최은성 석박통합과정, 이선경 박사과정, 최민진 석박통합과정, 박준 학사과정 EMNLP 2024 Findings 국제 학술대회 논문 채택

21 Sep 2024

DIAL 연구실 소속 최은성(석·박통합과정, 제1저자) 학생, 이선경(박사과정, 제2저자) 학생, 최민진(석·박통합과정, 제3저자) 학생, 박준(학사과정, 제4저자) 학생, 이종욱(교신저자) 교수가 참여한 “From Reading to Compressing: Exploring the Multi-document Reader for Prompt Compression” 논문이 자연어처리 분야 최우수 국제 학술대회인 Findings of the 2024 Conference on Empirical Methods in Natural Language Processing (EMNLP Findings)에 최종 게재가 승인되었으며 오는 11월에 발표될 예정입니다.

Abstract

Large language models (LLMs) have achieved significant performance gains using advanced prompting techniques over various tasks. However, the increasing length of prompts leads to high computational costs and obscures crucial information. Prompt compression methods have been proposed to alleviate these issues, but they encounter challenges in (i) capturing the global context and (ii) training the compressor effectively. To tackle these challenges, we introduce a novel prompt compression method, namely Reading To Compression (R2C), utilizing the Fusion-in-Decoder (FiD) architecture to discern the important information in the prompt. Specifically, the cross-attention scores of the FiD trained to answer the question are used to identify essential chunks and sentences from the prompt. R2C effectively captures the global context without compromising semantic consistency, while avoiding the necessity of pseudo-labels for training the compressor. Experimental results show that R2C retains key information, boosting the LLM performance by 10.5% in out-of-domain evaluations compared to the baselines.