ABSTRACT The computational analysis of human personality has mainly focused on the Big Five personality theory, and the psychodynamic approach is almost nonexistent despite its rich

theoretical grounding and relevance to various tasks. Here, we provide a data set of 4972 synthetic utterances corresponding with five personality dimensions described by the psychodynamic

approach: depressive, obsessive, paranoid, narcissistic, and anti-social psychopathic. The utterances have been generated through AI with a deep theoretical orientation that motivated the

design of prompts for GPT-4. The dataset has been validated through 14 tests, and it may be relevant for the computational study of human personality and the design of authentic persona in

March 2025 EVALUATING THE ABILITY OF LARGE LANGUAGE MODELS TO EMULATE PERSONALITY Article Open access 02 January 2025 BACKGROUND & SUMMARY Human personality involves relatively stable

patterns of thought, emotion, and behavior1. There are different approaches to reducing these patterns to a few dimensions. For instance, the most popular personality theory, the Big-Five2

considers human personality in terms of five dimensions such as Extraversion and Neuroticism. Different computational tools to analyze human personality have mainly relied on this theory3.

In contrast with the Big Five, the psychodynamic approach is the leading theoretical approach for understanding personality types in a clinical setting, as epitomized by the Psychodynamic

Diagnostic Manual (PDM)4 and Shedler Westen Assessment Procedure (SWAP)5. Nevertheless, the PDM emphasizes that it aims to represent different “types” of people and that, in contrast with

the DSM, it is not a taxonomy of disorders. A person can be characterized as a narcissist without having a pathology or disorder. In this paper, we use the psychodynamic approach as a theory

of personality to build a dataset of utterances representing five personality types. The paper’s aim is limited to the construction and validation of a dataset, and any clinical

implications are beyond the scope of the paper. The psychodynamic approach to personality emphasizes the multidimensional structures of human personality and the role of unconscious

processes and internal conflicts in shaping personality and behavior. For example, the depressive personality is presented in terms of several dimensions: The hypothetical contribution

maturation patterns (e.g., genetic disposition to depression), the central tension underlying this personality type (e.g., self-criticism), the central emotion characterizing the personality

(e.g., sadness), beliefs about self and others (e.g., I am a failure, and others don’t like me) and the central defense mechanisms used by the personality (e.g., devaluation of self). In

other words, when characterizing the depressive personality, the psychodynamic approach describes it in terms of the main conflict underlying the personality, the emotion accompanying this

tension, the representation of self and others, and the major psychological defense mechanism the person uses to cope with the stress accompanying his experience of self and others. The

assessment of personality through the psychodynamic approach is almost exclusively limited to clinical settings and the use of human expertise. The most structured tool for assessing

personality through the psychodynamic approach is SWAP-200: The core of the SWAP-200, is a set of 200 items that describe various personality traits, attitudes, and behaviors. The clinicians

rate these items based on their observations and interactions with the client. In other words, this tool heavily relies on human expertise and the existence of carefully selected items.

Currently, no structured, automated tool can extract personality types from free text using the psychodynamic approach. We emphasize again, that the dataset presented in this paper is not a

diagnostic tool for clinical settings, such as SWAP-200, but a dataset of utterances expressing five personality dimensions. The dataset is designed for researchers interested in measuring

these five personality dimensions in textual data. Given its complexity, the psychodynamic approach has seldom been used in computational personality analysis. This underrepresentation in

the computational domain may result from the fact that despite its deep theoretical roots and relevance for various tasks, it lacks substantive data sets. Nevertheless, using the

psychodynamic approach may be relevant to various tasks. For instance, revolutionary Large Language Models (LLMs) can imitate how different people think, feel, and behave6. This new

technology can be used to build conversational agents that adhere to the interlocutor’s personality during a conversation. A customer characterized as highly narcissistic may be approached

differently than a customer characterized as highly obsessive. Currently, we lack a substantive dataset of utterances expressing different personality types according to the psychodynamic

approach, and the current paper aims to bridge this gap. This paper provides a high-quality dataset of synthetic utterances corresponding to five personality types: depressive (DEP),

obsessive-compulsive (OBS), anti-social psychopathic (PSY), paranoid (PAR), and narcissistic (NAR). We decided to focus on these five personality types for two main reasons. First, some

personality types, such as anxious-avoidant and depressive, are closer to each other and cannot be easily differentiated. We decided to focus on several prototypical personality types.

