Emotion Ontology Studies: A Framework for Expressing Feelings Digitally and its Application to Sentiment Analysis

#paper/emotionOntology

摘要（Abstract）

情感本体（Emotion Ontologies）已被开发用于捕捉情感，这是一个涵盖离散情绪和感觉的概念，特别用于情感分析的研究，这一领域分析客户对公司或产品的态度。然而，关于如何调整和应用这些本体的努力仍然有限。本研究调查并综合了情感本体研究，开发了一个情感本体框架（Framework of Emotion Ontologies），该框架可以帮助用户选择或设计合适的情感本体，以支持情感分析，并增加用户对情感、上下文和行为信息在情感分析中的作用的理解。该框架源于情感本体、心理学和情感分析的研究，将情感本体分为离散情感本体或两种混合本体之一，这些混合本体结合了离散、维度性或成分性情感理论。为了展示框架的实用性，本文将该框架应用于情感分析应用程序中的情感本体开发。

通过本研究，提出的情感本体框架可以帮助情感分析中情感、上下文和行为信息的捕捉与表示，从而为情感分析应用提供理论支持和实践指导。

1. 介绍（INTRODUCTION）

本节介绍了情感分析的背景和重要性。社交媒体和在线内容能够提供有关用户态度和公司表现的宝贵信息。情感分析通过捕捉和分析用户的情感和观点，评估其对某个话题或产品的态度。情感本身包括情绪、心情以及对事物的正负亲和力，是情感分析中的核心部分。本研究旨在通过情感本体框架的开发，帮助用户更好地选择或设计情感本体，并加深对情感、上下文和行为信息在情感分析中作用的理解。

2. 评审范围及方法（REVIEW SCOPE AND METHODOLOGY）

2.1 Scope of Review

情感本体被定义为：与情感及其相关的上下文信息相关的明确概念化规范。一般而言，本体由一组正式构造及其相互关系组成。无论最终产生的结果数据类型如何（例如面部表情、声音、神经影像、文本、图片等），我们都专注于识别情感本体概念化规范及其发展中的模式，这些模式在不同领域中是共同的。从中，我们得出了一个情感本体框架，用于支持情感分析，通过回顾和综合相关文献。因此，我们审查了可重复使用的情感本体，这些本体支持情感分析，并排除了那些未开发与情感相关本体的研究 。

2.2 Research Methodology

介绍本研究的文献回顾方法，遵循了 Leidner 和 Kayworth（2006）提出的系统性文献回顾框架。研究过程包括四个主要步骤：制定纳入标准、开发文献检索策略、执行搜索以及记录和分析选定研究。通过使用关键词（如“情感本体”，“情感检测”等）在多个学术数据库中进行检索，筛选出明确描述情感本体结构和开发过程的研究。未详细说明本体结构、开发或评估的文献被排除，确保选出的文献具有较高的相关性和质量。 ![[Fig. 1. Overview of expanded literature selection process..png]]

3. RELATED RESEARCH

Table 1描述了本文使用的主要术语。 ![[Table 1. Terms and Descriptions.png]] ## 3.1 Sentiment Analysis and Challenges of Capturing Affect 情感分析是通过自然语言处理、计算语言学和文本分析来实现意见挖掘或主观性分析【28, 29】。它对公司尤其有用，可以帮助提取客户对其产品或服务的意见（无论是负面的还是正面的）【30】。

尽管情感分析方法取得了显著进展，但仍然面临许多挑战。首先，关于情感的构成没有达成共识【26, 27, 68】。核心概念（如情感、情绪）尚未得到很好的区分，这使得理解表达的细微差别变得困难【26】。一些研究未能识别并报告其在情感分类和检测方法中使用的心理学理论。

其次，针对情感的其他重要维度（如活跃度和支配性水平）的捕捉研究较少【31, 69】。然而，心理学研究【69-71】表明，可以通过多维度的组合（如愉快度、活跃度和支配性）来细致地捕捉情感信息【31】。单纯关注情感的极性（正面或负面）是不够的，因为这种方法本身具有主观性【27, 28】。

研究中还经常将情感归类为情绪类别（如愤怒、幸福、恐惧），忽视了触发情感的情境概念及其后果。很多研究侧重于捕捉与极性相关的信息（即情感的正负值）及情感的强度，却未能捕捉到情境信息，因此提取的数据缺乏必要的细节【72】。此外，尽管社交媒体分析和机器学习为提取情感信息、收集、分析和可视化数据提供了工具，但数据中仍存在许多技术性挑战，例如语义不一致性和社交媒体数据缺乏结构，这使得整合不同类型的数据变得困难【56, 74】。

因此，情感分析的一个重要挑战是如何有效捕捉情感的复杂性及其多维度属性，以便进行准确的分析和推断。

3.2 Ontology-based Sentiment Analysis

本节探讨了基于本体的情感分析方法的应用及其优势。本体通过表示特定领域的概念及其相互关系，为情感分析提供了更深层次的语义理解。相比传统的关键词和词频方法，基于本体的情感分析能够更好地捕捉自然语言中的隐含意义和情感特征。情感本体使得研究人员能够进行概念层次的分析，提取与情感相关的深层信息，并能够处理那些未明确表达情感但与情感相关的文本内容。

情感本体还能够捕捉情感的情境信息及其后果，提供对情感的更全面理解。心理学的情感理论也为情感本体的开发提供了理论基础。基于本体的情感分析能够有效地识别用户情感，预测潜在的风险和情感转变，为公司提供宝贵的客户态度和情感反馈。

Fig. 2. 展示了将在线评论映射到情感本体的示例。心理学中提出了许多理论来识别和解释情感，这些理论可以用于情感本体的开发【例如，79, 80–83】。Frijda【23, 84】提出情绪过程包括Fig. 2. 顶部高层次情感本体中展示的组成部分，底部则展示了用户在实际在线评论论坛中尝试以数字方式表达情感。

如果能够自动解释用户文本的情感，这将成为衡量客户对其问题态度的宝贵指标，并且能够预测问题升级的潜在风险。该评估要求将客户文本中的术语映射到情感本体进行语义解释【85】。在Fig. 2. 中，从句子“这简直荒谬！我感到愤怒！”中捕捉到负面情感。声明不打算再次购买公司产品可能对公司声誉管理构成潜在的担忧。 ![[Fig. 2. Mapping an online review to an emotion ontology for sentiment analysis..png]]

4. 情感本体框架（A FRAMEWORK OF EMOTION ONTOLOGIES）

This section presents a Framework of Emotion Ontologies derived from our review and synthesis of studies on emotion ontologies, psychology, and sentiment analysis. (See Figure 3.) The framework is intended to be used to address challenges of sentiment analysis and guide the development or selection of emotion ontologies for sentiment analysis. The framework is based on the three dominant paradigms of emotion: (1) discrete [86–89]; (2) dimensional [69–71, 90, 91]; and (3) componential process [79–83]. These form the basis for the framework because they: (1) indicate what concepts should be included in an emotion ontology; and (2) provide multiple, systematic ways to extract and represent affect. 本节介绍了从情感本体、心理学和情感分析研究中综述和综合得出的情感本体框架（见图3）。该框架旨在应对情感分析的挑战，并指导情感本体的开发或选择，以支持情感分析。该框架基于三种主要的情感范式：(1) 离散情感范式 [86–89]；(2) 维度情感范式 [69–71, 90, 91]；(3) 组合过程情感范式 [79–83]。这些范式构成了框架的基础，因为它们：(1) 指明了情感本体中应包含的概念；(2) 提供了多种系统化的方法来提取和表示情感。

The framework organizes emotion ontologies as: (1) DS (discrete), (2) hybrid of DS and DM (discrete and dimensional), and (3) hybrid of CM, DS, and DM (componential process, discrete, and dimensional) emotion ontologies. The hybrids are needed because a single emotion paradigm cannot capture all of the complexities of a real-world situation. Among 97 selected studies that we reviewed, 25.8% (25) of the studies are classified to use the discrete emotion ontology approach, 56.7% (55) of the studies for the hybrid of discrete and dimensional ontology, and 17.5% (17) of the studies for the hybrid of componential process, discrete, and dimensional ontology. (See Appendix B.) Table 2 summarizes these three types. Contextual extensions are additions necessary to capture real-world concepts for a given application. Their derivations are explained in detail below. 该框架将情感本体组织为：(1) 离散情感本体（DS）；(2) 离散与维度混合本体（DS & DM）；(3) 组合过程、离散与维度混合本体（CM, DS & DM）。混合本体的存在是因为单一的情感范式无法捕捉现实世界中的所有复杂性。在我们综述的97项研究中，25.8%（25项）的研究采用了离散情感本体方法，56.7%（55项）的研究采用了离散与维度混合本体方法，17.5%（17项）的研究采用了组合过程、离散与维度混合本体方法（见附录B）。表2总结了这三种类型的情感本体及其用途。上下文扩展是为了捕捉现实世界中的概念而进行的必要补充，其推导过程将在下文详细解释。

![[Table 2. Types of Emotion Ontologies and their Purposes.png]]

这一部分提出了一个情感本体框架，目的是帮助研究人员更好地理解和分析情感。这个框架基于三种主要的情感理论：

1. 离散情感理论：认为情感可以分为几种基本类型，比如快乐、悲伤、愤怒等。
2. 维度情感理论：认为情感可以通过几个维度来描述，比如情感的强度（强或弱）、情感的积极性（正面或负面）等。
3. 组合过程情感理论：认为情感是由多个部分组成的，比如某个事件引发了情感，情感又影响了行为。

这个框架把情感本体分为三种类型：

1. 离散情感本体：只关注基本的情感类型，比如快乐、悲伤等。
2. 离散与维度混合本体：不仅关注情感的类型，还关注情感的维度，比如情感的强度和积极性。
3. 组合过程、离散与维度混合本体：最复杂的一种，不仅关注情感的类型和维度，还关注情感是如何产生的（比如某个事件引发了情感）以及情感如何影响行为。

为什么需要这个框架？

因为情感分析很复杂，单一的理论无法完全解释情感的所有方面。通过结合不同的理论，这个框架可以帮助研究人员更全面地理解情感，并在情感分析中应用这些知识。

总结：

这个框架为情感分析提供了一个系统化的方法，帮助研究人员选择或设计适合的情感本体，从而更好地理解和分析情感。

4.1 Discrete Emotion Ontology

4.1.1 目标（Objective）

The development of a discrete ontology is motivated by the need to automatically classify different types of emotions, and components related to contexts (e.g., behaviors, environment, and systems related) with higher accuracy than humans. Based on emotion theories (from psychology studies) and empirical studies (from emotion ontology and sentiment analysis studies), concepts (types of emotion and contextual information), properties, and relationships are designed and used to classify text and find patterns. For example, to assess suicide risk, emotions of suicidal patients are inferred from suicide notes and captured in a suicide emotion ontology [92].

离散情感本体的开发动机在于需要自动分类不同类型的情感，以及与上下文相关的组件（如行为、环境和系统相关的内容），并且其准确性要高于人类。基于情感理论（来自心理学研究）和实证研究（来自情感本体和情感分析研究），设计并使用了概念（情感类型和上下文信息）、属性和关系来分类文本并发现模式。例如，为了评估自杀风险，从自杀遗书中推断出自杀患者的情感，并将其捕捉到自杀情感本体中 [92]。

4.1.2 理论基础（Theory Base）

The discrete emotion paradigm [86–89] focuses on a small set of fundamental emotions. They are ‘basic,’ in the sense of being both biologically and psychologically primitive [79]. Arnold [93] proposes 11 basic emotions (anger, aversion, courage, dejection, desire, despair, fear, hate, hope, love, and sadness). Tomkins [87] identifies eight core emotions (surprise, interest, joy, rage, fear, disgust, shame, and anguish). Izard [88] proposes 10 (anger, contempt, disgust, distress, fear, guilt, interest, joy, shame, and surprise). Weiner and Graham [94] claim only two (happiness and sadness). Examples are given in Table 3.

