NEUDM/acos

> 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](https://github.com/IsakZhang/ABSA-QUAD)
- Paper: [Aspect-Category-Opinion-Sentiment Quadruple Extraction with Implicit Aspects and Opinions](https://aclanthology.org/2021.acl-long.29.pdf)
- 说明：原始数据集由Restaurant-ACOS和Laptop-ACOS两个文件夹的数据组成，本次改造我将两个数据集的数据合并并区分为train、validation与test

#### 当前SOTA
*数据来自[论文](https://arxiv.org/abs/2305.09193)*

- 评价指标：F1 score
- SOTA模型：E2H-large (Restaurant-ACOS上F1 Score：**63.50** ， Laptop-ACOS上F1 Score：**44.51**)
- Paper：[Easy-to-Hard Learning for Information Extraction](https://arxiv.org/pdf/2305.09193.pdf)
- 说明：该论文来自[Google Scholar](https://scholar.google.com/scholar?as_ylo=2023&hl=zh-CN&as_sdt=2005&sciodt=0,5&cites=5295149944344120368&scipsc=)检索到的引用ACOS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。

The above dataset is from the ABSA (Aspect-Based Sentiment Analysis) domain. Its basic form consists of extracting from a sentence: aspect terms, aspect categories (term categories), sentiment polarity of the term in context, and opinion words targeting the term. Different datasets extract different information, as mentioned separately in the "instruction" key of each JSONL file. Here, I have transformed it into a generation task, requiring the model to generate extraction results in a specific format. #### Example of one data entry from the ACOS dataset in JSONL format: ``` { "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']]' " } ``` > Here, `label` and `extra` are not set. The `instruction` is provided as a string template as shown above, along with an example for one-shot learning. For the ABSA domain datasets (absa-quad, acos, arts, aste-data-v2, mams, semeval-2014, semeval-2015, semeval-2016, towe), each dataset shares the same instruction template with minor content differences. Moreover, some datasets have varying instruction content for different data entries within the same dataset. #### Original Dataset - Data [link](https://github.com/IsakZhang/ABSA-QUAD) - Paper: [Aspect-Category-Opinion-Sentiment Quadruple Extraction with Implicit Aspects and Opinions](https://aclanthology.org/2021.acl-long.29.pdf) - Description: The original dataset consists of two folders: Restaurant-ACOS and Laptop-ACOS. In this adaptation, I merged the data from both folders and split them into train, validation, and test sets. #### Current SOTA *Data from [paper](https://arxiv.org/abs/2305.09193)* - Evaluation Metric: F1 score - SOTA Model: E2H-large (F1 Score on Restaurant-ACOS: **63.50**, on Laptop-ACOS: **44.51**) - Paper: [Easy-to-Hard Learning for Information Extraction](https://arxiv.org/pdf/2305.09193.pdf) - Description: This paper is one of those citing the original ACOS paper, retrieved from [Google Scholar](https://scholar.google.com/scholar?as_ylo=2023&hl=zh-CN&as_sdt=2005&sciodt=0,5&cites=5295149944344120368&scipsc=). After comparing several papers from 2023, I selected the best-performing metric and model.

原始信息汇总

数据集概述

数据集类型

• 领域：Aspect-Based Sentiment Analysis (ABSA)
• 任务类型：Generation

数据集内容

• 数据结构：每条数据包含方面术语、方面类别、情感极性和观点词。
• 输入格式：句子
• 输出格式：4-tuple列表，每个tuple包含方面术语、方面类别、情感极性和观点词。

数据集示例

• 输入："the computer has difficulty switching between tablet and computer ."
• 输出：[[computer, laptop usability, negative, difficulty]]

数据集来源

• 原始数据集：由Restaurant-ACOS和Laptop-ACOS两个文件夹的数据组成，已合并并区分为train、validation与test。
• 相关论文：Aspect-Category-Opinion-Sentiment Quadruple Extraction with Implicit Aspects and Opinions

当前SOTA模型

• 评价指标：F1 score
• 模型：E2H-large
• 性能：在Restaurant-ACOS上F1 Score为63.50，在Laptop-ACOS上F1 Score为44.51。
• 相关论文：Easy-to-Hard Learning for Information Extraction

搜集汇总

背景与挑战

背景概述

NEUDM/acos是一个基于方面的情感分析（ABSA）数据集，专注于从产品评论句子中抽取方面术语、方面类别、情感极性和观点词组成的四元组。该数据集规模中等，包含约6,360条文本数据，划分为训练、验证和测试集，适用于生成式模型训练，旨在支持细粒度的情感信息提取任务。