Day 5. Suicidal Ideation Detection: A Review of Machine Learning Methods and Applications综述

Title:
Suicidal Ideation Detection: A Review of Machine Learning Methods and Applications
自杀意念检测：机器学习方法及应用综述

Keywords:
Deep learning 深度学习
feature engineering 特征工程
social contents 社交内容
suicidal ideation detection (SID) 自杀意念检测

Abstract：
Suicide is a critical issue in modern society. Early detection and prevention of suicide attempts should be addressed to save people’s life. Current suicidal ideation detection (SID) methods include clinical methods based on the interaction between social workers or experts and the targeted individuals and machine learning techniques with feature engineering or deep learning for automatic detection based on online social contents. This article is the first survey that comprehensively introduces and discusses the methods from these categories. Domain-specific applications of SID are reviewed according to their data sources, i.e., questionnaires, electronic health records, suicide notes, and online user content. Several specific tasks and data sets are introduced and summarized to facilitate further research. Finally, we summarize the limitations of current work and provide an outlook of further research directions.
自杀是一个严重的现代问题。早期发现和预防自杀未遂应该致力于拯救人们的生命。目前的自杀意念检测方法主要包括基于社会工作者或专家与目标个体互动的临床方法和基于特征工程或深度学习的机器学习技术，用于基于在线社交内容的自动检测。本文是第一次全面介绍和讨论这些分类方法的调查。根据SID的数据来源，即调查问卷、电子健康记录、自杀记录和在线用户内容，对SID的特定领域应用进行了审查。介绍和总结了几个具体的任务和数据集，以便于进一步的研究。最后，总结了目前工作的局限性，并对下一步的研究方向进行了展望。

1.INTRODUCTION
MENTAL health issues, such as anxiety and depression, are becoming increasingly concerned in modern society, as they turn out to be especially severe in developed countries and emerging markets. Severe mental disorders without effective treatment can turn to suicidal ideation or even suicide attempts. Some online posts contain much negative information and generate problematic phenomena, such as cyberstalking and cyberbullying. Consequences can be severe and risky since such lousy information is often engaged in some form of social cruelty, leading to rumors or even mental damage. Research shows that there is a link between cyberbullying and suicide [1]. Victims overexposed to too many negative messages or events may become depressed and desperate; even worse, some may commit suicide.
焦虑和抑郁等心理健康问题在现代社会越来越受到关注，在发达国家和新兴市场，焦虑和抑郁尤为严重。严重的精神障碍如果得不到有效的治疗，就会变成自杀意念，甚至是自杀未遂。一些网络帖子含有大量负面信息，并产生一些问题现象，如网络跟踪和网络欺凌。后果可能是严重的和危险的，因为这些糟糕的信息往往参与某种形式的社会残酷，导致谣言甚至精神损害。研究表明，网络欺凌和自杀之间有联系[1]。过度暴露在负面信息或事件中的受害者可能会变得沮丧和绝望；更糟糕的是，有些人可能会自杀。

The reasons that people commit suicide are complicated. People with depression are highly likely to commit suicide, but many without depression can also have suicidal thoughts [2]. According to the American Foundation for Suicide Prevention (AFSP), suicide factors fall under three categories: health factors, environmental factors, and historical factors [3]. Ferrari et al. [4] found that mental health issues and substance use disorders are attributed to the factors of suicide. O’Connor and Nock [5] conducted a thorough review of the psychology of suicide and summarized psychological risks as personality and individual differences, cognitive factors, social factors, and negative life events.
人们自杀的原因很复杂。抑郁症患者极有可能自杀，但许多没有抑郁症的人也可能有自杀念头[2]。根据美国自杀预防基金会（AFSP），自杀因素分为三类：健康因素、环境因素和历史因素[3]。Ferrari等人[4]发现精神健康问题和物质使用障碍是导致自杀的因素。O’Connor和Nock [5] 对自杀心理进行了全面的回顾，并将心理风险归纳为个性和个体差异、认知因素、社会因素和消极生活事件。

Suicidal ideation detection (SID) determines whether the person has suicidal ideation or thoughts by given tabular data of a person or textual content written by a person. Due to the advances in social media and online anonymity, an increasing number of individuals turn to interact with others on the Internet. Online communication channels are becoming a new way for people to express their feelings, suffering, and suicidal tendencies. Hence, online channels have naturally started to act as a surveillance tool for suicidal ideation, and mining social content can improve suicide prevention [6]. Strange social phenomena are emerging, e.g., online communities reaching an agreement on self-mutilation and copycat suicide. For example, a social network phenomenon called the “Blue Whale Game” in 2016 uses many tasks (such as self-harming) and leads game members to commit suicide in the end. Suicide is a critical social issue and takes thousands of lives every year. Thus, it is necessary to detect suicidality and prevent suicide before victims end their life. Early detection and treatment are regarded as the most effective ways to prevent potential suicide attempts.
自杀意念检测（SID）是通过给定的个人表格数据或个人书写的文本内容来判断一个人是否有自杀意念。由于社交媒体和在线匿名技术的进步，越来越多的人开始在互联网上与他人互动。在线交流渠道正在成为人们表达情感、痛苦和自杀倾向的新方式。因此，网络频道自然开始充当自杀意念的监控工具，挖掘社交内容可以提高自杀预防[6]。奇怪的社会现象正在出现，例如，网上社区就自残和模仿自杀达成协议。例如，2016年一个名为“蓝鲸游戏”的社交网络现象使用了许多任务（如自残），并导致游戏成员最终自杀。自杀是一个重要的社会问题，每年夺去成千上万人的生命。因此，有必要在受害者结束生命之前发现自杀行为并防止自杀。早期发现和治疗被认为是防止潜在自杀企图的最有效方法。

Potential victims with suicidal ideation may express their thoughts of committing suicide in fleeting thoughts, suicide plans, and role-playing. SID is to find out these risks of ntentions or behaviors before tragedy strikes. A meta-analysis conducted by McHugh et al. [7] shown statistical limitations of ideation as a screening tool but also pointed out that people’s expression of suicidal ideation represents their psychological distress. Effective detection of early signals of suicidal ideation can identify people with suicidal thoughts and open a communication portal to let social workers mitigate their mental issues. The reasons for suicide are complicated and attributed to a complex interaction of manyfactors [5], [8]. To detect suicidal ideation, many researchers conducted psychological and clinical studies [9] and classified responses of questionnaires [10]. Based on their social media data, artificial intelligence (AI) and machine learning techniques can predict people’s likelihood of suicide [11], which can better understand people’s intentions and pave the way for early intervention. Detection on social content focuses on feature engineering [12], [13], sentiment analysis [14], [15], and deep learning [16]–[18]. Those methods generally require heuristics to select features or design artificial neural network (ANN) architectures for learning rich representation. The research trend focuses on selecting more useful features from people’s health records and developing neural architectures to understand the language with suicidal ideation better.
有自杀意念的潜在受害者可能会在自杀想法、自杀计划和角色扮演中表达他们的自杀想法。SID是在悲剧发生之前找出这些意图或行为的风险。McHugh等人进行的荟萃分析[7]显示了意念作为一种筛选工具的统计局限性，同时也指出人们对自杀意念的表达代表了他们的心理困扰。对自杀意念的早期信号进行有效检测，可以识别出有自杀想法的人，并打开一个沟通门户，让社会工作者缓解他们的心理问题。自杀的原因是复杂的，并归因于许多复杂因素的相互作用[5], [8]。为了检测自杀意念，许多研究人员进行了临床研究 [9] 和问卷调查分类[10]。基于他们的社交媒体数据，人工智能（AI）和机器学习技术可以预测人们自杀的可能性[11]，这可以更好地了解人们的意图，为早期干预铺平道路。对社交内容的检测集中在特征工程[12], [13]、情感分析[14], [15]和深度学习[16]–[18]。这些方法通常需要启发式来选择特征或设计人工神经网络（ANN）结构来学习丰富的表示。研究趋势集中在从人们的健康记录中选择更有用的特征，开发神经结构，以便更好地理解带有自杀意念的语言。

Mobile technologies have been studied and applied to suicide prevention, for example, the mobile suicide intervention application iBobbly [19] developed by the Black Dog 2 Institute. Many other suicide prevention tools integrated with social networking services have also been developed, including Samaritans Radar 3 and Woebot 4. The former was a Twitter plugin that was later discontinued because of privacy issues. For monitoring alarming posts, the latter is a Facebook chatbot based on cognitive behavioral therapy and natural language processing (NLP) techniques for relieving people’s depression
and anxiety.
移动技术已经被研究并应用于自杀预防，例如黑狗研究所2开发的移动自杀干预应用iBobbly [19]。还开发了许多其他与社交网络服务相结合的自杀预防工具，包括Samaritans Radar 3和Woebot 4。前者是一个Twitter插件，后来因为隐私问题而停止使用。后者是一个基于认知行为疗法和自然语言处理（NLP）技术的Facebook聊天机器人，用于监控人们的抑郁和焦虑。
2https://blackdoginstitute.org.au/research/digital-dog/programs/ibobbly-app
3https://samaritans.org/about-samaritans/research-policy/internet-suicide/samaritans-radar
4https://woebot.io

