ABSA-Retail-Corpus
收藏资源简介:
ABSA-Retail-Corpus是一个包含10,814条多语言顾客评论的零售数据集,覆盖实体零售商店,并标注了八个方面类别及其情感。该数据集通过手动标注,从Google Maps中抓取数据,涵盖了不同国家和语言,为实体零售商店的顾客评论提供了深入的见解。数据集的创建旨在解决传统情感分析无法提供特定方面情感的问题,为实体零售商店的顾客评论提供了深入的见解。
ABSA-Retail-Corpus is a retail dataset containing 10,814 multilingual customer reviews covering physical retail stores, with annotations for eight aspect categories and their corresponding sentiments. Scraped from Google Maps and manually annotated, the dataset spans multiple countries and languages, delivering in-depth insights into customer reviews of physical retail stores. It was created to address the gap where traditional sentiment analysis cannot provide aspect-specific sentiment, further offering valuable in-depth insights into such customer reviews.
数据集概述
基本信息
- 数据集名称:A Retail-Corpus for Aspect-Based Sentiment Analysis with Large Language Models
- 数据规模:10,814条标注评论
- 标注类型:方面级情感分析(Aspect-Based Sentiment Analysis, ABSA)
- 语言:多语言(45种不同语言)
- 来源:实体零售店的客户评论
数据详情
- 评论数量:10,814条
- 多方面评论比例:44.7%(4,838条评论包含多于一个方面)
- 平均方面数:每条评论平均包含1.6个方面
- 评论长度:平均121个字符,范围从1个字符(主要为表情符号)到3,735个字符
- 覆盖国家:德国、法国、荷兰、意大利、西班牙、奥地利、比利时、葡萄牙、瑞士(按出现频率降序排列)
标注信息
- 方面类别:8个方面类别
- 情感标签:每个方面的情感极性
评估基准
- 评估模型:GPT-4和LLaMA-3-70B
- 准确率:两个模型均超过85%
- 性能比较:GPT-4在所有相关指标上整体优于LLaMA-3
代码实现
- GPT-4实现:使用Azure OpenAI的GPT-4模型进行ABSA,包含环境变量设置、令牌成本计算和少样本学习
- LLaMA实现:使用LLaMA-3-70B模型进行ABSA,包含详细文档、逐步解释和少样本学习
引用信息
bibtex @inproceedings{silcenco2025retail-corpus, title = {A Retail-Corpus for Aspect-Based Sentiment Analysis with Large Language Models}, author = {Şilcenco, Oleg and Machado, Marcos R. and Ugulino, Wallace C. and Braun, Daniel}, year = {2025}, booktitle = {Proceedings of the 8th International Conference on Natural Language and Speech Processing (ICNLSP 2025)} }




