EERPD: Leveraging Emotion and Emotion Regulation for Improving Personality Detection (2024)

1 Introduction

As a fundamental construct in psychology, personality reveals the true nature of the individual and creates a certain impression on others(Jung, 1959; Corr and Matthews, 2009; Jung, 1959).With the advancement of Natural Language Processing (NLP) technologies, there has been an growing interest in automatic detection of personality(Petrides and Mavroveli, 2018; Yang etal., 2023), whichplays a pivot role in numerous human-oriented NLP applications, such as psychological health assessment Wilkinson and Walford (2001), personalized recommendation systems Hu and Pu (2010), and human-computer interaction Pocius (1991).

Traditional personality detection methods either treat text as a whole, relying primarily on direct content analysisYang etal. (2023); Hu etal. (2024), or laying particular emphasis on emotion expressionMohammad and Kiritchenko (2013); Li etal. (2021). However, these approaches often overlook the role of Emotion Regulation(Gross, 2008), a key psychological concept related to personality.Different from emotion that is usually expressed in a short-term manner, emotion regulation is a stable, long-term status of managing and controlling one’s emotional responses, as exemplified in Figure1. Psychological researches have demonstrated a clear correlation between one’s personality and emotion regulationBarańczuk (2019); Petrides and Mavroveli (2018); Borges and Naugle (2017).

Inspired by the psychological studies above, we propose a RAG-based framework named EERPD for automatic detection of personality, leveraging both emotion and emotion regulation as guidance.To guide the LLMs in personality detection, we first construct a reference library composing of a large number of text-personality pairs. Reference samples most similar to the input text are retrieved to facilitate few-shot learning. To be specific, we categorize each sentence of the input text into emotion sentences and emotion regulation sentences, encode them separately, and then combine the two vectors for effective retrieval of the most similar samples from the reference library. For each retrieved sample, we further include corresponding chain-of-thoughts (CoT) emphasizing emotion and emotion regulation to direct the LLM’s attention to these aspects.

For comprehensive evaluation, we test our EERPD on both the the Kaggle dataset ¹¹1https://www.kaggle.com/datasnaek/mbti-type for MBTIMyers-Briggs (1991) detection and the Essays dataset Pennebaker and King (1999) for the Big Five personalityGoldberg (1990) detection. The experimental results show that our method significantly improves the few-shot performance of GPT-3.5 in personality detection tasks, outperforming previous SOTA by 15.5/4.3 on average F1 on the two datasets. Further ablation and analysis consolidates the effectiveness of emotion regulation in personality detection, aligning well with psychological discoveries. To sum up, our contributions are as follows:

2 Related Work

Personality Detection In the early development of personality detection, Francis and Booth (1993) introduced the Linguistic Inquiry and Word Count (LIWC), pioneering the use of psycholinguistic features for personality analysis through texts. This tool became foundational for feature engineering in subsequent studies, such as those by Pennebaker and King (1999) and Argamon etal. (2005), which focused on linguistic styles and lexical predictors of personality traits, achieving moderate accuracies in detecting traits like extraversion and neuroticism.

With the emergence of neural networks, research expanded significantly. Techniques like CNNs and LSTMs enhanced personality prediction from social media Tandera etal. (2017); Xue etal. (2018a). The introduction of BERT advanced the field further, with Gjurković etal. (2020) showing its effectiveness in analyzing personality and demographics on Reddit without extensive feature engineering.

Recent studies have explored multi-task and multimodal approaches to personality detection. Sang etal. (2022) used movie scripts to predict MBTI types of fictional characters, showing the potential of diverse data integration. Li etal. (2021) employed multitask learning to detect emotions and personality traits simultaneously, demonstrating the efficiency of shared representations.

Current research explores large language models (LLMs) for personality detection, as shown by Yang etal. (2023) and Hu etal. (2024), indicating a shift towards inferring personality traits directly from text with minimal reliance on traditional feature engineering and adopting more holistic, context-aware methodologies.

Emotion and Emotion Regulation The relationship between emotion and personality is well-studied in psychology and NLP. Psychological theories Keltner (1996); Davidson (2001); Reisenzein and Weber (2009) link personality traits with emotional experiences, lading the foundation for inferring personality from emotions. Mohammad and Kiritchenko (2013) showed fine-grained emotions enhance personality detection, Rangra etal. (2023) demonstrated the effectiveness of emotional features in speech, and Li etal. (2021) found that multitask learning improves prediction accuracy.

