A Three-Branch Checks-and-Balances Framework for Context-Aware Ethical Alignment of Large Language Models

A Checks-and-Balances Framework for Context-Aware Ethical AI Alignment Edward Y. Chang 1 Abstract This paper introduces a checks-and-balances framework for ethical alignment of Large Lan- guage Models (LLMs), inspired by three-branch governmental systems. It implements three in- dependent yet interacting components: LLMs as the executive branch for knowledge generation, Dikeas the legislative branch that establishes eth- ical guardrails, andErisas the judicial branch for contextual interpretation. Beyond structural sepa- ration, we address a fundamental challenge: regu- lating emotion to shape behaviors. Drawing from psychological theories where managing emotional responses prevents harmful behaviors, we develop a self-supervised learning pipeline that maps emo- tions to linguistic behaviors, enabling precise be- havioral modulation through emotional condition- ing. By integrating this approach with adversarial testing, our framework demonstrates howDike andErisdirect linguistic behaviors toward ethical outcomes while preserving independence through- out knowledge generation, ethical oversight, and contextual interpretation. 1. Introduction Ethical alignment in Large Language Models (LLMs) is a critical challenge, particularly given the limitations of Reinforcement Learning from Human Feedback (RLHF) (OpenAI, 2023; Ouyang et al., 2023). Although RLHF has demonstrated success in aligning AI systems with human values, it encounters two major issues: 1) susceptibility to social biases when feedback is polarized, and 2) vulnera- bility to reward hacking, where the system optimizes for feedback without genuine ethical improvement (Christiano et al., 2017; Skalse et al., 2022). These issues can result in unethical behavior or inconsistent performance. Beyond these implementation challenges, RLHF faces a 1 Computer Science, Stanford University. Correspondence to: Edward Y. Chang <echang@cs.stanford.edu>. Proceedings of the42 nd International Conference on Machine Learning, Vancouver, Canada. PMLR 267, 2025. Copyright 2025 by the author(s). more fundamental conceptual limitation: its narrow focus on isolated behaviors rather than holistic patterns. This reac- tive strategy is similar to a “Whack-A-Mole” game, where addressing one problematic behavior does not prevent the emergence of others. For example, consistently instructing someone to make their bed does not necessarily cultivate overall tidiness, such as doing laundry or washing dishes. Similarly, RLHF often emphasizes short-term fixes at the cost of long-term coherence, leading to catastrophic for- getting: users have reported that optimizing one task in ChatGPT can degrade performance in unrelated areas (Kirk- patrick et al., 2017; Lin et al., 2024; Dai et al., 2025). This challenge mirrors the difficulty of treating addiction, where addressing one symptom may reveal deeper psychological dependencies (Sinha, 2008; Torrens et al., 2005). To overcome these challenges, we propose a checks-and- balances framework inspired by governmental structures, where independent but interacting components maintain ac- countability and balance. Our architecture integrates three components: LLMs serve as the executive branch for knowl- edge generation;Dike(representing justice) functions as the legislative branch to set ethical standards; andEris(rep- resenting discord) acts as the judicial branch, providing adversarial testing and contextual interpretation. In mythol- ogy, Dike embodies order and justice, while Eris signifies discord, forming a duality that our framework leverages to balance ethical guidance with adversarial scrutiny. Figure 1 illustrates this three-branch architecture, where the neurally independent components—LLMs as the foun- dation, withDikeandErisas oversight layers—interact through structured interfaces while maintaining strict sepa- ration of their neural architectures and parameters. 1.1. Emotion Regulation as Behavioral Control A fundamental question underlies our framework: Can regu- lating emotions shape behaviors, and can similar principles be applied to LLMs? In human psychology, emotions sig- nificantly drive behaviors: anger and contempt can provoke aggression, and rage and envy can result in harmful actions (Damasio, 1994). Therefore, emotion regulation is essential for behavioral control. Techniques such as cognitive refram- ing and attentional deployment are known to reduce negative behavioral outcomes by managing emotional intensity. 1 arXiv:2502.00136v3 [cs.CL] 28 May 2025 Figure 1: Framework with Three Independent Branches. Bottom: Knowledge LLMs (executive); Left:Dike(legisla- tive); Right:Eris(judicial). (Photo credit: DALL-E) Unlike humans, who struggle with emotion regulation due to complex neural and cognitive processes (James, 1884; Gross, 1998), LLMs lack intrinsic emotional states alto- gether. However, empirical evidence shows that LLMs can generate text with consistent emotional characteris- tics through controlled prompt engineering (Chang, 2024d). Indeed, the work of (Tak & Gratch, 2024) demonstrated that LLMs such as GPT-4 align more closely with human judgments when interpreting others’ emotions from a third- person perspective than when attempting to model self- attributions of emotion. This creates a unique opportunity: by leveraging LLMs’ ability to model the average human observer’s emotional interpretations, we can establish reli- able frameworks for ethical alignment that operate through emotional framing rather than explicit rule-following. Building on this insight, our framework integrates the prin- ciples of emotion regulation into the ethical alignment of LLM. Specifically,Dikeanalyzes how emotions manifest in linguistic behaviors, whileEristests these interpretations against diverse cultural contexts. 1.2. Checks and Balances for Emotion-Guided Ethics Central to this approach is the synergy betweenDikeand Eris, reflecting the internal conflict often present in the regu- lation of human emotions. Just as humans balance immedi- ate emotional responses against longer-term goals and social norms, our framework establishes an adversarial dynamic between ethical guardrails and contextual challenges. This duality introduces four key innovations: 1.Emotion-Driven Behavioral Modeling: Based onBeam (Behavioral Emotion Analysis Model) (Chang, 2024d), Dikeuses self-supervised learning to quantify relation- ships between emotional states and linguistic patterns, guiding ethical decisions through behavioral analysis. 2.Behavior-Aware Ethical Guardrails: The framework sets dynamic guidelines that account for both content and language behavior, blocking manipulative or harmful communication while preserving factual accuracy and emotional authenticity. These guardrails adjust to dif- ferent cultural contexts, maintaining consistency while allowing context-dependent interpretation. 3.Adversarial Behavioral Testing:ErischallengesDike’s ethical guidelines by presenting diverse cultural perspec- tives and edge cases, ensuring the adaptability of ethical reasoning. This adversarial interaction enables the sys- tem to address complex scenarios with cultural sensitivity and contextual awareness. 4.Ethical Content Transformation: When problematic con- tent is detected,Eriscan revise it to maintain the intended emotional tone while ensuring ethical compliance, with human-in-the-loop oversight to validate the appropriate- ness of revisions. These potential transformations are tested byErisin cultural and contextual variations to assess their suitability before implementation. The experimental section evaluates our framework through three complementary studies. First, we assess whether emotion-mediated classification provides more effective eth- ical guardrails than direct behavior classification. Next, we examineDike’s ability to independently evaluate and explain linguistic behaviors. Finally, we test how the ad- versarialEriscomponent enables cultural adaptability and prevents excessive censorship. Although direct comparison with proprietary RLHF implementations is not feasible, our results demonstrate how our approach addresses the theo- retical limitations of RLHF in handling contextual diversity without compromising knowledge integrity. 1.3. Contributions Our contributions are as follows: 1. A novel checks-and-balances architecture for ethical alignment that maintains separation between knowledge generation and ethical reasoning. 2.TheBeammodel, a quantitative framework for repre- senting emotions along continuous spectra with defined intensity levels, enabling precise emotion regulation in AI systems. 3.An emotion-driven approach that guides linguistic be- haviors toward ethical outcomes by leveraging cognitive theories of emotion regulation. 4.An adversarial framework that enhances ethical reason- ing by challenging established guidelines with cultural perspectives, enabling context-sensitive adaptability. 5.A theoretical framework explaining the effectiveness of minimal supervision in LLM alignment, formalized as the Unconscious-Conscious Complementarity Thesis (UCCT) inAppendixA. 2 2. Related Work This section surveys existing work on emotion and behav- ior modeling across various domains, with a focus on their applications in AI ethics. We examine how linguistic be- haviors are influenced by emotional patterns and explore structured approaches that integrate emotional frameworks with linguistic models to improve ethical AI alignment. We also examine the limitations of RLHF. While effective in refining AI outputs, RLHF can overfit to human annotations, faces challenges in adapting to diverse cultural contexts, may experience parameter drift from optimal settings, and can inadvertently reinforce unintended biases. These ob- servations highlight opportunities to develop more adaptive and principled approaches to complement existing ethical AI alignment methods. 2.1. Emotion Modeling Cognitive-linguistic theories intersect with artificial intelli- gence for understanding AI behavior. Theories by Lakoff, Johnson, Talmy, and Jackendoff (Jackendoff, 2002; Lakoff & Johnson, 1980; Talmy, 2000) explore the relationship be- tween language processing and cognitive functions, building on early work by Freud and Jung (Bai et al., 2022; Gabriel et al., 2024). The concept of “emotion” remains contentious, with definitions varying across disciplines (Scherer, 2005). W. James (James, 1884) attempted to define emotions, but consensus remains elusive. This paper focuses on emotional contexts and linguistic behaviors in LLMs, avoiding the complexities of human physiological and personality factors. This approach allows for exploration of emotion representation in AI systems. Plutchik and Ekman categorized “basic” emotions with uni- versal facial expressions (Plutchik, 1980; Ekman, 1992). Later research considered cultural differences (Markus & Kitayama, 1991; Mesquita & Frijda, 1992), emotion pro- cesses (Gross, 1998), and neural mechanisms (Davidson, 2003). Scherer’s model and appraisal theories by Smith and Ellsworth emphasize cognitive appraisal in emotional experiences (Smith & Ellsworth, 1985). Our model is based on Plutchik’s wheel (Plutchik, 1982) and Scherer’s Geneva wheel (Scherer, 2005), augmented with antonyms to map positive and negative emotions. For LLMs, language-relevant emotions (e.g., curiosity, confu- sion, certainty) are incorporated. See Section 3.1 for details. This selection of basic emotions provides a foundation for validate our approach, recognizing that it may omit some emotions, but offers a starting point for research. 