Targeted Aspect-Based Sentiment Analysis via Embedding Commonsense Knowledge into an Attentive LSTM

Targeted Aspect-Based Sentiment Analysis via Embedding Commonsense Knowledge into an Attentive LSTM (AAAI-18)

原文下载

研究方向

情感分析

研究目的

分析人们对于特定属性的情感倾向是自然语言理解中的一项重要任务。在本文中,提出了一种对于实体属性情感分析的新颖的方法,通过加入常识知识来进行属性和实体的情感分析。

提出方案

1.提出了针对整个句子和实体的分层注意力模型
2.用外部知识来扩展LSTM
3.融合常识到深度神经网络

模型架构

Targeted Aspect-Based Sentiment Analysis via Embedding Commonsense Knowledge into an Attentive LSTM

LSTM

通常LSTM架构公式如下所示:
Targeted Aspect-Based Sentiment Analysis via Embedding Commonsense Knowledge into an Attentive LSTM
双向的时候,隐藏层为:
Targeted Aspect-Based Sentiment Analysis via Embedding Commonsense Knowledge into an Attentive LSTM

Target-Level Attention

计算实体的注意力权重向量:
Targeted Aspect-Based Sentiment Analysis via Embedding Commonsense Knowledge into an Attentive LSTM
Targeted Aspect-Based Sentiment Analysis via Embedding Commonsense Knowledge into an Attentive LSTM

Sentence-Level Attention Model

通过分析句子中的单个词来学习注意力:
Targeted Aspect-Based Sentiment Analysis via Embedding Commonsense Knowledge into an Attentive LSTM
Targeted Aspect-Based Sentiment Analysis via Embedding Commonsense Knowledge into an Attentive LSTM

常识知识

为了提高情感分类的准确性,使用常识库SenticNet作为知识资源embedding 到序列编码当中,不仅包含了概念表示,同时也蕴含着属性和情感之间的语义关系,如下图所示:Targeted Aspect-Based Sentiment Analysis via Embedding Commonsense Knowledge into an Attentive LSTM
然而,涵盖内容丰富的同时,也出现了SenticNet的维度很高的问题,例如:Targeted Aspect-Based Sentiment Analysis via Embedding Commonsense Knowledge into an Attentive LSTM
并且很多地方大多都是0,因此,需要在不损失原始空间语义关系的情况下,将其转换为低纬度嵌入,如下图所示:Targeted Aspect-Based Sentiment Analysis via Embedding Commonsense Knowledge into an Attentive LSTM

sentic LSTM

为了能够有效利用情感常识知识,我们提出了sentic LSTM。
作用:
1.协助从当前时刻到下一时间步过滤信息
2.为记忆细胞补充信息

Targeted Aspect-Based Sentiment Analysis via Embedding Commonsense Knowledge into an Attentive LSTM
Targeted Aspect-Based Sentiment Analysis via Embedding Commonsense Knowledge into an Attentive LSTM

数据集

1 SentiHood
Targeted Aspect-Based Sentiment Analysis via Embedding Commonsense Knowledge into an Attentive LSTM
2 SemEval-2015
Targeted Aspect-Based Sentiment Analysis via Embedding Commonsense Knowledge into an Attentive LSTM

注意力可视化

Targeted Aspect-Based Sentiment Analysis via Embedding Commonsense Knowledge into an Attentive LSTM
Targeted Aspect-Based Sentiment Analysis via Embedding Commonsense Knowledge into an Attentive LSTM

实验结果

最后,实验和几个模型分别在两个统一的数据集进行了对比,第二个 test 第一行和最后一行比较,增加了百分之二十,情感极性准确率也增加了,第二个数据的结果没有第一个高的原因可能是因为目标词用location表示。
Targeted Aspect-Based Sentiment Analysis via Embedding Commonsense Knowledge into an Attentive LSTM
Targeted Aspect-Based Sentiment Analysis via Embedding Commonsense Knowledge into an Attentive LSTM