Classic QA Models

• IR-absed approach
  1. Question Processing
  2. Passage Retrieval
  3. Answer Processing
• Knowledge-based

1️⃣ Question Processing

• Answer Type Detection
  • 정답이 어떤 Type이지? (person, place?)
• Query Formation
  • Query Keyword 선택
• Question Type Classification
• Focus Detection
• Reltation Extraction

<aside> ⭐ What are the two states you chould be reentering if you’re crossing Florida’s northern border?

• Question Type : What (Factoid)

• Answer Type: Location-State

• Query: two states, border, Florida

• Focus: the two states

• Relations: boarders </aside>

• Answer Type Taxonomy

  • Ambiguity가 있음
    • 박쥐는 뭐를 먹나요?
      • 음식? 식물? 동물?

2️⃣ Passage Retrieval

1. IR engine이 query terms를 바탕으로 document를 찾음
2. Document를 분리
3. Passage ranking

Machine Reading Comprehension

• MRC ?
  • 기계 ← 문서
  • 기계 → 질문에 대답
• 모델의 구조
  • Encoding Layer
    • 벡터표현
  • Co-attention layer
    • 상호 Attention으로 관계파악
  • Output Layer
    • 정답 단어의 시작과 끝 출력