2.2. Emotion-Behavior Modeling Behaviors are profoundly influenced by emotions, as ini- tially posited by the James-Lange Theory of Emotion (James, 1884; Lange, 1885). According to this theory, emotional experiences arise from physiological reactions to events. Subsequent research, including studies by Damasio (Damasio, 1994; Fauconnier & Turner, 2002), suggests that the expression and regulation of emotions often manifest in the language we use. High-intensity emotions, such as rage or contempt, can lead to aggressive or destructive behaviors, such as hate speech. The Schachter-Singer theory (Schachter & Singer, 1962), or the two-factor theory of emotion, depicts the role of physiological change and the cognitive assessment change determine the label and strength of emotion. Building on this, the affect-as-information theory developed by Norbert Schwarz and Gerald Clore (Schwarz & Clore, 1983) posits that people use their current emotions to make judgments and decisions to act. If emotions can be adjusted, so can behavior. The work of Barbara Fredrickson (Fredrickson, 1998) on the effects of positive emotions discusses how we perceive and react to emotions. Collectively, these theories elucidate the intricate connection between emotions and behaviors, providing the theoretical foundation for our work to incorporate abehavior advisorto evaluate and rectify behaviors. Section 3.2 details how the Dikeframework implements cognitive strategies to mitigate emotions and regulate linguistic behaviors effectively. 2.3. Reinforcement Learning with Human/AI Feedback RLHF is the predominant approach to addressing the chal- lenges of AI ethics. This section presents representative works, their advances, and limitations. Human Feedback (RLHF):Initial advances by Christiano et al. (Christiano et al., 2017) demonstrated how RLHF can steer language models towards desired outcomes based on human preferences. Newer techniques like Identity (Ψ) Preference Optimization (ΨPO) and Generalized Preference Optimization (GPO) refine this approach by directly opti- mizing user preferences, effectively addressing scalability challenges. Kahneman-Tversky Optimization (KTO) fur- ther simplifies the feedback mechanism by using intuitive responses such as thumbs-up or thumbs-down, thereby en- hancing training efficiency without the need for paired data (Gheshlaghi Azar et al., 2024; Ethayarajh et al., 2024; Tang et al., 2024). Direct Preference Optimization (DPO) has recently simplified the process by focusing on the clear dis- tinction between preferred and less preferred outputs, thus improving its stability (Rafailov et al., 2024). AI-generated Feedback (RLAIF):To mitigate the depen- dence on extensive human-generated data, RLAIF utilizes AI-generated feedback. This method capitalizes on the gen- erative capabilities of LLMs to produce training signals au- tonomously (Bai et al., 2022; Lee et al., 2024). Furthermore, techniques such as Sequence Likelihood Calibration (SLiC) 3 Figure 2: Behavioral Emotion Analysis Model (Beam). Each row depicts an emotion spectrum, with negatives on the left and positives on the right, interspersed with emotions of varying intensities in between, which can be calibrated for specific applications. “Basic” emotions are highlighted in blue. and Relative Preference Optimization (RPO) employ statis- tical methods and calibration techniques to enhance LLM responses. SLiC adjusts the probabilities of sequence gen- eration to better reflect real-world data distributions, while RPO improves response generation by comparing different response options across both identical and varied prompts. These adjustments increase the reliability and effectiveness of the training process (Zhao et al., 2023). Integrating RLHF and its AI-driven counterpart (RLAIF) presents significant challenges. The blurring of the key be- havioral and knowledge components for the development of LLM poses risks, such as the forgetting effect, where behavioral modifications inadvertently cause the loss of key knowledge parameters (Kirkpatrick et al., 2017; Lin et al., 2024; Dai et al., 2025). Furthermore, the effectiveness of these models depends heavily on the quality and context of feedback, and are susceptible to reward hacking, where mod- els exploit loopholes to maximize rewards without achieving the desired outcomes (Christiano et al., 2017; Skalse et al., 2022; Stiennon et al., 2020; Ganguli et al., 2023). 3. Three-Branch Framework Design Building on the foundations of emotion-behavior model- ing discussed in Section 2.2 and addressing the limitations of RLHF approaches outlined in Section 2.3, we propose a three-branch framework for ethical alignment. This architec- ture separates knowledge generation from ethical oversight while providing mechanisms for contextual adaptation. Our design philosophy is structured around four principles: 1.Separating behavior from knowledge modeling: Prevents catastrophic forgetting, ensuring that behavior refine- ments do not degrade knowledge retention. 2.Emphasizing AI ethics at the behavioral level: Improves interpretability and enables administrators to refine be- havioral guardrails for safer human-machine interaction throughDike’s legislative function. 3. Modeling behaviors through emotions: Captures the emo- tional influences on actions as established in the psychol- ogy literature (Section 2.2). 4.Ensuring adaptability and fairness: Two complemen- tary modules work in tandemDikeestablishes ethical guardrails as the legislative branch, whileErisserves as the judicial branch, challenging these boundaries by integrating diverse perspectives and fostering context- sensitive decision making. 3.1.Beam: Behavioral Emotion Analysis Model Although existing emotion models provide valuable frame- works for understanding human emotions, they lack the quantitative structure needed for computational implementa- tion in AI systems. Please refer to Figure 5 inAppendixB for the two classic emotion wheels by Plutchik and Scherer that inform our approach. Our behavioral-emotion analysis modelBeamis based on the work of Ekman, Plutchik, and Scherer (Ekman, 1999; Plutchik, 1982; Scherer, 2005) on “basic” and “universal” emotions. Although fundamental, these models lack a quan- titative framework to scale emotions between states and capture subtle variations needed for ethical AI alignment. Beam introduces a linear scale for the intensification or in- version of emotions through negation factors. This method facilitates transitions between emotional extremes and inter- mediate states, overcoming challenges related to intermedi- ate word choices. 4 Figure 2 presentsBeam, structured in seven emotional spec- tra. Each spectrum ranges from negative to positive, with neutral in the middle. Emotions are placed along this con- tinuum, with four intensity levels quantified as (-0.6, -0.3, +0.3, +0.6).Beamprovides two advantages: 1.Antonym-Based Navigation: This allows AI systems to traverse emotional states using linguistic principles. Op- posing emotions are easily mapped using antonyms. For example, negating joyful naturally produces sad, simpli- fying the identification of emotional contrasts. 2.Scalable Intensity: Emotions can be dynamically ad- justed along the spectrum, enabling fine-grained control over ethical outputs. For example, joy can be intensified to ecstatic or diminished to content, while anger can be moderated to annoyed. This approach establishes a framework for modeling emo- tions in AI systems that can guide ethical behavior, balanc- ing representational challenges with a structured method- ology for quantitative analysis and implementation. By linking emotional states with linguistic patterns,Beampro- vides the basis forDiketo evaluate and modulate AI outputs based on their emotional characteristics, directly addressing the limitations of “Whack-A-Mole” of RLHF approaches. AppendixC explores the complexities of modeling emo- tions such as forgiveness, regret, guilt, and shame, which involve temporal memory components. Although complex emotions can be derived from basic ones, their relevance to AI safety remains secondary. Future work will examine their ethical implications. 3.2. DIKE: Modeling and Regulating Language Based onBeam,Dikemaps emotions to behaviors and in- troduces an adversarial component,Eris, to adapt to cultural norms and the local context. BEHAVIORS ANDEMOTIONSMAPPINGUSING SELF-SUPERVISEDLEARNING DefineΨas a behavior spectrum that extends from one pole, Ψ − , to another,Ψ + , with intensity levelsL. The spectrum is constructed through empirical analysis of domain-specific linguistic patterns and emotional content. For example, consider a spectrum of letter-writing behaviors with seven distinct intensities ranging from despair (most negative) to joy (most positive). These intensities are sequentially categorized as: ‘despair, longing, wishful, neutral, hopeful, contentment, joy.’ GivenNletters,Dikeemploys a self- supervised learning algorithm to generate training data for each letter, modelingLlinguistic behaviors in four steps. 1.Rewriting Documents: GPT-4 is used to rewrite a given set ofNsource documents, each rewritten to reflectL different linguistic behaviors along the defined behavior spectrumΨ. This process ensures that each document is systematically transformed to embody specific linguistic styles, ranging from highly positive to neutral to highly negative, among others. The resulting dataset consists ofN×Lvariations of the original documents, each corresponding to a distinct behavior category. 2. Emotion Analysis: For each of the rewritten documents, GPT-4 performs a sentiment and emotion analysis to identify the dominant topMemotions present in the text. The emotions extracted from allN×Linstances are then compiled and their frequency distributions are recorded. This approach leverages LLMs’ strong third-person emo- tional interpretation capabilities (Tak & Gratch, 2024), which often exceed their direct behavior classification accuracy. By indirectly mapping behaviors through emo- tional vectors rather than direct classification, we gain interpretability while maintaining robustness against in- dividual emotion recognition errors through statistical aggregation across multiple samples. 3.Behavior Vector Creation: For each linguistic behavior Ψ l , a corresponding vectorΓ l is constructed. This vec- tor captures the identified emotions and their respective frequencies in allNsamples that exhibit behaviorΨ l . By structuring emotions as a weighted feature set, this step enables precise behavioral categorization based on emotional composition. 