {
        return new Promise(resolve => {
          this.detectionQueue.push({ resolve });
          worker.postMessage(aParams);
        });
      })
      .then(result => {
        // We have our asynchronous result from the worker.
        //
        // Determine if our input was large enough to trigger heap growth,
        // or if we're already waiting to destroy the worker when it's
        // idle. If so, schedule termination after the idle timeout.
        if (
          aParams.text.length >= this.LARGE_STRING ||
          this._idleTimeout != null
        ) {
          this.flushWorker();
        }

        return result;
      });
  },

  _worker: null,
  _workerReadyPromise: null,

  get workerReady() {
    if (!this._workerReadyPromise) {
      this._workerReadyPromise = new Promise(resolve => {
        let worker = new Worker(WORKER_URL);
        worker.onmessage = aMsg => {
          if (aMsg.data == "ready") {
            resolve(worker);
          } else {
            this.detectionQueue.shift().resolve(aMsg.data);
          }
        };
        this._worker = worker;
      });
    }

    return this._workerReadyPromise;
  },

  // Holds the ID of the current pending idle cleanup setTimeout.
  _idleTimeout: null,

  // Schedule the current worker to be terminated after the idle timeout.
  flushWorker() {
    if (this._idleTimeout != null) {
      clearTimeout(this._idleTimeout);
    }

    this._idleTimeout = setTimeout(
      this._flushWorker.bind(this),
      this.IDLE_TIMEOUT
    );
  },

  // Immediately terminate the worker, as long as there no pending
  // results. Otherwise, reschedule termination until after the next
  // idle timeout.
  _flushWorker() {
    if (this.detectionQueue.length) {
      this.flushWorker();
    } else {
      if (this._worker) {
        this._worker.terminate();
      }

      this._worker = null;
      this._workerReadyPromise = null;
      this._idleTimeout = null;
    }
  },
};

export var LanguageDetector = {
  /**
   * Detect the language of a given string.
   *
   * The argument may be either a string containing the text to analyze,
   * or an object with the following properties:
   *
   *  - 'text' The text to analyze.
   *
   *  - 'isHTML' (optional) A boolean, indicating whether the text
   *      should be analyzed as HTML rather than plain text.
   *
   *  - 'language' (optional) A string indicating the expected language.
   *      For text extracted from HTTP documents, this is expected to
   *      come from the Content-Language header.
   *
   *  - 'tld' (optional) A string indicating the top-level domain of the
   *      document the text was extracted from.
   *
   *  - 'encoding' (optional) A string describing the encoding of the
   *      document the string was extracted from. Note that, regardless
   *      of the value of this property, the 'text' property must be a
   *      UTF-16 JavaScript string.
   *
   * @returns {Promise<Object>}
   * @resolves When detection is finished, with a object containing
   * these fields:
   *  - 'language' (string with a language code)
   *  - 'confident' (boolean) Whether the detector is confident of the
   *      result.
   *  - 'languages' (array) An array of up to three elements, containing
   *      the most prevalent languages detected. It contains a
   *      'languageCode' property, containing the ISO language code of
   *      the language, and a 'percent' property, describing the
   *      approximate percentage of the input which is in that language.
   *      For text of an unknown language, the result may contain an
   *      entry with the languge code 'un', indicating the percent of
   *      the text which is unknown.
   */
  detectLanguage(aParams) {
    if (typeof aParams == "string") {
      aParams = { text: aParams };
    }

    return workerManager.detectLanguage(aParams);
  },

  /**
   * Attempts to determine the language in which the document's content is written.
   *
   * For the moment, while we investigate which language identification library
   * we would like to use, keep this logic in sync with language-id-engine.sys.mjs
   * @returns {string | null}
   */
  async detectLanguageFromDocument(aDocument) {
    // Grab a selection of text.
    let encoder = Cu.createDocumentEncoder("text/plain");
    encoder.init(aDocument, "text/plain", encoder.SkipInvisibleContent);
    let text = encoder
      .encodeToStringWithMaxLength(DOC_TEXT_TO_IDENTIFY_LENGTH)
      .replaceAll("\r", "")
      .replaceAll("\n", " ");

    const { language, confident } = await workerManager.detectLanguage({
      text,
    });

    workerManager.flushWorker();

    return confident ? language : null;
  },
};