多交互图卷积网络用于方面情感分析

王汝言; 陶中原; 赵容剑; 张普宁; 杨志刚

doi:10.11999/JEIT210459

多交互图卷积网络用于方面情感分析

doi: 10.11999/JEIT210459

1.
重庆邮电大学通信与信息工程学院重庆 400065
2.
先进网络与智能互联技术重庆市高校重点实验室重庆 400065
3.
泛在感知与互联重庆市重点实验室重庆 400065

基金项目: 国家自然科学基金(61901071, 61871062, 61771082)，重庆市高校创新团队建设计划(CXQT20017)，重庆市自然科学基金(cstc2019jcyj-zdxmX0024)

详细信息

作者简介:
王汝言：男，1969年生，教授，研究方向为泛在网络，多媒体信息处理等

陶中原：男，1996年生，硕士生，研究方向为自然语言处理，情感分析

赵容剑：男，1998年生，硕士生，研究方向为自然语言处理，情感分析

张普宁：男，1988年生，副教授，研究方向为物联网搜索等

杨志刚：男，1981年生，博士，研究方向为安全隐私保护等

通讯作者:
王汝言　wangry@cqupt.edu.cn

中图分类号: TP391.1
计量
- 文章访问数: 2055
- HTML全文浏览量: 855
- PDF下载量: 336
- 被引次数: 0
出版历程
- 收稿日期: 2021-05-25
- 修回日期: 2021-09-07
- 网络出版日期: 2021-09-17
- 刊出日期: 2022-03-28

Multi-interaction Graph Convolutional Networks for Aspect-level Sentiment Analysis

1.
School of Communication and Information Engineering, Chongqing University of Posts and Telecommunications, Chongqing 400065, China
2.
Advanced Network and Intelligent Connection Technology Key Laboratory of Chongqing Education Commission of China, Chongqing 400065, China
3.
Chongqing Key Laboratory of Ubiquitous Sensing and Networking, Chongqing 400065, China

Funds: The National Natural Science Foundation of China (61901071, 61871062, 61771082), The University Innovation Research Group Program of Chongqing (CXQT20017), The Natural Science Foundation of Chongqing (cstc2019jcyj-zdxmX0024)

摘要

摘要: 方面情感分析旨在识别句子中特定方面的情感极性，是一项细粒度情感分析任务。传统基于注意力机制方法，仅在单词之间进行单一的语义交互，没有建立方面词与文本词的语法信息交互，导致方面词错误地关注到与其语法无关的文本词信息。此外，单词的位置距离特征和语法距离特征，分别体现其在句子线性形式中和句子语法依存树中的位置关系，而基于图卷积网络处理语法信息的方法却忽略距离特征，使距方面词较远的无关信息对其情感分析造成干扰。针对上述问题，该文提出多交互图卷积网络(MIGCN)，首先将文本词位置距离特征馈入到每层图卷积网络，同时利用依存树中文本词的语法距离特征对图卷积网络的邻接矩阵加权，最后，设计语义交互和语法交互分别处理单词之间语义和语法信息。实验结果表明，在公共数据集上，准确率和宏F1值均优于基准模型。
- 方面情感分析 /
- 图卷积网络 /
- 依存树 /
- 注意力机制
Abstract: Aspect level sentiment analysis aims to identify the sentiment polarity of a specific aspect in a given context, and is a fine-grained sentiment analysis task. The traditional attention-based approach, which only performs the semantic interaction between words, does not establish the syntactic relation interaction between aspect words and text words, resulting in the aspect words incorrectly focusing on information about words that are irrelevant to their syntax. In addition, the positional distance feature and the syntactic distance feature of words, which reflect their relationships in the linear form of the sentence and in the syntactic dependency tree of the sentence, respectively, are ignored by the method of processing syntactic information using graph convolutional networks, allowing irrelevant information far from the aspect words to interfere with their sentiment analysis. To address this problem, a Multi-Interaction Graph Convolutional Network (MIGCN) is proposed. First, the context words positional distance features are fed into each layer of the graph convolutional network, while the adjacency matrix of the graph convolutional network is weighted by using the syntactic distance of context words in the dependency tree. Finally, semantic interaction and syntactic interaction are designed to process the semantic and syntactic information between words, respectively. The experimental results show the proposed model can outperform state-of-the-art baselines on the available datasets.
- Aspect-level sentiment analysis /
- Graph convolutional network /
- Dependency tree /
- Attention mechanisms