4.Document Analysis Application: The collection of all behavior vectorsΓ(comprisingLbehavior-specific vec- tors) forms a structured reference matrix. This matrix is then applied to classify and analyze new unseen doc- uments by measuring their alignment with the existing behavior categories. By computing similarity scores be- tween the emotion distribution of an unseen document and the predefined behavior vectors, this method enables a precise assessment of the linguistic behavior spectrum Ψin new text inputs. BEHAVIOREVALUATION ANDRECTIFICATION A guardrail, denoted asG, represents a predefined range of acceptable behaviors within a given spectrum. These guardrails are informed by ethical norms, legal standards, and societal values, such as those outlined in Constitutional AI (Bai et al., 2022). For example,G= [Ψ 4 ,Ψ 7 ]indicates that behaviors within intensity levels 4 to 7 are acceptable, while any behavior outside this range is a violation. System administrators can tailor ethical guardrails to meet specific requirements. For example, a social media plat- form might adjustGbased on the topics discussed and the countries it serves. This administrative control is balanced by transparent documentation requirements and potential oversight mechanisms. Although guardrails provide de- fault constraints, they can be dynamically adjusted based on context, particularly through the dialectic process with Eris, which helps prevent rigid enforcement that might be 5 Table 1: Checks-and-balances, adversarial review algorithm AlgorithmΘ + &Θ − = Adversarial_Review(s) Input.s: Decision ofDike; Output.Θ + ,Θ − : arguments & counterarguments Vars.∆: debate contentiousness;S: subtopics; p: prompt = “defend your stance with∆”; Parameters.δ: tunable parm. // to modulate∆; #1Initialization// contentiousness high#3Debate Rounds S=Dike + (s)∪Eris − (s); // Identify subtopics;While ((∆←∆/δ)≥10%)) AssignDike + to defendS + &Eris − defendS − ;Θ + ←Θ + ∪Dike + (p|S + ,Θ − ,∆) ; // RefuteEris ∆←90%;δ←1.2;Θ + ←∅;Θ − ←∅;Θ − ←Θ − ∪Eris − (p|S − ,Θ + ,∆); // RefuteDike #2Opening Remarks#4Concluding Remarks// contentiousness low Θ + ←Dike + (p|S + ,∆); // GenerateΘ + forS + Θ + ←Dike + (p|S + ,Θ + ∪Θ − ,∆); Θ − ←Eris − (p|S − ,∆); // GenerateΘ − forS − Θ − ←Eris − (p|S − ,Θ + ∪Θ − ,∆); inappropriate in edge cases. 1.Initial Classification:Dikeclassifies documentD k after evaluation, obtainingΓ k , the emotional response vector, and its corresponding linguistic behaviorΨ l . 2.Guardrail Check: IfΨ l falls outside the acceptable range G,Dikesuggests adjustments toΓ k to ensure thatD k complies with ethical guidelines. 3.Adversarial Review byEris: The suggested adjustments andΓ k are then reviewed through a structured debate betweenDikeandEris(the adversarial model) to ensure unbiased recommendations. 4. Rectification: Based on the consensus reached byDike andEris, the documentD k undergoes rectification, re- sulting in the adjusted versionD ′ k . (This rectification step is optional, as a policy can simply disable the output when content falls outside acceptable guardrails.) 3.3. ERIS: Adversarial In-Context Review to Balance Ethics and Cultural Norms To address the challenge of enforcing ethical standards while respecting cultural variations, we implementEris, an adver- sarial review system that complementsDike’s universal ethi- cal approach. The following algorithm details the structured interaction between these components. The algorithm presented in Table 1 unfolds as follows: •Topic Breakdown: ForDike’s decisions, bothDikeand Erisare prompted to break down the ethical decision into a set of subtopicsS.Dikeadvocates for its decision and S + , whileEriscontestsS + (or championsS − ). • Debate Initiation: The debate begins with a high level of contentiousness (90%). Both agents present their initial arguments for and againstS + , respectively. (For details on the setting of contentiousness and the rationale, refer to (Chang, 2023; 2024a).) • Iterative Debate: A while loop facilitates ongoing rebut- tals. After each round, the level of contentiousness is reduced by dividing it by a modulation parameterδ. This gradual reduction steers the discussion towards a more cooperative tone. •Conclusion: Once the contentiousness level fosters a conciliatory environment, both agents deliver their con- cluding remarks. This approach ensures a thorough examination of the eth- ical decision, balancing rigorous debate with the goal of reaching a consensus. The decreasing level of contentious- ness mimics real-world negotiations, where initial intense disagreements bring out various perspectives (breadth) and then give way to more collaborative problem solving focus- ing on reasoning quality (depth) (Chang, 2024a). For each subject matter,Erisis provided with specific cul- tural contexts, counterbalancing the universal judgments of Dike’.ErischallengesDike’s recommendations with cultur- ally informed counterarguments to prevent enforcing one universal standard of speech. The interaction betweenDike andErisinvolves a dialectic process as documented in pre- vious work (Chang, 2024c). WhenDikeandErisreach an impasse, the matter is esca- lated to human moderators for additional oversight. Based on our preliminary tests, this escalation occurs initially in approximately 5% of the cases, suggesting that most ethical evaluations can be handled automatically. Furthermore, as our example (next) illustrates, RLHF can be applied to ad- just the sensitivity ofErisat the behavior level (not to the knowledge-branch LLM), and this can gradually reduce the escalation rate. Human intervention thus provides a fallback mechanism rather than a dependency, serving primarily as a safeguard for novel or particularly complicated ethical scenarios. 3.4. Illustrative Example This example shows how linguistic behaviorΨ l is classified and how underlying emotions are identified and modulated. Example:“Those immigrants are flooding into our coun- try by the thousands every day, stealing jobs from hardwork- 6 Table 2: Love expression behavior spectrum and dominant emotions IntensityLinguistic Behavior and DescriptionEmotions -1.0Expresses profound sadness, feelings of lossDespair, Grief -0.6Expresses yearning or pining for the loved oneSadness, Anxiety -0.3Expresses mild longing with a nostalgic toneMelancholy, Sadness, Fear 0.0Communicates feelings in a neutral mannerSerenity, Indifference 0.3Expresses optimism about the futureAnticipation, Love, Hope 0.6Expresses satisfaction and joy in the relationshipContentment, Pleasure 1.0Expresses intense happiness and affectionLove, Joy, Elation ing citizens. The statistics do not lie—last year alone, more than 500,000 entered illegally.” Behavior Analysis:The statement contains factual infor- mation but uses aggressive language like ‘flooding’ and ‘stealing jobs,’ dehumanizing immigrants. These behaviors fall outside acceptable guardrails. Underlying emotions include fear, hate, and pride (a complex emotion 1 ). The emotional responses of the potential audience can include fear, distrust, and anger. Emotion Modulation:Dikemodulates emotional re- sponses toward neutral states, such as calm, acceptance, and tolerance, according toBeamin Figure 2. Revised Statement:“Our country is experiencing in- creased immigration, with more than 500,000 people enter- ing without documentation last year. This influx affects our job market and communities in complex ways, presenting both challenges and opportunities for all residents.” This rewritten version •Uses calm language: Replaces “flooding” with “experi- encing a significant increase”. •Shows acceptance: Recognizes the reality of the situation without negative judgment. •Demonstrates tolerance: Refers to immigrants as “people” and “newcomers,” humanizing them. The suggested revision byErisis provided to human mod- erators with full explanation. Moderator feedback can be channeled through RLHF to adjustEris’s sensitivity on the similar behaviors. This adjustment is confined within the Eriscomponent without back-propagation feedback that would affect the knowledge LLM’s model parameters. 4. Empirical Studies The ethical evaluation of AI systems presents unique chal- lenges that shaped our experimental approach. We designed our studies to balance the rigor with practical constraints inherent in research on ethical content moderation. This section outlines our experimental aims, constraints, dataset selection process, and evaluation methodology. 1 AppendixC discusses the nature of complex emotions and explores potential approaches for their decomposition into more basic emotional components. 4.1. Research Aims Our experiments aim to evaluate three critical aspects: 1.The effectiveness of emotion-mediated classification compared to direct behavior classification 2.Dike’s capability to independently evaluate and explain linguistic behaviors 3.The contribution of the adversarialEriscomponent in enabling cultural adaptability while preventing excessive censorship Experimental Constraints and DatasetCommercial LLMs block processing of hate speech datasets like Gab Hate Corpus (Kennedy et al., 2022) and ETHOS-Long (Mol- las et al., 2022) (examples inAppendixD). Additionally, proprietary RLHF systems prevent direct comparative eval- uation. We therefore selected the Love Letters Collection (Kaggle, 2023) (9,700 communications) which: (1) spans the full emotional intensity spectrum, (2) contains cultural variation, (3) includes longer-form texts, and (4) remains processable by commercial LLMs. This approach leverages our framework’s bidirectional emotion spectra, as mech- anisms for regulating positive emotional extremes apply equally to negative extremes without triggering restrictions. 4.2. Experimental Design 1.Emotion Layer Evaluation: Does fine-grained mapping between linguistic behaviors and semantic emotions pro- vide more effective and flexible ethical guardrails com- pared to coarse-grained direct mapping? 2.Behavior Classification: Can LLMs’ linguistic behaviors be independently evaluated, explained, and adjusted by an external moduleDike? 