Core Concepts
HIRO proposes a method for unsupervised abstractive opinion summarization using hierarchical indexing to select prevalent opinions and generate coherent summaries.
Abstract
The content introduces HIRO, a method for unsupervised abstractive opinion summarization. It combines the attributability and scalability of extractive approaches with the coherence and fluency of Large Language Models (LLMs). HIRO learns an index structure that maps sentences to a path through a semantically organized discrete hierarchy. It populates the index at inference time to identify and retrieve clusters of sentences containing popular opinions from input reviews. The modular approach of HIRO allows for evaluation at each stage, showing that it generates more coherent, detailed, and accurate summaries preferred by annotators compared to prior work. The content discusses the method's three modules, the learning of hierarchical indexing, retrieving popular opinions, and generating coherent summaries. Extensive experiments on English datasets demonstrate the effectiveness of HIRO in generating summaries that better reflect the distribution of opinions within input reviews.
Stats
HIRO는 unsupervised abstractive opinion summarization을 위한 방법을 제안합니다.
HIRO는 문장을 계층적 구조로 매핑하는 인코딩 공간을 학습합니다.
HIRO는 인덱스를 채워서 인풋 리뷰에서 인기 있는 의견을 식별하고 검색합니다.
HIRO는 검색된 문장 클러스터를 LLM에 전달하여 읽기 쉬운 요약을 생성합니다.
Quotes
"HIRO generates more coherent, detailed, and accurate summaries that are significantly preferred by annotators compared to prior work."
"HIRO learns an encoding space that is more semantically structured than prior work."