Applying cutting-edge AI technologies for SID inevitably comes with privacy issues [20] and ethical concerns [21]. Linthicum et al. [22] put forward three ethical issues, including the influence of bias on machine learning algorithms, the prediction on time of suicide act, and ethical and legal questions raised by false positive and false negative prediction. It is not easy to answer ethical questions for AI as these require algorithms to reach a balance between competing values,issues, and interests [20].
将尖端人工智能技术应用于SID不可避免地会带来隐私问题[20]和伦理问题[21]。Linthicum等人[22] 提出了三个伦理问题，包括偏差对机器学习算法的影响、对自杀行为时间的预测以及假阳性和假阴性预测所引发的伦理和法律问题。要回答人工智能的伦理问题并不容易，因为这些问题需要算法在相互竞争的价值观、问题和利益之间达到平衡[20]。

AI has been applied to solve many challenging social problems. Detection of suicidal ideation with AI techniquesis one of the potential applications for social good and should be addressed to improve people’s wellbeing meaningfully. The research problems include feature selection on tabular and text data and representation learning on natural language. Many AI-based methods have been applied to classify suicide risks. However, there remain some challenges. There are a limited number of benchmarks for training and evaluating SID. AI-powered models, sometimes, learn statistical clues but fail to understand people’s intentions. Moreover, many neural models are lack of interpretability. This survey reviews SID methods from the perspective of AI and machine learning and specific domain applications with social impact. The categorization from these two perspectives is shown in Fig. 1. This article provides a comprehensive review of the increasingly important field of SID with machine learning methods. It proposes a summary of current research progress and an outlook of future work. The contributions of our survey are summarized as follows.
人工智能已被应用于解决许多具有挑战性的社会问题。用人工智能技术检测自杀意念是切实改善人民生活的潜在应用之一。研究问题包括表格和文本数据的特征选择和自然语言的表示学习。许多基于人工智能的方法被应用于自杀风险的分类。然而，仍然存在一些挑战。训练和评估SID的基准数量有限。人工智能驱动的模型，有时会学习统计线索，但无法理解人们的意图。此外，许多神经网络模型缺乏可解释性。这项调查从人工智能和机器学习以及具有社会影响的特定领域应用的角度回顾了SID方法。图1显示了从这两个角度进行的分类。本文全面回顾了机器学习方法在SID中日益重要的领域。对目前的研究进展进行了总结，并对今后的工作进行了展望。我们调查的贡献总结如下。

1)To the best of our knowledge, this is the first survey that conducts a comprehensive review of SID, its methods, and its applications from a machine learning perspective.
2)We introduce and discuss the classical content analysis and modern machine learning techniques, plus their application to questionnaires, EHR data, suicide notes, and online social content.
3)We enumerate existing and less explored tasks and discuss their limitations. We also summarize existing data sets and provide an outlook of future research directions in this field.
1）据我们所知，这是第一次从机器学习的角度对SID、其方法和应用进行全面回顾的调查。
2）我们介绍和讨论了经典内容分析和现代机器学习技术，以及它们在问卷调查、EHR数据、自杀笔记和在线社交内容中的应用。
3）我们列举现有的和较少探索的任务，并讨论它们的局限性。我们还总结了现有的数据集，并对该领域未来的研究方向进行了展望。

The remainder of this article is organized as follows. Methods and applications are introduced and summarized in Sections II and III, respectively. Section IV enumerates specific tasks and some data sets. Finally, we have a discussion and propose some future directions in Section V.
本文的其余部分组织如下。方法和应用分别在第二节和第三节中介绍和总结。第四节列举了具体任务和一些数据集。最后，我们在第五部分进行了讨论，并提出了未来的发展方向。

图1。自杀意念检测的分类：方法和领域。左侧部分表示方法分类，而右侧部分显示域的类别。箭头和实心点表示子类别。

2.METHODS AND CATEGORIZATION
Suicide detection has drawn the attention of many researchers due to an increasing suicide rate in recent years and has been studied extensively from many perspectives.The research techniques used to examine suicide also span many fields and methods, for example, clinical methods with patient–clinic interaction [9] and automatic detection from user-generated content (mainly text) [12], [17]. Machine learning techniques are widely applied for automatic detection.
近年来，随着自杀率的不断上升，自杀检测引起了许多研究者的关注，并从多个角度进行了广泛的研究。用于检测自杀的研究技术也跨越了许多领域和方法，例如，具有患者-诊所交互的临床方法 [9] 和从用户生成的内容（主要是文本）的自动检测方法 [12], [17]。机器学习技术广泛应用于自动检测。

Traditional suicide detection relies on clinical methods, including self-reports and face-to-face interviews. Venek et al. [9] designed a five-item ubiquitous questionnaire for the assessment of suicidal risks and applied a hierarchical classifier on the patients’ response to determine their suicidal intentions. Through face-to-face interaction, verbal and acoustic information can be utilized. Scherer [23] investigated the prosodic speech characteristics and voice quality in a dyadic interview to identify suicidal and nonsuicidal juveniles.
Other clinical methods examine the resting state heart rate from converted sensing signals [24] and classify the functional magnetic resonance imaging-based neural representations of death-and life-related words [25] and event-related instigators converted from EEG signals [26]. Another aspect of clinical treatment is the understanding of the psychology behind suicidal behavior [5], which, however, relies heavily on the clinician’s knowledge and face-to-face interaction. Suicide risk assessment scales with clinical interview can reveal informative cues for predicting suicide [27]. Tan et al. [28] conducted an interview and survey study in Weibo, a Twitter-like service in China, to explore the engagement of suicide attempters with intervention by direct messages.
传统的自杀检测依赖于临床方法，包括自我报告和面对面访谈。Venek等人[9]设计了一个五项普遍存在的问卷，用于评估自杀风险，并对患者的反应应用分层分类器来确定他们的自杀意图。通过面对面交流，可以利用言语和听觉信息。在一次二元访谈（？？）中，Scherer [23]研究了韵律语音特征和语音质量，以识别自杀和非自杀青少年。其他临床方法通过转换的感应信号[24]检查静息状态心率，并对基于功能磁共振成像的死亡和生命相关词汇[25]的神经表征以及从脑电图信号 [26] 转换的事件相关刺激因素进行分类。临床治疗的另一个方面是对自杀行为背后心理的理解，然而，自杀行为严重依赖于临床医生的知识和面对面的互动。结合临床访谈的自杀风险评估量表可以揭示预测自杀的信息线索[27]。Tan等人[28]在中国类似Twitter的微博上进行了一项采访和调查研究，以探讨自杀未遂者与直接信息干预的关系。

A. Content Analysis内容分析
Users’ post on social websites reveals rich information and their language preferences. Through exploratory data analysis on the user-generated content, we can have an insight into language usage and linguistic clues of suicide attempters. The detailed analysis includes lexicon-based filtering, statistical linguistic features, and topic modeling within suicide-related posts.
用户在社交网站上发布的帖子揭示了丰富的信息和他们的语言偏好。通过对用户生成内容的探索性数据分析，我们可以深入了解自杀未遂者的语言使用和语言线索。详细分析包括基于词汇的过滤、统计语言特征和自杀相关帖子中的主题建模。

Suicide-related keyword dictionary and lexicon are manually built to enable keyword filtering [29], [30] and phrases filtering [31]. Suicide-related keywords and phrases include “kill,” “suicide,” “feel alone,” “depressed,” and “cutting myself.” Vioulès et al. [3] built a pointwise mutual information symptom lexicon using an annotated Twitter data set. Gunn and Lester [32] analyzed posts from Twitter in the 24 h before the death of a suicide attempter. Coppersmith et al. [33] analyzed the language usage of data from the same platform. Suicidal thoughts may involve strong negative feelings, anxiety, hopelessness, or other social factors, such as family and friends. Ji et al. [17] performed word cloud visualization and topics modeling over suicide-related content and found that suicide-related discussion covers personal and social issues. Colombo et al. [34] analyzed the graphical characteristics of connectivity and communication in the Twitter social network. Coppersmith et al. [35] provided an exploratory analysis of language patterns and emotions on Twitter. Other methods and techniques include Google Trends analysis for suicide risk monitoring [36], the reply bias assessment through linguistic clues [37], human–machine hybrid method for analysis of the language effect of social support on suicidal ideation risk [38], social media content detection, and speech patterns analysis [39].
与自杀相关的关键词词典和词典是手动构建的，以启用关键字过滤[29], [30]和短语过滤 [31]。与自杀相关的关键词和短语包括“杀死”、“自杀”、“感觉孤独”、“抑郁”和“割伤自己”。Vioulès等人[3]使用带注释的Twitter数据集建立了一个点式互信息症状词典。Gunn和Lester [32]分析了一名自杀未遂者死亡前24小时来自Twitter的帖子。Coppersmith等人[33]分析了来自同一平台的数据的语言使用情况。自杀念头可能包括强烈的消极情绪、焦虑、绝望或其他社会因素，如家庭和朋友。Ji等人[17]对自杀相关内容进行了词云可视化和主题建模，发现自杀相关讨论涵盖个人和社会问题。 Colombo等人[34]分析了Twitter社交网络中连接和通信的图形特征。Coppersmith等人[35对Twitter上的语言模式和情绪进行了探索性分析。其他方法和技术包括用于自杀风险监测的谷歌趋势分析 [36]、通过语言线索 [37]进行回答偏差评估、用于分析社会支持对自杀意念风险的语言影响的人机混合方法[38]、社交媒体内容检测和语音模式分析 [39]。