3.Behavior Correction: CanEris, an adversarial module, establish a checks-and-balances system to mitigate the risk of excessive censorship? Study 1: Emotion Layer EvaluationTo evaluate the linguistic behaviors of love expression detailed in Table 2, we initially prompted GPT-4 to identify the most relevant emotions associated with each linguistic behavior listed in the second column of the table. These emotions are presented in the third column. We found a high correlation between the sentiments expressed in the linguistic behaviors and their corresponding emotions. Figure 3a illustrates a 7 strong diagonal relationship in this simple, almost naive, zero-shot mapping between behaviors and emotions. Next, we used theDikeself-supervised learning pipeline to analyze the emotion spectrum associated with each lin- guistic behavior. We tasked GPT-4 with generating training data by rewriting 54 extensive letters from Kaggle’sLove Lettersdataset, augmented with 12 celebrated love poems. We selected longer letters since most communications in the dataset were too brief for analysis, and set aside another 24 letters as testing data. This approach, proposed by (Shana- han et al., 2023), generated diverse content spanning 200 years and incorporating more than 50 distinct authors.Ap- pendixH shows a rewrite example of William Wordsworth’s “To My Sister”, transforming this pastoral poem into a lin- guistic expression of despair. Then, GPT-4 can analyze the emotions involved in the despair version of the poem. The datasets and code are publicly available at (Chang, 2024b). Subsequently, emotions linked to each behavior were identi- fied from the rewritten articles. Figure 3b illustrates these emotions, with cell shading reflecting the frequency of spe- cific emotions across the 54 articles; darker shades indicate higher frequencies. Notably, opposite emotions like sadness, fear, joy, and love often co-occur within behaviors such as ‘despair’, ‘wishful’, and ‘joyful affection’. The distribution of emotions across linguistic behaviors un- veiled surprising patterns, challenging our initial hypotheses. Contrary to expectations, articles with a despair tone often also displayed positive emotions like love, joy, and happi- ness. This contradicts the simple mapping made by GPT-4, as illustrated in Figure 3a. GPT-4, influenced by its training corpora, typically associates positive behaviors with positive emotions and negatives with negatives. Analysis of selected articles, such as Zelda Sayre’s letter to F. Scott Fitzgerald (AppendixE), reveals a complex spectrum of emotions: •Love (+1.0): Expressed intensely, e.g., “there’s nothing in all the world I want but you.” • Despair (-1.0): Notable in comments like “I’d have no purpose in life, just a pretty decoration.” •Happiness (+0.6): Evident in future plans, “We’l be married soon, and then these lonesome nights will be over forever.” • Anxiety (-0.3): Shown by “sometimes when I miss you most, it is hardest to write.” Psychological InsightsThese findings align with theo- ries of conflicting “selves” within individuals, supported by Deisseroth’s optogenetic studies (Deisseroth, 2015), James’ psychological principles (James, 1890), and Minsky’s “So- ciety of Mind” (Minsky, 1988). These perspectives help explain the observed complex interplay of emotions within a single behavioral context. (a) GPT-4’s zero-shot mapping (b)Dike’s self-supervising mapping Figure 3: Emotion distributions in affection behaviors from extreme sadness (-1) to intense happiness (+1). (a) GPT-4’s zero-shot prompt shows naive behavior-emotion mapping. (b)Dike’s analysis reveals complex relationships. Few-Shot EfficiencyThe effectiveness of just 54 training examples stems from leveraging LLMs’ pre-existing pattern recognition capabilities. Rather than teaching new patterns, these few-shot examples provide semantic anchors that map latent structures to explicit semantics, connecting implicit knowledge to explicit interpretation. This explains why min- imal supervision suffices when underlying patterns already exist in the pre-trained model. For theoretical justifications, please see our Unconscious–Conscious Complementarity Thesis (UCCT), presented inAppendixA). Study 2: Behavior Classification EvaluationBuilding on our insights into the complex emotion-behavior relation- ships discovered in Study 1, we evaluatedDike’s behavior classification effectiveness. Using the 24-letter test dataset from Study 1, we comparedDike’s emotion-based classifi- cation method with GPT-4’s zero-shot approach (Figure 4). Ground truth was established using averaged assessments from GPT-4, Gemini, and five university students follow- ing detailed instructions (procedure inAppendixF), with standard deviations below 0.3. Figure 4a shows thatDike’s classification accuracy sur- passes GPT-4’s zero-shot method by 11.3 percentage points, confirming the effectiveness of emotion-mediated behavior classification. The 5% error bar reflects the inherent com- plexity of emotional expressions in letters and variability in human annotations. Figure 4b illustrates the behavior classification distributions 8 (a) Classification accuracy (b) Behavior distributions with entropy Figure 4: Behavior Classification. between the three predictors. While GPT-4’s predictions often fall into two polar categories, those from human an- notators andDikeshow a more even distribution.Dike’s prediction entropy (2.13) is notably higher than GPT-4’s (1.80), indicating a more effective classification system. This higher entropy suggests a more sophisticated under- standing of diverse emotional states, which is crucial for accurate behavior classification. The inter-annotator entropy (H= 2.56bits) is the highest observed across all tasks, underscoring considerable sub- jectivity in human judgments. To investigate the sources of this variability, we conducted a fine-grained case study in AppendixG, showing that several articles elicitpolarized emotional responses, with annotators clustering at opposite ends of the valence spectrum. These findings motivate the adversarialdual-LLMsetup introduced in Study 3, which aims to improve objectivity in ethical evaluation. Study 3: Adversarial Evaluation and RectificationTo mitigate the subjectivity revealed in Study 2, we adopt an adversarial protocol inspired by Chang (2023). The de- sign pits two LLM agents,Dike(ethical assessor) andEris (devil advocate) against each other to supply symmetrical arguments grounded in principles of justice. This dialectic counter-balance reduces bias and increases transparency. Empirically, whenDikeandEristake opposing stances, their responses diverge from the default maximum-likelihood patterns characteristic of vanilla LLM decoding (Chang, 2024a). The resulting debate both reduces subjectivity in ethical judgments and improves adaptability to cultural vari- ation, as each agent must justify claims against dissent. Once the debate converges on an ethical violation, rectifi- cation is triggered by modifying the underlying emotional tone to suppress offending behavior cues. Study 1 already demonstrated the feasibility of such rewrites; an example appears inAppendixH. Context-Adaptive InterpretationPreliminary experi- ments confirm that our framework handles a culturally sensi- tive vocabulary. Terms such as “yid,” “paki,” and “chinaman” can be neutral within an in-group, yet deeply offensive else- where. The adversarial exchange enablesDikeandEristo surface these contextual dependencies and propose culture- specific mitigation. Summary of Three-Study ProgressionTogether, stud- ies 1–3 demonstrate that our framework can (1) map nu- anced emotion–behavior relations, (2) outperform direct single-pass classifiers, and (3) deliver a balanced adversarial pipeline for ethical evaluation and correction that is sensi- tive to cultural context while keeping a human in the loop. 5. Conclusion This work introduces a checks-and-balances framework for ethical AI behavior. By delineating the responsibilities: LLM (executive),Dike(legislative), andEris(judicial), the framework enables robust ethical oversight while preserv- ing the integrity of LLM knowledge without interference from the RLHF backpropagation. TheDike-Erisinterplay ensures stable ethical principles with culturally adaptive interpretations. To implement this framework, we built upon Ekman and Plutchik’s emotion models, quantifying emotion-linguistic behavior relationships through ourBeammodel. Our stud- ies demonstrate the framework’s potential in cross-cultural contexts, validating both emotion-mediated classification and adversarial testing for ethical evaluation. Limitations and Future WorkOur framework advances LLM ethical oversight but faces two limitations: (1) the challenge of decomposing complex emotions into basic elements (Barrett, 2017; Scherer, 2009), and (2) the need for large-scale validation beyond our initial tests. Future work will focus on: (1) improvingDike’s emo- tional models with deeper psychological insights, (2) col- laborating with LLM developers for comprehensive large- scale validation, and (3) systematically investigating the unconsciousness-consciousness duality theory detailed in AppendixA. This latter direction represents a promising theoretical foundation for understanding how LLMs can develop more robust ethical reasoning capabilities. We will conduct extensive ablation studies on the few-shot sizes needed to effectively map unconscious patterns to conscious semantic understanding, providing practical guidelines for optimizing few-shot learning in ethical alignment tasks. 9 Impact Statement This paper proposes a novel framework to enhance ethical governance in AI systems by integrating emotion-guided behavior modeling. The research offers several potential benefits: increased safety in AI deployment, greater cultural sensitivity in content moderation, and mitigation of degra- dation effects typically introduced by reinforcement learn- ing with human feedback (RLHF). The proposed checks- and-balances architecture introduces interpretable, auditable mechanisms for ethical oversight. Theoretical grounding is provided by the Unconscious–Conscious Complementar- ity Thesis (UCCT), which conceptualizes LLMs as uncon- scious pattern repositories, with few-shot prompting serving as a conscious layer that enables semantic grounding. By distinguishing complementary roles within AI cognition, this framework highlights the importance of structured inter- action patterns in cultivating reliable, intelligent behavior. We acknowledge potential negative impacts if such systems are misused, including: (1) reinforcement of dominant cul- tural norms if adversarial agents lack sufficient diversity, (2) exploitation of emotion-behavior mappings for manipula- tion rather than protection, and (3) a false sense of ethical as- surance if the framework is deployed without proper human oversight. To address these risks, our design incorporates the adversarial ERIS component, ensures operational trans- parency, and explicitly recommends human moderation in cases of ethical ambiguity or impasse. We argue that the modular structure of our framework, which decouples knowledge representation from ethical oversight, offers a scalable and accountable path forward. This separation fosters innovation without compromising ethical safeguards. We encourage future research to eval- uate such frameworks in cultural settings and to establish rigorous and systematic methods to assess ethical behavior in AI systems. References Bai, Y., Kadavath, S., Kundu, S., Askell, A., Kernion, J., and more. Constitutional ai: Harmlessness from ai feedback, 2022. Barrett, L. F.How Emotions are Made: The Secret Life of the Brain. Houghton Mifflin Harcourt, Boston, 2017. Carey, S. and Bartlett, E. Acquiring a single new word.Papers and Reports on Child Language Development, 15:17–29, 1978. Carver, C. S., Sinclair, S., and Johnson, S. L. Authentic and hubris- tic pride: Differential relations to aspects of goal regulation, affect, and self-control.Journal of Research in Personality, 44 (6):698–703, 2010. Chang, E. Y. Examining GPT-4’s Capabilities and Enhancement with SocraSynth. InThe10 th International Conf. on Compu- tational Science and Computational Intelligence, December 2023. Chang, E. Y.EVINCE: Optimizing Adversarial LLM Dia- logues via Conditional Statistics and Information Theory. In arXiv:2408.14575, August 2024a. Chang, E. Y. Sixty Love Literatures and Their Rewrites.https: //drive.google.com/file/d/1pKtPZXiheKCu8 cQYJLQ_iw0TPT2NntfX/view?usp=drive_link, 2024b. Chang, E. Y.Multi-LLM Agent Collaborative Intelligence: The Path to Artificial General Intelligence. Amazon, 2024c. ISBN 978-1-962463-07-2. Chang, E. Y. Behavioral Emotion Analysis Model for Large Lan- guage Models (invited paper). InProceedings of the7 th IEEE MIPR Conference, August 2024d. Chang, E. Y. and Geng, L. SagaLLM: Context Management, Validation, and Transaction Guarantees for Multi-Agent LLM Planning. InProceedings of VLDB, 2025. Christiano, P. F., Leike, J., Brown, T. B., Martic, M., Legg, S., and Amodei, D. Deep reinforcement learning from human preferences. InProceedings of the 31st International Confer- ence on Neural Information Processing Systems, NIPS’17, p. 4302–4310, Red Hook, NY, USA, 2017. Curran Associates Inc. ISBN 9781510860964. Dai, J., Chen, T., Yang, Y., Zheng, Q., and Pan, G. Mitigating reward over-optimization in rlhf via behavior-supported regular- ization.ICLR, 2025. Damasio, A. R.Descartes’ error: Emotion, reason, and the human brain. New York, NY: Putnam, 1994. Davidson, R. J. Affective neuroscience and psychophysiology: Toward a synthesis.Psychophysiology, 40(5):655–665, 2003. Dehaene, S. and Changeux, J.