上述数据集为ABSA(Aspect-Based Sentiment Analysis)领域数据集,基本形式为从句子中抽取:方面术语、方面类别(术语类别)、术语在上下文中情感极性以及针对该术语的观点词,不同数据集抽取不同的信息,这点在jsonl文件的“instruction”键中有分别提到,在此我将其改造为了生成任务,需要模型按照一定格式生成抽取结果。
以acos数据集中抽取的jsonl文件一条数据举例:
{
"task_type": "generation",
"dataset": "acos",
"input": ["the computer has difficulty switching between tablet and computer ."],
"output": "[['computer', 'laptop usability', 'negative', 'difficulty']]",
"situation": "none",
"label": "",
"extra": "",
"instruction": "
Task: Extracting aspect terms and their corresponding aspect categories, sentiment polarities, and opinion words.
Input: A sentence
Output: A list of 4-tuples, where each tuple contains the extracted aspect term, its aspect category, sentiment polarity, and opinion words (if any). Supplement: \"Null\" means that there is no occurrence in the sentence.
Example:
Sentence: \"Also it's not a true SSD drive in there but eMMC, which makes a difference.\"
Output: [['SSD drive', 'hard_disc operation_performance', 'negative', 'NULL']]'
"
}
此处未设置label和extra,在instruction中以如上所示的字符串模板,并给出一个例子进行one-shot,ABSA领域数据集(absa-quad,acos,arts,aste-data-v2,mams,semeval-2014,semeval-2015,semeval-2016,towe)每个数据集对应instruction模板相同,内容有细微不同,且部分数据集存在同一数据集不同数据instruction内容不同的情况。
原始数据集
- 数据链接
- Paper: Tasty Burgers, Soggy Fries: Probing Aspect Robustness in Aspect-Based Sentiment Analysis
- 说明:原始数据集由laptop和restaurant两个领域的的json数据组成,本次改造我将两个数据集的数据合并并区分为train、validation与test,该数据的提出目的是测试模型鲁棒性,因此在引用该数据集的文章中多是通过在一个领域的数据上训练,在该数据集的另一个领域上测试。
当前SOTA
数据来自论文
- 评价指标:macro-averaged F1
- SOTA模型:CVIB
- 其他领域数据训练后在restaurant数据集上macro-averaged F1:70.29
- restaurant数据集上训练并测评的macro-averaged F1:82.03
- 其他领域训练后在laptop上测评的macro-averaged F1:69.39
- laptop数据集上训练并测评的macro-averaged F1:77.53 )
- Paper:Reducing Spurious Correlations for Aspect-Based Sentiment Analysis with Variational Information Bottleneck and Contrastive Learning
- 说明:该论文来自Google Scholar检索到的引用ARTS原论文的论文之一,我比较了2023年的一些论文工作后筛选了一个最优指标以及模型。