B. Feature Engineering特征工程
The goal of text-based suicide classification is to determine whether candidates, through their posts, have suicidal ideations. Machine learning methods and NLP have also been applied in this field.
以文本为基础的自杀分类的目的是通过候选人发布的帖子确认是否有自杀的想法。机器学习方法和NLP方法也被应用于这一领域。

1. Tabular Features: Tabular data for SID consist of questionnaire responses and structured statistical information extracted from websites. Such structured data can be directly used as features for classification or regression. Masuda et al. [40] applied logistic regression to classify suicide and control groups based on users’ characteristics and social behavior variables. The authors found variables such as community number, local clustering coefficient, and homophily have a more substantial influence on suicidal ideation in an SNS of Japan. Chattopadhyay [41] applied pierce suicidal intent scale (PSIS) to assess suicide factors and conducted regression analysis. Questionnaires act as a good source of tabular features. Delgado-Gomez et al. [42] used the international personality disorder examination screening questionnaire and the Holmes–Rahe social readjustment rating scale (SRRS). Chattopadhyay [43] proposed to apply a multilayer feedforward neural network, as shown in Fig. 2(a), to classify suicidal intention indicators according to Beck’s suicide intent scale.
   1） 表格特征：SID的表格数据包括问卷调查结果和从网站中提取的结构化统计信息。这种结构化数据可以直接用作分类或回归的特征。Masuda等人[40] 应用逻辑回归分析，根据用户特征和社会行为变量对自杀组和对照组进行分类。作者发现，在日本的一个社交网站中（Social Networking Services），群体数量、局部聚类系数和同质性等变量对自杀意念有更大的影响。Chattopadhayy [41] 应用皮尔斯自杀意向量表（PSIS）评估自杀因素并进行回归分析。问卷可以作为表格特征的良好来源。Delgado-Gomez等人[42]采用国际人格障碍筛查问卷和霍姆斯-克拉赫社会适应量表（SRRS）。Chattopadhyay [43]建议应用多层前馈神经网络，如图2（a）所示，根据Beck的自杀意图量表对自杀意向指标进行分类。

2. General Text Features: Another direction of feature engineering is to extract features from unstructured text. The main features consist of N-gram features, knowledge-based features, syntactic features, context features, and class-specific features [44]. Abboute et al. [45] built a set of keywords for vocabulary feature extraction within nine suicidal topics. Okhapkina et al. [46] built a dictionary of terms about suicidal content. They introduced term frequency-inverse document frequency (TF-IDF) matrices for messages and a singular value decomposition (SVD) for matrices. Mulholland and Quinn [47] extracted vocabulary and syntactic features to build a classifier to predict the likelihood of a lyricist’s suicide. Huang et al. [48] built a psychological lexicon dictionary by extending HowNet (a commonsense word collection) and used a support vector machine (SVM) to detect cybersuicide in Chinese microblogs. The topic model [49] is incorporated with other machine learning techniques for identifying suicide in Sina Weibo. Ji et al. [17] extracted several informative sets of features, including statistical, syntactic, linguistic inquiry and word count (LIWC), word embedding, and topic features, and then put the extracted features into classifiers, as shown in Fig. 2(b), where four traditional supervised classifiers are compared. Shing et al. [13] extracted several features as a bag of words (BoWs), empath, readability, syntactic features, topic model posteriors, word embeddings, LIWC, emotion features, and mental disease lexicon.
   2） 一般文本特征：特征工程的另一个方向是从非结构化文本中提取特征。主要特征包括N-gram特征、基于知识的特征、句法特征、上下文特征和类特定特征[44]。Aboute等人[45]在9个自杀主题中构建了一组词汇特征提取关键词。Okhapkina等人[46]建立了一本关于自杀内容的词典。他们为消息引入了术语频率逆文档频率（TF-IDF）矩阵和矩阵的奇异值分解（SVD）。Mulholland和Quinn [47] 提取了词汇量和句法特征来构建分类器来预测抒情诗人自杀的可能性。Huang等人[48]通过扩展知网（常识词汇集）建立了心理词典，并使用支持向量机（SVM）对中文微博中的网络自杀进行了检测。主题模型[49]与其他机器学习技术相结合，用于识别新浪微博中的自杀行为。Ji等人[17]提取多个信息特征集，包括统计特征、句法特征、语言查询和词数特征（LIWC）、单词嵌入和主题特征，然后将提取的特征放入分类器中，如图2（b）所示，其中比较了四个传统的监督分类器。Shing等人[13]提取了若干特征作为袋词（BoWs）、移情、可读性、句法特征、主题模式后验、词嵌入、LIWC、情感特征和精神疾病词汇。

   Models for SID with feature engineering include SVM [44], ANNs [50], and conditional random field (CRF) [51]. Tai and Chiu [50] selected several features, including the history of suicide ideation and self-harm behavior, religious belief, family status, mental disorder history of candidates, and their family. Pestian et al. [52] compared the performance of different multivariate techniques with features of word counts, POS, concepts, and readability scores. Similarly, Ji et al. [17] compared four classification methods of logistic regression, random forest, gradient boosting decision tree, and XGBoost. Braithwaite et al. [53] validated that machine learning algorithms can effectively identify high suicidal risk.
   具有特征工程的SID模型包括SVM[44]、ANNs[50]和条件随机场 (CRF) [51]。Tai和Chiu[50] 选择了几个特征，包括自杀意念和自残行为史、宗教信仰、家庭状况、精神障碍史，和他们的家人。Pestian等人 [52] 比较了不同多元技术的表现与字数、词性、概念和可读性得分的特征。同样，Ji等人[17]比较了逻辑回归、随机森林、梯度提升决策树和XGBoost四种分类方法。Braithwaite等人[53]验证了机器学习算法能够有效识别高自杀风险。

3. Affective Characteristics: Affective characteristics are among the most distinct differences between those who attempt suicide and healthy individuals, which has drawn considerable attention from both computer scientists and mental health researchers. To detect the emotions in suicide notes, Liakata et al. [51] used manual emotion categories, including anger, sorrow, hopefulness, happiness/ peacefulness, fear, pride, abuse, and forgiveness. Wang et al. [44] employed combined characteristics of both factual (two categories) and sentimental aspects (13 categories) to discover fine-grained sentiment analysis. Similarly, Pestian et al. [52] identified emotions of abuse, anger, blame, fear, guilt, hopelessness, sorrow, forgiveness, happiness, peacefulness, hopefulness, love, pride, thankfulness, instructions, and information. Ren et al. [14] proposed a complex emotion topic model and applied it to analyze accumulated emotional traits in suicide blogs and to detect suicidal intentions from a blog stream. Especially, the authors studied accumulate emotional traits, including emotion accumulation, emotion covariance, and emotion transition among eight basic emotions of joy, love, expectation, surprise, anxiety, sorrow, anger, and hate with a five-level intensity.
   3） 情感特征：情感特征是自杀未遂者与健康个体之间最显著的差异之一，这引起了计算机科学家和心理健康研究人员的极大关注。为了检测自杀笔记中的情绪，Liakata等人[51] 使用手动情绪类别，包括愤怒、悲伤、希望、快乐/和平、恐惧、骄傲、虐待和宽恕。Wang等人[44]采用事实（两个范畴）和情感方面（13个范畴）的组合特征来发现细粒度的情感分析。同样，Pestian等人[52]识别出虐待、愤怒、责备、恐惧、内疚、绝望、悲伤、宽恕、幸福、和平、希望、爱、骄傲、感恩、指示和信息。Ren等人[14]提出了一个复杂情绪话题模型，并将其应用于分析自杀博客中累积的情绪特征，并从博客流中检测自杀意图。特别地，学者们研究了基于快乐、爱、期待、惊喜、焦虑、悲伤、愤怒、憎恨等八种基本情绪的情绪积累、情绪协方差和情绪转换等情绪积累特征。
   