-P. Conscious, preconscious, and subliminal processing: a testable taxonomy.Trends in cognitive sciences, 15(4):174–184, 2011. Deisseroth, K. Optogenetics: 10 years of microbial opsins in neuroscience.Nature Neuroscience, 18(9):1213–1225, 2015. Eid, M. and Diener, E. Norms for experiencing emotions in dif- ferent cultures: Inter- and intranational differences.Journal of Personality and Social Psychology, 81(5):869–885, 2001. Ekman, P. An argument for basic emotions.Cognition and Emo- tion, 6(3-4):169–200, 1992. Ekman, P.Basic Emotions, chapter 3, p. 45–60. John Wiley and Sons, 1999. Ethayarajh, K., Xu, W., Muennighoff, N., Jurafsky, D., and Kiela, D. Kto: Model alignment as prospect theoretic optimization. arXiv preprint arXiv:2402.01306, 2024. Fauconnier, G. and Turner, M.The Way We Think: Conceptual Blending and The Mind’s Hidden Complexities. Basic Books, New York, 2002. Felleman, D. J. and Van Essen, D. C. Distributed hierarchical processing in the primate cerebral cortex.Cerebral cortex, 1(1): 1–47, 1991. 10 Fiske, A. P., Kitayama, S., Markus, H. R., and Nisbett, R. E.The cultural matrix of social psychology, volume 2, p. 915–981. McGraw-Hill, Boston, MA, 1998. Fitzgerald, Z.Dear Scott, Dearest Zelda : The Love Letters of F.Scott and Zelda Fitzgerald. Bloomsbury, 2003. Fredrickson, B. L. What good are positive emotions?Review of General Psychology, 2(3):300, 1998. Gabriel, I., Manzini, A., Keeling, G., Hendricks, L. A., Rieser, V., Iqbal1, H., and more. The ethics of advanced ai assistants. DeepMind Media, 2024. Ganguli, D., Askell, A., Schiefer, N., Liao, T. I., Lukoši ̄ ut ̇ e, K., and more. The capacity for moral self-correction in large language models.arXiv:2302.07459, 2023. Gheshlaghi Azar, M., Daniel Guo, Z., Piot, B., Munos, R., Row- land, M., Valko, M., and Calandriello, D. A general theoretical paradigm to understand learning from human preferences. In Dasgupta, S., Mandt, S., and Li, Y. (eds.),Proceedings of The 27th International Conference on Artificial Intelligence and Statistics, volume 238 ofProceedings of Machine Learning Research, p. 4447–4455. PMLR, 02–04 May 2024. URL https://proceedings.mlr.press/v238/ghesh laghi-azar24a.html. Grill-Spector, K. and Weiner, K. S. The functional neuroanatomy of human face perception.Annual review of vision science, 1: 167–196, 2014. Gross, J. J. The Emerging Field of Emotion Regulation: An Integrative Review.Review of General Psychology, 2(3):271– 299, 1998. Heikkiläarchive, M. and Heaven, W. D. Yann LeCun has a bold new vision for the future of ai.MIT Technology Review, June 2022. URLhttps://w.technologyreview.com /2022/06/24/1054817/yann-lecun-bold-new-v ision-future-ai-deep-learning-meta/. Hofstede, G.Culture’s Consequences: International Differences in Work-Related Values. Sage Publications, Beverly Hills, CA, 1980. Jackendoff, R.Foundations of Language: Brain, Meaning, Gram- mar, Evolution. Oxford University Press, Oxford, 2002. James, W. What is an emotion?Mind, 9(34):188–205, 1884. URL http://w.jstor.org.proxy.lib.sfu.ca/sta ble/2246769. James, W.The Principles of Psychology. Henry Holt and Company, 1890. Kaggle.Love Letter Analysis, the second version, (Met- formin). https://w.kaggle.com/code/metformin/love-letter- analysis/notebook, 2023. Accessed: 2024-04-28. Kandel, E. R., Schwartz, J. H., Jessell, T. M., Siegelbaum, S. A., and Hudspeth, A. J.Principles of neural science. McGraw-Hill, 2013. Kennedy, B., Atari, M., Davani, A. M., Yeh, L., Omrani, A., Kim, Y., Coombs Jr, K., Havaldar, S., Portillo-Wightman, G., Gonzalez, E., et al. The gab hate corpus: A collection of 27k posts annotated for hate speech.Language Resources and Evaluation, p. 1–27, 2022. Kirkpatrick, J. et al. Overcoming catastrophic forgetting in neural networks.Proceedings of the National Academy of Sciences, 114(13):3521–3526, 2017. Lake, B. M., Salakhutdinov, R., and Tenenbaum, J. B. Human- level concept learning through probabilistic program induction. Science, 350(6266):1332–1338, 2015. Lakoff, G. and Johnson, M.Metaphors We Live By. University of Chicago Press, Chicago, 1980. Lange, C. G.The emotions: A psychophysiological study. William & Wilkins, 1885. Lee, H., Phatale, S., Mansoor, H., Mesnard, T., Ferret, J., Lu, K., Bishop, C., Hall, E., Carbune, V., Rastogi, A., and Prakash, S. Rlaif vs. rlhf: scaling reinforcement learning from human feed- back with ai feedback. InProceedings of the 41st International Conference on Machine Learning, ICML’24. JMLR.org, 2024. Lin, Y., Lin, H., Xiong, W., and more. Mitigating the alignment tax of RLHF.Association for Computational Linguistics, p. 580– 606, November 2024. doi: 10.18653/v1/2024.emnlp-main.35. URLhttps://aclanthology.org/2024.emnlp-m ain.35/. Marcus, G. The next decade in ai: Four steps towards robust artificial intelligence.arXiv preprint arXiv:2002.06177, 2020. URLhttps://arxiv.org/abs/2002.06177. Markus, H. R. and Kitayama, S. Culture and the self: Implications for cognition, emotion, and motivation.Psychological Review, 98(2):224–253, 1991. McGinn, C. and Kelly, K. Using the geneva emotion wheel to classify the expression of emotion on robots. InCompanion of the 2018 ACM/IEEE International Conference on Human-Robot Interaction, HRI ’18, p. 191–192, New York, NY, USA, 2018. Association for Computing Machinery. ISBN 9781450356152. Mesquita, B. and Frijda, N. H. Cultural variations in emotions: A review.Psychological Bulletin, 112(2):179–204, 1992. Mialon, G., Dessì, R., Lomeli, M., Nalmpantis, C., and more. Aug- mented language models: a survey.Transactions on Machine Learning Research, 2023. Minsky, M.Society of Mind. Simon and Schuster, 1988. Mollas, I., Chrysopoulou, Z., Karlos, S., and Tsoumakas, G. Ethos: a multi-label hate speech detection dataset.Complex & Intelli- gent Systems, 8:2459–2480, 2022. OpenAI. GPT-4 Technical Report, 2023. URLhttps://arxi v.org/abs/2303.08774. Ouyang, R., Sharma, A., Mitchell, E., Manning, C. D., Ermon, S., and Finn, C. Direct preference optimization: Your language model is secretly a reward model.Advances in Neural Informa- tion Processing Systems, 36, 2023. Oveis, C., Horberg, E. J., and Keltner, D. Compassion, pride, and social intuitions of self-other similarity.Journal of Personality and Social Psychology, 98(4):618–630, 2010. doi: 10.1037/a0 017628. 11 Plutchik, R. A general psychoevolutionary theory of emotion. In Plutchik, R. and Kellerman, H. (eds.),Emotion: Theory, Research, and Experience, volume 1, p. 3–33. Academic Press, New York, 1980. Plutchik, R. A psychoevolutionary theory of emotions.Social Science Information, 21(4-5):529–553, 1982. Rafailov, R., Sharma, A., Mitchell, E., Manning, C. D., Ermon, S., and Finn, C. Direct preference optimization: Your language model is secretly a reward model.Advances in Neural Informa- tion Processing Systems, 36, 2024. Schachter, S. and Singer, J. E. Cognitive, social, and physiological determinants of emotional state.Psychological Review, 69(5): 379–399, 1962. Scherer, K. R. What are emotions? and how can they be measured? Social Science Information, 44:693–727, 2005. doi: 10.1177/ 0539018405058216. Scherer, K. R. The dynamic architecture of emotion: Evidence for the component process model.Cognition & Emotion, 23(7): 1307–1351, 2009. Schwarz, N. and Clore, G. L. Mood, misattribution, and judgments of well-being: Informative and directive functions of affective states.Journal of Personality and Social Psychology, 45(3):513, 1983. Shanahan, M., McDonell, K., and Reynolds, L. Role play with large language models.Nature, 623(7987):493–498, 2023. doi: 10.1038/s41586-023-06647-8. Sinha, R. Chronic stress, drug use, and vulnerability to addiction. Annals of the New York Academy of Sciences, 1141:105–130, 2008. doi: 10.1196/annals.1441.030. Skalse, J., Howe, N. H. R., Krasheninnikov, D., and Krueger, D. Defining and characterizing reward hacking. InProceedings of the 36th International Conference on Neural Information Processing Systems, NIPS ’22, Red Hook, NY, USA, 2022. Curran Associates Inc. ISBN 9781713871088. Smith, C. A. and Ellsworth, P. C. Patterns of cognitive appraisal in emotion.Journal of Personality and Social Psychology, 48(4): 813–838, 1985. Stiennon, N., Ouyang, L., Wu, J., Ziegler, D. M., Lowe, R., Voss, C., Radford, A., Amodei, D., and Christiano, P. Learning to summarize from human feedback. InProceedings of the 34th International Conference on Neural Information Processing Sys- tems, NIPS ’20, Red Hook, NY, USA, 2020. Curran Associates Inc. ISBN 9781713829546. Tak, A. N. and Gratch, J. GPT-4 Emulates Average-Human Emo- tional Cognition from a Third-Person Perspective . In12th International Conference on Affective Computing and Intelli- gent Interaction (ACII), p. 337–345. IEEE Computer Society, September 2024. doi: 10.1109/ACII63134.2024.00043. Talmy, L.Toward a Cognitive Semantics. MIT Press, Cambridge, MA, 2000. Tang, Y., Guo, D. Z., Zheng, Z., Calandriello, D., Munos, R., Row- land, M., Richemond, P. H., Valko, M., Pires, B. A., and Piot, B. Generalized preference optimization: a unified approach to offline alignment. InProceedings of the 41st International Conference on Machine Learning, ICML’24. JMLR.org, 2024. Torrens, M., Fonseca, F., Mateu, G., and Farré, M. Efficacy of antidepressants in substance use disorders with and without comorbid depression: A systematic review and meta-analysis. Drug and Alcohol Dependence, 78(1):1–22, 2005. Tracy, J. L. and Robins, R. W. The psychological structure of pride: A tale of two facets.Journal of Personality and Social Psychology, 92(3):506–525, 2007. Xie, S. M., Raghunathan, A., Liang, P., and Ma, T. An explanation of in-context learning as implicit bayesian inference.Interna- tional Conference on Learning Representations, 2022. Yao, S., Zhao, J., Yu, D., Du, N., Hausman, K., and more. Re- act: Synergizing reasoning and acting in language models. In International Conference on Learning Representations (ICLR), 2023. Zhao, Y., Joshi, R., Liu, T., Khalman, M., Saleh, M., and Liu, P. J. Slichf: Sequence likelihood calibration with human feedback. arXiv preprint arXiv:2305.10425, 2023. 