【作者】 张慧；

【导师】 王红；

【作者基本信息】 山东师范大学 ， 工程硕士（专业学位）， 2022， 硕士

【摘要】 抑郁症对人类的身心健康造成极大伤害,甚至危害社会。因此,尽早发现抑郁症早期症状并及时治疗极为重要。抑郁症的早期状态称为抑郁倾向。与大多数正常人一样,抑郁倾向患者会在社交媒体平台上分享自己的故事,表达情绪,以及寻求帮助和支持。因此,海量的社交平台数据为我们挖掘抑郁倾向的特征和发现抑郁倾向患者提供了可能。但是,如何更好地利用社交媒体平台数据,挖掘能够识别用户抑郁倾向的重要特征成为一项难题。尽管目前已有大量的基于文本的抑郁倾向检测方法,但是检测结果并不令人满意。首先,用于抑郁倾向挖掘的文本数据特征不够丰富,没有充分利用文本的多模态特征,存在语义信息缺失的问题;其次,缺乏对抑郁倾向检测起关键作用的单词或句子的关注,没有对重要特征进行重点学习;最后,目前大多数文本挖掘方法没有充分挖掘文本的句法结构特征,无法解决社交平台上文本中广泛存在的单词歧义问题,严重地影响了抑郁倾向检测效果。针对上述问题,本论文分析研究了社交媒体文本数据的特点、现有的文本挖掘方法以及抑郁倾向检测方法,提出了一系列基于文本预训练模型的抑郁倾向检测方法。本论文的主要贡献如下:(1)提出了基于多模态特征和文本预训练的抑郁倾向检测方法(MTDD),以解决文本数据的特征表示不够丰富,文本语义信息缺失,以致抑郁倾向检测效果不佳的问题。首先,MTDD模型是基于深度神经网络的混合模型,结合了卷积神经网络(Convolutional Neural Network,CNN)和双向长短时记忆网络(Bidirectional Long Short-Term Memory,Bi LSTM)网络,避免了抑郁倾向识别的单一模型存在的泛化能力不强问题;其次,MTDD模型基于文本的多模态特征进行向量表示学习,包括文本特征、语义特征和领域知识,使得模型更加健壮。(2)提出了基于嵌入语言模型、分层注意力网络和文本预训练的抑郁倾向检测方法(E-HAN),以解决对抑郁倾向检测起关键作用的单词或句子关注不足,没有对重要特征进行重点学习的问题。首先,利用预训练模型和嵌入语言模型(Embedding from Language Models,ELMO)获取词嵌入,并融合单词的情感特征及词性特征,组成词的多粒度特征,获得丰富的词特征表示;其次,从单词级和句子级分别进行特征提取,并引入注意力机制。在捕获单词和句子特征的同时,赋予其不同的权重,突出重要特征信息,提高模型的检测性能。(3)提出了基于文本预训练和依存句法分析的抑郁倾向检测方法(PMDT),以解决单词语义歧义、句法结构特征挖掘不充分的问题。首先,提出分段思想,将长文本分成若干短文本,每段短文本的词嵌入向量加权求平均得到整个长文本的特征向量表示;其次,根据依存句法分析结果获得句子中各个成分之间的二元依存关系,构建依存关系矩阵,利用句法信息构建每个单词的特征表示,并通过图卷积网络(Graph Convolution Network,GCN)学习单词之间的关系特征;再次,融合BERT(Bidirectional Encoder Representation from Transformers)词嵌入特征及句子语法结构特征作为文本的特征表示,结合双层递归神经网络(Bidirectional Recurrent Neural Network,Bi RNN)模型提取特征并输出预测结果。在多个公开数据集进行了大量实验,并使用多项指标评价论文工作的预测性能。结果表明,本论文提出的三种方法具有更优的性能,检测结果优于目前主流的基于文本的抑郁倾向检测方法。更多还原

【Abstract】 Depression causes great harm to human physical and mental health and even endangers society.Therefore,it is essential to find the early symptoms of depression and treat them in time.An earlier state of people with depression is called depressive tendency.Like most normal people,people with depressive tendency share their stories,express their emotions,and seek help and support on social media platforms.Therefore,the massive social platform data allows us to mine the characteristics of depression tendency and discover the patients with depression tendency.However,how to effectively utilize social media platform data to find important features to identify users’ depressive tendency has become a challenging problem.Although there are many text-based methods for detecting depression tendency,the detection results are not satisfactory.First,the feature representation of the text data published by the user is not rich enough,the multimodal features of the text are not fully utilized,resulting in the lack of semantics information of the text;Second,some methods do not attach sufficient attention to the words or sentences that play a vital role in the detection of depression tendency,and do not focus on learning their features;Finally,most of the current text mining methods do not fully mine the syntactic structure features of the text.They cannot solve the semantic ambiguity of words in the text on social platforms,which seriously reduces the effect of detecting depression tendency.Pointing to the above problems,this thesis studies the characteristics of social media texts,text-based mining methods,and depression tendency detection methods.It proposes a series of depression tendency detection methods based on text pre-training models.The main contributions of this thesis are as follows:(1)Propose a multimodal feature and text pre-training based method for depressive tendency detection(MTDD)to solve the problem of insufficient feature representation of text data and lack of text semantic information,resulting in poor detection of depression tendency.First,the MTDD model is a hybrid model based on a deep neural network,combined with CNN(Convolutional Neural Network)and Bi LSTM(Bidirectional Long Short-Term Memory)network,to avoid the problem of the poor generalization ability of a single model for depression tendency identification;Second,the MTDD model performs vector representation learning based on multimodal features of text,including text features,semantic features,and domain knowledge,making the model more robust.(2)Present a depression tendency detection method based on embedding from language models,hierarchical attention networks,and text pre-training(E-HAN)to solve the problem of insufficient attention to keywords or sentences and no focus on learning their features.First,use the pre-training model and Embedding from Language Models(ELMO)to obtain word embeddings,and fuse the emotional features and part-of-speech features of words to form multi-granularity features of words,and obtain rich word feature representations;Second,feature extraction is performed separately from the word-level and sentence-level.An attention mechanism is introduced to capture the features of words and sentences and give them different weights to highlight important feature information and improve the detection performance of the model.(3)Propose a depression tendency detection method based on text pre-training and dependency parsing(PMDT)to solve the problems of semantic ambiguity of words and insufficient syntactic structure features.First,the idea of segmentation is proposed.The lengthy text is divided into several short texts,and the word embedding vector of each short text is weighted and averaged to obtain the feature vector representation of the entire long text;Second,the binary dependencies between the components in the sentence are obtained according to the results of the dependency syntax analysis to construct the dependency matrix.The feature representation of each word is constructed using syntactic information,and the relational features between words are learned through Graph Convolution Network(GCN);Third,the BERT(Bidirectional Encoder Representation from Transformers)word embedding feature and the sentence grammatical structure feature are combined as the feature representation of the text,combined with the Bidirectional Recurrent Neural Network(Bi RNN)model to extract features and output prediction results.We conduct extensive experiments on multiple public datasets and use various metrics to evaluate the predictive performance of our works.The results show that the three methods proposed in this thesis work better than the state-of-the-art text-based depression tendency detection approaches.更多还原