일 | 월 | 화 | 수 | 목 | 금 | 토 |
---|---|---|---|---|---|---|
1 | ||||||
2 | 3 | 4 | 5 | 6 | 7 | 8 |
9 | 10 | 11 | 12 | 13 | 14 | 15 |
16 | 17 | 18 | 19 | 20 | 21 | 22 |
23 | 24 | 25 | 26 | 27 | 28 | 29 |
30 | 31 |
- 백준
- 완전탐색
- DeepLearning
- GitHub Action
- FDS
- leetcode
- autoencoder
- Matplotlib
- 네이버AItech
- NLP
- GIT
- datascience
- Kaggle
- vscode
- github
- 코딩테스트
- rnn
- 프로그래머스
- FastAPI
- Kubernetes
- docker
- NaverAItech
- wandb
- GCP
- PytorchLightning
- 알고리즘
- pep8
- torchserve
- python
- pytorch
- Today
- Total
목록NLP (11)
Sangmun

https://arxiv.org/abs/1912.13318 LayoutLM: Pre-training of Text and Layout for Document Image Understanding Pre-training techniques have been verified successfully in a variety of NLP tasks in recent years. Despite the widespread use of pre-training models for NLP applications, they almost exclusively focus on text-level manipulation, while neglecting layout and arxiv.org Introduction Document A..

https://arxiv.org/abs/2103.10213 ICDAR2019 Competition on Scanned Receipt OCR and Information Extraction Scanned receipts OCR and key information extraction (SROIE) represent the processeses of recognizing text from scanned receipts and extracting key texts from them and save the extracted tests to structured documents. SROIE plays critical roles for many doc arxiv.org Introduction SROIE는 Scanne..

https://arxiv.org/abs/2102.01373 An Improved Baseline for Sentence-level Relation Extraction Sentence-level relation extraction (RE) aims at identifying the relationship between two entities in a sentence. Many efforts have been devoted to this problem, while the best performing methods are still far from perfect. In this paper, we revisit two pro arxiv.org 1. Introduction 본 논문은 RE task에 대한 논문이며..

1. Introduction 그동안 많은 self-supervised 방법론들이 제시가 되었고 NLP task에서 많은 성과를 얻었다. 그리고 특히 성과가 좋았던 것은 Cloze Tasks에서 영감을 받아 제안되었던 Masked Laguage model이었다. 하지만 이러한 방법론들은 특정 태스크에만 집중을 하여 범용성이 떨어졌다고 본 논문은 주장을 하고 있다. 따라서 본 논문은 bert의 encoder와 gpt의 decoder 구조를 합한 bart를 제안하였으며 pre-train 과정은 원본 데이터를 bert와 유사하게 mask를 씌워서 encoder에 input으로 넣어주면 decoder가 그것을 복원하는 것이다. bert와의 mask 기법의 차이점은 bert는 하나의 토큰당 하나의 mask가 생기지..

https://arxiv.org/abs/1901.11196 EDA: Easy Data Augmentation Techniques for Boosting Performance on Text Classification Tasks We present EDA: easy data augmentation techniques for boosting performance on text classification tasks. EDA consists of four simple but powerful operations: synonym replacement, random insertion, random swap, and random deletion. On five text classificati arxiv.org 1.Int..
https://arxiv.org/pdf/1508.07909.pdf https://huggingface.co/docs/transformers/tokenizer_summary#subword-tokenization Summary of the tokenizers Reinforcement learning models huggingface.co 개요 단어 단위로 vocab을 구성을 하면 임베딩을 사용할 때 임베딩의 매개변수가 많이 커지게 되는 단점이 있다. 이는 RNN층의 매개변수의 수보다 압도적으로 많게 됩니다. 이런 매개변수 비중의 비대칭성을 해결하기 위해 문자 단위 토큰화가 주목을 받았으나 너무 길어진 sequence와 성능 저하의 문제점이 발생하게 됩니다. 따라서 서브워드 단위로 토큰화를 진행하게 되었는데 ..

http://colah.github.io/posts/2015-08-Understanding-LSTMs/

https://web.stanford.edu/class/cs224n/slides/cs224n-2019-lecture08-nmt.pdf 보통 NLG 같은 task에서는 자연어를 생성할 때 매 타입스텝마다 decoder를 거쳐서 나온 단어가 다음 타임스텝의 input으로 들어가게 된다. 그리고 decoder에서는 매 타임스텝마다 가장 확률이 높은 단어를 선택해서 출력을 하게 되는데 이것을 Greedy decoding이라고 한다. 하지만 Greedy decoding의 문제점은 매 타임스텝의 최대 확률 값만을 고려하는 것이 전체 타임스텝으로 보면 적절하지 않은 문장을 출력으로 가지게 될 수 있다는 점이다. 그렇다면 greedy decoding에서 하는 방식인 현재 타입스텝만의 확률을 고려하는 방식이 아닌 전체 ..