Feynman Technique နှင့် Learning Notes

// Toy heuristic, not real NLP: flags jargon used without a nearby plain-language cue.
function findExplanationGaps(explanation, jargonTerms, windowChars = 60) {
  const lower = explanation.toLowerCase();
  const plainCueMarkers = ["means", "in other words", "i.e.", "that is", "which is", "is a ", "is the ", "(", "or simply"];
  const gaps = [];

  for (const term of jargonTerms) {
    const termLower = term.toLowerCase();
    let idx = lower.indexOf(termLower);
    if (idx === -1) continue;

    const windowStart = Math.max(0, idx - windowChars);
    const windowEnd = Math.min(lower.length, idx + term.length + windowChars);
    const nearby = lower.slice(windowStart, windowEnd);

    const hasCue = plainCueMarkers.some((marker) => nearby.includes(marker));
    if (!hasCue) gaps.push(term);
  }

  return gaps;
}

const jargon = ["embedding", "tokenization", "gradient descent", "overfitting"];

const explanationA =
  "An embedding is a list of numbers that represents meaning, so similar " +
  "words end up with similar numbers. Tokenization is the step that splits " +
  "text into small pieces the model can read. Gradient descent is the process " +
  "that nudges the model's numbers downhill toward fewer mistakes.";

const explanationB =
  "The model uses embedding and tokenization before gradient descent " +
  "kicks in, and overfitting can happen if you are not careful.";

console.log("Explanation A gaps:", findExplanationGaps(explanationA, jargon));
console.log("Explanation B gaps:", findExplanationGaps(explanationB, jargon));