离散情感范式[86-89]关注一小部分基本情感。它们是“基本的”，即在生物学和心理上都是原始的[79]。Arnold[93]提出了11种基本情感（愤怒、厌恶、勇气、沮丧、欲望、绝望、恐惧、仇恨、希望、爱和悲伤）。Tomkins[87]识别了八种核心情感（惊讶、兴趣、喜悦、愤怒、恐惧、厌恶、羞耻和痛苦）。Izard[88]提出了10种情感（愤怒、轻蔑、厌恶、痛苦、恐惧、内疚、兴趣、喜悦、羞耻和惊讶）。Weiner和Graham[94]则只提出了两种情感（快乐和悲伤）。表3中给出了一些示例。 ![[Table 3. Illustrative Discrete Emotions, their Functions, and their Behaviors (Izard 1993).png]]

4.1.3 分类的先前情感本体研究（Classified Prior Emotion Ontology Studies）

In our review of prior studies, the discrete ontology is the second most preferred approach (26.6%). Emotion ontologies adopt a discrete approach when they use theories from a discrete emotion paradigm and focus on capturing types of emotions from blog posts or text [9, 92, 95], as shown in Appendix B. The simple approach uses a set of basic emotions to classify affect. Pestian et al.’s [92] suicide ontology captures discrete emotions of patients from suicide notes, such as guilt, hostility, hopelessness, helplessness, and despair. A complex approach [e.g., 96], includes additional types of emotions appearing in a hierarchical structure. Balahur et al. [9, 95] combine the emotion sets from Parrot’s (2001) tree-structured list of emotions, and Plutchik’s (2001) wheel of emotion to derive seven primary emotions (anger, fear, disgust, shame, sadness, joy, and guilt). Li et al.’s [97] ontology consists of Ekman’s six basic emotions [98] with subdimensions under each emotion that contain relevant Chinese phrases. 在我们对先前研究的综述中，离散本体是第二受欢迎的方法（26.6%）。情感本体采用离散方法时，通常使用离散情感范式的理论，并专注于从博客文章或文本中捕捉情感类型[9, 92, 95]，如附录B所示。简单的方法使用一组基本情感来分类情感。Pestian等人[92]的自杀本体从自杀遗书中捕捉患者的离散情感，如内疚、敌意、绝望、无助和绝望。复杂的方法[如96]则包括出现在层次结构中的其他情感类型。Balahur等人[9, 95]结合了Parrot（2001）的树状情感列表和Plutchik（2001）的情感轮，推导出七种主要情感（愤怒、恐惧、厌恶、羞耻、悲伤、喜悦和内疚）。Li等人[97]的本体由Ekman的六种基本情感[98]组成，每种情感下包含相关的中文短语。

These studies are useful, but limited. They often do not capture the contexts related to affect [10, 95]. Semantic relationships, such as is-a-sub/superclass-of, and other relationships between two classes (affect and a contextual extension), can capture temporal sequences and necessary conditions [20, 99]. However, such relationships are frequently neglected in emotion ontology development and evaluation. 这些研究虽然有用，但也有局限性。它们通常没有捕捉到与情感相关的上下文[10, 95]。语义关系，如“是子类/超类”以及情感与上下文扩展概念之间的其他关系，可以捕捉时间序列和必要条件[20, 99]。然而，这些关系在情感本体的开发和评估中经常被忽视。

Additionally, some studies ignore references to emotion theories in their development. Although it is possible to infer patterns of outcomes (e.g., behaviors) of affect [95], prior research often classifies affect, but does not interpret the outcomes or infer additional meaning. For example, it might be possible to infer a potential future behavior (e.g., move away) from an extracted term(s), (e.g., anger, a type of emotion) and a product’s malfunction (environmental context information). 此外，一些研究在开发过程中忽略了情感理论的引用。尽管可以从提取的术语（如愤怒，一种情感类型）和产品故障（环境上下文信息）中推断出潜在的未来行为（如远离），但先前的研究通常只分类情感，而不解释结果或推断额外的含义。

4.1.4 离散本体的设计与开发（Design and Development of Discrete Ontology）

By combining the discrete emotion paradigm and prior emotion ontology studies, discrete ontologies can be developed for affect classifications, as illustrated in Figure 4. The combination provides the ontologies with a strong theory base from psychology and practicality from existing studies. Two types of discrete ontologies can support either a complex or simple classification of affect. These are illustrated in Figure 4 on the left-hand side and right-hand side, respectively. The discrete ontologies in Figure 4 are extendable. With theoretical support from the discrete emotion paradigm, researchers and practitioners can include additional concepts and relationships among them, such as superclass-subclass. 通过结合离散情感范式和先前的情感本体研究，可以开发用于情感分类的离散本体，如图4所示。这种结合为情感本体提供了来自心理学的强大理论基础和现有研究的实用性。两种类型的离散本体可以支持简单或复杂的情感分类。图4左侧和右侧分别展示了复杂和简单的情感分类。图4中的离散本体是可扩展的。在离散情感范式的理论支持下，研究人员和从业者可以包括额外的概念及其之间的关系，如超类-子类关系。

Discrete ontologies can be augmented with contextual extensions that provide rich, insightful information associated with affect. They could be behavioral, individual, environmental, or system related, as shown in Table 4. For example, in addition to types of emotion, an ontology could capture the characteristics of both users and products. 离散本体可以通过上下文扩展来增强，这些扩展提供了与情感相关的丰富且具有洞察力的信息。它们可以是行为、个体、环境或系统相关的，如表4所示。例如，除了情感类型外，本体还可以捕捉用户和产品的特征。

The discrete ontology can be used alone or as a module in a larger ontology. For a simple classification, the discrete ontology includes subclasses (e.g., Anger, Fear, Disgust, OtherBasicEmotions) of the BasicEmotionCategory class. Since the discrete approach identifies basic emotions, a decision on what emotions to include depends on a study’s objective. A set of basic emotions might be all that is needed to capture customers’ satisfaction, anger, disappointment, frustration, and so forth. For a complex classification, an ontology could support a hierarchical classification, where each class of the ComplexEmotionCategory class has additional subclasses. For example, subclasses of Fear can be Scared, Afraid, Panicky, Nervous, Worried, and Tense [100]. Subclasses of Anger include Angry, Frustrated, Irritated, Unfulfilled, Discontented, Envious, and Jealous [100]. 离散本体可以单独使用，也可以作为更大本体的模块。对于简单分类，离散本体包括BasicEmotionCategory类的子类（如Anger, Fear, Disgust, OtherBasicEmotions）。由于离散方法识别基本情感，选择哪些情感取决于研究的目标。一组基本情感可能足以捕捉客户的满意度、愤怒、失望、沮丧等。 对于复杂分类，本体可以支持层次分类，其中ComplexEmotionCategory类的每个类都有额外的子类。例如，Fear的子类可以是Scared, Afraid, Panicky, Nervous, Worried, 和 Tense[100]。Anger的子类包括Angry, Frustrated, Irritated, Unfulfilled, Discontented, Envious, 和 Jealous[100]。

![[Fig. 4. Illustrations of discrete emotion ontology.png]] ![[Table 4. Examples of Contextual Extensions for Discrete Ontologies.png]] ### 4.1.5 示例案例：EmotiNet本体（An Example Case: EmotiNet Ontology） To overcome subjectivity, Balahur et al. [95] successfully detect emotions in textual expressions. The EmotiNet ontology, consisting of emotion and contextual extensions (e.g., action, and family relations ontologies), is based on appraisal [101] and the wheel of emotion [102]. It is semi-automatically populated and extended with real examples (e.g., the self-reported affect data bank by the International Survey of Emotional Antecedents and Reactions (ISEAR)). During the population process, the major tasks were: identify expressions associated with the six types of emotions from 175 examples; identify semantic roles; design situations based on the ‘actor-action-object-emotional reaction’ model; and model emotions and their relationships based on theories. 为了克服主观性，Balahur等人[95]成功检测了文本表达中的情感。EmotiNet本体由情感和上下文扩展（如动作和家庭关系本体）组成，基于评估理论[101]和情感轮[102]。它是半自动填充的，并通过真实示例（如国际情感前因和反应调查（ISEAR）的自我报告情感数据库）进行扩展。在填充过程中，主要任务是：从175个示例中识别与六种情感类型相关的表达；识别语义角色；基于“行动者-动作-对象-情感反应”模型设计情境；并根据理论建模情感及其关系。

4.1.6 评估（Evaluation）

Studies show the potential and credibility of the discrete emotion approach [9, 92, 95]. Balahur et al. [95] evaluate the EmotiNet ontology by verifying its structure and contents, and validating its method for extracting emotion from text. The evaluation is conducted with multiple test sets in real-world examples, including ISEAR. In comparison, Danisman and Alpkocak [103] find much higher precision and recall rates. Pestian et al.’s [92] use the suicide ontology with machine learning methods to show much higher accuracy (78%) than trainees (49%) and mental health professionals (63%) when classifying suicide notes. 研究表明，离散情感方法具有潜力和可信度[9, 92, 95]。Balahur等人[95]通过验证EmotiNet本体的结构和内容，并验证其从文本中提取情感的方法，对该本体进行了评估。评估是在多个测试集中进行的，包括ISEAR的真实示例。相比之下，Danisman和Alpkocak[103]发现了更高的精确率和召回率。Pestian等人[92]使用自杀本体与机器学习方法，显示在分类自杀遗书时的准确率（78%）远高于受训者（49%）和心理健康专业人员（63%）。

4.1.7 贡献（Contribution）

Discrete ontologies are particularly useful for capturing simple or extended sets of emotions. The discrete approach provides a well-established theory base to capture discrete emotions and associated characteristics, such as environmental stimuli, behavioral outcomes, and subjective experiences. These can be represented by affect, its subconcepts, contextual extensions, and relationships (Figure 4). Discrete ontologies identify candidate concepts and relationships that can be customized. The hasRelationship between affect and contextual extension concepts can be further specified (e.g., triggers, resultsIn, leadsTo) to capture relationships between affect and other important concepts, such as behaviors and products. 离散本体特别适用于捕捉简单或扩展的情感集合。离散方法提供了一个完善的理论基础，用于捕捉离散情感及其相关特征，如环境刺激、行为结果和主观体验。这些可以通过情感、其子概念、上下文扩展和关系来表示（图4）。离散本体识别了可以定制的候选概念和关系。情感与上下文扩展概念之间的“hasRelationship”可以进一步指定（如“触发”、“导致”），以捕捉情感与其他重要概念（如行为和产品）之间的关系。

4.2 离散与维度混合情感本体（Hybrid of Discrete and Dimensional Emotion Ontology）

4.2.1 目标（Objective）

The objective of the discrete emotion and dimensional hybrid ontology is to provide a formal specification of shared knowledge: affect (including discrete emotions) and its dimensions (valence, activation, etc.); and contextual information and relationships. These support automatic classifications of various dimensions and contextual data associated with affect. 离散情感与维度混合本体的目标是提供共享知识的正式规范：情感（包括离散情感）及其维度（效价、激活等）；以及上下文信息和关系。这些支持自动分类与情感相关的各种维度和上下文数据。

