【作者】 王辉；

【导师】 张忠宝；

【作者基本信息】 北京邮电大学 ， 计算机科学与技术， 2025， 硕士

【摘要】 情感分析旨在构建文本情感语义的计算表征,是自然语言处理领域的核心研究方向之一。作为其重要分支,方面级情感分析聚焦于细粒度的情感语义解析。其中,方面级情感分类作为该体系的基础模块,目标是精准识别特定对象或属性的情感极性。该任务的关键在于建立特定对象与其情感表达之间的语义映射关系,从而支撑细粒度的情感理解。尽管基于预训练与提示学习的方法已在该领域取得显著进展,但仍面临双重挑战:其一,数据稀疏性和类别分布失衡制约了特定对象与其情感倾向之间关联的建立,导致模型难以充分学习和迁移知识。其二,提示模板的隐含知识未被充分挖掘,导致模型难以构建多提示间的知识协同机制,从而制约其在应对多样化文本时的表现。本文针对以上两个关键问题提出了两种方法:(1)针对数据稀疏性和类别分布失衡制约模型知识迁移这一问题,本文提出了一种基于单提示学习的方法。首先,为了增强原始数据的语义信息,该方法通过引入外部知识来获取特定对象的概念信息从而提高模型对上下文语境的敏感性以及对情感极性的识别能力。其次,该方法设计了一个双阶段训练方案,通过利用知识增强的数据与软标签中的类间信息有效建立特定对象与其情感倾向之间的关联,进一步提升模型在少样本场景下的表现。(2)针对不同提示之间未能充分利用隐含知识这一问题,本文提出了一种基于多提示学习的方法。该方法首先设计了多种提示模版并通过训练评估其性能表现,然后通过教师-学生模型的知识蒸馏技术来利用表现较好的模型引导表现较差的模型从而挖掘不同提示的隐含知识,帮助模型更好地学习提示间的潜在关联以增强其对不同文本的适应性。基于多个公开数据集上的实验结果表明,本文提出的基于两种提示学习的方法在F1分数上显著优于基准模型,有效克服了现有方法的局限性并提升了模型性能,为方面级情感分类及实际应用提供了创新解决方案和技术支持。更多还原

【Abstract】 Sentiment analysis aims to construct computational representations of emotional semantics in text,serving as one of the core research directions in natural language processing.As an important subfield,aspect-based sentiment analysis focuses on fine-grained sentiment understanding.Within this framework,aspect-level sentiment classification plays a fundamental role,aiming to accurately identify the sentiment polarity toward specific objects or attributes.The key to this task lies in establishing a semantic mapping between target objects and their associated sentiment expressions,which underpins fine-grained sentiment comprehension.Despite significant progress achieved through pre-training and prompt learning-based approaches,two major challenges remain.First,data sparsity and class imbalance hinder the establishment of robust associations between aspect terms and their sentiment tendencies,making it difficult for models to fully acquire and transfer knowledge.Second,the implicit knowledge embedded in prompt templates has not been sufficiently explored,limiting the model’s ability to construct collaborative mechanisms across multiple prompts,which in turn constrains its performance when handling diverse textual inputs.To address these two critical issues,this thesis proposes two methods:(1)To tackle the challenge of limited knowledge transfer caused by data sparsity and class imbalance,a single-prompt learning approach is introduced.This method first enriches the semantic information of the original data by incorporating external knowledge to retrieve conceptual information related to specific targets,thereby enhancing the model’s sensitivity to contextual cues and its ability to detect sentiment polarity.Furthermore,a two-stage training scheme is designed,leveraging knowledge-augmented data and inter-class information in soft labels to effectively establish associations between aspects and sentiment tendencies,thus improving performance in few-shot scenarios.(2)To address the insufficient utilization of implicit knowledge across different prompts,a multi-prompt learning method is proposed.This approach first designs multiple prompt templates and evaluates their performance through training.It then employs a teacher-student knowledge distillation framework,where stronger models guide weaker ones to extract the hidden knowledge across prompts.This helps the model better learn the latent relationships among different prompts,thereby enhancing its adaptability to diverse textual inputs.Experimental results on multiple public datasets demonstrate that the two prompt learning-based methods proposed in this thesis significantly outperform baseline models in terms of F1 score.These methods effectively overcome the limitations of existing approaches and enhance model performance,providing innovative solutions and technical support for aspect-level sentiment classification and real-world applications.更多还原