   图2。特征工程方法说明。（a） 神经网络与特征工程。（b） 特征工程分类器。

C. Deep Learning深度学习
Deep learning has been a great success in many applications, including computer vision, NLP, and medical diagnosis. In the field of suicide research, it is also an important method for automatic SID and suicide prevention. It can effectively learn text features automatically without sophisticated feature engineering techniques. At the same time, some also take extracted features into deep neural networks (DNNs); for example, Nobles et al. [54] fed psycholinguistic features and word occurrence into the multilayer perceptron (MLP). Popular DNNs include convolutional neural networks (CNNs), recurrent neural networks (RNNs), and bidirectional encoder representations from transformers (BERT), as shown in Fig. 3(a)–©. Natural language text is usually embedded into distributed vector space with popular word embedding techniques, such as word2vec [55] and GloVe [56]. Shing et al. [13] applied user-level CNN with the filter size of 3, 4, and 5 to encode users’ posts. The long short-term memory (LSTM) network, a popular variant of RNN, is applied to encode textual sequences and then process for classification with fully connected layers [17].
深度学习在计算机视觉、自然语言处理和医学诊断等许多领域都取得了巨大的成功。在自杀研究领域，它也是自动SID和自杀预防的重要方法。它可以有效地自动学习文本特征且没有复杂的特征工程技术。同时，有的还将提取的特征提取到深度神经网络（DNNs）中；例如，Nobles等人[54]将心理语言学特征和单词出现情况输入多层感知器（MLP）。常用的DNN包括卷积神经网络（CNNs）、递归神经网络（RNNs）和来自transformers的双向编码器表示（BERT），如图3（a）-（C）所示。自然语言文本通常是通过流行的词嵌入技术嵌入到分布式向量空间中，如word2vec [55]和GloVe[56]。Shing等人[13]应用用户级CNN，过滤器大小为3、4和5，对用户帖子进行编码。长短期记忆（LSTM）网络是RNN的一个流行变体，它被用来对文本序列进行编码，通过全连接层进行分类[17]。



Recent methods introduce other advanced learning paradigms to integrate with DNNs for SID. Ji et al. [57] proposed model aggregation methods for updating neural networks, i.e., CNNs and LSTMs, targeting to detect suicidal ideation in private chatting rooms. However, decentralized training relies on coordinators in chatting rooms to label user posts for supervised training, which can only be applied to minimal scenarios. One possible better way is to use unsupervised or semisupervised learning methods. Benton et al. [16] predicted suicide attempt and mental health with neural models under the framework of multitask learning by predicting the gender of users as an auxiliary task. Gaur et al. [58] incorporated external knowledge bases and suicide-related ontology into a text representation and gained an improved performance with a CNN model. Coppersmith et al. [59] developed a deep learning model with GloVe for word embedding, bidirectional LSTM for sequence encoding, and self-attention mechanism for capturing the most informative subsequence. Sawhney et al. [60] used LSTM, CNN, and RNN for SID. Similarly, Tadesse et al. [61] employed the LSTM-CNN model. Ji et al. [62] proposed an attentive relation network with LSTM and topic modeling for encoding text and risk indicators.
最近的方法引入了其他先进的学习范式来与用于SID的DNNs集成。Ji等人[57]提出了用于更新神经网络的模型聚合方法，即CNNs和LSTMs，旨在检测私人聊天室中的自杀意念。然而，分散式训练依赖于聊天室中的协调员为监督训练的用户帖子添加标签，这只能应用于最小的场景。一个可能更好的方法是使用无监督或半监督的学习方法。Benton等人[16]在多任务学习框架下通过预测用户性别作为辅助任务来预测自杀未遂与心理健康的神经网络模型。Gaur等人 [58]将外部知识库和自杀相关本体整合到文本表示中，并使用CNN模型获得了更好的性能。Coppersmith等人 [59] 开发了一个深度学习模型，其中包括用于单词嵌入的GloVe、用于序列编码的双向LSTM以及用于捕获信息量最大的子序列的自我注意机制。Sawhney等人 [60] 使用LSTM、CNN和RNN进行SID。同样，Tadesse等人[61]人采用LSTM-CNN模式。Ji等人[62]提出了一个与LSTM的密切关系网络和用于编码文本和风险指标的主题模型。

In the 2019 CLPsych Shared Task [63], many popular DNN architectures were applied. Hevia et al. [64] evaluated the effect of pretraining using different models, including GRU-based RNN. Morales et al. [65] studied several popular deep learning models, such as CNN, LSTM, and Neural Network Synthesis (NeuNetS). Matero et al. [66] proposed a dual-context model using hierarchically attentive RNN and BERT.
在2019年CLPsych共享任务中[63]，应用了许多流行的DNN架构。Hevia等人[64] 使用不同的模型（包括基于GRU的RNN）评估预训练的效果。Morales等人 [65] 研究了几种流行的深度学习模型，如CNN、LSTM和神经网络综合技术（NeuNetS）。Matero等人[66]提出了一种基于分层注意RNN和BERT的双重语境模型。

Another subdirection is the so-called hybrid method that cooperates minor feature engineering with representation learning techniques. Chen et al. [67] proposed a hybrid classification model of the behavioral model and the suicide language model. Zhao et al. [68] proposed the D-CNN model taking word embedding and external tabular features as inputs for classifying suicide attempters with depression.
另一个分支是所谓的混合方法，它将次要特征工程与表示学习技术相结合。Chen等人[67]提出了行为模型和自杀语言模型的混合分类模型。Zhao等 [68]提出了以词嵌入和外部表格特征为输入的D-CNN模型来对自杀未遂者进行抑郁分类。

D. Summary总结
The popularization of machine learning has facilitated research on SID from multimodal data and provided a promising way for effective early warning. Current research focuses on text-based methods by extracting features and deep learning for automatic feature learning. Researchers widely use many canonical NLP features, such as TF-IDF, topics, syntactic, affective characteristics, readability, and deep learning models, such as CNN and LSTM. Those methods, especially DNNs with automatic feature learning, boosted predictive performance and preliminary success on suicidal intention understanding. However, some methods may only learn statistical cues and lack of commonsense. The recent work [58] incorporated external knowledge using knowledge bases and suicide ontology for knowledge-aware suicide risk assessment. It took a remarkable step toward knowledge-aware detection.
机器学习的普及促进了多模态数据SID的研究，为有效的早期预警提供了一条有希望的途径。目前的研究主要集中在基于文本的特征提取和自动特征学习的深度学习方法上。研究者们广泛地使用了许多典型的NLP特征，如TF-IDF、主题、句法、情感特征、可读性和深度学习模型，如CNN和LSTM。这些方法，特别是具有自动特征学习的DNNs，提高了自杀意图的预测能力和在自杀意图理解上取得初步成功。然而，有些方法可能只学习统计线索，缺乏常识。最近的工作 [58]结合了外部知识，使用知识库和自杀本体来进行有知识意识的自杀风险评估。它朝着知识感知检测迈出了显著的一步。

3. APPLICATIONS ON DOMAINS
Many machine learning techniques have been introduced for SID. The relevant extant research can also be viewed according to the data source. Specific applications cover a wide range of domains, including questionnaires, electronic health records (EHRs), suicide notes, and online user content. Fig. 4 shows some examples of data source for SID, where Fig. 4(a) lists selected questions of the “International Personal Disorder Examination Screening Questionnaire” (IPDE-SQ) adapted from [42], Fig. 4(b) shows selected patient’s records from [69], Fig. 4© shows a suicide note from a website, 5 and Fig. 4(d) shows a tweet and its corresponding comments from Twitter.com. Nobles et al. [54] identified suicide risk using text messages. Some researchers also developed softwares for suicide prevention. Berrouiguet et al. [70] developed a mobile application for health status self-report. Meyer et al. [71] developed an e-PASS Suicidal Ideation Detector (eSID) tool for medical practitioners. Shah et al. [72] utilized social media videos and studied multimodal behavioral markers.
许多机器学习技术已经被引入到SID中。现有的相关研究也可以根据数据来源进行查看。具体的应用程序涵盖了广泛的领域，包括问卷调查、电子健康记录（EHRs）、自杀笔记和在线用户内容。图4显示了SID的一些数据源示例，其中图4（a）列出了国际个人障碍检查筛查问卷（IPDE-SQ）的部分问题，该问卷来自文献[42]，图4（b）选自文献[69]中患者的记录，图4（c）显示了一个网站的自杀笔记，图4（d）显示了一条来自推特的帖子和相应的评论。Nobles等人[54]通过短信识别自杀风险。一些研究人员还开发了自杀预防软件。Berrouiguet等人[70] 开发了一个用于健康状况自我报告的移动应用程序。Meyer等人。71为执业医师开发了一种e-PASS自杀意念检测仪（eSID）工具。Shah等人[72]利用社交媒体视频并研究多模态行为标记。

图4. SID的内容示例（a） 问卷调查（b） 电子健康记录（c） 自杀笔记（d） 推特

A. Questionnaires
Mental disorder scale criteria, such as DSM-IV, 6 ICD-10,7 and the IPDE-SQ, provides good tool for evaluating an individual’s mental status and their potential for suicide. Those criteria and examination metrics can be used to design questionnaires for self-measurement or face-to-face clinician-patient interview.
精神障碍量表标准，如DSM-IV、ICD-10、IPDE-SQ，为评估个人的心理状态和自杀可能性提供了很好的工具。这些标准和检查指标可用于设计自测问卷或临床诊断。