![[Fig. 5. Two dimensional circumplex model for affect..png]]

4.2.2 理论基础（Theory Base）

The discrete emotion and dimensional paradigms are complementary [104]. The dimensional paradigm [e.g., 69–71, 90, 91] is based on multiple bipolar dimensions such as valence, activation, or dominance. Valence captures the affect of pleasantness or unpleasantness (i.e., positive or negative) [70]. Activation indicates whether one’s affect is activated or deactivated [70] and is a continuum. Dominance, also a continuum, captures the affect of control over a situation [105]. Figure 5 shows a sample circumplex model of affect [69]. The horizontal dimension represents valence, the continuum from displeasure (e.g., distressed) to pleasure (e.g., happy). The vertical dimension is the activation dimension and a continuum from deactivation (e.g., calm) to activation (e.g., excited). This model has been widely adapted in emotion ontology development [106, 107]. 离散情感和维度范式是互补的 [104]。维度范式 [例如69, 70, 71, 90, 91] 基于多个双极维度，如效价、激活或支配性。效价捕捉情感的愉悦或不愉悦（即正面或负面）[70]。激活表明一个人的情感是被激活还是被抑制 [70]，并且是一个连续体。支配性也是一个连续体，捕捉对情境的控制感 [105]。图5展示了一个情感的双维度环状模型 [69]。水平维度代表效价，从不满（如痛苦）到愉悦（如快乐）的连续体。垂直维度是激活维度，从抑制（如平静）到激活（如兴奋）的连续体。该模型在情感本体开发中被广泛采用 [106, 107]。

Plutchik [102] also uses a circumplex model with additional types of dimensions. The vertical dimension of his corn shape model indicates the intensity of emotion, the circle represents degrees of commonality of emotion, and the eight sections capture basic emotions, presented as four pairs of opposite emotions. The blank spaces between sections show emotions that are combinations of two of the basic ones. To support sentiment analysis, Plutchik’s model is reinterpreted and an hourglass of emotions developed “by organizing primary emotions around four independent but concomitant dimensions, whose different levels of activation make up the total emotional state of the mind” [108, p. 148–9]. Subsequently, a revised model, called the “hourglass model revisited” is proposed by Susanto et al. [68, p. 100]. It is motivated by addressing such issues as “uncanny color associations, presence of neutral emotions, absence of some polar emotions, wrong associations of antithetic emotions, low polarity scores for compound emotions, and absence of self-conscious or moral emotions” [68, p. 97]. Plutchik [102] 也使用了一个环状模型，并增加了其他类型的维度。他的玉米形状模型的垂直维度表示情感强度，圆圈表示情感的普遍程度，八个部分捕捉基本情感，呈现为四对相反的情感。部分之间的空白显示了由两种基本情感组合而成的情感。为了支持情感分析，Plutchik的模型被重新解释，并开发了“情感沙漏模型”，通过围绕四个独立但并存的维度组织主要情感，其不同的激活水平构成了心理的总体情感状态 [108, p. 148-9]。随后，Susanto等人 [68, p. 100] 提出了一个修订模型，称为“沙漏模型修订版”，旨在解决诸如“不自然的颜色关联、中性情感的存在、某些极性情感的缺失、复合情感的低极性分数以及自我意识或道德情感的缺失”等问题 [68, p. 97]。

A major benefit of this hybrid is that it has discrete emotions as primary factors, with various dimensions, secondary. Together, these make measures reliable [104]. For example, in a hierarchical model [100], shown in Figure 6, basic emotions are extended and organized into three levels and combined with a dimensional approach. At the superordinate level, a dimensional approach (a valence dimension) differentiates positive and negative affect. At the mid-level, a discrete emotion approach (eight basic emotions) is applied with four types of negative and four types of positive emotions. At the subordinate level, an additional 42 emotions are classified based on Richins’s [109] clusters of emotion descriptors. 这种混合的主要好处在于，它以离散情感为主要因素，各种维度为次要因素。两者结合使测量更加可靠 [104]。例如，在分层模型 [100] 中，如图6所示，基本情感被扩展并组织为三个层次，并与维度方法结合。在上级层次，维度方法（效价维度）区分了正面和负面情感。在中级层次，离散情感方法（八种基本情感）应用于四种负面和四种正面情感。在下级层次，基于Richins的 [109] 情感描述符集群，进一步分类了42种情感。

![[Fig. 6. Hierarchy of emotions (Laros and Steenkamp 2005)..png]]

4.2.3 分类的先前情感本体研究（Classified Prior Emotion Ontology Studies）

Each emotion paradigm, individually, has limitations. The discrete emotion paradigm focuses a limited set of basic emotions [100, 110]. The dimensional approach can lose information because some types of emotions cannot easily be mapped into dimension spaces (e.g., surprise). Other emotions are difficult to distinguish [96]. Fear, anger, disgust, and stress can all have the same characteristics of an unpleasant high-activation state, and are thus located close to one another. The dimensions can be ambiguous. For example, it may not be clear whether the valence dimension (positive vs. negative) indicates the appraisal of an event or the affect triggered by the appraisal [111]. An activation dimension could be a perceived activation in an event, or physiological arousal triggered by the event. 每种情感范式单独使用时都有局限性。离散情感范式关注有限的基本情感集 [100, 110]。维度方法可能会丢失信息，因为某些类型的情感不容易映射到维度空间（如惊讶）。其他情感难以区分 [96]。恐惧、愤怒、厌恶和压力都可能具有不愉快的高激活状态特征，因此它们的位置彼此接近。维度可能是模糊的。例如，效价维度（正面 vs. 负面）可能不清楚是表示对事件的评估还是由评估触发的情感 [111]。激活维度可能是事件中的感知激活，也可能是由事件触发的生理唤醒。

More than half the studies (56.4%) both employ emotion paradigms as a theory base and develop a hybrid ontology approach. We classified these studies as a hybrid of a discrete and dimensional ontology (Appendix B). Prior research has developed both simple forms of the hybrid ontology approach (e.g., [96, 107, 112, 113]) and more complex ones (e.g., [13, 106, 114–117]). For the simple hybrid approach, García-Rojas et al. [112] study virtual humans in animations, using six basic emotions to group twenty-five archetypal facial expressions. Each emotion is characterized by two dimensions (valence and activation) [91]. Dellandréa et al. [96] use neural networks to map images into a two dimensional (valence and activation) emotional space. Evidence theory maps the images into four emotion categories (anger, sadness, jubilation, and wonder), so the hybrid helps deal with the ambiguous and subjective nature of emotions. To develop social robots, Graterol et al. [113] develop an emotion ontology that can classify affect into eleven emotions (with their intensity) from multimedia contents including text, images, speech, or videos. A neural network is used to label the emotions. Contextual information including person, object, and modality is captured to provide rich information related to emotions. Dragoni et al. [76] improve OntoSenticNet [118, 119] and propose OntoSenticNet 2, in which an emotion ontology is used to extract the granularity of affective information embedded explicitly and implicitly in multimodal resources (e.g., documents, images, and videos) “based on SenticNet, a commonsense knowledge base for sentiment analysis” [76, p. 109]. The valence (polarity) of emotion is captured. Contextual information such as SenticConcept, domain, resource, and SenticReasoning can also be captured. 超过一半的研究（56.4%）同时采用情感范式作为理论基础，并开发了混合本体方法。我们将这些研究分类为离散与维度混合本体（附录B）。先前的研究开发了简单形式的混合本体方法（如 [96, 107, 112, 113]）和更复杂的形式（如 [13, 106, 114–117]）。对于简单的混合方法，García-Rojas等人 [112] 研究了动画中的虚拟人类，使用六种基本情感对25种原型面部表情进行分组。每种情感由两个维度（效价和激活）表征 [91]。Dellandréa等人 [96] 使用神经网络将图像映射到二维（效价和激活）情感空间。证据理论将图像映射到四种情感类别（愤怒、悲伤、欢欣和惊奇），因此混合方法有助于处理情感的模糊和主观性。为了开发社交机器人，Graterol等人 [113] 开发了一种情感本体，可以从多媒体内容（包括文本、图像、语音或视频）中将情感分类为11种情感（及其强度）。捕捉包括人物、对象和模态在内的上下文信息，以提供与情感相关的丰富信息。Dragoni等人 [76] 改进了OntoSentiNet [118, 119] 并提出了OntoSentiNet 2，其中使用情感本体从多模态资源（如文档、图像和视频）中提取嵌入的显式和隐式情感信息的粒度，“基于SentiNet，一个用于情感分析的常识知识库” [76, p. 109]。捕捉情感的效价（极性）。还可以捕捉上下文信息，如SentiConcept、领域、资源和SentiReasoning。

Complex forms of the hybrid approach often use a hierarchical structure (similar to Figure 6), with various dimensions (e.g., [13, 106, 114–117, 120, 121]). Francisco et al. [116] create a hierarchical structure with two levels [122, 123] for text mining. The basic level includes sadness, happiness, surprise, fear, and anger. The next level includes sub-types of each basic emotion. These emotions are based on three dimensions (valence, activation, and dominance). To analyze sentiment, Cambria et al. [115] develop a hybrid approach of an emotion ontology based on the hourglass of emotion. Cambria et al. [114, 117] and Dragoni et al. [76] employ the hourglass model revisited [68] to develop an emotion ontology. Both the original hourglass model and the hourglass model revisited are based on Plutchik’s [102] wheel of emotion (Figure 5(b)). However, there are some differences. In the original hourglass of emotions employed in Cambria et al. [115], basic emotions are characterized by four affective dimensions: aptitude, pleasantness, sensitivity, and attention. These are further classified based on six levels of activation, indicating the intensity of the emotions. Compound emotions (e.g., love, aggressiveness, disappointment) are created by combining different affect dimensions. For example, love is generated from a combination of the positive pleasantness and positive aptitude dimensions. In the hourglass model revisited, in Cambria et al. [114, 117] and Dragoni et al. [76], notable differences include: new affective dimensions (attitude, introspection, sensitivity, and temper); significant polar emotions (e.g., calmness, eagerness); self-conscious or moral emotions (e.g., pride, prejudice, guilt), and others. 复杂形式的混合方法通常使用分层结构（类似于图6），具有各种维度（如 [13, 106, 114–117, 120, 121]）。Francisco等人 [116] 创建了一个具有两个层次 [122, 123] 的分层结构，用于文本挖掘。基本层次包括悲伤、快乐、惊讶、恐惧和愤怒。下一层次包括每种基本情感的子类型。这些情感基于三个维度（效价、激活和支配性）。为了分析情感，Cambria等人 [115] 开发了一种基于情感沙漏模型的混合情感本体方法。Cambria等人 [114, 117] 和Dragoni等人 [76] 采用了沙漏模型修订版 [68] 来开发情感本体。原始的情感沙漏模型和沙漏模型修订版都基于Plutchik的 [102] 情感轮（图5(b)）。然而，两者之间存在一些差异。在Cambria等人 [115] 采用的原始情感沙漏模型中，基本情感由四个情感维度表征：能力、愉悦性、敏感性和注意力。这些进一步基于六个激活水平进行分类，表示情感的强度。复合情感（如爱、攻击性、失望）通过组合不同的情感维度生成。例如，爱是由积极的愉悦性和积极的能力维度组合生成的。在Cambria等人 [114, 117] 和Dragoni等人 [76] 的沙漏模型修订版中，显著差异包括：新的情感维度（态度、内省、敏感性和脾气）；显著的极性情感（如平静、渴望）；自我意识或道德情感（如骄傲、偏见、内疚）等。