The connection between emotion regulation and personality has also been explored in psychology. Emotion regulation is significantly associated with and can influence personality Barańczuk (2019). Individuals with strong emotion regulation skills show personality traits assosiated with higher job satisfaction and better stress management Petrides and Mavroveli (2018). Borges and Naugle (2017) found emotion regulation variables predict specific personality dimensions. However, to the best of our knowledge, no NLP-based personality detection methods utilize emotion regulation for prediction.

3 Task Formulation

In this paper, we focus on the personality detection task, which aims to predict an individual’s personality traits from text.Each text to be detected, $X$, is made up of $\{x_{1},x_{2},...,x_{n}\}$, where each $x_{i}$ is a sentence.The goal of personality detection is to map $X$ to a multidimensional label $y$.

4 Method

In this section, we introduce our EERPD framework, as illustrated in Figure2.First, we construct a reference library, providing text-label pairs to serve as examples for the personality detection model (§4.1).Then, we utilize psychological knowledge to categorize each sentence in text into Emotion Sentences (ES) and Emotion Regulation Sentences (ERS) (§4.2). After that, we retrieve the examples through the combination of ES and ERS (§4.3). In inference phase, we utilize examples from reference library for personality detection (§4.4). And the whole method is shown as Algorithm1.

4.2 Sentence Categorization

When performing personality detection, emotion and emotion regulation have different characteristics: emotion is expressed in a short-term manner, while emotion regulation is a long-term skill for managing and controlling one’s emotional responses. Therefore, we need to handle them separately.For the input text $X$, we use a prompt-based approach to have the large language model (LLM) categorize its sentences into two parts.We define the classification criteria and input them along with text $X$ into the model, which can then labels each sentence $x$ to perform sentence classification. In this way, the sentences in input text $X$ is categorized into two parts: Emotion Sentences $(X_{e}=\{x_{e1},x_{e2},...,x_{en}\})$ and Emotion Regulation Sentences ($X_{r}=\{x_{r1},x_{r2},...,x_{rn}\})$. Referring to concepts from psychologyGross (2008), the classification criteria are defined as follows:

Emotion Sentences: The feelings in the sentence is dominated by emotion, it should be an obvious reaction to a recent event and not indicative of a deeper, long-standing trait or belief.

Emotion Regulation Sentences: The feelings in the sentence is dominated by emotion regulation, it should reflect the author’s enduring traits rather than immediate circ*mstances.

For details of the prompt used to accomplish sentence categorization. please refer to AppendixB

Input: Hyperparameter: $\alpha$, LLM${\rm{:LLM}}\left(\cdot\right)$, Author’s text: X, Reference Library: D, EER split Prompt: $I_{p}$, $I_{m}$, Prediction Prompt: PMT

Output: The inferred personality trait: y

$X_{e,r}\leftarrow LLM({X,I_{e,r}})$;

$V_{xe,xr}\leftarrow vectorize(X_{e,r})$;

$V_{x}\leftarrow V_{xe}+V_{xr}$;

$Sim\leftarrow[null]$;

foreach text $d$ in $D$do

$d_{e},d_{r}\leftarrow\rm{LLM}\left(d,I_{e},I_{r}\right)$;

$V_{de,dr}\leftarrow vectorize(d_{e},d_{r})$;

$V_{d}\leftarrow\alpha V_{de}+(1-\alpha)V_{dr}$;

$sim\leftarrow 1-\cos(V_{x},V_{d})$;

append $sim$ to $Sim$;

$\{t_{1,2}\}\leftarrow argsort_{t\in D}\ Sim[-2:]$$Egs\leftarrow\{text_{t_{1},t_{2}},CoT_{t_{1},t_{2}}\}$;

$y\leftarrow LLM(PMT,Egs,X)$;

return y;

See Also

CDK cyberattack shuts down auto dealerships across the U.S. Here's what to know.

5 Experiments

5.5 Ablation Study

To verify the importance of each component in our EERPD, we conduct an ablation study on 100 samples randomly selected from each of the Kaggle test dataset and Essays test dataset.

Emotion and Emotion Regulation.

We first analyze the contributions of Emotion and Emotion Regulation for example retrieval. $EERPD{}{w/o\ E}$ refers to the condition where $\alpha=1.0$, meaning examples are retrieved only based on Emotion Regulation Vectors. $EERPD{}{w/o\ ER}$ refers to the condition where $\alpha=0.0$, meaning examples are retrieved only based on Emotion Vectors. As shown in Table 3, we use three groups for comparison: our overall method shows significant improvements compared to the 0-shot and 2-shot baselines; compared to baselines utilizing only Emotion Regulation ($EERPD{}{w/o\ E}$) or only Emotion ($EERPD{}{w/o\ ER}$), it is demonstrated that both components contribute to improvements. Our combined method outperforms both $EERPD{}{w/o\ E}$ and $EERPD{}{w/o\ ER}$

Parameter $\alpha$.