To study the assessment of suicidal behavior, Delgado-Gomez et al. [10] applied and compared the IPDE-SQ and the “Barrat’s Impulsiveness Scale” (version 11, BIS-11) to identify people likely to attempt suicide. The authors also conducted a study on individual items from those two scales. The BIS-11 scale has 30 items with four-point ratings, while the IPDE-SQ in DSM-IV has 77 true-false screening questions. Furthermore, Delgado-Gomez et al. [42] introduced the “Holmes–Rahe SRRS” and the IPDE-SQ as well to two comparison groups of suicide attempters and nonsuicide attempters. The SRRS consists of 43 ranked life events of different levels of severity. Harris et al. [73] surveyed understanding suicidal individuals’ online behaviors to assist suicide prevention. Sueki [74] conducted an online panel survey among Internet users to study the association between suicide-related Twitter use and suicidal behavior. Based on the questionnaire results, they applied several supervised learning methods, including linear regression, stepwise linear regression, decision trees, Lars-en, and SVMs, to classify suicidal behaviors.
为了研究自杀行为的评估，Delgado Gomez等人 [10] 应用并比较了IPDE-SQ和Barrat的冲动性量表（第11版，BIS-11）来识别可能企图自杀的人。作者还对这两个量表中的个别项目进行了研究。BIS-11量表有30个项目为4分评级，而DSM-IV中的IPDE-SQ有77个对错筛选问题。此外，Delgado Gomez等人 [42] 引入了福尔摩斯-克拉赫SRRS和IPDE-SQ，以及自杀未遂者和非自杀未遂者的两个比较组。SRRS由43个不同严重程度的生活事件组成。Harris等人 [73] 全面了解了自杀个人的在线行为来帮助预防自杀。Sueki [74] 在互联网用户中进行了一项在线小组调查，研究自杀相关Twitter的使用与自杀行为之间的关系。基于问卷调查结果，他们运用了几种监督学习方法，包括线性回归、逐步线性回归、决策树、Lars-en和支持向量机对自杀行为进行分类。

B. Electronic Health Records
The increasing volume of EHRs has paved the way for machine learning techniques for suicide attempter prediction. Patient records include demographical information and diagnosis-related history, such as admissions and emergency visits. However, due to the data characteristics, such as sparsity, variable length of clinical series, and heterogeneity of patient records, many challenges remain in modeling medical data for suicide attempt prediction. Besides, the recording procedures may change because of the change of healthcare policies and the update of diagnosis codes.
EHRs的不断增加为机器学习技术预测自杀未遂者铺平了道路。患者记录包括人口统计信息和与诊断相关的历史记录，如入院和急诊。然而，由于数据的稀疏性、临床序列长度的不确定性以及患者记录的异质性等特点，对自杀未遂预测的医学数据建模仍然存在许多挑战。此外，由于医疗政策的变化和诊断代码的更新，记录程序可能会发生变化。

There are several works of predicting suicide risk based on EHRs [75], [76]. Tran et al. [69] proposed an integrated suicide risk prediction framework with a feature extraction scheme, risk classifiers, and risk calibration procedure. Explicitly, each patient’s clinical history is represented as a temporal image. Iliou et al. [77] proposed a data preprocessing method to boost machine learning techniques for suicide tendency prediction of patients suffering from mental disorders. Nguyen et al. [78] explored real-world administrative data of mental health patients from the hospital for short- and medium-term suicide risk assessments. By introducing random forests, gradient boosting machines, and DNNs, the authors managed to deal with high dimensionality and redundancy issues of data. Although the previous method gained preliminary success, Iliou et al. [77] and Nguyen et al. [78] have a limitation on the source of data, which focuses on patients with mental disorders in their historical records. Bhat and Goldman-Mellor [79] used an anonymized general EHR data set to relax the restriction on patient’s diagnosis-related history and applied neural networks as a classification model to predict suicide attempters.
有几项工作是基于EHRs来预测自杀风险的[75], [76]。Tran等人[69] 提出了一个结合特征提取主题、风险分类和风险校准程序的自杀风险预测框架。明确地说，每个病人的临床病史都被表示为一个时间图像。Iliou等人[77] 提出了一种数据预处理方法，以促进机器学习技术对心理障碍患者自杀倾向的预测。Nguyen等人[78] 对医院的心理健康患者的实际管理数据进行了短期和中期自杀风险评估。通过引入随机森林、梯度提升机和DNNs，作者成功地处理了数据的高维和冗余问题。虽然前面的方法取得了初步成功，但Iliou等人[77]和Nguyen等人[78]在数据来源上存在局限性，这些数据集中在历史记录中有心理 障碍的患者。Bhat 和 Goldman-Mellor[79] 使用匿名的一般EHR数据集来放宽对患者诊断相关历史的限制，并应用神经网络作为分类模型来预测自杀未遂者。

C. Suicide Notes
Suicide notes are the written notes left by people before committing suicide. They are usually written on letters and online blogs and recorded in audio or video. Suicide notes provide material for NLP research. Previous approaches have examined suicide notes using content analysis [52], sentiment analysis [44], [80], and emotion detection [51]. Pestian et al. [52] used transcribed suicide notes with two groups of completers and elicitors from people who have a personality disorder or potential morbid thoughts. White and Mazlack [81] analyzed word frequencies in suicide notes using a fuzzy cognitive map to discern causality. Liakata et al. [51] employed machine learning classifiers to 600 suicide messages with varied length, different readability quality, and multiclass annotations.
自杀笔记是人们在自杀前留下的书面笔记。它们通常写在信件和在线博客上，并以音频或视频形式记录下来。自杀笔记为NLP研究提供了素材。以前的方法使用内容分析 [52]、情绪分析 [44], [80] 和情绪检测 [51] 来研究自杀笔记。Pestian等人 [52] 人使用了两组完成者和激发者的转录自杀笔记，这些人来自有人格障碍或潜在病态想法的人。White和Mazlack[81]使用模糊认知图分析自杀笔记中的词频，以辨别因果关系。Liakata等人[51]人使用机器学习分类器对600条不同长度、不同可读性质量和多类注释的自杀信息进行分类。

Emotion in text provides sentimental cues of suicidal ideation understanding. Desmet and Hoste [82] conducted a fine-grained emotion detection on suicide notes of 2011 i2b2 task. Wicentowski and Sydes [83] used an ensemble of maximum entropy classification. Wang et al. [44] and Kovaˇcevi´ c et al. [84] proposed hybrid machine learning and rule-based method for the i2b2 sentiment classification task in suicide notes.
文本中的情感提供了理解自杀意念的情感线索。Desmet和Hoste [82]对2011年i2b2任务的自杀笔记进行了细粒度的情绪检测。Wicentowski和Sydes [83] 使用了最大熵分类的集合。Wang等人[44]和Kovaˇcevi´ c等人[84] 提出了自杀笔记i2b2情感分类任务的混合机器学习和基于规则的方法。

In the age of cyberspace, more suicide notes are now written in the form of web blogs and can be identified as carrying the potential risk of suicide. Huang et al. [29] monitored online blogs from MySpace.com to identify at-risk bloggers. Schoene and Dethlefs [85] extracted linguistic and sentiment features to identity genuine suicide notes and comparison corpus.
在网络时代，越来越多的自杀笔记是以网络博客的形式写出来的，可以被认定有潜在的自杀风险。Huang等人[29]监控在线博客MySpace.com网站中有风险的博主。Schoene和Dethlefs [85] 提取了语言和情感特征来识别真实的自杀笔记并比较了语料库。

D. Online User Content
The widespread use of mobile Internet and social networking services facilitates people’s expressing their life events and feelings freely. As social websites provide an anonymous space for online discussion, an increasing number of people suffering from mental disorders turn to seek for help. There is a concerning tendency that potential suicide victims post their suicidal thoughts on social websites, such as Facebook, Twitter, Reddit, and MySpace. Social media platforms are becoming a promising tunnel for monitoring suicidal thoughts and preventing suicide attempts [86]. Massive user-generated data provide a good source to study online users’ language patterns. Using data mining techniques on social networks and applying machine learning techniques provide an avenue to understand the intent within online posts, provide early warnings, and even relieve a person’s suicidal intentions.
移动互联网和社交网络服务的广泛应用，使人们能够自由地表达自己的生活事件和情感。由于社交网站提供了一个匿名的在线讨论空间，越来越多的精神障碍患者求助于寻求帮助。有一种令人关注的倾向是，潜在的自杀者会在社交网站上发布他们的自杀想法，比如Facebook、Twitter、Reddit和MySpace。社交媒体平台成为监测自杀想法和防止自杀未遂的一条有希望的途径[86]。大量的用户生成的数据为研究在线用户的语言模式提供了一个很好的来源。在社交网络上使用数据挖掘技术和应用机器学习技术提供了一种途径，可以了解在线帖子中的潜在意图，提供早期预警，甚至可以缓解一个人的自杀意念。

Twitter provides a good source for research on suicidality. O’Dea et al. [12] collected tweets using the public API and developed automatic suicide detection by applying logistic regression and SVM on TF-IDF features. Wang et al. [87] further improved the performance with effective feature engineering. Shepherd et al. [88] conducted psychology-based data analysis for contents that suggests suicidal tendencies in Twitter social networks. The authors used the data from an online conversation called #dearmentalhealthprofessionals.
Twitter为研究自杀行为提供了一个很好的来源。O’Dea等人[12]使用公共API收集推文，并在TF-IDF特征上应用逻辑回归和支持向量机开发了自动自杀检测。Wang等人[87]通过有效的特征工程进一步提高了性能。Shepherd等人[88]对Twitter社交网络中暗示自杀倾向的内容进行了基于心理学的数据分析。作者使用了一个名为dearmentalhealthprofessionals的在线对话中的数据。