The dimensions, shown in Table 5, are commonly found in emotion ontology studies. 表5中列出了情感本体研究中常见的维度。

Although some studies [106, 113, 120, 121, 124] attempt to capture contextual components, others do not. Behavioral concepts are mostly ignored. Semantic relationships between concepts, such as affect and behavior, are rarely examined, so information about necessary conditions and temporal sequences is lost [99]. Further specification and design of such components and relationships, based on a strong theory base, are needed. 尽管一些研究 [106, 113, 120, 121, 124] 尝试捕捉上下文组件，但其他研究没有。行为概念大多被忽略。概念之间的语义关系，如情感和行为，很少被研究，因此丢失了关于必要条件和时间序列的信息 [99]。基于强大的理论基础，进一步规范和设计这些组件和关系是必要的。

![[Table 5. Typical Dimensions used in Emotion Ontology Studies.png]] ### 4.2.4 离散与维度混合情感本体的设计与开发（Design and Development of the Hybrid of Discrete and Dimensional Emotion Ontology） The discrete and dimensional hybrid, as illustrated in Figure 7, is developed from the discrete and dimensional emotion paradigms and prior studies. It comprises discrete emotion concepts, affective dimensions, contextual extensions, and relevant relationships and supports both simple and complex affect classifications.2 The concepts and relationships can be extended further. 离散与维度混合本体，如图7所示，是从离散和维度情感范式及先前研究中开发出来的。它由离散情感概念、情感维度、上下文扩展和相关关系组成，支持简单和复杂的情感分类²。这些概念和关系可以进一步扩展。

² 这包括图4中所示的离散情感方法，但在本图中进行了简化。

The BasicEmotionCategory class and associated classes (e.g., Anger, Fear, Disgust) support a simple affect classification. The ComplexEmotionCategory class and relevant sublevel emotion classes, organized in a hierarchical structure, support complex affect classification. For the dimensional properties, common ones, such as Valence, Activation, and Dominance properties (e.g., [125]), characterize types of affect, supporting the extraction of rich and fine-grained sentiment information. Other dimensions can be added: Attention, Sensitivity, and Aptitude from the Hourglass of Emotions model (e.g., [114, 115, 117, 126]) or Intensity and Degree of Commonality from the Wheel of Emotion [102] (Figure 5(b)). BasicEmotionCategory类及其相关类（如Anger、Fear、Disgust）支持简单的情感分类。ComplexEmotionCategory类及其相关的子层次情感类，组织在分层结构中，支持复杂的情感分类。对于维度属性，常见的属性（如Valence、Activation和Dominance属性（如 [125]））表征情感类型，支持提取丰富且细粒度的情感信息。可以添加其他维度：来自情感沙漏模型的Attention、Sensitivity和Aptitude（如 [114, 115, 117, 126]）或来自情感轮的Intensity和Degree of Commonality [102]（图5(b)）。

For semantic relationships, the discrete emotion and dimensional paradigms provide the theory bases. Designing relationships are not confined to hierarchical or superclass-subclass relationships. More complex relationships among emotions, dimensions, and concepts from contextual extensions can be designed to capture meaningful information. 对于语义关系，离散情感和维度范式提供了理论基础。设计关系不仅限于层次或超类-子类关系。可以设计更复杂的关系，以捕捉情感、维度和上下文扩展概念之间的有意义信息。

![[Fig. 7. An illustration of hybrid of discrete and dimensional emotion ontology..png]] ### 4.2.5 示例案例：用于情感分析的OntoEmotion本体（An Example Case: OntoEmotion Ontology for Sentiment Analysis） Baldoni et al. [6] identify the important role of affective computing in the social web, which focuses on analyzing users’ tagging and automatically identifying emotions. They develop an ArsEmotica application that can analyze tagged artworks in a social tagging site and capture emotional concepts. Sentiment lexicons are used and an ontology, OntoEmotion, developed. A complex form of the hybrid of the discrete and dimensional ontology captures a hierarchical structure of emotions and valence (positiveness, negativeness, and neutrality), with tags directly indicating recognized ontological concepts. Emotions are ranked to identify the ones most relevant to artwork. The result is rich emotional semantics for the tagged emotions and valences [6]. Baldoni等人 [6] 识别了情感计算在社交网络中的重要作用，社交网络专注于分析用户的标签并自动识别情感。他们开发了一个ArsEmotica应用程序，可以分析社交标签网站中的标记艺术品并捕捉情感概念。使用情感词典并开发了一个本体OntoEmotion。复杂形式的离散与维度混合本体捕捉了情感的分层结构和效价（正面性、负面性和中性），标签直接指示识别的本体概念。情感被排名以识别与艺术品最相关的情感。结果为标记的情感和效价提供了丰富的情感语义 [6]。

4.2.6 评估（Evaluation）

Several studies show strong potential for the hybrid of the discrete and dimensional ontology (e.g., [6, 13, 113–115, 120]). Baldoni et al. [6] conduct an application-level evaluation by applying ArsEmotica application to the artwork community [99]. Garcia-Crespo et al. [13] develop an application to analyze customer social networks by mining customer emotion, which shows promising results for detecting emotions, based on user survey results. From experiments with three scenarios, a combination of basic emotion and valence scenario achieves the highest precision and recall rates [13]. 几项研究表明，离散与维度混合本体具有强大的潜力（如 [6, 13, 113–115, 120]）。Baldoni等人 [6] 通过将ArsEmotica应用程序应用于艺术品社区进行了应用级评估 [99]。Garcia-Crespo等人 [13] 开发了一个应用程序，通过挖掘客户情感来分析客户社交网络，基于用户调查结果显示，检测情感的结果很有希望。通过三个场景的实验，基本情感和效价场景的组合达到了最高的精确度和召回率 [13]。

4.2.7 贡献（Contribution）

Benefits emerge from the combined theoretical support of both emotion paradigms. The discrete emotion approach identifies which set of emotions (basic or complex) to select for a given purpose and context. The dimensional components capture information on the position of an emotion and its traits across different dimensions, so normalized scores can be used. This makes it possible to capture and analyze emotion from digital data at fine-grained levels (e.g., negative ∼ positive (−5 ∼ +5), passive∼ active (−5 ∼ +5)) [127]. The contextual extensions capture rich contextual information about affect. Behavioral concepts are particularly useful when predicting customers’ behaviors (Table 4). 结合两种情感范式的理论支持带来了好处。离散情感方法确定了选择哪种情感集（基本或复杂）以用于特定目的和上下文。维度组件捕捉了情感的位置及其在不同维度上的特征，因此可以使用标准化分数。这使得可以从数字数据中捕捉和分析情感，达到细粒度水平（如负面～正面（-5～+5），被动～主动（-5～+5））[127]。上下文扩展捕捉了与情感相关的丰富上下文信息。行为概念在预测客户行为时特别有用（表4）。

4.3 组合过程、离散与维度混合情感本体（Hybrid of Componential Process, Discrete, and Dimensional Emotion Ontology）

4.3.1 目标（Objective）

The combination of componential process, discrete and dimensional ontology was developed to overcome the limitations of polarity classifications and to capture the components related to an individual’s emotion generation. For example, in sentiment analysis for marketing purposes, capturing the emotion generation process of customers is very important. What made a smartphone user angry or happy? How did the user appraise a smartphone issue or a fancy function? What kinds of behavioral intentions are expressed by users who experience a positive or negative emotion? The objective of the ontology is to capture interrelated components associated with emotion generation (e.g., event, appraisal, emotion, behavioral intention) as well as various dimensions (e.g., valence, activation) of a set of emotions to support automatic classification from a huge amount of unstructured data. 组合过程、离散与维度混合本体的开发旨在克服极性分类的局限性，并捕捉与个体情感生成过程相关的组件。例如，在市场营销的情感分析中，捕捉客户的情感生成过程非常重要。是什么让智能手机用户感到愤怒或快乐？用户如何评估智能手机问题或花哨的功能？经历正面或负面情感的用户表达了哪些行为意图？该本体的目标是捕捉与情感生成相关的相互关联的组件（如事件、评估、情感、行为意图）以及一组情感的各个维度（如效价、激活），以支持从大量非结构化数据中进行自动分类。

4.3.2 理论基础（Theory Base）

All three emotion paradigms serve as the theory base for the hybrid of the componential process, discrete emotion and dimensional ontologies, with the componential process paradigm being dominant. The componential process paradigm [e.g., 79, 80–83] views emotion as a dynamic, multi-componential phenomenon. Frijda [23, 84] suggests that an emotion process is comprised of several different components, starting when someone (e.g., a customer) encounters a significant event and appraises (or evaluates) that event. The appraisals, in turn, trigger affective responses and action readiness. These components influence actual behavior. Feedback loops are also possible and can influence the resulting behavior from an event or appraisal [23]. 所有三种情感范式都作为组合过程、离散情感和维度本体的理论基础，其中组合过程范式占主导地位。组合过程范式 [79, 80, 81, 82, 83, 6] 将情感视为一种动态的、多组件的现象。Frijda [84, 23] 提出，情感过程由几个不同的组件组成，当某人（如客户）遇到一个重要事件并评估该事件时，情感过程开始。评估反过来触发情感反应和行动准备。这些组件影响实际行为。反馈循环也是可能的，并且可以影响事件或评估的最终行为 [23]。

Other studies also propose multiple components. Mesquita et al. [128] identify a broader set of components: (a) antecedent event, (b) emotional experience, (c) appraisal, (d) physiological change, (e) change in action readiness, (f) behavior, (g) change in cognitive functioning and beliefs, and (h) processes, suggesting that most emotional instances entail all of these. Scherer [111] identifies five components: cognitive (appraisal); neurophysiological (bodily symptoms); motivational (action tendencies); motor expression (facial and vocal expression); and subjective feeling (emotional experience). However, the most common emotion process components that constitute the hybrid of componential process, discrete emotion, and dimensional ontology, are as follows. 其他研究也提出了多个组件。Mesquita等人 [128] 识别了一组更广泛的组件：(a) 前因事件，(b) 情感体验，(c) 评估，(d) 生理变化，(e) 行动准备的变化，(f) 行为，(g) 认知功能和信念的变化，以及 (h) 过程，表明大多数情感实例都涉及所有这些。Scherer [111] 识别了五个组件：认知（评估）；神经生理（身体症状）；动机（行动倾向）；运动表达（面部和声音表达）；以及主观感受（情感体验）。然而，构成组合过程、离散情感和维度混合本体的最常见情感过程组件如下。

Event. An event, such as product malfunction, stimulates a response after it is evaluated [111]. An emotionally significant event triggers one’s appraisal and sequential emotion process components, such as an affective response, bodily change, action readiness, or behavior [23]. 事件：一个事件，如产品故障，在评估后引发反应 [111]。一个情感上重要的事件触发一个人的评估和连续的情感过程组件，如情感反应、身体变化、行动准备或行为 [23]。

Appraisal. Appraisal is an essential cognitive component [23]. One appraises an event and its consequences on multiple dimensions and processes these appraisals sequentially: (1) relevance, (2) implication, (3) coping potential, and (4) normative significance [81, 129], although some may overlap. Major appraisal types appear in Table 6. 评估：评估是一个重要的认知组件 [23]。一个人评估一个事件及其在多维度上的后果，并依次处理这些评估：(1) 相关性，(2) 含义，(3) 应对潜力，以及 (4) 规范性意义 [129, 81]，尽管有些可能重叠。主要的评估类型见表6。