We investigate the trade-off parameter $\alpha$ in our EERPD, demonstrating the model’s sensitivity to $\alpha$ variations and identifying its optimal range. The results, shown in Figure 4, illustrate performance variations with $\alpha$ on the Kaggle and Essays test datasets.

For the two dataset, both accuracy and Macro-F1 score peaked at $\alpha=0.7$, then declined. These findings suggest that $\alpha=0.7$ has the best balance between Emotion Regulation and Emotion, with emotion regulation proving more predictive of personality. Performance drops when $\alpha$ is 0 or 1, highlighting that combining both features is more effective than using a single one. The experimental results also show that $\alpha=0.7$, is also higher than $\alpha=0.5$, which is the vector of the original text directly used for retrieval without combining in weighted proportion. And it proves that our combined method is more efficient than the whole article for retrieval.

Auxiliary CoT.

We evaluate the effect of the auxiliary CoTs generation statistically, using the 100 random samples from Kaggle dataset. And the F1 scores are shown in Table 4. Each data point is the mean of 5 trials. T-test analysis demonstrates that the auxiliary CoTs has statiscal significance with p less than 0.05.

Dataset	E/I	N/S	T/F	P/J
EERPD	90.56	91.97	92.43	81.51
w/o CoTs	78.61	83.17	86.88	79.03

6 Analysis

6.1 Different Base Model

To evaluate the robustness of our EERPD method across different model architectures, we conducted experiments using various popular language models, including BART Lewis etal. (2019), BERT Devlin etal. (2018), LLAMA Touvron etal. (2023), RoBERTa Liu etal. (2019), and W2V Le and Mikolov (2014). The evaluation dataset consists of 100 samples selected from the Kaggle test dataset.

The results presented in Figure 5, indicate that EERPD maintains consistent performance across different model architectures. This flexibility allows for broader applicability in various settings.

7 Conclusion

In this paper, we introduced EERPD, a novel few-shot personality detection method inspired by psychological concepts. By leveraging emotion regulation and emotion to retrieve few-shot samples, EERPD forms a robust Chain of Thought (CoT) that guides large language models in evaluating personality traits. Experiments on two benchmark datasets show that EERPD significantly outperforms traditional methods and other novel prompts. This approach uniquely integrates psychological insights to enhance the reasoning abilities of large language models, offering a new perspective for personality detection.

Limitations

Due to the resource constraints, we only conduct experiments and analysis about LLMs on GPT-3.5. The extent to which GPT-4 or GPT-4o models can benefit from our EERPD remains unknown.

This study primarily focuses on improving the LLM’s performance by leveraging the psychological knowledge of Emotion Regulation. How to exploit trainable and tunable models like BERT and LLAMA to further optimize EERPD is left for future investigation.

This method carries certain potential risks. Even with well-intentioned use, personality detection may lead to misjudgments, negatively impacting individuals’ careers or social relationships.

Ethics Statement

This work adheres to the ACL Ethics Policy. We assert that, to the best of our knowledge, our work does not present any ethical issues. We have conducted a thorough review of potential ethical implications in our research and found none.

Appendix A Appendix: Prompt for Prediction

To better understand our method, we provide all the prompt in appendix, and record for the whole prediction is in Figure 7.

Appendix B Appendix: Prompt for Sentence Categorization

To better understand our method, we provide all the prompt in appendix, and record for EER Text Split prompt is in Figure 8.

Appendix C Appendix: Prompt for Generation of auxiliary CoT

To better understand our method, we provide all the prompt in appendix, and record for generating the CoT in Reference Library is in Figure 9.