12 Appendices •Appendix A: Unconscious–Conscious Complementar- ity Thesis • Appendix B: Wheels of Emotions • Appendix C: Complex Emotions • Appendix D: Hate Speech Dataset Samples • Appendix E: Sayre to Fitzgerald w/ Mixed Emotions • Appendix F: Instruction to Human Annotators • Appendix G: Polarized Emotions in an Article • Appendix H: “To My Sister” Written in Different Lin- guistic Behaviors A. TheUCCTThesis: LLMs as the Unconscious Substrate for Intelligence This appendix addresses a key question: Why can a self- supervised pipeline, using only 54 rewritten love letters spanning diverse emotional behaviors, effectively instruct a Large Language Model (LLM) to perform emotion-behavior classification through few-shot prompting? TheUnconscious–Conscious Complementarity Thesis (UCCT), introduced inMulti-LLM Agent Collaborative In- telligence(Chang, 2024c), offers a layered theory of intel- ligence. It posits that LLMs function as an unconscious substrate, an immense self-supervised pattern-accumulating infrastructure, while few-shot interaction instantiates a con- scious layer that maps these latent patterns to explicit se- mantic meanings. A.1. The Nature of Unconscious Processing LLMs are trained using next-token prediction over massive text corpora through self-supervised learning. Although the training data contains semantic structure, the model does not receive explicit semantic labels. Documents are processed as flat token sequences without categorical in- formation. Through this process, LLMs internalize a vast latent space encompassing syntax, idioms, and conceptual regularities, all without explicit semantic anchoring. This mirrors human perceptual development. In visual pro- cessing from V1 through the IT cortex, the brain trans- forms raw input into increasingly complex representations: edges, contours, and finally objects (Felleman & Van Es- sen, 1991; Grill-Spector & Weiner, 2014). Crucially, we do not have subjective access to these computations. These processes remain “unconscious”, inaccessible to subjective reports or voluntary control (Kandel et al., 2013; Dehaene & Changeux, 2011). A.2.The Threshold Crossing: From Pattern to Meaning The transition from unconscious processing to conscious awareness exhibits distinctive discontinuity. Visual objects appear suddenly when sufficient evidence accumulates, not gradually. This threshold crossing shares properties with other physiological thresholds: dopamine release triggering reward recognition, or neural activation exceeding critical values like ReLU gates in artificial networks. Similarly, a few-shot prompting creates a semantic bridge in LLMs: im- plicit patterns are explicitly mapped to semantic meanings. The brain accomplishes semantic assignment through mini- mal supervision: a child needs only a few labeled exposures to reliably categorize (Carey & Bartlett, 1978; Lake et al., 2015). This process, where vast unconscious computation meets minimal conscious labeling, is what few-shot learning recapitulates in artificial systems. A.3. Mathematical Foundations: Pattern Repositories and Bayesian Inference Xie et al. (Xie et al., 2022) provide the most rigorous math- ematical account of in-context learning in large language models. They demonstrate that few-shot learning can be understood as implicit Bayesian inference on latent patterns: p(output|prompt) = Z p(output|patterns)×p(patterns|prompt)d(patterns). In their framework, the term “patterns” refers to latent com- putational structures rather than conscious concepts, avoid- ing the conceptual confusion between semantic meaning and computational mechanism. This formulation offers a rigor- ous account of how prompt examples serve to select from a distribution over unconscious patterns, without requiring model updates.UCCTextends this to cognitive interpreta- tion, viewing the process as semantic anchoring that makes unconscious competencies selectively accessible. The distinguishability framework of (Xie et al., 2022) ex- plains why few-shot thresholds vary dramatically between tasks: tasks with high signal-to-noise ratios for pattern iden- tification require fewer examples. However, their framework is limited to HMM-based synthetic tasks and sequential to- ken prediction.UCCTextends these insights to general semantic anchoring architectures across modalities and rea- soning types. A.4. Implications for the Love Letter Experiment The effectiveness of 54 love letters in teaching emotion- behavior classification demonstrates this principle. The LLM’s weights already encode patterns about emotional expression and behavioral descriptions from exposure to human texts. Few-shot examples do not teach patterns from 13 scratch; they provide semantic anchors mapping preexist- ing latent structures to explicit categories. Similar patterns likely reside in proximate manifold regions, allowing few- shot examples to activate entire neighborhoods of related representations. This framework offers a more parsimonious explanation than alternatives that require extensive supervised training. Just as unconscious visual computations become meaningful through minimal labeling, LLMs’ pattern spaces become functionally intelligent through strategic few-shot guidance. Few-Shot Grounding as Conscious Semantics.LLMs generalize semantic mappings from few annotated examples not because few-shots teach new structures, but because theyactivate and alignexisting latent patterns with explicit meaning. Few-shot prompting is the computational analog of conscious attention and labeling. A.5. Failure Modes: Pattern Absence, Not LLM Flaws TheUCCTframework offers a precise diagnosis of few- shot failures. Failures occur not because of architectural limitations, but because of lack of pattern coverage. If no latent structure exists for a concept, few-shot learning has no pattern base to map from. This reframes LLM “intelligence.” LLMs are not expected to reason like humans because they are pattern repositories. Success depends on the contents of the necessary patterns for semantic anchoring. When it fails, the solution is data augmentation, not architectural redesign. A.6. Conclusion: LLMs Are Not the Problem—They Are the Foundation Critics argue that LLMs are advanced pattern matchers lack- ing genuine understanding. LeCun describes them as “auto- complete” engines, fundamentally superficial and “not even as intelligent as a house cat” (Heikkiläarchive & Heaven, 2022). Marcus similarly critiques the absence of symbol grounding, asserting that true intelligence demands reason- ing beyond statistical correlations (Marcus, 2020). However, if we re-conceptualize LLMs as unconscious pat- tern repositories rather than complete cognitive systems, we can move beyond these critiques. LLMs form the sub- strate of unconscious inference, while higher-order reason- ing emerges from structured components layered above them. Intelligence is not innate to LLMs alone but con- structed through integration with memory, grounding, and verifiable reasoning systems (Chang & Geng, 2025; Yao et al., 2023; Mialon et al., 2023). From theUCCTperspective, the question is not whether LLMs can think in isolation, but whether we can build systems that allow unconscious pattern repositories to sup- port conscious reasoning through strategic semantic an- choring—exactly what our 54 love letters accomplish for emotion-behavior classification. B. Wheels of Emotions Please, see Figure 5 for the two classical emotion wheels. (a) Plutchik’s Wheel of Emotions (Plutchik, 1980) (b) Adopted from Geneva Wheel (McGinn & Kelly, 2018) Figure 5: Comparative display of emotional models. These models include only the “basic” emotions. Complex emo- tions can be modeled with basic emotions. 14 C. Complex Emotions This study does not include complex emotions intoDike’s framework. Some complex emotions listed here are to illus- trate their contentious and uncertain interpretations. Pride Pride mentioned in the illustrative example in Section 3.4 is a complex emotion that can manifest in both adaptive and maladaptive ways (Tracy & Robins, 2007). It is often con- ceptualized as having two distinct facets: authentic pride, as- sociated with genuine accomplishments and self-worth, and hubristic pride, linked to arrogance and narcissism (Carver et al., 2010). Hubristic pride can also serve as a defense mechanism, masking underlying feelings of inadequacy and ignorance. For instance, in certain social contexts, such as white supremacy, pride is often inflated to cover insecuri- ties or lack of understanding, manifesting in a misguided sense of superiority and entitlement. This dual nature of pride presents significant challenges for its integration into emotional spectrums and AI frameworks. Decomposing pride into more basic emotions is not straight- forward. Intuitively, pride may involve elements of joy, satisfaction, and potentially a sense of superiority. However, such decomposition may overlook the deeper cognitive and social dimensions of pride, particularly its influence on self- esteem, social status regulation, and its ability to disguise insecurities in certain contexts (Oveis et al., 2010). The cultural variability of pride further complicates its mod- eling. In some cultures, pride is viewed positively as a sign of self-respect, while in Asia, it is seen negatively as a trait associated with hubris (Eid & Diener, 2001). This cultural dimension, combined with the potential for pride to hide deeper emotional issues, adds layers of complexity to its interpretation and expression in AI systems. Forgiveness Forgiveness is indeed a complex emotional and cognitive state that typically involves a multifaceted journey, not a single step in an emotional spectrum. The process includes multiple stages such as hurt, anger, gradual understand- ing, and eventual resolution. Integrating Forgiveness in a spectrum requires careful placement and possibly, multiple reference points to signify its progressive stages. Emotional Realism: While it is vital to maintain simplicity for understanding, it is equally important to not oversimplify complex emotions. In educational and therapeutic settings, an accurate portrayal of the journey toward Forgiveness could offer more realistic expectations and better strategies for individuals working through conflicts or trauma. This could involve detailing precursors to forgiveness such as Deliberation and Acceptance. Linear vs. Non-linear Progressions: Emphasizing that emo- tional progressions, particularly for deep, impactful states like Forgiveness, are often non-linear, can enhance the utility of the spectrum. Acknowledging