Affect. This is a subjective feeling component [81]. One’s appraisal of an event triggers affective responses [23]. Appraisal gives cognitive character, whereas affect deals with feelings. For feeling states, discrete emotions are generated by a specific cause and are short-lived [5]. For feeling traits, one experiences positive or negative affect, from positive or negative emotions, respectively, with levels ranging from activated to inactivated [5]. 情感：这是一个主观感受组件 [81]。一个人对事件的评估触发情感反应 [23]。评估赋予认知特征，而情感则处理感受。对于感受状态，离散情感由特定原因生成，并且是短暂的 [5]。对于感受特质，一个人经历正面或负面情感，分别来自正面或负面情感，其水平从激活到未激活 [5]。

Action readiness. This is “an individual’s tendency to interact with their environment” [24, p. 213]. Action readiness variables could be: moving toward (e.g., I want to stay close or approach something); moving away (e.g., I want to protect myself or stay away from something); and moving against (e.g., I want to oppose or go against something, such as not buying it). Action readiness can result from a particular emotion. For example, anger and rage can result in moving against. 行动准备：这是“个体与环境互动的倾向” [24, p. 213]。行动准备变量可以是：接近（如我想保持亲近或接近某物）；远离（如我想保护自己或远离某物）；以及反对（如我想反对或对抗某物，如不购买它）。行动准备可能由特定情感引发。例如，愤怒和愤怒可能导致反对。

Behavior. One takes action to change one’s surroundings or maintain a desired situation. The inclination of behavior is determined by the negativity (positivity) and the intensity of all other emotional components including event, appraisal, affect, and action readiness [84]. 行为：一个人采取行动以改变其周围环境或维持期望的情境。行为的倾向由所有其他情感组件的负面性（正面性）和强度决定，包括事件、评估、情感和行动准备 [84]。

![[Table 6. Major Appraisal Types and their Dimensions.png]]

4.3.3 分类的先前情感本体研究（Classified Prior Emotion Ontology Studies）

Bianchi-Berthouze and Lisetti [78] develop an emotion ontology to capture affect from human facial expressions. The components include affect, facial expression, various types of appraisal (e.g., agency, novelty, controllability, modifiability, and external norms), and action tendency. Their taxonomy of affect includes feeling states (emotion and mood) and feeling traits (e.g., pleasantness vs. unpleasantness). For the discrete emotion approach, they use a set of six basic emotions (happy, sad, surprised, disgusted, fearful, and angry), with two contextual extensions (neutral and unspec) for capturing facial expressions. They also include an extended set of emotions (called ‘emotion-label’). For the dimensional approach, they capture affect, based on valence (positive or negative) and intensity (high, medium, or low) dimensions. Bianchi-Berthouze和Lisetti [78] 开发了一个情感本体，以捕捉人类面部表情中的情感。组件包括情感、面部表情、各种类型的评估（如代理、新颖性、可控性、可修改性和外部规范）以及行动倾向。他们的情感分类包括感受状态（情感和情绪）和感受特质（如愉悦 vs. 不愉悦）。对于离散情感方法，他们使用了一组六种基本情感（快乐、悲伤、惊讶、厌恶、恐惧和愤怒），并包括两个上下文扩展（中性和未指定）以捕捉面部表情。他们还包含了一组扩展的情感（称为“情感标签”）。对于维度方法，他们基于效价（正面或负面）和强度（高、中或低）维度捕捉情感。

Grassi’s [7] human emotion ontology (HEO) annotates emotions in complicated multimedia data (e.g., voice, facial expression, text, gesture). Based on appraisal theory [129] and the emotion process approach [80], this ontology includes componential components (appraisal dimensions, emotions, action tendencies, and regulation). The emotions are classified using both the discrete emotion and dimensional approaches. For the discrete emotion approach, the ontology includes: a class for six basic emotions [130]; and a class for an extended set of 48 emotions. For the dimensional approach, the ontology captures the status of emotion based upon valence, activation, and dominance dimensions. HEO has been adopted in other research [8, 126, 131]. Grassi的 [7] 人类情感本体（HEO）注释了复杂多媒体数据（如声音、面部表情、文本、手势）中的情感。基于评估理论 [129] 和情感过程方法 [80]，该本体包括组合组件（评估维度、情感、行动倾向和调节）。情感使用离散情感和维度方法进行分类。对于离散情感方法，本体包括：六种基本情感 [130] 的类；以及一组48种扩展情感的类。对于维度方法，本体捕捉了基于效价、激活和支配维度的情感状态。HEO已被其他研究采用 [8, 126, 131]。

Several studies [8, 78, 126, 131] identify a behavioral related concept, Action Tendency. This is a critical concept since emotion is a strong trigger of Action Tendency, which links an individual’s appraisal and actual actions [7]. There have also been efforts to exploit the benefit of designing relationships in an ontology; for example, ConceptNet [132], and others [7, 8]. The componential process approach suggests a temporal sequence among major concepts with significant relationships between Appraisal and Affect, between Appraisal and Action Readiness, and between these concepts and Behavior. 几项研究 [8, 78, 126, 131] 识别了与行为相关的概念，行动倾向。这是一个关键概念，因为情感是行动倾向的强烈触发因素，它将个体的评估与实际行动联系起来 [7]。还有一些努力利用了设计本体中关系的好处；例如，ConceptNet [132] 和其他 [7, 8]。组合过程方法建议主要概念之间的时间序列，并具有评估与情感、评估与行动准备以及这些概念与行为之间的显著关系。

4.3.4 组合过程、离散与维度混合本体的设计与开发Design and Development of Hybrid of Componential Process, Discrete Emotion, and Dimensional Ontology）

The hybrid of componential process, discrete emotion, and dimensional ontology, as illustrated in Figure 8, is developed from a combination of the three emotion paradigms and prior studies, whose concepts and relationships can be further extended. 组合过程、离散与维度混合本体，如图8所示，是从三种情感范式和先前研究的组合中开发出来的，其概念和关系可以进一步扩展。

The discrete emotion and dimensional paradigms, as theory bases, are applied to Affect. The ontology is largely comprised of the (sub)concepts of Event, Appraisal, Affect, Action Readiness, Behavior, and their contextual extensions, supporting simple and complex affect classifications. The major concepts or classes (Event, Appraisal, Affect, Action Readiness, and Behavior), and their relationships, are based on theories of the componential process paradigm [23, 24, 81]. The discrete emotion approach is applied to the BasicEmotionCategory and ComplexEmotionCategory classes and their subclasses. The hybrid also supports a simple and complex affect classification. The dimensional approach is applied to Dimension and its subclasses, serving as properties that characterize affect. The Contextual Extensions capture context (Individual, Environmental and System Related concepts). Here, the contextual extensions can be added to any of the major concepts in the ontology (shown in the dash-lined hasRelationship relation). Also, hasRelationship can be customized and re-named based upon the purpose of a study. Behavioral related concepts can include ActionReadiness (e.g., behavioral intention) and Behavior (e.g., actual behaviors). Behavior can be captured by: (1) extracting relevant terms directly from user-generated text based on a lexicon; and (2) inferring the inclination of a behavior based on the negativity (or positivity) and intensity of other remaining components, taking advantage of an established theory base from the emotion paradigms and findings from prior psychology studies [23, 84]. 离散情感和维度范式作为理论基础应用于情感。本体主要由事件、评估、情感、行动准备、行为及其上下文扩展的（子）概念组成，支持简单和复杂的情感分类。主要概念或类（事件、评估、情感、行动准备和行为）及其关系基于组合过程范式的理论 [23, 24, 81]。离散情感方法应用于BasicEmotionCategory和ComplexEmotionCategory类及其子类。混合本体还支持简单和复杂的情感分类。维度方法应用于Dimension及其子类，作为表征情感的属性。上下文扩展捕捉上下文（个体、环境和系统相关概念）。在这里，上下文扩展可以添加到本体中的任何主要概念（如虚线hasRelationship关系所示）。此外，hasRelationship可以根据研究目的进行定制和重命名。行为相关概念可以包括ActionReadiness（如行为意图）和Behavior（如实际行为）。行为可以通过以下方式捕捉：(1) 基于词典从用户生成的文本中提取相关术语；(2) 基于其他剩余组件的负面性（或正面性）和强度推断行为的倾向，利用情感范式的既定理论基础和心理学研究的发现 [23, 84]。

In the development of semantic relationships, three emotion paradigms provide the theory bases. Various types of between-concept relationships can also be designed, including relationships between superclass-subclass and relationships between classes (e.g., emotions, dimensions, appraisals, behaviors, and other concepts from contextual extensions) [133]. 在语义关系的开发中，三种情感范式提供了理论基础。还可以设计各种类型的概念间关系，包括超类-子类关系和类之间的关系（如情感、维度、评估、行为和其他上下文扩展概念）[133]。

![[Fig. 8. An illustration of hybrid of componential process, discrete, and dimensional emotion ontology..png]] ### 4.3.5 案例示例：基于人类情感本体的客户意见挖掘（An Example Case: Human Emotion Ontology for Customer’s Opinion Mining） Cambria et al. [8] use the Grassi’s [7] human emotion ontology (HEO) in sentiment analysis for marketing and product positing. The process of inferring affect in text follows. First, in the natural language process (NLP) module, the indicators of the valence of affect are interpreted and classified (e.g., negations, emoticons, adverbs, exclamation words). Second, a semantic parser breaks the text into concepts and relevant information (e.g., frequency, valence, activation) from the classified text in the NLP module. Third, in the AffectiveSpace phase, four dimensional vectors are created, which contain information on four affective dimensions: pleasantness, attention, sensitivity, and aptitude. Next, the extracted affective information is encoded with RDF/XML, based on HEO, and stored in a database, allowing powerful queries. HEO is developed using OWL description logic (OWL DL), which allows a taxonomical organization of emotions, their properties, and other relevant concepts. Cambria等人 [8] 在情感分析中使用了Grassi的 [7] 人类情感本体（HEO）进行市场营销和产品定位。推断文本中情感的过程如下。首先，在自然语言处理（NLP）模块中，解释并分类情感效价的指标（如否定、表情符号、副词、感叹词）。其次，语义解析器将文本分解为概念和相关信息（如频率、效价、激活）从NLP模块中的分类文本中提取。第三，在AffectiveSpace阶段，创建了四个维度向量，其中包含四个情感维度的信息：愉悦性、注意力、敏感性和能力。接下来，提取的情感信息使用HEO编码为RDF/XML，并存储在数据库中，允许强大的查询。HEO使用OWL描述逻辑（OWL DL）开发，允许对情感、其属性和其他相关概念进行分类组织。

4.3.6 评估（Evaluation）

Several studies [8, 78, 126, 131] demonstrate the potential of a hybrid of the componential process, discrete, and dimensional ontology approach. Cambria et al.’s [8] prototype assesses the accuracy of polarity detection in YouTube video reviews. They also use a corpus of LiveJournal.com and evaluate the capability of affect recognition. Both evaluations show high precision and recall rates. Bianchi-Berthouze & Lisetti [78] evaluate their system where an emotion ontology is realized to extract affective information from human facial expressions. The system’s precision performance in the filtering process outperforms Altavista.com. 几项研究 [8, 136, 131, 78] 展示了组合过程、离散与维度混合本体方法的潜力。Cambria等人 [8] 的原型评估了YouTube视频评论中极性检测的准确性。他们还使用了LiveJournal.com的语料库，并评估了情感识别的能力。两项评估都显示出高精确度和召回率。Bianchi-Berthouze & Lisetti [78] 评估了他们的系统，其中情感本体被实现为从人类面部表情中提取情感信息。系统在过滤过程中的精确性能优于Altavista.com。

