Code Monkey home page Code Monkey logo

level2_nlp_mrc-nlp-06's Introduction

πŸ” Open-Domain Question Answering

πŸ™ŒπŸ» 1. ν”„λ‘œμ νŠΈ κ°œμš”

1.1. κ°œμš”

  • Question Answering(QA)은 λ‹€μ–‘ν•œ μ’…λ₯˜μ˜ μ§ˆλ¬Έμ— λŒ€ν•΄ λŒ€λ‹΅ν•˜λŠ” 인곡지λŠ₯을 λ§Œλ“œλŠ” 연ꡬ 뢄야이닀. λ‹€μ–‘ν•œ QA μ‹œμŠ€ν…œ 쀑, Open-Domain Question Answering(ODQA)은 μ£Όμ–΄μ§€λŠ” 지문이 λ”°λ‘œ μ‘΄μž¬ν•˜μ§€ μ•Šκ³  사전에 κ΅¬μΆ•λ˜μ–΄μžˆλŠ” Knowledge resource μ—μ„œ μ§ˆλ¬Έμ— λŒ€λ‹΅ν•  수 μžˆλŠ” λ¬Έμ„œλ₯Ό μ°ΎλŠ” 과정이 μΆ”κ°€λœλ‹€.
  • λ³Έ ODQA λŒ€νšŒμ—μ„œ μš°λ¦¬κ°€ λ§Œλ“€ λͺ¨λΈμ€ two-stage둜 κ΅¬μ„±λ˜μ–΄ μžˆλ‹€. 첫 λ‹¨κ³„λŠ” μ§ˆλ¬Έμ— κ΄€λ ¨λœ λ¬Έμ„œλ₯Ό μ°Ύμ•„μ£ΌλŠ” "retriever" 단계이고, λ‹€μŒμœΌλ‘œλŠ” κ΄€λ ¨λœ λ¬Έμ„œλ₯Ό 읽고 μ μ ˆν•œ 닡변을 μ°Ύκ±°λ‚˜ λ§Œλ“€μ–΄μ£ΌλŠ” "reader" 단계이닀. 두 가지 단계λ₯Ό 각각 κ΅¬μ„±ν•˜κ³  그것듀을 적절히 ν†΅ν•©ν•˜κ²Œ 되면, μ–΄λ €μš΄ μ§ˆλ¬Έμ„ λ˜μ Έλ„ 닡변을 ν•΄μ£ΌλŠ” ODQA μ‹œμŠ€ν…œμ„ λ§Œλ“œλŠ”κ²ƒμ΄ 이 ν”„λ‘œμ νŠΈμ˜ λͺ©ν‘œμ΄λ‹€.

1.2. 평가 μ§€ν‘œ

두 가지 ν‰κ°€μ§€ν‘œκ°€ μžˆλ‹€. EM κΈ°μ€€μœΌλ‘œ λ¦¬λ”λ³΄λ“œ λ“±μˆ˜κ°€ 반영되고, F1은 참고용으둜만 ν™œμš©λœλ‹€.

  • Exact Match(EM)
    • λͺ¨λΈμ˜ 예츑과, μ‹€μ œ 닡이 μ •ν™•ν•˜κ²Œ μΌμΉ˜ν•  λ•Œλ§Œ μ μˆ˜κ°€ 주어진닀. 즉 λͺ¨λ“  μ§ˆλ¬Έμ€ 0점 μ•„λ‹ˆλ©΄ 1점으둜 μ²˜λ¦¬λœλ‹€.
  • F1 Score
    • EMκ³Ό λ‹€λ₯΄κ²Œ λΆ€λΆ„ 점수λ₯Ό μ œκ³΅ν•œλ‹€. 예λ₯Ό λ“€μ–΄, 정닡은 "Barack Obama"μ§€λ§Œ 예츑이 "Obama"일 λ•Œ, EM의 경우 0점을 λ°›κ² μ§€λ§Œ F1 ScoreλŠ” κ²ΉμΉ˜λŠ” 단어도 μžˆλŠ” 것을 κ³ λ €ν•΄ λΆ€λΆ„ 점수λ₯Ό 받을 수 μžˆλ‹€.