HTML全文

图 1 方面情感分析任务样例

下载: 全尺寸图片幻灯片

图 2 包含相反极性方面词的句子依存树

下载: 全尺寸图片幻灯片

图 3 MIGCN模型框架

下载: 全尺寸图片幻灯片

图 4 方面为短语的依存树样例

下载: 全尺寸图片幻灯片

图 5 GCN网络深度与准确率的关系

下载: 全尺寸图片幻灯片

图 6 GCN网络深度与宏F1值的关系

下载: 全尺寸图片幻灯片

图 7 方面为单词“staff”的可视化结果

下载: 全尺寸图片幻灯片

图 8 方面为短语“Indian food”的可视化结果

下载: 全尺寸图片幻灯片

表 1 基于语法加权的邻接矩阵算法(算法1)

输入：$T$：句子依存树；$a$：句子中方面词；$N$：句子序列长度；
输出：$\boldsymbol{A}$: 语法加权邻接矩阵；
(1) 初始化${\boldsymbol{A} } \in {\boldsymbol{R}^{N \times N} }$中所有元素为0;
(2) 从依存树$T$的根部遍历每个节点$i$:
(3) 　　　设置矩阵的主对角线元素${{A}_{ii} } = 1$;
(4) 　　　遍历根为节点$i$的子树中的所有节点$j$;
(5) 　　　　　令${{A}_{ij} } = 1$和${{A}_{ji} } = 1$;
(6) 　　　　　计算节点$i$与句子中方面词$a$的语法距离$d$;
(7) 　　　　　令${{A}_{i{a_i} } } = { {\rm{Weight} } }(d)$和${{A}_{ {a_i}i} } = { {\rm{Weight} } }(d)$。

下载: 导出CSV

表 2 数据集的统计

数据集	积极		中性		消极
数据集	训练集	测试集	训练集	测试集	训练集	测试集
Lap14	994	341	464	169	870	128
Twitter	1561	173	3127	346	1560	173
Rest14	2164	728	637	196	807	196
Rest15	912	326	36	34	256	182
Rest16	1260	469	69	30	439	117

下载: 导出CSV

表 3 不同模型结果对比(%)

类别	模型	Twitter		Lap14		Rest14		Rest15		Rest16
类别	模型	准确率	宏F1	准确率	宏F1	准确率	宏F1	准确率	宏F1	准确率	宏F1
基线	SVM	63.40	63.30	70.49	–	80.16	–	–	–	–	–
基线	LSTM^[2]	69.56	67.70	69.28	63.09	78.13	67.47	77.37	55.17	86.80	63.88
交互模型	IAN^[3]	72.50	70.81	72.05	67.38	79.26	70.09	78.54	52.65	84.74	55.21
	MGAN^[4]	72.54	70.81	75.27	70.81	81.25	71.94	–	–	–	–
	AOA^[17]	72.30	70.20	76.62	67.52	79.97	70.42	78.17	57.02	87.50	66.21
	AEN-GloVe^[10]	72.83	69.81	73.51	69.04	80.98	72.14	–	–	–	–
GCN 模型	ASGCN ^[6]	72.15	70.40	75.55	71.05	80.77	72.02	79.89	61.89	88.99	67.48
	TD-GAT^[7]	72.20	70.45	75.63	70.74	81.32	71.72	80.38	60.50	87.71	67.87
	BiGCN^[18]	74.16	73.35	74.59	71.84	81.97	73.48	81.16	64.79	88.96	70.84
	kumaGCN^[19]	72.45	70.77	76.12	72.42	81.43	73.64	80.69	65.99	89.39	73.19
本文模型	MIGCN	73.31	72.12	76.59	72.44	82.32	74.31	80.81	64.21	89.50	71.97

下载: 导出CSV

表 4 消融实验结果(%)

模型	Twitter		Lap14		Rest14		Rest15		Rest16
模型	准确率	宏F1	准确率	宏F1	准确率	宏F1	准确率	宏F1	准确率	宏F1
MIGCN	73.31	72.12	76.59	72.44	82.32	74.31	80.81	64.21	89.50	71.97
w/o se	71.34	69.54	74.97	70.89	80.21	71.85	78..60	59.01	88.20	69.20
w/o sy	72.45	70.64	75.34	71.03	81.55	73.29	79.40	62.22	89.02	66.72
w/o we	72.88	70.88	76.49	72.28	81.73	73.64	79.95	64.00	88.58	71.12
w/o ga	72.93	71.45	75.91	71.85	81.85	73.53	79.52	63.92	88.58	68.98
w/o sy + ga	72.98	71.40	75.08	70.84	81.19	72.74	78.17	58.24	88.26	68.21
w/o se + ga	72.16	70.33	74.71	70.54	79.82	71.09	79.46	61.55	88.80	67.97

下载: 导出CSV

参考文献(19)

