GithubHelp home page GithubHelp logo

bannng / 2021-dialogue-summary-competition Goto Github PK

View Code? Open in Web Editor NEW

This project forked from cosmoquester/2021-dialogue-summary-competition

0.0 0.0 0.0 304 KB

[2021 훈민정음 한국어 음성•자연어 인공지능 경진대회] 대화요약 부문 알라꿍달라꿍 팀의 대화요약 학습 및 추론 코드를 공유하기 위한 레포입니다.

Python 98.87% Dockerfile 1.13%

2021-dialogue-summary-competition's Introduction

2021 Dialogue Summary Competition

Code style: black Imports: isort cosmoquester codecov

[2021 훈민정음 한국어 음성•자연어 인공지능 경진대회] 대화요약 부문 알라꿍달라꿍 팀의 대화요약 학습 및 추론 코드를 공유하기 위한 레포입니다.

팀원: 박상준(cosmoquester), 최기원(ckw1140), 오혜린(Hyerin-oh)

저희 팀은 이렇게 세 명으로 구성되어 있습니다.

최종 리더보드는 아래와 같고, 대화요약 부문에서 1등을 해내 네이버 대표상을 수상하였습니다.

Rank Name Score
1 알라꿍달라꿍 0.34286246
2 ... 0.3330445
3 ... 0.33254071

Directory Structure

# run 디렉토리에는 실행할 수 있는 스크립트들이 들어있습니다.
run
├── inference.py
├── interactive.py
├── train.py
└── train_tokenizer.py

# summarizer 디렉토리에는 데이터로더와 학습로직이 들어있습니다.
summarizer
├── data.py
├── method
│   ├── default.py
│   ├── r3f.py
│   ├── rdrop.py
│   └── rl.py
├── scheduler.py
└── utils.py

Process

최종제출에 사용했던 대화요약 모델을 학습하는 절차는 다음과 같습니다.

1. Tokenizer training

저희 팀은 토크나이저를 대회에서 제공한 학습데이터로 직접 학습해 사용하였습니다. 토크나이저의 학습은 학습데이터의 발화와 요약문을 mecab으로 1차분절한 후에 unigram 방법으로 학습해 사용하였습니다.

2. BART pretraining

저희는 모델 아키텍쳐로는 BART를 사용했습니다. 외부 데이터를 사용할 수 없기 때문에 학습 데이터셋의 대화 텍스트로 사전학습을 수행했으며 노이즈는 Text infilling과 Sentence Permutation을 적용했습니다. Sentence Permutation은 턴 단위로 수행하였습니다.

3. Dialogue Summarization finetune (R3F)

사전학습 후에는 Dialogue Summarization task를 학습시켰습니다. 이때 Abstract Summarization에서 좋은 효과를 보이는 R3F 기법을 적용하였습니다.

4. Dialogue Summarization finetune (RL)

마지막으로 학습의 목표를 대회의 평가지표인 ROUGE-L F1 score와 align시키기 위해서 고전적인 RL을 적용했습니다. target metric은 모델이 생성한 요약문과 실제 요약문간의 mecab분절 기준 ROUGE-L F1 score를 사용하였습니다.

References

2021-dialogue-summary-competition's People

Contributors

cosmoquester avatar

Recommend Projects

  • React photo React

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

  • Vue.js photo Vue.js

    🖖 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. 📊📈🎉

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.