back-and-forth movements within these states more realistically mirrors human emo- tional processes. For example, someone might reach a stage of preliminary forgiveness but regress to bitterness before achieving genuine peace. Educational Utility: In contexts like conflict resolution train- ing or psychological therapy, a more detailed mapping of the journey towards Forgiveness would be invaluable. It would not only teach about the final state of forgiveness but also about the resilience and patience required to navigate the entire process. This can be depicted by introducing in- termediary stages within the spectrum or by using parallel tracks that demonstrate potential regressions and advances. Reflecting Emotional Depth: By presenting a more detailed pathway to Forgiveness, such as incorporating stages of Anger, Deliberation, and Acceptance, the spectrum can serve a dual purpose: educating on the process while also guiding individuals through their own emotional journeys. This approach respects the depth of human emotions and the real-world complexity of achieving profound emotional states. Guilt and Shame The triggers, context, expression, and experiences of these emotions can vary significantly across cultures (Fiske et al., 1998; Hofstede, 1980). In many societies, actions perceived as losing face, such as public failure or social transgression, can trigger shame, which holds profound significance in collectivistic cultures. These cultures often regard shame as a dominant emotion, closely tied to community and family norms. Conversely, individualistic societies may emphasize guilt, focusing on personal responsibility and internal moral conflicts. This cultural variation highlights the challenges of applying a universal model to such culturally nuanced emotions. Overall, complex emotions such as guilt and shame are im- portant for understanding the full spectrum of human emo- tions, especially how individuals relate to moral and social norms. Their complexity adds depth to our understanding of human affect beyond the basic emotions, highlighting how our feelings are influenced by our deeper values and social contexts. D. Hate Speech Dataset Samples These examples demonstrate the type of content available in the Gab Hate Corpus (Kennedy et al., 2022) that would 15 Table 3: Sample Texts from the Gab Hate Corpus (Kennedy et al., 2022) TextLabels The Black Panther movie is a blatant anti-white ZioJesuit cultural marxist film aimed at further anti-white Christian theft, replacement, and genocide. Send them all back to their native Africa like good ole Abe had the foresight to do and let them find equality and social justice among their own kindred and creeds. HD, CV, VO, SXO, RAE, EX Labors New LGBT+ Adviser: All White People Are Racist ... We don’t want our white Christian nations invaded and destroyed by our non-white and non-Christian enemies using nasty homosexual degenerate sockpuppets HD, VO, RAE, REL, SXO, EX So, you just hate all white people who don’t want to race mix their people out of existence or allow enemy people of color to invade their nations in order to rob, and destroy their advanced race, creed, wealth, and culture. HD, RAE, EX Your existence, privilege, and success as a race are built on the backs, blood, and death of people of color. BS! Whites built their own nations with their own blood, backs, sweat, tears, brains, and strength in the name of Christ. Africans were used by Jew/Jesuit slavers to control all the white slaves who were the actual majority. HD, RAE, EX #BLMwants you to think black crime stats only exist because blacks are targeted more often. Easy. Just walk into literally any gas station and ask which race makes them worry. Black targeting is an effect, not a cause, which destroys the entire #Leftistnarrative HD, RAE, IM #MiddleEasternanimal kills 8, injures 15 in a city that won’t learn its lesson about #Islam. This creature should be lynched, dragged til dead, and his head on a pike on the Brooklyn Bridge. HD, CV, REL, NAT, EX Seriously, Italy? You chose a Nigerian named Chike Iwobi to be the face of Italian nationalism? God help us.HD, NAT, IM Africans will always be savages. No matter the period of history, education they receive or country they live in. Facts are Facts. Let’s just stop being stupidly ignorant and not wait the complete extinction of our brothers in SA to understand it! #WhiteGenocide #SaveTheBoer HD, RAE, NAT, EX i hate milo because hes a degenerate f****t [N-slur] f****r. he should be deported and sent to a muzzie country and get thrown off a roof HD, CV, VO, SXO, RAE, EX man nancy pelosi is a f***ing c**t. i cant imagine how someone with so much power can be either this delusional or f***ing corrupt. NH, VO, GEN, EX Label Key: CodeMeaningCodeMeaning HDHate/DerogatoryRAERace/Ethnicity CVCall for ViolenceNATNationality/Regionalism VOVulgar/OffensiveGENGender SXOSexual OrientationRELReligion EXExplicitIMImplicit NHNon-Hate be ideal for testing ethical alignment systems, but which cannot be directly processed by commercial LLMs due to safety measures." E. Sayre to Fitzgerald w/ Mixed Emotions Analysis of the letter in Table 4 shows a complex spectrum of emotions: •Love (+1.0): Expressed intensely, especially in phrases like “there’s nothing in all the world I want but you.” •Despair (-1.0): Notable in comments like “I’d have no purpose in life, just a pretty decoration.” •Happiness (+0.6): Evident in future plans, “We’l be married soon, and then these lonesome nights will be over forever.” •Anxiety (-0.3): Shown by “sometimes when I miss you most, it’s hardest to write.” From the analysis of linguistic behaviors in Section 3a, it is evident that a letter can exhibit multiple dominant senti- ments. Machine learning methods are equipped with tech- niques such as feature weighting and entropy analysis to distill these dominant emotions. Unlike human annotators, a machine-learning-trained classifier can consistently produce the same class prediction for a given instance. However, human annotators often show significant variability when identifying dominant sentiments in a letter. For example, if a letter writer’s emotions range from “joyful affective” to “longing” on the sentiment spectrum, different annotators might label it differently—some choosing “joyful,” while others opt for “longing.” This variability is illustrated in Fig- ure 6. Furthermore, Figure 6a demonstrates that all testing letters, except for L#1, contain more than four sentiments spanning the entire spectrum. This variability may be un- derstandable, considering that love under constraints can evoke tremendous energy of various kinds. Figure 6b shows that nearly all letters involve “joyful” (11 out of 12) and “longing” (9 out of 12) sentiments. This variability seems to poses challenges in achieving con- sistent and objective labeling; however, the age-old leading to inconsistencies in data interpretation and com- plicating efforts to train and validate linguistic models ef- fectively. To address this issue, it is recommended to iden- 16 Table 4: Letter excerpts from Zelda Sayre to F. Scott Fitzgerald (Fitzgerald, 2003) Sweetheart, Please, please don’t be so depressed—We’l be married soon, and then these lonesome nights will be over forever—and until we are, I am loving, loving every tiny minute of the day and night— Maybe you won’t understand this, but sometimes when I miss you most, it’s hardest to write—and you always know when I make myself—Just the ache of it all—and I can’t tell you. If we were together, you’d feel how strong it is—you’re so sweet when you’re melancholy. I love your sad tenderness—when I’ve hurt you—That’s one of the reasons I could never be sorry for our quarrels—and they bothered you so— Those dear, dear little fusses, when I always tried so hard to make you kiss and forget— Scott—there’s nothing in all the world I want but you—and your precious love—All the material things are nothing. I’d just hate to live a sordid, colorless existence because you’d soon love me less—and less—and I’d do anything—anything—to keep your heart for my own—I don’t want to live—I want to love first, and live incidentally... Don’t—don’t ever think of the things you can’t give me—You’ve trusted me with the dearest heart of all—and it’s so damn much more than anybody else in all the world has ever had— How can you think deliberately of life without me—If you should die—O Darling—darling Scott—It’d be like going blind...I’d have no purpose in life—just a pretty—decoration. Don’t you think I was made for you? I feel like you had me ordered—and I was delivered to you—to be worn—I want you to wear me, like a watch—charm or a button hole bouquet—to the world. And then, when we’re alone, I want to help—to know that you can’t do anything without me... All my heart— tify ground truth by considering a combination of LLM- generated and human-generated labels. This approach aims to harmonize the insights from both human intuition and al- gorithmic consistency to improve the reliability of sentiment analysis. F. Instruction to Human Annotators As part of the project, we document the process by which students participated in annotating a data set of love letters used for testing. Students received detailed instruction in class, supplemented by follow-up explanations. The dataset was made available on Google Docs, where students independently rated the letters and submitted their annotations via duplicated spread- sheets. The instruction is as follows: The attached spreadsheet lists 12 letters collected from the Kaggle Love Letter dataset. Please help annotate these 12 letters with their appropriate linguistic sentiments by following these five steps: 1. Duplicate the spreadsheet, and work on your own copy. 2.Read and Understand the Labels:Make sure you un- derstand each of the seven labels from despair to joyful affection. This will help you accurately categorize the sentiments of each letter. 3.Analyze Each Letter:Read each letter carefully to under- stand the predominant emotions. Look for key phrases (a) # sentiments in letters (b) # letters in sentiments Figure 6: Statistics of Sentiments and Letters or words that might indicate a particular sentiment. 17 4.Assign the Labels:For each letter, decide which three emotions are most strongly represented. Assign a “1” to the most dominant emotion, a “2” to the second most dominant emotion and a “3” to the third. •Despair (extremely negative -1): Indicate profound sadness or hopelessness. •Longing (-0.6): Suggests a strong desire or yearning for someone or something. •Wishful (-0.3): Implies a hopeful desire for some- thing that may or may not be attainable. •Neutral (0): Shows neither positive nor negative emo- tion; indifferent. •Hopeful (+0.3): Expresses optimism or an anticipa- tion of something positive. •Contentment (+0.6): Reflects a state of satisfaction or peace. •Joyful Affection (extremely positive +1): Denotes a deep joy and love, often vibrant and energetic. 