[1]	JIANG Long, YU Mo, ZHOU Ming, et al. Target-dependent twitter sentiment classification[C]. The 49th Annual Meeting of the Association for Computational Linguistics: Human Language Technologies, Portland, USA, 2011: 151–160.
[2]	TANG Duyu, QIN Bing, FENG Xiaocheng, et al. Effective LSTMs for target-dependent sentiment classification[C]. The 26th International Conference on Computational Linguistics, Osaka, Japan, 2016: 3298–3307.
[3]	MA Dehong, LI Sujian, ZHANG Xiaodong, et al. Interactive attention networks for aspect-level sentiment classification[C]. The 26th International Joint Conference on Artificial Intelligence, Melbourne, Australia, 2017: 4068–4074.
[4]	FAN Feifan, FENG Yansong, and ZHAO Dongya Multi-grained attention network for aspect-level sentiment classification[C]. The 2018 Conference on Empirical Methods in Natural Language Processing, Brussels, Belgium, 2018: 3433–3442.
[5]	KIPF T N and WELLING M. Semi-supervised classification with graph convolutional networks[C]. The 5th International Conference on Learning Representations, Toulon, France, 2017.
[6]	ZHANG Chen, LI Qiuchi, and SONG Dawei. Aspect-based sentiment classification with aspect-specific graph convolutional networks[C]. The 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing, Hong Kong, China, 2019: 4568–4578.
[7]	HUANG Binxuan and CARLEY K M. Syntax-aware aspect level sentiment classification with graph attention networks[C]. The 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing, Hong Kong, China, 2019: 5469–5477.
[8]	PHAN M H and OGUNBONA P O. Modelling context and syntactical features for aspect-based sentiment analysis[C]. The 58th Annual Meeting of the Association for Computational Linguistics, Online, 2020: 3211–3220.
[9]	XUE Wei and LI Tao. Aspect based sentiment analysis with gated convolutional networks[C]. The 56th Annual Meeting of the Association for Computational Linguistics, Melbourne, Australia, 2018: 2514–2523.
[10]	SUN Kai, ZHANG Richong, MENSAH S, et al. Aspect-level sentiment analysis via convolution over dependency tree[C]. The 2019 Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing, Hong Kong, China, 2019: 5679–5688.
[11]	BAI Xuefeng, LIU Pengbo, and ZHANG Yue. Investigating typed syntactic dependencies for targeted sentiment classification using graph attention neural network[J]. IEEE/ACM Transactions on Audio, Speech, and Language Processing, 2021, 29: 503–514. doi: 10.1109/TASLP.2020.3042009
[12]	ZHOU Li, WANG Tingyu, QU Hong, et al. A weighted GCN with logical adjacency matrix for relation extraction[C]. The 24th European Conference on Artificial Intelligence, Santiago de Compostela, Spain, 2020: 2314–2321.
[13]	DONG Li, WEI Furu, TAN Chuanqi, et al. Adaptive recursive neural network for target-dependent twitter sentiment classification[C]. The 52nd Annual Meeting of the Association for Computational Linguistics, Baltimore, USA, 2014: 49–54.
[14]	PONTIKI M, GALANIS D, PAVLOPOULOS J, et al. Semeval-2014 task 4: Aspect based sentiment analysis[C]. The 8th International Workshop on Semantic Evaluation, Dublin, Ireland, 2014: 27–35.
[15]	PONTIKI M, GALANIS D, PAPAGEORGIOU H, et al. Semeval-2015 task 12: Aspect based sentiment analysis[C]. The 9th International Workshop on Semantic Evaluation, Denver, USA, 2015: 486–495.
[16]	PONTIKI M, GALANIS D, PAPAGEORGIOU H, et al. Semeval-2016 task 5: Aspect based sentiment analysis[C]. The 8th International Workshop on Semantic Evaluation, San Diego, USA, 2016: 19–30.
[17]	SONG Youwei, WANG Jiahai, JIANG Tao, et al. Targeted sentiment classification with attentional encoder network[C]. 28th International Conference on Artificial Neural Networks and Machine Learning, Munich, Germany, 2019: 93–103.
[18]	ZHANG Mi and QIAN Tieyun. Convolution over hierarchical syntactic and lexical graphs for aspect level sentiment analysis[C]. The 2020 Conference on Empirical Methods in Natural Language Processing, Online, 2020: 3540–3549.
[19]	CHEN Chenhua, TENG Zhiyang, and ZHANG Yue. Inducing target-specific latent structures for aspect sentiment classification[C]. The 2020 Conference on Empirical Methods in Natural Language Processing, Online, 2020: 5596–5607.