👻已复现👻Utilizing BERT for Aspect-Based Sentiment Analysis via Constructing Auxiliary Sentence

2022-09-16 阅读量

Abstract & Introduction

What ABSA?
- 传统的情感分析任务假设是整个文本是具有整体极性，也就是一个句子就表示一种情感倾向。但实际工作中，用户的评论可能包含不同的方面，例如：“这本书是精装版，但价格有点高。” 书的“外观”的极性为正，“价格”的极性为负。因此诞生了一个新的任务,基于方面的情绪分析（ABSA）。ABSA任务的操作步骤是 (1)确定表达情感的每个目标; (2)确定每个目标的情感极性。
本文贡献？
- 基于ABSA任务转换成句子对任务
  - 例如其中一种转换方式：
    - 原句：“LOCATION1 is central London so extremely expensive”
    - 可能的一个辅助句：“what do you think of the price of LOCATION1”
- 微调BERT模型，提升在SentiHood 和 SemEval2014 Task 4 数据集上的效果。

作者已开源：点我🤐

本人复现了部分实验：点我🤐

Methodology

任务描述

TABSA

$$
给定有m个单词的句子s= {w_{1}, w{2},…,w_{m}}，句子s中包括k个目标实体T={t_{1},t_{2},…,t_{k}}
$$

$$
以及固定的属性集合 A = {general, price, transit-location, safety}
$$

$$
预测每一个情感实体T映射到每一个属性集合A的情感极向y \in {positive, negative, none}
$$

例如：

ABSA

semeval2014\task4\subtask3

$$
给定有m个单词的句子s= {w_{1}, w{2},…,w_{m}}, 以及属性集合A={a_{1}, a_{2},…,a_{k}}，预测句子s中有哪些属性
$$

For example, given the set of aspect categories {food, service, price, ambience, anecdotes/miscellaneous}:

“The restaurant was too expensive” → {price}
“The restaurant was expensive, but the menu was great” → {price, food}

semeval2014\task4\subtask4

$$
给定有m个单词的句子s= {w_{1}, w{2},…,w_{m}}, 以及属性集合A={a_{1}, a_{2},…,a_{k}}
$$

$$
以及情感极向Y={y_{1}, y_{2},…,y_{k}},预测句子s中有哪些属性以及这些属性的情感极向
$$

Given a set of pre-identified aspect categories (e.g., {food, price}), determine the polarity (positive, negative, neutral or conflict) of each aspect category.

For example:

“The restaurant was too expensive” → {price: negative}
“The restaurant was expensive, but the menu was great” → {price: negative, food: positive}

构造辅助句子

QA-M (三分类):
- 对每个属性生成一个句子，预测情感极向 positive、negative或none
- pair (LOCATION1, safety), the sentence we generate is “what do you think of the safety of location - 1 ?”
NLI-M (三分类):
- 对每个属性生成一个句子，预测情感极向 positive、negative或none
- pair (LOCATION1, safety), the sentence we generate is “location - 1 - safety”.
QA-B (二分类):
- 对每个属性生成三个情感极向的句子，预测yes 或 no
- pair (LOCATION1, safety)
  - the polarity of the aspect safety of location - 1 is positive
  - the polarity of the aspect safety of location - 1 is negative
  - the polarity of the aspect safety of location - 1 is none
NLI-B (二分类):
- 对每个属性生成三个情感极向的句子，预测yes 或 no
- pair (LOCATION1, safety)
  - location - 1 - safety- positive
  - location - 1 - safety- negative
  - location - 1 - safety- none