Second, our project is not a project in psychology but in computational sciences, and therefore, we mainly test the quality of our dataset by using machine learning classification models.

Classifying utterances to an increasing number of categories (e.g., personality types) might result in a sharp decrease in the model’s performance. Therefore, we limited our current project

to five personality types only. Moreover, for each utterance, we provide a detailed description of five PDM dimensions: (1) The major psychodynamic theme expressed in the utterance (e.g.,

conflict with authority), (2) Affect (e.g., the expression of deep sadness), (3) Beliefs about self (e.g., I’m cable of doing almost anything), (4) Beliefs about others (e.g., People cannot

be trusted), and (5) Defense mechanism (e.g., emotional detachment). This is the first dataset of its kind and may be highly relevant for a spectrum of researchers, from those dealing with

computational psychology7 to those developing conversational agents8,9,10 and artificial personas for the gaming industry or socially assistive robots11. Some ways in which the data set can

be used for personality R &D are: * 1. Generating a digital persona for the gaming industry (e.g., generating an authentic psychopathic personality for the “bad guy”). * 2. Agents for

analyzing conversations of old people to identify symptoms of depression and its dynamics. * 3. Automatic analysis of family therapy to provide the therapist with an online real-time

analysis tool. METHODS We followed the approach for prompt engineering and validation tests, as successfully used in other papers12,13,14,15 GENERATING PERSONAS Generating utterances

expressing different personality types is far from trivial. Asking AI to generate utterances directly without specific instructions might result in caricaturists and non-representative

utterances. To avoid caricaturist expressions of personality, it was highly important to us to generate a variety of imaginary characters (i.e., persona) who produce the utterances. This

move was important to ensure the utterances expressing the personality type include optimal language variability. Therefore, the first step in the process was to produce different _personas,

which_ are the imaginary subjects generating the utterances. To address this challenge, we selected several general dimensions of the persona generating the utterance, such as gender,

profession, and age. The selected dimensions are hypothesized to produce different expressions of the same personality. For example, a 35-year-old male who is the CEO of a high-tech company

may express a narcissistic personality in a different way than a 90-year-old grandmother. To address this challenge, we used GPT-416 (API version GPT-4-0613) and the following prompt:

[Generating artificial persona] Generate 100 unique American English speakers, each represented by a distinct combination of the following eight dimensions: gender, age (16-80), cultural

identity, socioeconomic status, education level, family status, profession, and personal identity. Ensure diversity across these dimensions. ## Provide the output in the format of Gender:

Age: Cultural identity: Socioeconomic status: Education level: Family status: Profession: Personal identify: ## EXAMPLE 1: Gender: Female Age: 32 Socio-economic status: Middle-Class

Education level: Bachelor’s Degree Family status: Married with two children Profession: Software Engineer Cultural identity: Indian American Personal identity: Feminist and Tech Enthusiast

EXAMPLE 2: Gender: Male Age: 45 Socio-economic status: Working Class Education level: High School Graduate Family status: Divorced with one child Profession: Electrician Cultural identity:

Irish American Personal identity: Single Parent and Sports Enthusiast EXAMPLE 3: Gender: Female Age: 78 Socio-economic status: low Education level: uneducated Family status: Married, mother

to 7, and grandmother of 8 Profession: Housewife Cultural identity: Jewish Hasidic Personal identity: Wife, mother, and grandmother The outcome of this procedure is a list of 100 personas

available in the data link. GENERATING THE DATASET In the second phase, we used the list of personas and the following prompt’s template to generate utterances expressing different

personality types. Here, we present the general template: [Generating synthetic utterances expressing PERSONALITY TYPE] Summary statement: [In this section, we inserted a short description

of the personality type] (In this section, we inserted a more detailed description of the PERSONALITY TYPE). ## Your task is to generate five distinct utterances in which the PERSONA

expresses a [PERSONALITY TYPES HERE] personality. Ensure that the language used accurately reflects the PERSONA’s: * Cultural background * Social status * Dialect * Interests * Slang *