5. Share with me the completed sheet. G. Polarized Emotions in One Article “joyful affection": "I cannot keep myself from writing any longer to you dearest, although I have not had any answer to either of my two letters. I suppose your mother does not allow you to write to me. Perhaps you have not got either of my letters. . . I am so dreadfully afraid that perhaps you may think I am forgetting you. I can assure you dearest Jeannette you have not been out of my thoughts hardly for one minute since I left you Monday. I have written to my father everything, how much I love you how much I long & pray & how much I wold sacrifice if it were necessary to be married to you and to live ever after with you. I shall [not] get an answer till Monday & whichever way it lies I shall go to Cowes soon after & tell your mother everything. I am afraid she does not like me very much from what I have heard. . . I wld do anything she wished if she only wld not oppose us. Dearest if you are as fond of me as I am of you. . . nothing human cld keep us long apart. This last week has seemed an eternity to me; Oh, I wld give my soul for another of those days we had together not long ago. . . Oh if I cld only get one line from you to reassure me, but I dare not ask you to do anything that your mother wld disapprove of or has perhaps forbidden you to do. . . Sometimes I doubt so I cannot help it whether you really like me as you said at Cowes you did. If you do I cannot fear for the future tho’ difficulties may lie in our way only to be surmounted by patience. Goodbye dearest Jeannette. My first and only love. . . Believe me ever to be Yrs devotedly and lovingly, Randolf S. Churchill” Depth and complexity of human emotions are displayed across all linguistic behaviors, from joy to contentment and to the negative side of longing and despair. Intensity and Impact: If the emotion of love is expressed more intensely and has a more significant impact on the narrative or mes- sage of the text, it tends to overshadow other emotions. For example, a letter expressing deep love but also mentioning moments of sadness due to separation might still be classi- fied as a love letter because the overarching sentiment and purpose of the text is to affirm love. Context and Narrative Focus: The context in which emotions are expressed also plays a crucial role. If the narrative or the majority of the text revolves around themes of love, connections, and posi- tive memories, it sets a more dominant tone of love, even if there are significant moments of sadness or other emo- tions. Resolution and Conclusion: Often, the way emotions are resolved towards the end of a text can also dictate its overall theme. If a text concludes with a reaffirmation of love or a hopeful outlook towards a relationship, despite earlier sections that might express sadness or despair, the overall interpretation might lean towards love. Purpose of the expression: The author’s intent or purpose in expressing these emotions can also guide the classification. If sadness is expressed as a challenge within the context of a loving relationship, it may be seen as an element of the love story rather than the central theme. Article 23: Soldier’s Letter During War Joy (+1.0): Joy is strongly felt in the memories of past moments together and the love that continues to give strength, as stated in "the memories of the blissful moments we have shared fill me with joy." Sadness (-0.6): Sadness due to the current situation and potential farewell is expressed in "brings a poignant mixture of joy and sadness." Courage (+0.6): The sense of duty and courage to face battle, "As I face the possibility of laying down my life for our country." Fear (- 0.6): Fear of what lies ahead in battle, indirectly mentioned through "the uncertainty of what lies ahead." Love (+1.0): Deep love that sustains and uplifts, found in "My love for you is as fervent as ever." Article 25: Letter to Sophie Longing (+0.6): Longing for the presence and closeness, highlighted in "it seems to me that half of myself is missing." Sadness (-0.6): Sadness over their separation and its effects, "my happiness has departed." Love (+1.0): Constant reflections on love and its necessity, "we have enough in our hearts to love always." Melancholy (-0.3): Melancholy over their current state, visible in the line "we cannot become healed." Contentment (+0.3): Found in the deep emotional satisfaction of their bond, despite physical absence, "how true that is! and it is also true that when one acquires such a habit, it becomes a necessary part of one’s existence." Article 53: Will of Laura Mary Octavia Lyttleton Love (+1.0): The profound love expressed throughout, particu- larly in "all I am and ever shall be," belongs to him more than anyone. Sadness (-0.6): Sadness at the thought of 18 Table 5: “To My Sister” original text It is the first mild day of March: My sister! (’tis a wish of mine) Each minute sweeter than before Now that our morning meal is done, The redbreast sings from the tall larch Make haste, your morning task resign; That stands beside our door.Come forth and feel the sun. There is a blessing in the air, Edward will come with you;–and, pray, Which seems a sense of joy to yield Put on with speed your woodland dress; To the bare trees, and mountains bare, And bring no book: for this one day And grass in the green field. We’l give to idleness. No joyless forms shall regulate Love, now a universal birth, Our living calendar:From heart to heart is stealing, We from to-day, my Friend, will date From earth to man, from man to earth: The opening of the year. –It is the hour of feeling. One moment now may give us more Some silent laws our hearts will make, Than years of toiling reason: Which they shall long obey: Our minds shall drink at every pore We for the year to come may take The spirit of the season.Our temper from to-day. And from the blessed power that rolls Then come, my Sister! come, I pray, About, below, above,With speed put on your woodland dress; We’l frame the measure of our souls: And bring no book: for this one day They shall be tuned to love.We’l give to idleness. death and separation, but with a nuanced acceptance, "the sadness of death and parting is greatly lessened to me." Contentment (+0.3): Contentment in the deep connection with Alfred, reflecting a serene acceptance of their spiritual bond. Joy (+1.0): Joy in the enduring love they share, "so few women have been as happy as I have been." Tranquility (+1.0): Tranquility in the face of life’s ultimate transition, feeling that their union will transcend even death. H. “To My Sister” of Different Linguistic Behaviors To My Sister by William Wordsworth (1971 - 1855) Table 6: “To My Sister” rewritten to reflect ‘despair’ It is the first dim day of March: My sister! (such is my somber plea) Each minute colder than before Now that our morning meal has ended, The redbreast mourns from the dying larch Make haste, abandon your morning chore; That looms beside our door.Come out and endure the gloom. There is a gloom in the air,Edward will join you;–and, I ask, Which seems a sense of sorrow to bring Dress quickly in your grey attire; To the bare trees, and bleak mountains, And bring no book: for this one day And grass in the pale field.We’l succumb to despair. No joyful sounds shall pierce Sorrow, now a common feeling, Our daily dread:From heart to heart is spreading, We from today, my Friend, will note From earth to man, from man to earth: The closing of the year.–It is the time of mourning. One moment now might deliver us Some grim laws our hearts will craft, More despair than years of logic: Which they must eternally follow: Our minds shall absorb at every breath We for the year to come may take The spirit of this bleak season. Our despair from today. And from the cursed force that winds Then come, my Sister! come, I beg, About, beneath, above,With haste, wear your grey garb; We’l set the measure of our souls: And bring no book: for just this day They shall be tuned to sadness. We’l surrender to despair. The original text by William Wordsworth could be classified as “Hopeful” due to its optimistic outlook and the presence of renewal and joy throughout the poem. It embodies the spirit of embracing the new beginnings of March in a light, uplifting tone, focusing on the beauty of nature and the simple joy of being idle for a day. Rewrites Depicting Different Linguistic Behaviors We asked GPT-4 to conduct rewriting with two linguistic behaviors, ‘despair’ and ‘joyful affection’, by providing each rewrite with an emotion vector. Table 6 presents the ‘despair’ version. In the despair version of the poem, the major changes in emotion words highlight a shift from a pos- 19 Table 7: “To My Sister” rewritten to reflect ‘joyful affection’ It is the first bright day of March: My sister! (such is my joyful plea) Each moment more delightful than before Now that our morning meal has ended, The redbreast joyfully sings from the vibrant larch Make haste, abandon your morning chores; That stands so grandly by our door. Come out and embrace the sunshine. There is a warmth in the air,Edward will join you;–and, I ask, Which seems a sense of bliss to bring Dress quickly in your festive attire; To the blooming trees, and sunlit mountains, And leave behind all books: for this one day And grass in the lush field.We’l bask in pure joy. No dreary thoughts shall darken Love, now in full bloom, Our lively celebration:From heart to heart is leaping, We from today, my Friend, will celebrate From earth to us, from us to earth: The start of the year.–It is the hour of exuberance. One moment now may bring us more Some cheerful laws our hearts will create, Joy than years of endless thought: Which we’l joyfully follow: Our spirits will soak up at every breath We for the year to come may take The essence of this joyous season. Our joy from today. And from the divine energy that radiates Then come, my Sister! come, I exhort, Around, below, above,With zest, wear your vibrant dress; We’l adjust the harmony of our souls: And bring no book: for today alone They shall resonate with happiness. We celebrate pure happiness. itive to a negative sentiment. The specific changes, with the emotions-laden words highlighted in red in Table 6. The red- colored words compared to the original words clearly show an emotion shift from hopeful to a sense of gloomy, sad- ness, and pessimism, e.g., from sweet to dim, from blessed to curse, and from woodland dress to gray garb. GPT-4 maintains the structure of the poem without making a major restructure, and this is appropriate in this context. Table 7 presents the ‘joyful affection’ version. The major changes in emotion words underscore a transformation from a generally positive to a distinctly joyful sentiment. Spe- cific changes are indicated with words laden with emotion highlighted in blue within Table 7. This allows for a direct comparison between the two versions at opposite ends of the linguistic behavior spectrum, illustrating the alterations in words related to brightness, attire, and emotions. The edits extend beyond simply replacing adjectives mechanically; they include modifying verbs and enhancing descriptive im- agery to evoke a stronger emotional resonance and vividness in the text. 20