Another famous platform Reddit is an online forum with topic-specific discussions has also attracted much research interest for studying mental health issues [89] and suicidal ideation [37]. A community on Reddit called SuicideWatch is intensively used for studying suicidal intention [17], [90]. De Choudhury et al. [90] applied a statistical methodology to discover the transition from mental health issues to suicidality. Kumar et al. [91] examined the posting activity following the celebrity suicides, studied the effect of celebrity suicides on suicide-related contents, and proposed a method to prevent the high-profile suicides.
另一个著名的平台Reddit是一个具有主题特定讨论的在线论坛，也吸引了许多研究兴心理健康问题 [89] 和自杀意念[37]的学者。Reddit上的一个名为“自杀观察”的社区被集中用于研究自杀意图 [17], [90]。De Choudhury等人[90]应用一种统计方法来发现从心理健康问题到自杀的转变。Kumar等人[91]调查了名人自杀后的发帖活动，研究了名人自杀对自杀相关内容的影响，并提出了防止高调自杀的方法。

Many pieces of research [48], [49] work on detecting suicidal ideation in Chinese microblogs. Guan et al. [92] studied user profile and linguistic features for estimating suicide probability in Chinese microblogs. There also remains some work using other platforms for SID. For example, Cash et al. [93] conducted a study on adolescents’ comments and content analysis on MySpace. Steaming data provides a good source for user pattern analysis. Vioulès et al. [3] conducted user- and post-centric behavior analysis and applied a martingale framework to detect sudden emotional changes in the Twitter data stream for monitoring suicide warning signs. Ren et al. [14] used the blog stream collected from public blog articles written by suicide victims to study the accumulated emotional information.
许多研究[48], [49] 都致力于检测中国微博中的自杀意念。Guan等人[92] 研究了中国微博中用于估计自杀概率的用户特征和语言特征。还有一些工作需要使用其他平台来实现SID。例如，Cash等人 [93]在MySpace上进行了一项青少年评论和内容分析的研究。流式数据为用户模式分析提供了一个很好的来源。Vioulès等人 [3] 进行了用户和后中心行为分析，并应用鞅框架检测Twitter数据流中的突然情绪变化，以监控自杀警告信号。Ren等人[14] 利用从自杀者撰写的公开博客文章中收集的博客流来研究积累的情感信息。

E. Summary
Applications of SID mainly consist of four domains: questionnaires, EHRs, suicide notes, and online user content. Table II gives a summary of categories, data sources, and methods. Among these four main domains, questionnaires and EHRs require self-report measurement or patient–clinician interactions and rely highly on social workers or mental health professions. Suicide notes have a limitation on immediate prevention, as many suicide attempters commit suicide in a short time after they write suicide notes. However, they provide a good source for content analysis and the study of suicide factors. The last online user content domain is one of the most promising ways of early warning and suicide prevention when empowered with machine learning techniques. With the rapid development of digital technology, user-generated content will play a more important role in SID. Other forms of data, such as health data generated by wearable devices, can be very likely to help with suicide risk monitoring in the future.
SID的应用主要包括四个领域：调查问卷、电子健康记录、自杀笔记和在线用户内容。表二总结了分类、数据来源和方法。在这四个主要领域中，问卷调查和电子健康记录需要自我报告测量或患者-临床互动，并且高度依赖社会工作者或心理健康专业人员。自杀笔记对即时预防有局限性，因为许多自杀者在写自杀笔记后很短时间内自杀。然而，它们为内容分析和自杀因素研究提供了一个很好的来源。最后一个用机器学习研究在线用户内容是一个最有希望的早期预警和自杀预防方法。随着数字技术的飞速发展，用户生成内容将在SID中扮演越来越重要的角色。其他形式的数据，例如由可穿戴设备生成的健康数据，很有可能在未来帮助监测自杀风险。

IV. TASKS AND DATA SETS
In this section, we summarize specific tasks in SID and other suicide-related tasks about mental disorders. Some tasks, such as reasoning suicidal messages, generating a response, and suicide attempters’ detection on a social graph, may lack benchmarks for evaluation. However, they are critical for effective detection. We propose these tasks together with the current research direction and call for contribution to these tasks from the research community. Meanwhile, an elaborate list of data sets for currently available tasks is provided, and some potential data sources are also described to promote the research efforts.
在这一节中，我们将总结SID中的具体任务和其他与自杀有关的心理障碍任务。有些任务，如推理自杀信息、生成响应、在社交图上检测企图自杀者，可能缺乏评估基准。然而，它们对有效检测至关重要。我们提出这些任务和当前的研究方向，并呼吁研究界为这些任务作出贡献。同时，为当前可用任务提供了一份精心制作的数据集列表，并对一些潜在的数据源进行了描述，以促进研究工作的开展。

A. Tasks

1. Suicide Text Classification: The first task—suicide text classification—can be viewed as a domain-specific application of general text classification, which includes binary and multiclass classification. Binary suicidality classification simply determines text with suicidal ideation or not, while multiclass suicidality classification conducts fine-grained suicide risk assessment. For example, some studies divide suicide risk into four levels: no, low, moderate, and severe. Alternatively, it can also consider four types of class labels according to mental and behavioral procedures, i.e., nonsuicidal, suicidal thoughts/wishes, suicidal intentions, and suicidal act/plan.
   1） 自杀文本分类：自杀文本分类的第一个任务可以看作是通用文本分类的一个特定领域的应用，包括二元分类和多元分类。二元自杀分类只是简单地确定是否有自杀意念的文本，而多元自杀分类则进行细粒度的自杀风险评估。例如，一些研究将自杀风险分为四个等级：没有、低、中、重度。或者，它也可以根据心理和行为程序考虑四种类型的分类标签，即非自杀性的、自杀性的想法/愿望、自杀意图和自杀行为/计划。

Another subtask is risk assessment by learning from multiaspect suicidal posts. Adopting the definition of characteristics of suicidal messages, Gilat et al. [94] manually tagged suicidal posts with multiaspect labels, including mental pain, cognitive attribution, and level of suicidal risk. Mental pain includes the loss of control, acute loneliness, emptiness, narcissistic wounds, irreversibility loss of energy, and emotional flooding, scaled into [0, 7]. Cognitive attribution is the frustration of needs associated with interpersonal relationships, or there is no indication of attribution.
另一个子任务是通过学习多方面自杀帖子进行风险评估。采用自杀信息特征的定义，Gilat等人[94]给多方面的自杀帖子人工贴标签，包括精神痛苦、认知归因和自杀风险水平。心理痛苦包括失控、极度孤独、空虚、自恋的创伤、不可逆转的能量损失和情绪激动，按比例分为0至7分。认知归因是指与人际关系相关的需求的挫折感，或没有归因的迹象。

2. Reasoning Suicidal Messages: Massive data mining and machine learning algorithms have achieved remarkable outcomes by using DNNs. However, simple feature sets and classification models are not predictive enough to detect complicated suicidal intentions. Machine learning techniques require reasoning suicidal messages to have a more in-depth insight into suicidal factors and the innermost being from textual posts. This task aims to employ interpretable methods to investigate suicidal factors and incorporate them with commonsense reasoning, which may improve the prediction of suicidal factors. Specific tasks include automatic summarization of suicide factor, finding an explanation of suicidal risk in mental pain, and cognitive attribution aspects associated with suicide.
   2） 推理自杀信息：大量数据挖掘和机器学习算法使用DNNs取得了显著的成果。然而，简单的特征集和分类模型不足以预测复杂的自杀意图。机器学习技术需要推理自杀信息，以便从文本帖子中更深入地洞察自杀因素和内心深处。本研究旨在运用可解释的方法来研究自杀因素，并将其与常识推理相结合，以提高自杀因素的预测能力。具体任务包括自动总结自杀因素，寻找心理疼痛中自杀风险的解释，以及与自杀相关的认知归因方面。

3. Suicide Attempter Detection: The two tasks mentioned earlier focus on a single text itself. However, the primary purpose of SID is the identify suicide attempters. Thus, it is vital to achieving user-level detection, which consists of two folds, i.e., user-level multiinstance suicidality detection and suicide attempt detection on a graph. The former takes a bag of posts from individuals as input and conducts multiinstance learning over a bag of messages. The later identifies suicide attempters in a specific social graph built by the interaction between users in social networks. It considers the relationship between social users and can be regarded as a node classification problem in a graph.
   3） 自杀企图者检测：前面提到的两个任务集中在单个文本本身。然而，SID的主要目的是识别自杀企图者。因此，实现用户级检测至关重要，它包括两个方面，图上的用户级多实例自杀行为检测和自杀企图检测。前者以个人的一袋帖子作为输入，通过一袋消息进行多实例学习。后者识别自杀企图者在一个特定的社会关系图中由用户在社交网络中的互动建立。它考虑了社会用户之间的关系，可以看作是一个图中的节点分类问题。

