【作者】 李超；

【导师】 沈波；

【作者基本信息】 北京交通大学 ， 信息与通信工程， 2024， 硕士

【摘要】 随着互联网技术的快速发展以及相关应用的普及,越来越多的人喜欢在社交媒体上分享自己的生活以及讨论热点事件,进而产生了海量的文本数据。这些数据大都蕴含着人们对于相关商品、服务和实事政治的观点,具有某种特定的情感极性。如何让机器自动识别出这些文本所表达的情感倾向,一直是学术界和工业界的研究热点。在深度神经网络的推动下,深度学习模型在文本情感分类任务中取得很好的效果。但是深度学习模型复杂度高、参数多导致其透明度低,人们无法理解这种“端到端”模型做出决策的过程。研究者提出了一些解释方法,来解释文本情感分类模型的预测行为,但这些方法没有考虑单词之间的语义信息,导致解释效果并不理想。此外,模型的鲁棒性较低以及缺乏对扰动语义的学习,会导致其在遇到微小扰动的样本时,预测行为不一致。此时,模型不仅容易遭受攻击,还会使得其决策过程变得更加难以理解和解释。本论文的工作得到了国家自然科学基金项目(No.62376018)的支持。论文的主要工作和创新点如下:(1)提出了基于方面情感词聚合的增强解释方法。针对现有解释方法在解释模型决策过程时,没有考虑单词之间语义信息的问题,本文提出了一种基于方面情感词聚合的增强解释方法,来提升解释方法的有效性。具体而言,为了学习单词之间的语义信息,本文设计了一个SKEP+CRF模型来实现方面情感词对的抽取,从而捕获了方面词与情感词之间的关联信息。本文提出了一个重要性分数加权策略,该策略将方面情感词与解释方法相融合,从而能够得到更加合理的解释。实验结果表明,本文提出的方法在合理性与忠诚性的评估指标上都优于所对比的其他解释方法。(2)提出了基于多维度扰动的对抗训练方法。针对模型鲁棒性低以及对扰动语义的学习较少的问题,本文提出了一种基于多维度扰动的对抗训练方法,来增强模型鲁棒性的同时,提升解释方法在扰动样本上的效果。该方法采用SemAttack算法在三种维度上生成了对抗样本,并对模型进行了对抗训练,从而能够让模型学习到丰富多样的扰动语义。实验结果表明,模型在经过对抗训练后被攻击成功率显著下降,其鲁棒性得到增强。此外,解释方法在合理性与忠诚性的评估指标上都得到了提升,其中合理性在总体上最多能提升6%左右。正文中共使用图20幅,表29个,参考文献75篇。更多还原

【Abstract】 With the rapid development of Internet technology and the popularization of related applications,more and more people prefer to share their lives and discuss hot topics on social media,leading to a massive amount of textual data.These data mostly contain people’s opinions on related goods,services,and current affairs,with a certain specific sentiment polarity.How to automatically identify the sentiment tendencies expressed in these texts has been a research focus in both academia and industry.Driven by deep neural networks,deep learning models have achieved excellent performance in text sentiment classification tasks.However,the high complexity and numerous parameters of deep learning models result in low transparency,making it difficult for people to understand the decision-making process of such "end-to-end" models.Researchers have proposed some explanation methods to interpret the prediction behavior of text sentiment classification models,but these methods do not consider the semantic information between words,leading to unsatisfactory explanation effects.Additionally,the low robustness of the model and the lack of learning perturbation semantics can result in inconsistent prediction behavior when encountering samples with minor perturbations.At this point,the model is not only prone to be attacked but also makes its decision-making process more difficult to understand and explain.This work was supported by National Natural Science Foundation of China(Grant No.62376018).The main work and innovations of this paper are as follows:(1)A reinforced explanation method based on aspect sentiment word aggregation is proposed.To address the issue that existing explanation methods do not consider the semantic information between words when explaining the model’s decision-making process,this paper proposes an enhanced explanation method based on aspect sentiment word aggregation to improve the effectiveness of explanation methods.Specifically,to learn the semantic information between words,this paper designs an SKEP+CRF model to extract aspect sentiment word pairs,thereby capturing the associative information between aspect words and sentiment words.This paper proposes an importance score weighted strategy that integrates aspect sentiment words with explanation methods,thereby obtaining more reasonable explanations.Experimental results show that the proposed method outperforms other compared explanation methods in terms of rationality and fidelity evaluation metrics.(2)An adversarial training method based on multidimensional perturbations is proposed.To address the problems of low model robustness and insufficient learning of perturbation semantics,this paper proposes an adversarial training method based on multidimensional perturbations to enhance model robustness and improve the effectiveness of explanation methods on perturbed samples.This method utilizes the SemAttack to generate adversarial samples in three dimensions and conducts adversarial training on the model,enabling the model to learn a diverse range of perturbation semantics.Experimental results show that the model’s success rate in being attacked significantly decreases after adversarial training,and its robustness is enhanced.In addition,the evaluation metrics of rationality and fidelity of the explanation method are improved,with rationality being able to improve by up to about 6% overall.This paper includes 20 figures,29 tables and 74 references.更多还原