Распределение обращений по теме

Функция обучения

function trainTextModel(samples) {
  if (!Array.isArray(samples) || samples.length < 4 || samples.length > 30) throw new Error('CLASSIFIER_COUNT');
  const classes = new Map();
  const vocabulary = new Set();
  for (const sample of samples) {
    if (!sample || typeof sample.label !== 'string' || !sample.label.trim() ||
        typeof sample.text !== 'string' || !sample.text.trim()) throw new Error('TRAINING_FORMAT');
    const label = sample.label.trim();
    const key = label.toLowerCase();
    const tokens = tokenizeText(sample.text);
    if (!tokens.length) throw new Error('TRAINING_WORDS');
    if (!classes.has(key)) classes.set(key, {label, documents: 0, tokens: 0, counts: new Map()});
    const target = classes.get(key);
    target.documents += 1;
    for (const token of tokens) {
      vocabulary.add(token);
      target.tokens += 1;
      target.counts.set(token, (target.counts.get(token) || 0) + 1);
    }
  }
  if (classes.size < 2) throw new Error('CLASSIFIER_LABELS');
  if (Array.from(classes.values()).some(target => target.documents < 2)) throw new Error('CLASSIFIER_EXAMPLES');
  return {kind: 'classifier', classes: Array.from(classes.values()), vocabulary, documents: samples.length};
}

Функция проверки запроса

function rankText(model, text) {
  if (!model || model.kind !== 'classifier') throw new Error('NOT_TRAINED');
  if (typeof text !== 'string' || text.length > 500) throw new Error('QUERY_LIMIT');
  const tokens = tokenizeText(text);
  const known = tokens.filter(token => model.vocabulary.has(token));
  const coverage = tokens.length ? known.length / tokens.length : 0;
  if (!known.length) return {review: true, reason: tokens.length ? 'UNKNOWN_WORDS' : 'QUERY_WORDS', label: null, scores: [], known: 0, total: tokens.length, coverage, gap: 0};

  const counts = new Map();
  known.forEach(token => counts.set(token, (counts.get(token) || 0) + 1));
  const logs = model.classes.map(target => {
    let value = Math.log(target.documents / model.documents);
    counts.forEach((count, token) => {
      value += count * Math.log(((target.counts.get(token) || 0) + 1) / (target.tokens + model.vocabulary.size));
    });
    return {label: target.label, value};
  });
  const maximum = Math.max(...logs.map(item => item.value));
  const total = logs.reduce((sum, item) => sum + Math.exp(item.value - maximum), 0);
  const scores = logs.map(item => ({label: item.label, score: Math.exp(item.value - maximum) / total}))
    .sort((left, right) => right.score - left.score);
  const gap = scores[0].score - scores[1].score;
  const reason = tokens.some(token=>token.startsWith('neg:')&&!model.vocabulary.has(token)) ? 'NEGATION_REVIEW' : coverage < 0.2 ? 'LOW_COVERAGE' : gap < 0.12 ? 'CLOSE_SCORES' : '';
  return {review: !!reason, reason, label: reason ? null : scores[0].label, scores, known: known.length, total: tokens.length, coverage, gap};
}

Показаны функции из этой демонстрации. Полная реализация также включает подготовку текста и интерфейс.