4.3.7 贡献（Contribution）

The hybrid of componential process, discrete, and dimensional emotion ontology is the most sophisticated, enabling researchers to capture contextual, cognitive, affective, and behavioral concepts. Theory-driven relationships, between the major concepts, capture information derived from the systematic connections and sequences among the concepts. The sequences are critical, so using relationships can lead to automatically extracting logical conclusions about data that has been annotated with the hybrid ontology terms, thus providing meaningful insights into the annotations [134, 135]. In contrast to the discrete and the hybrid of discrete and dimensional emotion ontologies, its concepts and relationships are interconnected, and operate as a unit (versus being combined freely). Rich and fine-grained information can be obtained for affect. Affect and its subclasses, supported by the discrete emotion approach, capture basic or complex categories of emotions. Various dimensions of affect can be captured with contextual extensions to customize this hybrid. 组合过程、离散与维度混合情感本体是最复杂的，使研究人员能够捕捉上下文、认知、情感和行为概念。理论驱动的关系，在主要概念之间，捕捉了从概念之间的系统连接和序列中得出的信息。这些序列至关重要，因此使用关系可以导致自动提取逻辑结论，关于已用本体术语标记的数据，从而提供对底层注释的有用见解 [134, 135]。与离散和离散与维度混合情感本体相比，其概念和关系相互关联，并作为一个单元运行（而不是自由组合）。可以获得丰富且细粒度的情感信息。情感及其子类，由离散情感方法支持，捕捉基本或复杂的情感类别。可以通过上下文扩展捕捉情感的各种维度，以定制此混合本体。

5. 情感本体的开发用于情感分析（DEVELOPMENT OF EMOTION ONTOLOGIES FOR SENTIMENT ANALYSIS）

Emotion ontologies can be created using general ontology development methodologies: (1) stagebased, (2) evolving prototype, and (3) ontology learning [14, 136]. We show the distinct features of each methodology and how to use them in the application of our framework. 情感本体可以使用一般的本体开发方法来创建：(1) 基于阶段的方法，(2) 演化原型方法，以及 (3) 本体学习方法 [14, 136]。我们展示了每种方法的独特特征，以及如何将它们应用于我们的框架。

Stage-based ontology development methodologies. These methodologies develop ontologies based on specified stages [136]. Uschold and King [137] suggest a four-staged methodology: (1) identify purpose; (2) build the ontology (ontology capture, coding, and integrating existing ontologies); (3) evaluate; and (4) document. In TOVE (TOronto Virtual Enterprise) project, Grüninger and Fox [138] propose a six-staged methodology: (1) motivating scenario, (2) informal competency questions, (3) terminology (first-order logic), (4) formal competency questions, (5) axioms (first-order logic), and (6) completeness theorems. Both methodologies emphasize the usefulness of scenarios to guide the process of ontology construction [139]. Here, scenarios, as a means of benchmarking various design solutions, help practitioners model, validate information, and communicate with stakeholders. Thus, stage-based methodologies are suitable for scenario-based ontology development, in which the objectives and requirements are apparent [136]. 基于阶段的本体开发方法。这些方法基于指定的阶段开发本体 [136]。Uschold和King [137] 提出了一个四阶段的方法：(1) 确定目的；(2) 构建本体（本体捕获、编码和集成现有本体）；(3) 评估；(4) 文档化。在TOVE（多伦多虚拟企业）项目中，Gruninger和Fox [138] 提出了一个六阶段的方法：(1) 动机场景，(2) 非正式能力问题，(3) 术语（一阶逻辑），(4) 正式能力问题，(5) 公理（一阶逻辑），以及 (6) 完备性定理。这两种方法都强调了使用场景来指导本体构建过程的有用性 [139]。在这里，场景作为基准各种设计解决方案的手段，帮助从业者建模、验证信息并与利益相关者沟通。因此，基于阶段的方法适用于基于场景的本体开发，其中目标和需求是明确的 [136]。

Evolving-prototype ontology development methodologies. These methodologies develop ontologies from scratch, support knowledge acquisition activities, and use incremental development processes based on evolving prototypes [e.g., 136, 140, 141]. For example, Fernández-López et al. [142] propose an ontology life cycle, starting from specification, conceptualization, formalization, integration, implementation, and maintenance. Concurrently, knowledge acquisition, ontology evaluation, and documentation are performed and support the ontology life cycle. Noy and McGuinness [143] suggest an iterative ontology development process that will continue until the ontology life cycle ends: (1) determine the domain and scope of the ontology; (2) consider reusing existing ontologies; (3) enumerate important terms in the ontology; (4) define the classes and the class hierarchy; (5) define the properties of classes-slots; (6) define the facets of the slots; and (7) create instances. The evolving-prototype methodologies are very useful when ontology requirements are not easily identified at the beginning and need to be refined as time progresses. 演化原型本体开发方法。这些方法从零开始开发本体，支持知识获取活动，并使用基于演化原型的增量开发过程（例如 [136, 140, 141]）。例如，Fernandez-Lopez等人 [142] 提出了一个本体生命周期，从规范、概念化、形式化、集成、实现和维护开始。同时，知识获取、本体评估和文档化被执行并支持本体生命周期。Noy和McGuinness [143] 提出了一个迭代的本体开发过程，该过程将持续到本体生命周期结束：(1) 确定本体的领域和范围；(2) 考虑重用现有本体；(3) 枚举本体中的重要术语；(4) 定义类和类层次结构；(5) 定义类的属性-槽；(6) 定义槽的方面；(7) 创建实例。当本体需求在开始时不易识别并需要随时间细化时，演化原型方法非常有用。

Ontology learning methodologies. These methodologies can automatically develop an ontology by extracting concepts and relationships from data sources [30]. They can also develop an ontology from scratch [14] or reuse existing ontologies and further enrich or adapt them. Unstructured, semi-structured, or structured data can be used for ontology learning by employing various technologies such as NLP (Natural Language Processing) tools, machine learning, data mining information retrieval, and statistics [14, 144]. 本体学习方法。这些方法可以自动从数据源中提取概念和关系来开发本体 [30]。它们还可以从零开始开发本体 [14] 或重用现有本体并进一步丰富或调整它们。可以使用非结构化、半结构化或结构化数据进行本体学习，通过使用各种技术，如自然语言处理（NLP）工具、机器学习、数据挖掘信息检索和统计 [144, 14]。

In Text2Onto, a framework of ontology learning from textual data, Cimiano and Völker [145] develop a POM (Probabilistic Ontology Model) that retains the outcomes of various ontology learning algorithms, whose purposes include: learning about changes to the corpus, comparing to a reference repository, and generating results. They also integrate both machine learning and NLP techniques and learn patterns to capture taxonomic relationships. To develop an ontology for sentiment analysis in Twitter, Kontopoulos et al. [30] propose OntoGen [146], a semi-automatic, data-driven ontology editor supported by machine learning and text mining techniques. OntoGen [146] can provide users with potential concepts and relationships and automatically assign instances to the concepts by employing both unsupervised and supervised learning. The benefits of using ontology learning methodologies include avoiding potential subjectivity in manual ontology development; and reducing the cost and time of the development [14]. 在Text2Onto中，一个从文本数据中学习本体的框架，Cimiano和Volker [145] 开发了一个POM（概率本体模型），保留了各种本体学习算法的结果，其目的包括：学习语料库的变化，与参考存储库进行比较，并生成结果。他们还集成了机器学习和NLP技术，并学习模式以捕捉分类关系。为了开发用于Twitter情感分析的本体，Kontopoulos等人 [30] 提出了OntoGen [146]，一个半自动的、数据驱动的本体编辑器，由机器学习和文本挖掘技术支持。OntoGen [146] 可以为用户提供潜在的概念和关系，并通过使用无监督和监督学习自动将实例分配给概念。使用本体学习方法的好处包括避免手动本体开发中的潜在主观性；并减少开发成本和时间 [14]。

5.1 情感本体框架在开发方法中的应用（Application of Framework of Emotion Ontologies to Development Methodologies）

According to Cimiano et al. [144], regardless of the ontology development methodology, generally the ontology life cycle includes the following phases: feasibility study, requirements analysis, conceptualization, maintenance, evaluation, and application. In the conceptualization phase, a model is formalized and implemented with various ontology languages. Our framework (Figure 3 and Table 2) can be applied to this phase in a stand-alone ontology development approach, which means that a chosen emotion ontology serves as the main ontology, with additional ontologies and/or contextual extensions added as supporting concepts. A chosen ontology of our framework can be included as an emotion ontology module in a much larger ontology development effort. 根据Cimiano等人 [144] 的说法，无论本体开发方法如何，通常本体生命周期包括以下阶段：可行性研究、需求分析、概念化、维护、评估和应用。在概念化阶段，模型被形式化并使用各种本体语言实现。我们的框架（图3和表2）可以应用于此阶段的独立本体开发方法，这意味着选择的情感本体作为主要本体，其他本体和/或上下文扩展作为支持概念添加。我们框架中选择的本体可以作为情感本体模块包含在更大的本体开发工作中。

Our framework provides ontology developers with role models of emotion ontologies, which can be adapted and extended to develop customized emotion ontologies. It can be applied to the ontology development methodologies as shown below. 我们的框架为本体开发者提供了情感本体的角色模型，可以适应和扩展以开发定制的情感本体。它可以应用于本体开发方法，如下所示。

When using the stage-based methodologies, it is important to identify a good fit between a type of ontology in our framework and the scenarios associated with a target project. For instance, in Uschold and King’s [137] methodology, once a decision on a type of emotion ontology is made, the chosen ontology approach provides knowledge bases for the ‘developing ontology’ stage where an ontology is captured and coded and existing ontologies are integrated. Adaptation and/or extension of the chosen emotion ontology occurs in this stage. 当使用基于阶段的方法时，重要的是确定我们框架中的本体类型与目标项目相关场景之间的良好匹配。例如，在Uschold和King的 [137] 方法中，一旦确定了情感本体的类型，选择的本体方法为“开发本体”阶段提供了知识基础，其中本体被捕获和编码，并集成现有本体。在此阶段发生选择情感本体的适应和/或扩展。

Similarly, with the evolving-prototype methodologies, a chosen emotion ontology approach can guide the specification, conceptualization, formalization, and integration processes of ontology development as suggested by Fernández-López et al. [142]. Identifying additional contextual extensions and adapting the chosen ontology can occur through ontology requirement specification [142] and the stage of ‘determining the domain and scope of ontology’ in Noy and McGuinness’ [143] iterative development process. 同样，使用演化原型方法时，选择的情感本体方法可以指导本体开发的规范、概念化、形式化和集成过程，如Fernández-López等人 [142] 所建议的。通过本体需求规范 [142] 和Noy和McGuinness [143] 迭代开发过程中的“确定本体的领域和范围”阶段，可以识别额外的上下文扩展并适应选择的本体。

When using the ontology learning methodologies, types of structured data (such as an “already defined knowledge model including existing ontologies and database schema”) are used as important knowledge resources [14, p. 17]. Thus, a chosen emotion ontology approach from our framework can provide knowledge resources and additional concepts and structures to add to an existing taxonomy. Supervised and unsupervised methods can be used to achieve this objective [144]. 当使用本体学习方法时，结构化数据（如“已经定义的知识模型，包括现有本体和数据库模式”）被用作重要的知识资源 [14, p. 17]。因此，我们框架中选择的情感本体方法可以提供知识资源和额外的概念和结构，以添加到现有分类中。可以使用监督和无监督方法来实现这一目标 [144]。