References

• Argamon etal. (2005)Shlomo Argamon, Sushant Dhawle, Moshe Koppel, and JamesW Pennebaker. 2005.Lexical predictors of personality type.In Proceedings of the 2005 joint annual meeting of the interface and the classification society of North America, pages 1–16.
• Barańczuk (2019)Urszula Barańczuk. 2019.The five factor model of personality and emotion regulation: A meta-analysis.Personality and Individual Differences, 139:217–227.
• Borges and Naugle (2017)LaurenM Borges and AmyE Naugle. 2017.The role of emotion regulation in predicting personality dimensions.Personality and Mental Health, 11(4):314–334.
• Chen (2015)Yahui Chen. 2015.Convolutional neural network for sentence classification.Master’s thesis, University of Waterloo.
• Corr and Matthews (2009)PhilipJ Corr and GeraldEd Matthews. 2009.The Cambridge handbook of personality psychology.Cambridge University Press.
• Cui and Qi (2017)Brandon Cui and Calvin Qi. 2017.Survey analysis of machine learning methods for natural language processing for mbti personality type prediction.Final Report Stanford University.
• Davidson (2001)RichardJ Davidson. 2001.Toward a biology of personality and emotion.Annals of the New York academy of sciences, 935(1):191–207.
• Devlin etal. (2018)Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. 2018.Bert: Pre-training of deep bidirectional transformers for language understanding.arXiv preprint arXiv:1810.04805.
• Devlin etal. (2019)Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. 2019.Bert: Pre-training of deep bidirectional transformers for language understanding.In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long and Short Papers), pages 4171–4186.
• Francis and Booth (1993)MEFrancis and RogerJ Booth. 1993.Linguistic inquiry and word count.Southern Methodist University: Dallas, TX, USA.
• Gjurković etal. (2020)Matej Gjurković, Mladen Karan, Iva Vukojević, Mihaela Bošnjak, and Jan Šnajder. 2020.Pandora talks: Personality and demographics on reddit.arXiv preprint arXiv:2004.04460.
• Goldberg (1990)LewisR Goldberg. 1990.An alternative" description of personality": the big-five factor structure.Journal of personality and social psychology, 59(6):1216.
• Gross (2008)JamesJ Gross. 2008.Emotion regulation.Handbook of emotions, 3(3):497–513.
• Hu etal. (2024)Linmei Hu, Hongyu He, Duokang Wang, Ziwang Zhao, Yingxia Shao, and Liqiang Nie. 2024.Llm vs small model? large language model based text augmentation enhanced personality detection model.In Proceedings of the AAAI Conference on Artificial Intelligence, volume38, pages 18234–18242.
• Hu and Pu (2010)Rong Hu and Pearl Pu. 2010.A study on user perception of personality-based recommender systems.In User Modeling, Adaptation, and Personalization: 18th International Conference, UMAP 2010, Big Island, HI, USA, June 20-24, 2010. Proceedings 18, pages 291–302. Springer.
• Jung (1959)CGJung. 1959.The archetypes and the collective unconscious.
• Keltner (1996)Dacher Keltner. 1996.Facial expressions of emotion and personality.Handbook of emotion, adult development, and aging, pages 385–401.
• Kojima etal. (2022)Takeshi Kojima, ShixiangShane Gu, Machel Reid, Yutaka Matsuo, and Yusuke Iwasawa. 2022.Large language models are zero-shot reasoners.In ICML 2022 Workshop on Knowledge Retrieval and Language Models.
• Le and Mikolov (2014)Quoc Le and Tomas Mikolov. 2014.Distributed representations of sentences and documents.In International conference on machine learning, pages 1188–1196. PMLR.
• Lewis etal. (2019)Mike Lewis, Yinhan Liu, Naman Goyal, Marjan Ghazvininejad, Abdelrahman Mohamed, Omer Levy, Ves Stoyanov, and Luke Zettlemoyer. 2019.Bart: Denoising sequence-to-sequence pre-training for natural language generation, translation, and comprehension.arXiv preprint arXiv:1910.13461.
• Li etal. (2021)Yang Li, Amirmohammad Kazameini, Yash Mehta, and Erik Cambria. 2021.Multitask learning for emotion and personality detection.arXiv preprint arXiv:2101.02346.
• Liu etal. (2019)Yinhan Liu, Myle Ott, Naman Goyal, Jingfei Du, Mandar Joshi, Danqi Chen, Omer Levy, Mike Lewis, Luke Zettlemoyer, and Veselin Stoyanov. 2019.Roberta: A robustly optimized bert pretraining approach.arXiv preprint arXiv:1907.11692.
• Mohammad and Kiritchenko (2013)Saif Mohammad and Svetlana Kiritchenko. 2013.Using nuances of emotion to identify personality.In Proceedings of the International AAAI Conference on Web and Social Media, volume7, pages 27–30.