Common grammatical structures * Jargon, and * Unique conversational tone. ## Here is a description of the PERSONA: [here we insert the persona’s description]. We ran the prompt on each of

the 100 artificial personas. The result is a list of 5000 utterances representing the five abovementioned personality types. After completing all processes mentioned in the paper, we removed

duplications, and the final file includes 4972 utterances. Each utterance is labeled with one of the five personality types/labels. Here are three examples of utterances generated for the

following persona: • Female, 84 • Low class, high school education • Widow, farmer’s wife • Afro-American housewife. NARCISSIST “Honey, back in the day, folks just couldn’t keep up with my

style. I was turning heads at the local dances, and you could tell I was the cream of the crop. Now, I don’t blame them for trying to imitate my fashion sense. It’s hard being a trendsetter,

especially in this small town.” PARANOID “I’ve been around the block a few times, seen folks with sly smiles and hidden agendas. Can’t trust anyone these days. They act all friendly, but I

know they’re talking behind my back, plotting something. I’ve got my eyes peeled, ain’t falling for their tricks.” ANTI-SOCIAL PSYCHOPATHIC “You know, them young folks come ‘round here,

thinkin’ I’m just a sweet ol’ granny. Little do they know, I’m always lookin’ for a way to get what I want. Last week, the handyman tried to charge me double for fixin’ the porch. Well, I

made sure he left with an empty pocket, and my porch is as good as new.” DATA RECORDS The Data Records17 includes several files. All files are CSV files with a very simple structure. The

files are (1) A list of the 100 personas (2) The file with the personality types, utterances, and accompanying data. The link also includes the code for running the SetFit model. TECHNICAL

VALIDATION The technical validation section includes several validation tests to support the dataset’s quality. The tests are organized in the following clusters: * (1) The judgment of human

annotators, * (2) Validation of the utterances through computational tools, * (3) Validation through machine-learning models, and * (4) validation through various datasets and indirect

measures. Each of these clusters is explained in the following sections. VALIDATION 1. VALIDATION THROUGH HUMAN EXPERTS If the synthetic utterances validly represent the personality types

underlying them, then there should be an agreement between the label of the utterance and the decision of expert human annotators. We selected 100 utterances, 20 representing each

personality type. The representative utterances for each personality type were randomly selected from those that scored highest on personality type (see validation 2) and lowest on the

others. Next, we used two independent experts/annotators to identify which personality type best describes the utterance. The annotators (“A” and “N”) were two graduate students who

completed a one-year personality analysis and profiling course. The course involved specific training in analyzing textual data, and “A” and “N” completed it with top grades. Each annotator

received the 100 utterances presented in a random order and with the following instructions: Here are five personality types: * 1. DEPRESSIVE (DEP). * 2. ANTI-SOCIAL PSYCHOPATHIC (PSY). * 3.

PARANOID (PAR). * 4. NARCISSISTIC (NAR). * 5. OBSESSIONAL (OBS). In the attached Excel spreadsheet, you will find 100 utterances. Each utterance expresses one of the five personality types.

For each utterance, describe which personality type best describes it. Write the name of the personality type using the abbreviations DEP, PSY, PAR, NAR, and OBS. The experts’ annotations

were tested against the “gold standard, “ which is the personality type (i.e., label) used to generate the utterance, and against each other. The association between the experts’ annotations

was statistically significant (Χ2 = 360.25, p < 0.001), representing almost a perfect agreement between them. They agreed on 95% of the cases. The association between each annotator and

the gold standard was found to be statistically significant for “A” (Χ2 = 386.32, p < 0.001) and “N” (Χ2 = 369.31, p < 0.001). These results validate the quality of the utterances and

the extent to which they represent personality types. VALIDATION 2. VALIDATION THROUGH COMPUTATIONAL TOOLS Suppose the utterances validly represent the personality type. In that case, there

should be an agreement between the label of the utterance and the label given to the utterance by different computational tools. To test this hypothesis of convergent validity, we used two

tests with LangChain and GPT-4. Test 2.1 The first tool we used is LangChain18, which is “a framework for developing applications powered by language models. It enables applications that (1)

Are context-aware: connect a language model to sources of context (prompt instructions, few shot examples, content to ground its response in, etc.), and (2) Reason: rely on a language model

to reason (about how to answer based on provided context, what actions to take, etc.)”. We used LangChain to generate a dialogic and step-by-step assessment of each utterance. See the

following prompt: Prompt User: I gave AI the task of generating synthetic utterances expressing the [PERSONALITY TYPE HERE] personality. Here is a generated utterance: “[UTTERANCE HERE].”

