基于多层局部信息融合的在线论坛用户心理危机识别

心理健康问题已经成为当今社会关注的焦点,它严重威胁着家庭和睦与社会稳定.有心理危机的用户经常通过特定的社区论坛或者社交媒体来求助或倾述,这为用户心理危机识别开辟了一个新的途径.论坛帖子长短不一,但判断心理危机的核心信息往往体现在局部内容上,基于此特点,本文构建了一个结合分层长短记忆网络和卷积神经网络的多层局部信息融合模型(Multi-layer Partial Information Fusion model,MPIF),利用论坛用户发布的帖子,检测用户的心理危机严重程度.模型的特点在于:1)利用预训练语言模型BERT对用户帖子中的句子进行向量化表示,充分考虑词语在不同语境中的不同含义表达;2)分别从词、短语、以及句子层面挖掘反映用户心理危机状态的信息,采用深度分层LSTM网络和注意力机制相结合的方式来获取待分类帖子中词语层面以及句子层面的局部信息,利用CNN网络中多种大小不同的卷积核来提取帖子中短语层面的局部信息;3)采用注意力机制和最大池化层,使得模型不仅能够有效地利用局部信息给出心理危机程度的判断,同时可以将这些局部信息展示给心理专家,辅助专家更快了解患者.基于CLPsych2019 Shared Task评测任务的实验结果显示,与评测时排名第一的模型相比,MPIF模型的官方评测指标All-F1值(自杀风险程度a,b,c,d 4个类别的F1值取平均)高出3.9%.经消融实验发现,去除LSTM词语层、CNN短语层、LSTM句子层,All-F1分别下降4%、4.3%、2.4%.

Multi-layer Partial Information Fusion Model for Psychological Crisis Identification of Online Forum Users

Mental health problems have become the focus of social concern,which seriously threaten family harmony and social stability.Users with mental crisis often ask for help through specific community forums or social media,which opens up a new way to identify mental crisis problem.The forum posts vary in length,but the core information of mental crisis is reflected in the local content.Hereby,this dissertation constructs a Multi-layer Partial Information Fusion model MPIF combining hierarchical long short memory network and convolutional neural network.The model uses the forum users’ posts to detect the severity of users’ mental crisis.MPIF has the following features.1) Using the pre-trained language model BERT to express the vectorization of sentences in users’ posts,fully considering the different meanings of words in different contexts;2) Mining information reflecting users’ mental crisis from words,phrases,and sentences level.Through the combination of deep layered LSTM network and attention mechanism,the local information at the word level and sentence level is obtained.Using convolution kernels of different sizes in CNN network to extract the local information of phrase level in posts;3) Using attention mechanism and maximum pooling layer,the model can not only effectively use local information to judge the degree of psychological crisis,but also show the local information to psychological experts to help them understand patients more quickly.The experiments based on the CLPsych2019 Shared Task evaluation show that MPIF model’s All-F1(The F1 of suicide risk degree a,b,c,d is taken as the average value) is 3.9% higher than the top ranked team.After ablation experiment,we found that removing LSTM word layer,CNN phrase layer and LSTM sentence layer,All-F1 decreased by 4%,4.3% and 2.4%.