结合词性、位置和单词情感的内存网络的方面情感分析
|
摘要
在基于方面的情感分析中,上下文中不同单词对于推测方面情感的重要性是不同的,关注机制是非常适合计算每个单词的重要性的,但现有的关注机制没有利用单词词性,单词和方面间的位置信息,这些信息能够帮助计算每个单词的重要性,另外单词情感信息在推测方面情感时也非常重要.所以本文在端对端内存网路(ETEMN)的基础上提出了一种结合词性,位置和单词情感的内存网络来进行基于方面的情感分析,该网路首先在词向量的基础上融入单词情感信息,然后利用该网络中提出的结合词性、位置信息的CNN关注机制(POSP-CNNAM)分析上下文中每个单词在推测方面情感的重要性并生成上下文向量,最后利用上下文向量进行情感分析.通过在2个来自于SemEval 2014任务4的数据集上进行对比实验,结果表明POSP-CNNAM能够有效的分析上下文中每个单词在推测方面情感的重要性,并且在基于方面的情感分析中和LSTM、TD-LSTM、SVM、ETEM N相比,本文提出的内存网路能够获得更好的结果.
The importance of each word in the context is different when inferring the sentiment polarity of an aspect in aspect level sentiment analysis. The attention mechanism is suitable for calculating the importance of each word in the context. However,the current attention mechanism doesn't use the information of Part of Speech of each word and the position between each word and aspect. The information can help calculate the importance of each word. In addition,the sentiment polarities of words in the context are also important when inferring the sentiment polarity of an aspect. So,we introduce a memory network with Part of Speech,position and polarity of word based on End to End Memory Network( ETEMN) for aspect level sentiment analysis. First,the memory network incorporates the polarity of word into word embedding. Then using CNN attention mechanism with Part of Speech and position( POSP-CNNAM)proposed by the memory network to analyze the importance of each word in context when inferring the sentiment polarity of an aspect and generate a vector for the context. Finally,using the vector for sentiment analysis. Through results of an experiment on 2 datasets from SemEval 2014 task 4,we can knowthat POSP-CNNAMcan effectively analyze the importance of each word in context when inferring the sentiment polarity of an aspect. And compared with LSTM,TD-LSTM,SVMand ETEMN,the proposed memory network can achieve better performance for aspect level sentiment analysis.
The importance of each word in the context is different when inferring the sentiment polarity of an aspect in aspect level sentiment analysis. The attention mechanism is suitable for calculating the importance of each word in the context. However,the current attention mechanism doesn't use the information of Part of Speech of each word and the position between each word and aspect. The information can help calculate the importance of each word. In addition,the sentiment polarities of words in the context are also important when inferring the sentiment polarity of an aspect. So,we introduce a memory network with Part of Speech,position and polarity of word based on End to End Memory Network( ETEMN) for aspect level sentiment analysis. First,the memory network incorporates the polarity of word into word embedding. Then using CNN attention mechanism with Part of Speech and position( POSP-CNNAM)proposed by the memory network to analyze the importance of each word in context when inferring the sentiment polarity of an aspect and generate a vector for the context. Finally,using the vector for sentiment analysis. Through results of an experiment on 2 datasets from SemEval 2014 task 4,we can knowthat POSP-CNNAMcan effectively analyze the importance of each word in context when inferring the sentiment polarity of an aspect. And compared with LSTM,TD-LSTM,SVMand ETEMN,the proposed memory network can achieve better performance for aspect level sentiment analysis.
引文
[1]Kiritchenko S,Zhu X,Cherry C,et al.NRC-Canada-2014:detecting aspects and sentiment in customer review s[J].Proceedings of the 8th International Workshop on Semantic Evaluation(SemEval2014),2014,1(1):437-442.
[2]Yi Shun-ming,Yi Hao,Zhou Guo-dong.Twitter sentiment classification w ith sentimental feature vector[J].Journal of Chinese Computer Systems,2016,37(11):2454-2458.
[3]Socher R,Perelygin A,Wu J Y,et al.Recursive deep models for semantic compositionality over a sentiment treebank[J].Proceedings of the 2013 Conference on Empirical M ethods in Natural Language Processing,2013,1(1):1631-1642.
[4]Nguyen T H,Shirai K.PhraseRNN:phrase recursive neural network for aspect-based sentiment analysis[C].Conference on Empirical M ethods in Natural Language Processing,2015:2509-2514.
[5]Tang D,Qin B,Liu T.Document modeling with gated recurrent neural netw ork for sentiment classification[J].Proceedings of the2015 Conference on Empirical M ethods in Natural Language Processing,2015,3(12):1422-1432.
[6]Tang D,Qin B,Feng X,et al.Effective LSTMs for target-dependent sentiment classification[J].Proceedings of COLING 2016,the26th International Conference on Computational Linguistics,2016,4(12):3298-3307.
[7]Wang Y,Huang M,Zhao L.Attention-based lstm for aspect-level sentiment classification[J].Proceedings of the 2016 Conference on Empirical M ethods in Natural Language Processing,2016,1(5):606-615.
[8]Ruder S,Ghaffari P,Breslin J G.A hierarchical model of reviews for aspect-based sentiment analysis[J].Proceedings of the 2016Conference on Empirical M ethods in Natural Language Processing,2016,5(7):999-1005.
[9]Kim Y.Convolutional neural networks for sentence classification[J].Proceedings of the 2014 Conference on Empirical Methods in Natural Language Processing,2014,3(1):1746-1751.
[10]Poria,Soujanya,Erik Cambria,et al.Aspect extraction for opinion mining w ith a deep convolutional neural netw ork[J].Know ledgeBased Systems,2016,108(1):42-49.