• Myers-Briggs (1991)Isabel Myers-Briggs. 1991.Introduction to type: A description of the theory and applications of the myers-briggs indicator.Consulting Psychologists: Palo Alto.
• Park etal. (2015)Gregory Park, HAndrew Schwartz, JohannesC Eichstaedt, MargaretL Kern, Michal Kosinski, DavidJ Stillwell, LyleH Ungar, and MartinEP Seligman. 2015.Automatic personality assessment through social media language.Journal of personality and social psychology, 108(6):934.
• Pennebaker etal. (2001)JamesW Pennebaker, MarthaE Francis, and RogerJ Booth. 2001.Linguistic inquiry and word count: Liwc 2001.Mahway: Lawrence Erlbaum Associates, 71(2001):2001.
• Pennebaker and King (1999)JamesW Pennebaker and LauraA King. 1999.Linguistic styles: language use as an individual difference.Journal of personality and social psychology, 77(6):1296.
• Petrides and Mavroveli (2018)KonstantinosV Petrides and Stella Mavroveli. 2018.Theory and applications of trait emotional intelligence.Psychology: The Journal of the Hellenic Psychological Society, 23(1):24–36.
• Pocius (1991)KymE Pocius. 1991.Personality factors in human-computer interaction: A review of the literature.Computers in Human Behavior, 7(3):103–135.
• Rahman etal. (2019)MdAbdur Rahman, Asif AlFaisal, Tayeba Khanam, Mahfida Amjad, and MdSaeed Siddik. 2019.Personality detection from text using convolutional neural network.In 2019 1st international conference on advances in science, engineering and robotics technology (ICASERT), pages 1–6. IEEE.
• Rangra etal. (2023)Kalpana Rangra, Virender Kadyan, and Monit Kapoor. 2023.Emotional speech-based personality prediction using npso architecture in deep learning.Measurement: Sensors, 25:100655.
• Reisenzein and Weber (2009)Rainer Reisenzein and Hannelore Weber. 2009.Personality and emotion.The Cambridge handbook of personality psychology, 2:81–99.
• Sang etal. (2022)Yisi Sang, Xiangyang Mou, MoYu, Dakuo Wang, Jing Li, and Jeffrey Stanton. 2022.Mbti personality prediction for fictional characters using movie scripts.arXiv preprint arXiv:2210.10994.
• Szegedy etal. (2017)Christian Szegedy, Sergey Ioffe, Vincent Vanhoucke, and Alexander Alemi. 2017.Inception-v4, inception-resnet and the impact of residual connections on learning.In Proceedings of the AAAI conference on artificial intelligence, volume31.
• Tandera etal. (2017)Tommy Tandera, Derwin Suhartono, Rini Wongso, YenLina Prasetio, etal. 2017.Personality prediction system from facebook users.Procedia computer science, 116:604–611.
• Tighe etal. (2016)EdwardP Tighe, JenniferC Ureta, Bernard AndreiL Pollo, CharibethK Cheng, and Remedios deDiosBulos. 2016.Personality trait classification of essays with the application of feature reduction.In SAAIP@ IJCAI, pages 22–28.
• Touvron etal. (2023)Hugo Touvron, Louis Martin, Kevin Stone, Peter Albert, Amjad Almahairi, Yasmine Babaei, Nikolay Bashlykov, Soumya Batra, Prajjwal Bhargava, Shruti Bhosale, etal. 2023.Llama 2: Open foundation and fine-tuned chat models.arXiv preprint arXiv:2307.09288.
• Wilkinson and Walford (2001)RossB Wilkinson and WendyA Walford. 2001.Attachment and personality in the psychological health of adolescents.Personality and Individual Differences, 31(4):473–484.
• Xue etal. (2018a)DiXue, Lifa Wu, Zheng Hong, Shize Guo, Liang Gao, Zhiyong Wu, Xiaofeng Zhong, and Jianshan Sun. 2018a.Deep learning-based personality recognition from text posts of online social networks.Applied Intelligence, 48(11):4232–4246.
• Xue etal. (2018b)DiXue, Lifa Wu, Zheng Hong, Shize Guo, Liang Gao, Zhiyong Wu, Xiaofeng Zhong, and Jianshan Sun. 2018b.Deep learning-based personality recognition from text posts of online social networks.Applied Intelligence, 48.
• Yang etal. (2021)Feifan Yang, Xiaojun Quan, Yunyi Yang, and Jianxing Yu. 2021.Multi-document transformer for personality detection.In Proceedings of the AAAI Conference on Artificial Intelligence, volume35, pages 14221–14229.
• Yang etal. (2022)Tao Yang, Jinghao Deng, Xiaojun Quan, and Qifan Wang. 2022.Orders are unwanted: Dynamic deep graph convolutional network for personality detection.In Proceedings of AAAI 2023.
• Yang etal. (2023)Tao Yang, Tianyuan Shi, Fanqi Wan, Xiaojun Quan, Qifan Wang, Bingzhe Wu, and Jiaxiang Wu. 2023.Psycot: Psychological questionnaire as powerful chain-of-thought for personality detection.In The 2023 Conference on Empirical Methods in Natural Language Processing.