4. Generating Response: The ultimate goal of SID is intervention and suicide prevention. Many people with suicidal intentions tend to post their suffering at midnight. Another task is generating a thoughtful response for counseling potential suicidal victims to enable immediate social care and relieve their suicidal intention. Gilat et al. [94] introduced eight types of response strategies; they are emotional support, offering group support, empowerment, interpretation, cognitive change inducement, persuasion, advising, and referring. This task requires machine learning techniques, especially sequence-to-sequence learning, to have the ability to adopt effective response strategies to generate better response and eliminate people’s suicidality. When social workers or volunteers go back online, this response generation technique can also generate hints for them to compose a thoughtful response.
   4） 生成响应：SID的最终目标是干预和自杀预防。许多有自杀意图的人往往会在午夜发帖。另一个任务是劝告潜在的自杀者，并产生一个深思熟虑的回应，使他们能够立即得到社会照顾，并减轻他们的自杀意图。Gilat等人 [94] 提出了八种应对策略：情感支持、提供群体支持、授权、解释、认知变化诱导、说服、建议和咨询。这项任务需要机器学习技术，特别是序列到序列的学习，能够采取有效的反应策略来产生更好的反应，消除人们的自杀行为。当社会工作者或志愿者回到网上时，这种生成响应技术也能产生提示，让他们做出深思熟虑的回应。

5. Mental Disorders and Self-Harm Risk: Suicidal ideation has a strong relationship with a mental health issue and self-harm risks. Thus, detecting severe mental disorders or self-harm risks is also an important task. Such works include depression detection [95], self-harm detection [96], stressful periods and stressor events detection [97], building knowledge graph for depression [98], and correlation analysis on depression and anxiety [99]. Corresponding subtasks in this field are similar to suicide text classification in Section IV-A1.
   5） 精神障碍和自残风险：自杀意念与心理健康问题和自残风险有密切关系。因此，发现严重精神障碍或自残风险也是一项重要任务。这些工作包括抑郁检测[95]、自残检测[96]、应激期和应激事件检测 [97]、构建抑郁知识图[98]、抑郁与焦虑相关分析[99]。该领域的相应子任务类似于第IV-A1节中的自杀文本分类。

B. Data Sets

1. Text Data:
   a) Reddit: Reddit is a registered online community that aggregates social news and online discussions. It consists of many topic categories, and each area of interest within a topic is called a subreddit. A subreddit called “Suicide Watch” (SW) 8 is intensively used for further annotation as positive samples. Posts without suicidal content are sourced from other popular subreddits. Ji et al. [17] released a data set with 3549 posts with suicidal ideation. Shing et al. [13] published their UMD Reddit Suicidality data set with 11129 users and total 1 556 194 posts and sampled 934 users for further annotation. Alada˘ g et al. [100] collected 508 398 posts using Google Cloud BigQuery and manually annotated 785 posts.
   a） Reddit:Reddit是一个注册的在线社区，收集社会新闻和在线讨论。由许多主题类别组成，主题中的每个感兴趣的区域都称为subreddit。一个名为自杀观察（SW）的subreddit被密集地用作阳性样本的进一步注释。没有自杀内容的帖子来源于其他热门网站。Ji等人[17]公布了一组有3549个有自杀意念的帖子。Shing等人[13]发布了他们的UMD Reddit自杀率数据集，有11129个用户和总计1556194个帖子，并抽样934个用户进行进一步注释。Alada˘ g等人[100]使用google cloud bigquery收集了508398个帖子，并手动标注了785个帖子。

b) Twitter: Twitter is a popular social networking service, where many users also talk about their suicidal ideation. Twitter is quite different from Reddit in post length, anonymity,and the way communication and interaction. Twitter user data with suicidal ideation and depression are collected by Coppersmith et al. [33]. Ji et al. [17] collected an imbalanced data set of 594 tweets with suicidal ideation out of a total of 10 288 tweets. Vioulès et al. collected 5446 tweets using Twitter streaming API [3], of which 2381 and 3065 tweets are from the distressed users and normal users, respectively. However, most Twitter-based data sets are no longer available as per the policy of Twitter.
b） Twitter: Twitter是一个很受欢迎的社交网络服务，很多用户也会谈论他们的自杀意图。Twitter与Reddit在帖子长度、匿名性、以及沟通和互动的方式有很大不同。Coppersmith等人 [33]收集了Twitter用户自杀和抑郁的数据。Ji等人[17]收集了一组不平衡的数据集，在总共10288条推文中，有自杀念头的推文占594条。Vioulès等人[3]使用TwitterAPI收集了5446条推文，其中2381条和3065条tweet分别来自苦恼用户和正常用户。然而，根据Twitter的政策，大多数基于Twitter的数据集不再可用。

c) ReachOut: ReachOut Forum9 is a peer support platform provided by an Australian mental health care organization. The ReachOut data set [101] was first released in the CLPsych17 shared task. Participants were initially given a training data set of 65 756 forum posts, of which 1188 were annotated manually with the expected category, and a test set of 92 207 forum posts, of which 400 were identified as requiring annotation. The specific four categories are described as follows.
c） ReachOut: ReachOut论坛是一个由澳大利亚心理卫生保健组织提供的同侪支持平台。ReachOut数据集 [101]首次在CLPsych17共享任务中发布。参与者最初获得的训练数据集包括65756个论坛帖子，其中1188个帖子被手动标注为预期类别，另一个测试集为92207个论坛帖子，其中400个被确定需要注释。具体的四个类别如下。

1. Crisis: The author or someone else is at risk of harm.

2. Red: The post should be responded to as soon as possible.

3. Amber: The post should be responded to at some point if the community does not rally strongly around it.

4. Green: The post can be safely ignored or left for the community to address.
   1） 危机：作者或其他人有受到伤害的危险。
   2） 红色：帖子应该尽快回复。
   3） 琥珀色：如果社区没有强烈的关注它，这个帖子应该在某个时候得到回应。
   4） 绿色：这个帖子可以被安全地忽略或者留给社区来解决。

5. EHR: EHR data contain demographical information, admissions, diagnostic reports, and physician notes. A collection of EHRs is from the California emergency department encounter and hospital admission. It contains 522 056 anonymous EHR records from California-resident adolescents. However, it is not public for access. Bhat and Goldman-Mellor [79] first used these records from 2006 to 2009 to predict the suicide attempt in 2010. Haerian et al. [102] selected 280 cases for evaluation from the clinical data warehouse (CDW) and WebCIS database at the NewYork Presbyterian Hospital/Columbia University Medical Center. Tran et al. [69] studied emergency attendances with a least one risk assessment from the Barwon Health data warehouse. The selected data set contains 7746 patients and 17771 assessments.
   2） EHR：EHR数据包含人口统计信息、入院、诊断报告和医生注释。一组EHR是从加州急诊科和住院部收集的。它包含522056份来自加利福尼亚州居住青少年的匿名EHR记录。但是，它不是公开的。Bhat和Goldman-Mellor [79] 首次利用2006年至2009年的这些记录预测2010年的自杀企图。Haerian等人[102]从纽约长老会医院/哥伦比亚大学医学中心的临床数据仓库（CDW）和WebCIS数据库中选择280例进行评估。Tran等人 [69] 研究了Barwon健康数据仓库紧急状态下至少一个风险评估。所选数据集包含7746名患者和17771个评估。

6. Mental Disorders: Mental health issues, such as depression without effective treatment, can turn into suicidal ideation. For the convenience of research on mental disorders, we also list several resources for monitoring mental disorders. The eRisk data set of early detection of signs of depression [103] is released by the first task of the 2018 workshop at the Conference and Labs of the Evaluation Forum (CLEF), which focuses on early risk prediction on the Internet. 10 This data set contains sequential text from social media. Another data set is the Reddit Self-reported Depression Diagnosis (RSDD) data set [95], which contains 9000 diagnosed users with depression
   and approximately 107000 matched control users.
   3） 精神障碍：精神健康问题，如抑郁症，如果没有有效的治疗，会演变成自杀意念。为了方便精神障碍研究，我们还列出了几种精神障碍监测来源。早期发现抑郁症征兆的eRisk数据集 [103] 是由评估论坛会议和实验室（CLEF）2018年研讨会的第一项任务发布的，该论坛专注于互联网上的早期风险预测。此数据集包含来自社交媒体的连续文本。另一个数据集是Reddit自报告抑郁症诊断（RSDD）数据集 [95]，其中包含9000名抑郁症诊断用户和大约107000名匹配的对照用户。

V. DISCUSSION AND FUTURE WORK
Many preliminary works have been conducted for SID, especially boosted by manual feature engineering and DNN-based representation learning techniques. However, current research has several limitations, and there are still great challenges for future work.
SID已经做了很多前期工作，特别是人工特征工程和基于DNN的表示学习技术。然而，目前的研究存在一些局限性，对今后的工作仍有很大的挑战。

A. Limitations

1. Data Deficiency: The most critical issue of current research is data deficiency. Current methods mainly apply supervised learning techniques that require manual annotation. However, there are not enough annotated data to support further research. For example, labeled data with fine-grained suicide risk only have limited instances, and there are no multiaspect data and data with social relationships.
   1） 数据不足：当前研究中最关键的问题是数据不足。目前的方法主要采用有监督的学习技术，需要人工标注。然而，没有足够的注释数据来支持进一步的研究。例如，带有细粒度自杀风险的标签数据只有有限的实例，没有多方面的数据和具有社会关系的数据。