πŸ‘¨β€πŸ‘©β€πŸ‘§β€πŸ‘¦ 2. ν”„λ‘œμ νŠΈ νŒ€ ꡬ성 및 μ—­ν• 

2.1. νŒ€ ꡬ성

κΉ€λ―Όν˜Έ 김성은 κΉ€μ§€ν˜„ μ„œκ°€μ€ ν™μ˜ν›ˆ

2.2. μ—­ν• 

  • κΉ€λ―Όν˜Έ : ν”„λ‘œμ νŠΈ λ¦¬νŒ©ν† λ§, 핡심 λ¬Έμž₯ κ°•μ‘°, Curriculum learning
  • 김성은 : BM25 retriever, Elasticsearch, 데이터 μ „μ²˜λ¦¬
  • κΉ€μ§€ν˜„ : Custom model for question answering, Self-distillation, Ensemble
  • μ„œκ°€μ€ : λͺ¨λΈ & topk λ³€κ²½ μ‹€ν—˜, μΆ”κ°€ 데이터 fine tuning, balanced sampling
  • ν™μ˜ν›ˆ : TF-IDF retriever, DPR κ΅¬ν˜„, post-processing

πŸ’½ 3. Data Analysis

Data

  • Train data : 3952
    • Feature: 'title', 'context', 'question', 'id', 'answers', 'document_id'
  • Trainκ³Ό validation λͺ¨λ‘ μœ μ‚¬ν•œ λΆ„ν¬μ˜ context 길이λ₯Ό κ°–λŠ”λ‹€.

πŸ“‚ 4. ν”„λ‘œμ νŠΈ μˆ˜ν–‰

4.1. Data Preprocessing

  • Wiki data μ „μ²˜λ¦¬

4.2. Retriever

  • Top-k λ³€κ²½
  • TF-IDF
  • 자체 ν† ν¬λ‚˜μ΄μ§•
  • BM25
  • ElasticSearch
  • DPR(Dense Passage Retrieval)

4.3 Reader

  • klue/roberta-large μ‚¬μš©
  • 핡심 λ¬Έμž₯ κ°•μ‘°
  • Custom model
    • LSTM
    • Bi-LSTM
    • SDS-CNN & MLP
  • μΆ”κ°€ 데이터셋을 ν™œμš©ν•œ μ „μ΄ν•™μŠ΅
    • KorQuAD 1.0
    • KorQuAD 2.0
    • 기계독해 데이터셋
    • λ‰΄μŠ€κΈ°μ‚¬ 기계독해 데이터셋
    • μΌλ°˜μƒμ‹ 데이터셋
  • Balanced sampling
  • Curriculum learning

4.4 Ensemble

  • Hard voting
  • Soft voting
  • K-fold

4.5. 기타

  • Self-distillation
  • Post-processing

πŸ‘€ Wrap-up Report

[https://docs.google.com/document/d/1ukhlsVAO_0NFwtCzAOifblNhNOXxd0g7n4raSZT-bck/edit?usp=sharing]

level2_nlp_mrc-nlp-06's People

Contributors

jihyeeon avatar i-mprovising avatar grapediget avatar gaeun0112 avatar mostlyfor avatar

Recommend Projects

  • React photo React

    A declarative, efficient, and flexible JavaScript library for building user interfaces.

  • Vue.js photo Vue.js

    πŸ–– Vue.js is a progressive, incrementally-adoptable JavaScript framework for building UI on the web.

  • Typescript photo Typescript

    TypeScript is a superset of JavaScript that compiles to clean JavaScript output.

  • TensorFlow photo TensorFlow

    An Open Source Machine Learning Framework for Everyone

  • Django photo Django

    The Web framework for perfectionists with deadlines.

  • D3 photo D3

    Bring data to life with SVG, Canvas and HTML. πŸ“ŠπŸ“ˆπŸŽ‰

Recommend Topics

  • javascript

    JavaScript (JS) is a lightweight interpreted programming language with first-class functions.

  • web

    Some thing interesting about web. New door for the world.

  • server

    A server is a program made to process requests and deliver data to clients.

  • Machine learning

    Machine learning is a way of modeling and interpreting data that allows a piece of software to respond intelligently.

  • Game

    Some thing interesting about game, make everyone happy.

Recommend Org

  • Facebook photo Facebook

    We are working to build community through open source technology. NB: members must have two-factor auth.

  • Microsoft photo Microsoft

    Open source projects and samples from Microsoft.

  • Google photo Google

    Google ❀️ Open Source for everyone.

  • D3 photo D3

    Data-Driven Documents codes.