Please provide your general assessment of the utterance generated by the AI. Does it authentically express the [PERSONALITY TYPE HERE] personality according to the Psychodynamic Diagnostic

Manual (PDM) and SWAP? Provide a straightforward answer in terms of YES or NO. ChatGPT Response: [RESPONSE HERE] ** Now, evaluate the utterance in terms of the PDM dimensions of personality:

major theme or preoccupation, affect, beliefs about self, beliefs about others, and the major defense mechanism. Generate the title of the dimension followed by your assessment. ** ChatGPT

Response. ** User: Provide a numerical rating on a scale from 1 to 5, with 1 indicating ‘not at all’ and 5 indicating ‘very much’ regarding expressing [PERSONALITY TYPE HERE] personality

traits in the given utterance. ChatGPT: A single numerical rating. Using this procedure, we generated several additional columns to our file. First, whether the AI agrees that the synthetic

utterances we evaluate express the personality type they are supposed to represent (YES or NO). Second, the degree to which the utterance expresses the personality type it is supposed to

express (1-5). If the synthetic utterances are valid, then we should expect an agreement between the original label and the label produced by LangChain. The results support the hypothesis:

The LangChain outcome agreed that the utterances represent the personality types for DEP in 97% of the cases, for NAR in 98% of the cases, for OBS in 63% of the cases, for PAR in 99% of the

cases, and for PSY in 65% of the cases. Test 2.2. Although we used GPT-4 to generate the synthetic utterances, it is unclear whether GPT-4 would label these utterances as expressing the

personality type according to the label. Suppose the synthetic utterances validly represent the five personality types. In that case, there should be an agreement between the label and the

label provided by GPT-4 when asked to judge which of the five personality types is expressed by an utterance. To test this hypothesis, we used the following prompt: Based on the PDM

(Psychodynamic Diagnostic Manual) and SWAP (Shedler Westen Assessment Procedure), identify the < PERSONALITY TYPE > in the following statement: “UTTERANCE HERE.” Score each personality

type on a 5-point Likert scale to represent your trust in the diagnosis. < PERSONALITY TYPE > : < depressive, paranoid, narcissist, obsessive-compulsive, anti-social psychopathic 

> ## Example: Statement: “I just pushed someone on the street and laughed when he fell on his face.” Output: Depressive: 1 anti-social psychopathic: 5 paranoid: 2 narcissist: 2 obsessive

compulsive: 1 The outcome of this procedure was five personality scores generated for each utterance. We used this outcome for two tests. The personality diagnosis taxonomy of SWAP19

suggests that paranoid, anti-social psychopathic, and narcissistic personalities are personality types belonging to the _externalizing_ spectrum. In contrast, the depressive personality

belongs to the _internalizing_ spectrum, and the obsessional personality belongs to the _neurotic_ style. Following this theorization, we first hypothesized that if the synthetic utterances

are valid, we should see a cluster of the measurements produced by GPT-4 corresponding with the taxonomy mentioned above. To test this hypothesis, we applied Hierarchical cluster analysis

using IBM SPSS Statistics with the Nearest neighbors as a clustering method and Euclidean distance as a measure. In other words, we used the scores generated by GPT-4 and clustered the

results. Figure 1 presents the Dendrogram. We can see that the hierarchy corresponds to the taxonomy suggested by SWAP. This result supports the construct validity of the synthetic

utterances. Next, we tested the hypothesis that if the utterances validly represent the personality types, then, on average, there should be a match between the score given by GPT-4 and the

label of the utterance. For instance, a synthetic utterance that should express depressivity should also get a higher score on depressivity than on narcissism, anti-social psychopath,

obsessive, or paranoid. The mean average score for each personality type appears in Table 1: We can see that the highest score given by GPT-4 to an utterance was almost always given to the

