| 000 | 00000nam c2200205 c 4500 | |
| 001 | 000046048349 | |
| 005 | 20230530104414 | |
| 007 | ta | |
| 008 | 200629s2020 ulkd bmAC 000c eng | |
| 040 | ▼a 211009 ▼c 211009 ▼d 211009 | |
| 085 | 0 | ▼a 0510 ▼2 KDCP |
| 090 | ▼a 0510 ▼b 6D36 ▼c 1117 | |
| 100 | 1 | ▼a 고미영, ▼g 高美煐 |
| 245 | 1 0 | ▼a Analyzing position bias in question answering system / ▼d Miyoung Ko |
| 260 | ▼a Seoul : ▼b Graduate School, Korea University, ▼c 2020 | |
| 300 | ▼a iv, 22장 : ▼b 도표 ; ▼c 26 cm | |
| 500 | ▼a 지도교수: 강재우 | |
| 502 | 0 | ▼a 학위논문(석사)-- ▼b 고려대학교 대학원, ▼c 컴퓨터·전파통신공학과, ▼d 2020. 8 |
| 504 | ▼a 참고문헌: 장 18-22 | |
| 530 | ▼a PDF 파일로도 이용가능; ▼c Requires PDF file reader(application/pdf) | |
| 653 | ▼a Question Answering | |
| 776 | 0 | ▼t Analyzing Position Bias in Question Answering System ▼w (DCOLL211009)000000232146 |
| 900 | 1 0 | ▼a Ko, Mi-young, ▼e 저 |
| 900 | 1 0 | ▼a 강재우, ▼g 姜在雨, ▼d 1969-, ▼e 지도교수 ▼0 AUTH(211009)151698 |
| 945 | ▼a KLPA |
소장정보
| No. | 소장처 | 청구기호 | 등록번호 | 도서상태 | 반납예정일 | 예약 | 서비스 |
|---|---|---|---|---|---|---|---|
| No. 1 | 소장처 과학도서관/학위논문서고/ | 청구기호 0510 6D36 1117 | 등록번호 123064864 | 도서상태 대출가능 | 반납예정일 | 예약 | 서비스 |
| No. 2 | 소장처 과학도서관/학위논문서고/ | 청구기호 0510 6D36 1117 | 등록번호 123064865 | 도서상태 대출가능 | 반납예정일 | 예약 | 서비스 |
컨텐츠정보
초록
Many extractive question answering models are trained to predict start and end positions of answers.
The choice of predicting answers as positions is mainly due to its simplicity and effectiveness.
In this study, we hypothesize that when the distribution of the answer positions is highly skewed in the training set (e.g., answers lie only in the $k$-th sentence of each passage), QA models predicting answers as positions learn spurious positional cues and fail to give answers in different positions.
We first illustrate this \textit{position bias} in popular extractive QA models such as BiDAF and BERT and thoroughly examine how position bias propagates through each layer of BERT.
To safely deliver position information without position bias, we train models with various de-biasing methods including entropy regularization and randomized position.
We found that reducing correlation between word positions and answers helps us to resolve position bias.
목차
Abstract Contents i List of Figures iii List of Tables iv 1 Introduction 1 2 Related Work 4 3 Analyzing Position Bias 6 3.1 Position Bias on Synthetic Datasets . . . . . . . . . . . . . . . . . . . . .6 3.2 Visualization of Position Bias . . . . . . . . . . . . . . . . . . . . . . . . .7 3.3 Generalizing to Different Positions . . . . . . . . . . . . . . . . . . . . . .11 4 De-biasing Position Bias12 4.1 Method . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .12 4.1.1Entropy Regularization . . . . . . . . . . . . . . . . . . . . . . . .12 4.1.2Randomized Position . . . . . . . . . . . . . . . . . . . . . . . . . .13 4.2 Experiments . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .13 4.2.1Implementation Details . . . . . . . . . . . . . . . . . . . . . . . .13 4.2.2Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .13 5 Discussion 15 6 Conclusion 17 Bibliography 18