5.2 情感本体开发的示例（Illustration of Emotion Ontology Development）

We illustrate emotion ontology development using an evolving-prototype ontology development methodology as suggested by Noy and McGuinness [143]. The hybrid of componential process, discrete, and dimensional ontology is selected from the framework (Figure 3 and Table 2), developed as a stand-alone emotion ontology, and further extended by identifying contextual extensions. The objective of the ontology development is to support sentiment analysis in project reviews by customers in a company’s social media sites. We incorporate the framework by adding a “select an emotion ontology” step to choose a suitable ontology development approach. 我们使用Noy和McGuinness [143] 建议的演化原型本体开发方法来说明情感本体的开发。从框架中选择组合过程、离散与维度混合本体（图3和表2），开发为独立情感本体，并通过识别上下文扩展进一步扩展。本体开发的目的是支持公司社交媒体网站中客户项目评论的情感分析。我们通过添加“选择情感本体”步骤来结合框架。

[Step 1] Identify the domain and scope of the emotion ontology [步骤1] 确定情感本体的领域和范围
Pragmatic questions guide researchers in defining the domain and scope of the emotion ontology: (1) What is the target environment? (2) Who are the target users? (3) What is the target domain? (4) What is the target information? (5) Which characteristics should an ontology developer consider? We illustrate these for a project to capture sentiments from customer feedback, while mining online technology support forums. 实用问题指导研究人员定义情感本体的领域和范围：(1) 目标环境是什么？(2) 目标用户是谁？(3) 目标领域是什么？(4) 目标信息是什么？(5) 本体开发者应考虑哪些特征？我们为捕捉客户反馈情感的项目进行了说明，同时挖掘在线技术支持论坛。

• The target environment is online technology support forums.
• 目标环境是在线技术支持论坛。
• The target users are customers who experience technology malfunctions and staff and voluntary experts who support users in resolving customers’ issues as they appear in support forums.
• 目标用户是经历技术故障的客户，以及支持用户在支持论坛中解决问题的员工和自愿专家。
• Target domains are products such as notebooks and desktops.
• 目标领域是产品，如笔记本电脑和台式机。
• Target information is the sentiments (e.g., thoughts, affect) and behavioral intentions of customers in the company’s forums.
• 目标信息是公司论坛中客户的情感（如想法、情感）和行为意图。

[Step 2] Select an emotion ontology [步骤2] 选择情感本体
To select an ontology from the framework (Figure 3), consider the objective of the sentiment analysis project. If it is to capture basic or complex discrete emotions, select the discrete emotion ontology. If it is to additionally capture the dimensional nature of affect, select the hybrid of discrete and dimensional ontology. If it is to capture the interrelated components associated with the process of emotion generation and the impact of affect as well as the dimensional nature of the affect, then select the hybrid of componential process, discrete, and dimensional ontology.

为了从框架中选择本体（图3），考虑情感分析项目的目标。如果目标是捕捉基本或复杂的离散情感，选择离散情感本体。如果目标是额外捕捉情感的维度性质，选择离散与维度混合本体。如果目标是捕捉与情感生成过程相关的组件以及情感的影响以及情感的维度性质，则选择组合过程、离散与维度混合本体。

To capture sentiments from customer feedback, suppose we are interested in capturing information on overall affect generating elements such as technology issues, customers’ appraisals of a situation, affect (a basic set of emotions of customers), and behavioral intentions in order to infer actual behaviors (e.g., no/repurchase a product, continue to use product, or switch products). Additionally, the polarity (positive or negative valence) of affect needs to be captured. Therefore, the hybrid of the componential process, discrete, and dimensional ontology is appropriate. Furthermore, the results from Step 1 (domain and scope) show the need for contextual extensions to capture contextual information, such as product and user profiles. 为了捕捉客户反馈中的情感，假设我们有兴趣捕捉整体情感生成元素的信息，如技术问题、客户对情况的评估、情感（客户的基本情感集）和行为意图，以推断实际行为（如不再/再次购买产品、继续使用产品或切换产品）。此外，需要捕捉情感的极性（正面或负面效价）。因此，组合过程、离散与维度混合本体是合适的。此外，步骤1（领域和范围）的结果显示需要上下文扩展来捕捉上下文信息，如产品和用户配置文件。

[Step 3] Define classes and class hierarchy [步骤3] 定义类和类层次结构
The hybrid of componential process, discrete, and dimensional ontology is customized by adding and refining (sub)classes, as shown in Figure 9. For instance, in the Affect class (concept), a basic emotion category is defined to capture users’ emotions (e.g., Anger, Disappointment, Joy, and others). The ActionReadiness class is designed to extract users’ behavioral intentions: Move Toward (e.g., continuance use, repurchase), Move Away (e.g., no repurchase), and others. Corresponding to the ActionReadiness subclasses, the Behavior class is defined to infer the inclinations of actual behaviors based on the remaining components since text does not often include terms that indicate actual behaviors: Move Toward, Move Away, and Others. As contextual extensions, Product, as well as User and its subclasses (Age, Gender, and others), are added. Figure 9 is an example, which can be further customized and extended. 组合过程、离散与维度混合本体通过添加和细化（子）类进行定制，如图9所示。例如，在Affect类（概念）中，定义了基本情感类别以捕捉用户的情感（如Anger、Disappointment、Joy等）。ActionReadiness类设计用于提取用户的行为意图：Move Toward（如继续使用、再次购买）、Move Away（如不再购买）等。对应于ActionReadiness子类，Behavior类定义为基于剩余组件推断实际行为的倾向，因为文本通常不包括指示实际行为的术语：Move Toward、Move Away等。作为上下文扩展，添加了Product以及User及其子类（Age、Gender等）。图9是一个示例，可以进一步定制和扩展。

![[Fig. 9. Example of hybrid of componential process, discrete, and dimensional emotion ontology..png]]

[Step 4] Define the properties of classes
[步骤4] 定义类的属性
For the Event class, properties such as EventID and TitleOfEvent are identified and can be added to the ontology. These properties are the elements that form each class and result from further searches and exploration. Similarly, for the other classes (Appraisal, Affect, etc.). 对于Event类，识别了诸如EventID和TitleOfEvent等属性，并可以添加到本体中。这些属性是形成每个类的元素，并来自进一步的搜索和探索。同样，对于其他类（Appraisal、Affect等）。

[Step 5] Define the facets of the slots
[步骤5] 定义槽的方面 Properties (or slots) can have rules governing their values, called facets. The EventID property, for instance, can have only one value, so ‘1’ is assigned to :MAXIMUM-CARDINALITY. 属性（或槽）可以具有管理其值的规则，称为方面。例如，EventID属性只能有一个值，因此将“1”分配给:MAXIMUM-CARDINALITY 。

[Step 6] Create instances
[步骤6] 创建实例
Ontology development software can support the instantiation of the emotion ontology. It can work with different types of ontology designing and supporting languages, including OWL, RDF, RDF schema, and XML [147]. 本体开发软件可以支持情感本体的实例化。它可以与不同类型的本体设计和支持语言一起工作，包括OWL、RDF、RDF模式和XML [147]。

5.3 框架在情感分析中的应用（Application of the Framework to Sentiment Analysis）

After an emotion ontology is developed, it can be applied to sentiment analysis. Figure 10 illustrates a process for doing, based on a set of steps identified in sentiment analysis research [e.g., 25, 148, 149]. 在开发情感本体后，可以将其应用于情感分析。图10展示了一个基于情感分析研究中识别的一组步骤的过程 [例如25, 148, 149]。

In the data input step, user-generated data are collected from online social network sites. For data extraction and classification, natural language processing (NLP) and text analytics are often employed [148]. In the sentiment detection step, the emotion ontology (chosen or developed) guides which pieces of sentences in user reviews should be retained (e.g., those suggesting subjective affect) or removed [148]. 在数据输入步骤中，从在线社交网络站点收集用户生成的数据。对于数据提取和分类，通常使用自然语言处理（NLP）和文本分析 [148]。在情感检测步骤中，情感本体（选择或开发的）指导应保留用户评论中的哪些句子（如那些暗示主观情感的句子）或删除 [148]。

In the sentiment classification step, sentiment is classified at the document, sentence, or aspect levels [148]. Documents and sentences bearing sentiments can be classified based on a polarity (e.g., positive, negative, neutral). In the aspect level, the sentiment can be classified based on specific aspects of entities (e.g., a product or service that triggered affect) [150]. The most commonly used sentiment classification methods for this task are [149, 151]: (1) lexicon-based, which uses manual annotation to create a set of seed words and then bootstrap them via synonym detection from online dictionaries; and (2) statistical methods, which use automated tools (e.g., point-wise mutual information, chi-square, latent semantic indexing). An emotion ontology encoded in ontology designing languages (e.g., OWL and RDF schema) can support a lexicon-based approach. Using an emotion ontology in connection with a lexicon(s) could be very promising for classifying sentiment at the aspect-level (versus document or sentence level) [152]. 在情感分类步骤中，情感在文档、句子或方面级别进行分类 [148]。带有情感的文档和句子可以基于极性（如正面、负面、中性）进行分类。在方面级别，情感可以基于实体的特定方面（如触发情感的产品或服务）进行分类 [150]。最常用的情感分类方法用于此任务的是 [149, 151]：(1) 基于词典的方法，使用手动注释创建一组种子词，然后通过在线词典中的同义词检测引导它们；(2) 统计方法，使用自动化工具（如点互信息、卡方检验、潜在语义索引）。使用本体设计语言（如OWL和RDF模式）编码的情感本体可以支持基于词典的方法。将情感本体与词典结合使用在方面级别（相对于文档或句子级别）分类情感可能非常有前途 [152]。

The emotion ontologies in the framework (Figure 3 and Table 2) can support a lexicon-based approach because it provides a formal representation of aspects and semantic associations between them. It can also capture associated concepts and enable reasoning, which is better than a taxonomy or relational database-based approach [25]. This should lead to a high level of sentiment classification accuracy, compared to non-ontology-based approaches [153]. 框架中的情感本体（图3和表2）可以支持基于词典的方法，因为它提供了方面的正式表示和它们之间的语义关联。它还可以捕捉相关概念并启用推理，这比分类或基于关系数据库的方法更好 [25]。这应该导致与非本体方法相比的高水平情感分类准确性 [153]。

The interpretation and inference of output step is newly added to the general sentiment analysis steps. Prior sentiment analysis studies rarely go beyond the outputs generated from the sentiment classification. Furthermore, user-generated data from social network sites often omit sentences that indicate affect or actual behavior. Instead, they only include facts and opinions, even though the context captured can imply that users are emotional, suggesting that further interpretation of sentiment would be helpful. A simple illustration of sentiment analysis, with the hybrid of componential process, discrete, and dimensional emotion ontology, is shown in Table 7. The interpretation and inferences are based on theories of the componential process paradigm, where each component (Event, Appraisal, Affect, Action Readiness, and Behavior) is interrelated and synchronized. The combination of all components influences the inclination of behavior [23, 84]. 解释和推断输出步骤是新添加到一般情感分析步骤中的。先前的情感分析研究很少超越情感分类生成的输出。此外，来自社交网络站点的用户生成数据通常省略指示情感或实际行为的句子。相反，它们只包括事实和意见，即使捕捉到的上下文暗示用户是情感的，这表明进一步解释情感将是有帮助的。表7展示了一个简单的组合过程、离散与维度混合情感本体的情感分析示例。解释和推断基于组合过程范式的理论，其中每个组件（事件、评估、情感、行动准备和行为）相互关联并同步。所有组件的组合影响行为的倾向 [23, 84]。