[11]Chen T,Xu R,He Y,et al.Improving sentiment analysis via sentence type classification using BiLSTM-CRF and CNN[J].Expert Systems w ith Applications,2017,72(1):221-230.
[12]Jebbara,Soufian,Philipp Cimiano.Aspect-based relational sentiment analysis using a stacked neural netw ork architecture[J].ECAI 2016-22nd European Conference on Artificial Intelligence,2016,14(8):43-52.
[13]Bahdanau D,Cho K,Bengio Y.Neural machine translation by jointly learning to align and translate[J].ar Xiv preprint ar Xiv:1409.0473,2014.
[14]Rush A M,Chopra S,Weston J.A neural attention model for abstractive sentence summarization[J].Proceedings of the 2015 Conference on Empirical M ethods in Natural Language Processing,2015,3(4):379-389.
[15]Sukhbaatar S,Weston J,Fergus R.End-to-end memory networks[J].Advances in Neural Information Processing Systems,2015,21(8):2440-2448.
[16]Tang D,Qin B,Liu T.Aspect level sentiment classification with deep memory netw ork[J].In Proceedings of the 2016 Conference on Empirical M ethods in Natural Language Processing,2016,26(8):214-224.
[17]Duchi,John,Elad Hazan,et al.Adaptive subgradient methods for online learning and stochastic optimization[J].Journal of M achine Learning Research 12.Jul,2011,4(9):2121-2159.
[18]Kai Sheng Tai,Richard Socher,Christopher D.Improved semantic representations from tree-structured long short-term memory netw orks[J].In Proceedings of the 53rd Annual M eeting of the Association for Computational Linguis-tics,2015,1(1):1556-1566.
[2]易顺明,易昊,周国栋.采用情感特征向量的Twitter情感分类方法研究[J].小型微型计算机系统,2016,37(11):2454-2458.
[2]Yi Shun-ming,Yi Hao,Zhou Guo-dong.Twitter sentiment classification w ith sentimental feature vector[J].Journal of Chinese Computer Systems,2016,37(11):2454-2458.
[3]Socher R,Perelygin A,Wu J Y,et al.Recursive deep models for semantic compositionality over a sentiment treebank[J].Proceedings of the 2013 Conference on Empirical M ethods in Natural Language Processing,2013,1(1):1631-1642.
[4]Nguyen T H,Shirai K.PhraseRNN:phrase recursive neural network for aspect-based sentiment analysis[C].Conference on Empirical M ethods in Natural Language Processing,2015:2509-2514.
[5]Tang D,Qin B,Liu T.Document modeling with gated recurrent neural netw ork for sentiment classification[J].Proceedings of the2015 Conference on Empirical M ethods in Natural Language Processing,2015,3(12):1422-1432.
[6]Tang D,Qin B,Feng X,et al.Effective LSTMs for target-dependent sentiment classification[J].Proceedings of COLING 2016,the26th International Conference on Computational Linguistics,2016,4(12):3298-3307.
[7]Wang Y,Huang M,Zhao L.Attention-based lstm for aspect-level sentiment classification[J].Proceedings of the 2016 Conference on Empirical M ethods in Natural Language Processing,2016,1(5):606-615.
[8]Ruder S,Ghaffari P,Breslin J G.A hierarchical model of reviews for aspect-based sentiment analysis[J].Proceedings of the 2016Conference on Empirical M ethods in Natural Language Processing,2016,5(7):999-1005.
[9]Kim Y.Convolutional neural networks for sentence classification[J].Proceedings of the 2014 Conference on Empirical Methods in Natural Language Processing,2014,3(1):1746-1751.
[10]Poria,Soujanya,Erik Cambria,et al.Aspect extraction for opinion mining w ith a deep convolutional neural netw ork[J].Know ledgeBased Systems,2016,108(1):42-49.
[11]Chen T,Xu R,He Y,et al.Improving sentiment analysis via sentence type classification using BiLSTM-CRF and CNN[J].Expert Systems w ith Applications,2017,72(1):221-230.
[12]Jebbara,Soufian,Philipp Cimiano.Aspect-based relational sentiment analysis using a stacked neural netw ork architecture[J].ECAI 2016-22nd European Conference on Artificial Intelligence,2016,14(8):43-52.
[13]Bahdanau D,Cho K,Bengio Y.Neural machine translation by jointly learning to align and translate[J].ar Xiv preprint ar Xiv:1409.0473,2014.
[14]Rush A M,Chopra S,Weston J.A neural attention model for abstractive sentence summarization[J].Proceedings of the 2015 Conference on Empirical M ethods in Natural Language Processing,2015,3(4):379-389.
[15]Sukhbaatar S,Weston J,Fergus R.End-to-end memory networks[J].Advances in Neural Information Processing Systems,2015,21(8):2440-2448.
[16]Tang D,Qin B,Liu T.Aspect level sentiment classification with deep memory netw ork[J].In Proceedings of the 2016 Conference on Empirical M ethods in Natural Language Processing,2016,26(8):214-224.
[17]Duchi,John,Elad Hazan,et al.Adaptive subgradient methods for online learning and stochastic optimization[J].Journal of M achine Learning Research 12.Jul,2011,4(9):2121-2159.
[18]Kai Sheng Tai,Richard Socher,Christopher D.Improved semantic representations from tree-structured long short-term memory netw orks[J].In Proceedings of the 53rd Annual M eeting of the Association for Computational Linguis-tics,2015,1(1):1556-1566.
[2]易顺明,易昊,周国栋.采用情感特征向量的Twitter情感分类方法研究[J].小型微型计算机系统,2016,37(11):2454-2458.