ASC (TIL, 19 tasks) (Task Incremental Aspect Sentiment Classification)

一组 19 个 ASC 数据集（对 19 个产品的评论）产生了 19 个任务的序列。每个数据集代表一个任务。数据集来自 4 个来源： (1) HL5Domains (Hu and Liu, 2004)，包含 5 个产品的评论； (2) Liu3Domains (Liu et al., 2015) 对 3 个产品进行了评论； (3) Ding9Domains (Ding et al., 2008) 有 9 个产品的评论； (4) SemEval14 对 2 个产品进行评论 - SemEval 2014 任务 4，适用于笔记本电脑和餐厅。对于（1）、（2）和（3），我们分割了大约 10% 的原始数据作为验证数据，另外大约 10% 的原始数据作为测试数据。对于 (4)，我们使用训练集中的 150 个示例进行验证。为了与现有研究保持一致（Tang et al., 2016），没有使用属于冲突极性的示例（关于一个方面术语表达正面和负面情绪）。 19 个数据集的统计数据和详细信息在页面 https://github.com/ZixuanKe/PyContinual 上给出。

A collection of 19 ASC datasets (reviews for 19 products) forms a sequence of 19 tasks, where each dataset corresponds to one individual task. All datasets are sourced from four origins: (1) HL5Domains (Hu and Liu, 2004), which contains reviews for 5 products; (2) Liu3Domains (Liu et al., 2015), with reviews for 3 products; (3) Ding9Domains (Ding et al., 2008), covering reviews for 9 products; (4) SemEval14, i.e., SemEval 2014 Task 4, which includes reviews for 2 products (laptop and restaurant domains). For datasets (1), (2) and (3), we split approximately 10% of the original data as the validation set, and another ~10% as the test set. For (4), we use 150 examples from the training set as validation data. In line with prior research (Tang et al., 2016), examples with conflicting polarities (expressing both positive and negative sentiments towards a single aspect term) are excluded. Statistics and detailed information of all 19 datasets are provided at https://github.com/ZixuanKe/PyContinual.

搜集汇总

背景与挑战

背景概述

该数据集包含19个任务，每个任务对应一个产品的评论数据，用于任务增量学习的情感分析研究。数据来源于四个不同渠道，其详细统计信息可在GitHub页面获取。