In the final step, the presentation of output, the results of the sentiment analysis are organized into various formats (charts, graphs, tables, or texts). Statistical data (e.g., means, frequency) and textual data from the interpretation and inference step can be organized and displayed. 在最后一步，输出展示中，情感分析的结果被组织成各种格式（图表、图形、表格或文本）。来自解释和推断步骤的统计数据（如均值、频率）和文本数据可以被组织和显示。

![[Table 7. Illustration of Captured Data from Previous Steps and Results of Interpretations.png]]

6. 讨论（DISCUSSION ）

6.1 对研究的启示（Implications for Research）

This research has derived a Framework of Emotion Ontologies from research on emotion ontology, psychology, and sentiment analysis (Figure 3 and Table 2). Besides its intended uses, the framework could facilitate active communication among researchers in sentiment analysis, emotion ontologies, and psychology studies. 本研究通过综合情感本体、心理学和情感分析领域的研究，提出了情感本体框架（图3和表2）。除了其既定用途外，该框架还有助于促进情感分析、情感本体和心理学研究者之间的交流。

A Framework of Emotion Ontologies. The framework can provide guidance for developing and supporting emotion ontologies for sentiment analysis. Affect is an important concept in sentiment analysis, but a great deal of work has focused only on polarity (e.g., [27, 28, 154]). The emotion ontologies can capture and represent richer and more fine-grained information on affect. In sentiment analysis, there is inconsistent use of terminology [26, 27], which this research addresses by differentiating core concepts (affect, emotion). Technology challenges arise when dealing with large amounts of social data that have semantic inconsistencies and a lack of structure [74]. Systematic mapping of terms to an ontology can help to obtain consistent interpretations [3, 16] from reusable emotion ontologies. 情感本体框架。该框架为开发和支持情感分析中的情感本体提供了指导。尽管情感是情感分析的核心概念，但许多研究仅关注情感极性（如[27, 28, 154]）。情感本体能够捕捉更丰富、更细粒度的情感信息。情感分析中术语使用的不一致问题[26, 27]在本研究中通过区分核心概念（如“情感”与“情绪”）得以解决。此外，社交媒体数据的语义不一致性和结构缺失等技术挑战[74]可通过系统化的本体映射来缓解，从而提升数据解释的一致性[3, 16]。

For a given application the following questions arise.
针对具体应用时，以下问题值得关注：

1. Which emotion theory and paradigm are the most suitable for a given purpose and context? 哪些情感理论和范式最适合特定应用场景？
2. What are the representational and theoretical reasoning requirements and alternatives for an ontology development effort? 情感本体的表征和推理需要哪些要求？
3. Where do researchers find the data (e.g., emotion descriptors, stable behavioral patterns associated with particular emotions, criteria for measuring affect) required for emotion ontology design and development? 如何获取情感描述词、行为模式或情感测量标准等数据？

Our framework addresses these questions by providing emotion ontologies, based on both emotion theories and empirical studies. Since the selection of an appropriate emotion ontology is important, Figure 11 provides a decision tree that could aid in the selection process. Table 8 summarizes the benefits of each ontology in the framework and its application. 本框架通过整合情感理论和实证研究回答了这些问题。图11提供了选择情感本体的决策树，表8总结了框架中各类本体的优势和应用场景。

![[Fig. 11. Decision tree for emotion ontology selection..png]]

![[Table 8. Potential Users and Benefits of the Emotion Ontologies in the Framework.png]]

Context aware sentiment analysis. Capturing contextual information is essential for analyzing interactions between humans and environments [10] but has received less attention in sentiment analysis. Our review of emotion ontology studies shows that understanding the context relevant to the generation of affect could make it easier to offer proper products or services to customers (e.g., [78, 95, 121]). Contextual information, such as person and object related to affect, allows sentiment analysis researchers to recognize and interpret individuals’ affect in-depth in terms of analyses and uses of data [113]. We, therefore, integrated ‘Contextual Extensions’ into the emotion ontologies of the framework as one of the main components, which support the capture of contextual information such as individual, environmental, and system related concepts (Table 4 and Figures 4, 7, 8, and 9). 情境感知情感分析：捕捉上下文信息对于分析人与环境的交互至关重要[10]，但在情感分析中较少被关注。本框架通过“情境扩展”（Contextual Extensions）支持捕捉个体、环境和系统相关的上下文信息（表4及图4、7、8、9）。例如，结合用户和产品的特征，可更深入地理解用户情感[78, 95, 121]。

Behavior identifying sentiment analysis. Customers’ behavioral information (e.g., repurchase, switch, discontinuance) is very valuable information in sentiment analysis. It is of practical interest of firms and allows them to prepare proper strategies to resolve relevant issues. Despite its importance, capturing behavioral information has been previously missing in sentiment analysis [17]. Action readiness or behavioral intention (e.g., will not purchase again, will buy a different computer) can be extracted directly from user-generated text based on a lexicon [17]. On the other hand, actual behavior or potential future behaviors can be inferred by interpreting patterns of relevant data such as negativity and intensity of Appraisal and Affect [95], thereby harnessing prior emotion theories and findings from psychology [23, 84]. Depending upon the situation, our review of emotion ontology studies can provide guidance to researchers for sentiment analysis in three ways [155]: (1) developing a controlled vocabulary that defines Behavior, Action Readiness, Behavioral Concepts, and other existing classes (Table 4 and Figures 4, 7, 8, and 9); (2) specifying the relationships among Behavior, Action Readiness, Behavioral Concepts, and other classes; and (3) identifying the emotion paradigms and psychology studies that provide theoretical support and a knowledge base. 行为识别情感分析：用户行为（如复购、换品）对企业策略制定具有重要价值。本框架通过整合“行为意图”（Action Readiness）和“实际行为”（Behavior）等概念，结合情感理论和心理学研究，支持从文本中直接提取或间接推断用户行为倾向（如负面评价可能预示“不再购买”）[23, 84]。

Sequencing. The importance of sequence features in ontology has been recognized and applied in such fields as biology [134] and electrical engineering [135], but received little attention in sentiment analysis, even though they can provide insightful information (e.g., associations, sequences, and necessary conditions) when firms try to capture information on the emotion generating process of customers. Examples include: Which product or service creates issues and why? How do customers appraise the issues? How do the customers feel based on an issue? And, then, what are (or will be) their behavioral responses? The sequence of concepts (or classes) in emotion ontology is important. In a given situation, the hybrid of componential process & discrete & dimensional emotion ontology would be a valuable. The componential process emotion paradigm, a major theory base, provides information on the sequence of concepts [23, 84], so the hybrid ontology can specify precisely the features of sequences among concepts and relationships [134]. The benefits include: reusable knowledge representation of annotations that collectively define inter-related multi-components, the sequence, and relationships [135]. Also, using the hybrid ontology allows researchers to “automatically draw logical conclusions about data that has been labelled” with the ontology terms, “capture important trends in the data,” and hence “provide useful insights into the underlying annotations” [134, p. 11]. 序列化建模：情感生成过程中的概念序列（如“事件→评估→情感→行为”）能为企业提供关键洞察。基于情感过程范式（Componential Process Paradigm）的混合本体（CM & DS & DM）支持序列关系的逻辑推理，从而自动生成注释数据的趋势分析[134, 135]。 ## 6.2 对实践的启示（Implications for Practice） We provided and illustrated guidelines for how to design and develop emotion ontologies for each development step. Use of the Framework of Emotion Ontologies illustrates how a selected ontology can be customized for one’s purpose and used in various phases of sentiment analysis. Through the contextual extensions, interoperability among ontologies can be facilitated by reusing and integrating existing ontologies [121]. It can also be useful to connect an existing ontology to the customized ontology to obtain the contextual information needed to understand affect. An example is Cambria et al.’s [8] opinion mining of customers in social network sites. They reuse OMR (ontology for media resources) and FOAF (friend of a friend) ontologies and integrate them with their HEO (human emotion ontology). The OMR provides the vocabulary for media resources; FOAF captures information on people. This facilitates the mining of opinions, affect, and affect concepts from users who interact with multimedia content. 本研究提供了情感本体开发的逐步指南，并通过案例展示了框架的定制化应用。通过“情境扩展”，本体的互操作性得以增强，例如Cambria等人[8]通过整合OMR（媒体资源本体）和FOAF（社交关系本体），实现了多媒体内容中用户情感和观点的挖掘。 ## 6.3 Future Research Additional recommendations for when emotion ontologies could be helpful are given in Table 9. 表9总结了情感本体的潜在应用方向

Prior research on emotion ontologies cannot completely capture the vast amount of affect embedded in virtual characters or human expressions (e.g., body, facial, or speech). Doing so would require a rather large ontology of emotions, backed by strong theoretical support and proven successful applications. Of course, it is very difficult to argue for completeness since ontologies usually evolve. Furthermore, assessing the quality of ontologies is difficult for many reasons, including completeness and usefulness [156]. Addressing these issues should aid in emotion ontology development and application to other fields. Other potential areas are noted below.

Digital media. Emotion ontologies can support the extraction of affect in digital media (e.g., images, photos, animation, videos). Online museum portals and art exhibitions, for example, have features that allow participants to express their feelings and emotions by writing reviews [6]. Emotion ontology studies also attempt to capture participants’ affect from facial, vocal, and body expressions in multimedia with animation [7, 112]. 数字媒体。情感本体可以支持数字媒体（例如图像、照片、动画、视频）中的情感提取。例如，在线博物馆门户网站和艺术展览有允许参与者通过写评论来表达他们的感受和情绪的功能[6]。情感本体研究还试图从动画多媒体中的面部、声音和身体表情中捕捉参与者的情感[7,112]。

NeuroIS. Affective neuroscience studies use functional neuroimaging techniques, with data that must be integrated from experiments recorded in multiple databases [157]. An emotion ontology could be used to annotate the data (e.g., BrainMap database) [121]. NeuroIS。情感神经科学研究使用功能性神经成像技术，数据必须从多个数据库中记录的实验中整合[157]。情感本体可用于标注数据（例如，BrainMap数据库）[121]。

Social applications. An emotion ontology can become part of a wide range of applications (e.g., social games, educational applications, military training simulations), because it provides a structure for affect. Researchers can select target emotions [124]. 社会应用。情感本体可以成为广泛应用的一部分（例如，社交游戏，教育应用，军事训练模拟），因为它提供了情感的结构。研究者可以选择目标情绪[124]。

Customer relationship management (CRM). An emotion ontology could help classify customers’ affect with potential applications residing in customers’ data from chat services, company’s blogs, etc., to support customer relationship management [13]. It may be appropriate to respond differently, depending upon the reactions from a customer. 客户关系管理（CRM）。情感本体可以帮助将客户的情感与聊天服务、公司博客等客户数据中的潜在应用进行分类，以支持客户关系管理[13]。根据客户的反应，采取不同的反应可能是合适的。 # 7. 总结（CONCLUSION ） This paper has reviewed and synthesized research in emotion ontologies, psychology, and sentiment analysis, and proposes a Framework of Emotion Ontologies that can be used to guide sentiment analysis projects that require the support of an emotion ontology. The framework provides a standard way of representing affect, based on emotion theories, and augmented with contextual and behavioral information related to affect. The framework should be useful to assist researchers and practitioners in identifying and developing an ontology for their specific application and understanding the theories upon which the ontologies are based. Further work is required to apply the framework in different application domains.