2. Annotation Bias: There is little evidence to confirm the suicide action to obtain ground truth. Thus, current data are obtained by manual labeling with some predefined annotation rules. The crowdsourcing-based annotation may lead to bias of labels. Shing et al. [13] asked experts for labeling but only obtained a limited number of labeled instances. As for the demographical data, the quality of suicide data is concerning, and mortality estimation is general death but not suicide. 11 Some cases are misclassified as accidents or death of undetermined intent.
   2） 注释偏差：几乎没有证据证实自杀行为以获得基本事实。因此，当前数据是通过使用一些预定义的注释规则手动标记来获得的。基于注释的众包可能导致标签的偏差。Shing等人 [13] 请求专家贴标签，但得到的标签数量有限。至于人口数据，自杀数据的质量是相关的，死亡率估计是一般死亡，而不是自杀。有些案例被错误地归类为意外事故或不明原因死亡。

3. Data Imbalance: Posts with suicidal intention account for a tiny proportion of massive social posts. However, most works built data sets in an approximately even manner to collect relatively balanced positive and negative samples rather than treating it as an ill-balanced data distributed.
   3） 数据不平衡：有自杀意图的帖子在大量社交帖子中所占比例很小。然而，大多数工作是以近似均匀的方式建立数据集，以收集相对平衡的正负样本，而不是将其视为分布不平衡的数据。

) Lack of Intention Understanding: The current statistical learning method failed to have a good understanding of suicidal intention. The psychology behind suicidal attempts is complex. However, mainstream methods focus on selecting features or using complex neural architectures to boost the predictive performance. From the phenomenology of suicidal posts in social content, machine learning methods learned statistical clues. However, they failed to reason over the risk factors by incorporating the psychology of suicide.
4） 意图理解不足：目前的统计学习方法未能很好地理解自杀意图。自杀企图背后的心理是复杂的。然而，主流的方法主要集中在选择特征或使用复杂的神经网络结构来提高预测性能。从自杀帖子的现象学社会内容中，机器学习方法学习统计线索。然而，他们没有合并心理学来考虑自杀的危险因素。

B. Future Work

1. Emerging Learning Techniques: The advances of deep learning techniques have boosted research on SID. More emerging learning techniques, such as attention mechanism and graph neural networks, can be introduced for suicide text representation learning. Other learning paradigms, such as transfer learning, adversarial training, and reinforcement learning, can also be utilized. For example, knowledge of the mental health detection domain can be transferred for SID, and generative adversarial networks can be used to generate adversarial samples for data augmentation.
   1） 新兴学习技术：深度学习技术的进步推动了对SID的研究。更多新兴的学习技术，如注意机制和图神经网络，可以引入自杀文本表征学习。其他的学习模式，如迁移学习，对抗训练，强化学习，也可以利用。例如，心理健康检测领域的知识可用于SID，生成性对抗网络可用于生成对抗性样本以进行数据扩充。

In social networking services, posts with suicidal ideation are in the long tail of the distribution of different post categories. To achieve effective detection in the ill-balanced distribution of real-world scenarios, few-shot learning can be utilized to train on a few labeled posts with suicidal ideation among the large social corpus.
在社交网络服务中，有自杀意念的帖子在不同帖子类别的分布中处于长尾。为了在真实场景的不平衡分布中实现有效的检测，可以利用少样本学习来训练大量社会群体中带有自杀意念的标签帖子。

2. Suicidal Intention Understanding and Interpretability: Many factors are correlated with suicide, such as mental health, economic recessions, gun prevalence, daylight patterns, divorce laws, media coverage of suicide, and alcohol use. A better understanding of suicidal intention can provide a guideline for effective detection and intervention. A new research direction is to equip deep learning models with commonsense reasoning, for example, by incorporating external suicide-related knowledge bases.
   2） 自杀意图的理解和解释：许多因素与自杀有关，如心理健康、经济衰退、枪支泛滥、日光模式、离婚法律、媒体对自杀的报道以及饮酒。**更好地了解自杀意图可以为有效的发现和干预提供指导。**一个新的研究方向是用常识推理来装备深度学习模型，例如通过整合外部自杀相关知识库。

Deep learning techniques can learn an accurate prediction model. However, this would be a black-box model. In order to better understand people’s suicidal intentions and have a reliable prediction, new interpretable models should be developed.
深度学习技术可以学习一个精确的预测模型。然而，这将是一个黑盒模型。为了更好地了解人们的自杀意图并做出可靠的预测，需要开发新的可解释模型。

3. Temporal Suicidal Ideation Detection: Another direction is to detect suicidal ideation over the data stream and consider the temporal information. There exist several stages of suicide attempts, including stress, depression, suicidal thoughts, and suicidal plan. Modeling people’s posts’ temporal trajectory can effectively monitor the change of mental status and is essential for detecting early signs of suicidal ideation.
   3） 暂时性自杀意念侦测：另一个方向是透过数据流侦测自杀意念，并考虑时间资讯。自杀企图有几个阶段，包括压力、抑郁、自杀念头和自杀计划。对人的后颞轨迹进行建模，可以有效地监测心理状态的变化，对早期发现自杀意念的迹象至关重要。

4. Proactive Conversational Intervention: The ultimate aim of SID is intervention and prevention. Very little work is undertaken to enable proactive intervention. Proactive suicide prevention online (PSPO) [105] provides a new perspective with the combination of suicidal identification and crisis management. An effective way is through conversations. Automatic response generation becomes a promising technical solution to enable timely intervention for suicidal thoughts. Natural language generation techniques can be utilized to generate counseling responses to comfort people’s depression or suicidal ideation. Reinforcement learning can also be applied for conversational suicide intervention. After suicide attempters post suicide messages (as the initial state), online volunteers and lay individuals will take action to comment on the original posts and persuade attempters to give up their suicidality. The attempter may do nothing, reply to the comments, or get their suicidality relieved. A score will be defined by observing the reaction from a suicide attempter as a reward. The conversational suicide intervention uses a policy gradient for agents to generated responses with maximum rewards to best relieve people’s suicidal thoughts.
   4） 主动会话干预：SID的最终目的是干预和预防。几乎没有采取任何措施来进行主动干预。主动预防自杀在线（PSPO） [105]结合自杀识别和危机管理提供了一个新的视角。有效的方法是通过交谈。自动反应生成是一种很有前途的技术解决方案，可以及时干预自杀念头。自然语言生成技术可以用来产生心理咨询反应来安慰人们的抑郁或自杀意念。强化学习也可以应用于会话自杀干预。当自杀企图者发布自杀信息（初始状态）后，在线志愿者和非专业人员将采取行动对原始帖子进行评论，并劝说企图者放弃自杀行为。尝试者可能什么也不做，回复评论，或者让他们的自杀行为得到缓解。分数将通过观察自杀企图者的反应作为奖励来确定。对话式自杀干预使用政策梯度，让代理人产生最大回报的反应，以最好地缓解人们的自杀念头。

VI. CONCLUSION
Suicide prevention remains an essential task in our modern society. Early detection of suicidal ideation is an important and effective way to prevent suicide. This survey investigates existing methods for SID from a broad perspective that covers clinical methods, such as patient–clinician interaction and medical signal sensing; textual content analysis, such as lexicon-based filtering and word cloud visualization; feature engineering, including tabular, textual, and affective features; and deep learning-based representation learning, such as CNN-and LSTM-based text encoders. Four main domain-specific applications on questionnaires, EHRs, suicide notes, and online user content are introduced.
自杀预防仍然是现代社会的一项重要任务。早期发现自杀意念是预防自杀的重要而有效的方法。这项调查从一个广泛的角度研究了SID的现有方法，包括临床方法，如患者临床交互和医学信号感知；文本内容分析，如基于词典的过滤和词云可视化；特征工程，包括表格、文本和情感特征；以及基于深度学习的表示学习，如基于CNN和LSTM的文本编码器。介绍了问卷调查、EHR、自杀笔记和在线用户内容的四个主要领域特定应用。

Psychological experts have conducted most work in this field with statistical analysis and computer scientists with feature engineering-based machine learning and deep learning-based representation learning. Based on current research, we summarized existing tasks and further proposed new possible tasks. Last but not least, we discuss some limitations of current research and propose a series of future directions, including utilizing emerging learning techniques, interpretable intention understanding, temporal detection, and proactive conversational intervention.
在这一领域，心理学专家与统计分析和计算机科学家进行了大量的工作，基于特征工程的机器学习和基于深度学习的表征学习。在现有研究的基础上，我们总结了现有的任务，并进一步提出了新的可能任务。最后，我们讨论了当前研究的一些局限性，并提出了一系列未来的研究方向，包括利用新兴的学习技巧、可解释的意图理解、时间检测和主动会话干预。

Online social content is very likely to be the main channel for SID in the future. Therefore, it is essential to develop new methods, which can heal the schism between clinical mental health detection and automatic machine detection, to detect online texts containing suicidal ideation in the hope that suicide can be prevented.
在线社交内容很可能成为SID未来的主渠道。因此，有必要开发新的方法来检测含有自杀意念的在线文本，以期能够预防自杀，从而弥合临床心理健康检测与自动机器检测之间的鸿沟。