xinxngxin commited on
Commit
a871796
1 Parent(s): 4ecd3ba

Create README.md

Browse files
Files changed (1) hide show
  1. README.md +24 -0
README.md ADDED
@@ -0,0 +1,24 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ > 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。
2
+
3
+ #### 以acos数据集中抽取的jsonl文件一条数据举例:
4
+
5
+ ```
6
+ {
7
+ "task_type": "generation",
8
+ "dataset": "acos",
9
+ "input": ["the computer has difficulty switching between tablet and computer ."],
10
+ "output": "[['computer', 'laptop usability', 'negative', 'difficulty']]",
11
+ "situation": "none",
12
+ "label": "",
13
+ "extra": "",
14
+ "instruction": "
15
+ Task: Extracting aspect terms and their corresponding aspect categories, sentiment polarities, and opinion words.
16
+ Input: A sentence
17
+ 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.
18
+ Example:
19
+ Sentence: \"Also it's not a true SSD drive in there but eMMC, which makes a difference.\"
20
+ Output: [['SSD drive', 'hard_disc operation_performance', 'negative', 'NULL']]'
21
+ "
22
+ }
23
+ ```
24
+ > 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。