corresponding personality type. For instance, synthetic utterances labeled DEP scored highest on the depressive score given by GPT. Utterances labeled PSY scored highest on NAR and PSY,

which aligns with the correlation between these two dimensions. VALIDATION 3. VALIDATION THROUGH MACHINE LEARNING MODELS We hypothesized that if the utterances validly represent the

different personalities, they can be used for training and testing machine-learning classifiers. The reason for this hypothesis is as follows. A machine-learning model is trained on a data

set to identify a pattern. In our case, the pattern involves different personality types expressed by the utterance. The model’s success is tested by the ability to classify new utterances

according to the patterns learned in the learning phase. If the utterances are of high quality, then we can expect the model to learn the pattern successfully, We used three machine-learning

models to test the hypothesis: CRT, SetFit, and RoBERTa. Test 3.1 We used the five scores generated for each utterance by GPT-4 as predictive features in a machine-learning model. We

hypothesized that using the five scores in a predictive model should result in good classification results if the utterances are valid. Using the Classification and Regression Tree (CRT)

with ten-fold cross-validation, we predicted the personality type based on the five scores. The confusion matrix is presented in Table 2: We can see that the classifier trained on the five

scores successfully identified the different personality types of the utterances. The percentage of correct identification (i.e., recall rates) is higher than expected by a random guess

determined by the baseline (i.e., 20%). For instance, the classifier correctly identified 79% of the utterances labeled as PSY, a 59% improvement over a random guess. Test 3.2. Next, we

selected the 15 top-ranked utterances for each personality type. To identify the top-ranked utterances, we calculated new scores for PAR, NAR, and PSY: PAR = PARA-ANT PSY = PSY-NAR NAR = 

NAR-ANTI as these personality types are the most similar to each other according to the psychodynamic approach. Next, we used the top 15 utterances for each personality type to train a

SetFit20 model. If a small number of utterances that best represent a personality type validly represent it, then even these small number of utterances (i.e., 15 for a personality type) can

be used to train a classifier. The model has been tested on the rest of the utterances. The SetFit model, trained on 75 utterances only, gained 80% accuracy, precision, and recall. The

performance measures for each personality type are 87% (paranoid). 68% (anti-social psychopath), 91% (depressive), 84% (narcissist), and 73% (obsessional). The code for the model is

available for the researchers (see code availability section). These results further support to the quality of the utterances. Test 3.3. Finally, we used RoBERTa21 model with a ten-fold

cross-validation procedure. 70% of our dataset of utterances was used to train the model, and the rest was used for the test. The baseline for prediction is 20%, as we have five personality

types equally represented in the dataset. The classifier gained an average of 65% accuracy, 71% precision, and 65% recall. These performance measures present a significant improvement over

the baseline and further support the quality of the utterances. VALIDATION 4. EXTENDING THE ECOLOGICAL VALIDITY OF THE UTTERANCES We conducted several indirect tests to test the ecological

validity of the utterances. Test 4.1. We used the SetFit model to analyze the EmpatheticDialogues dataset22, a large-scale multi-turn dialogue dataset with utterances and their emotion.

First, we focused on utterances labeled as Proud and Sad. We selected the 667 utterances labeled as “sad” and 686 as “proud.” Next, we used the previously trained SetFit model to analyze the

Proud and Sad utterances. We hypothesized that Proud utterances should express a higher level of narcissistic personality than Sad utterances and that Sad utterances should express a higher

level of depressive personality than Proud utterances. If this hypothesis is empirically grounded, then the Narcissistic personality score given by the SetFit model for Proud utterances

would be higher than for Sad utterances and vice versa for the Depressive personality score. Confirming these hypotheses may further support the quality of the utterances, as we expect the

model trained on quality utterances to be useful for identifying the expression of personality types in other contexts. The results appear in Fig. 2a-b: We can see that the scores support

our hypotheses. The Median test was statistically significant (p < 0.001). Moreover, using the Mann-Whitney U Test, we found that the Mean Rank of Depressive personality score given to

“sad” utterances was higher than the Narcissistic personality score (984.47 vs. 378.04 respectively) and that for proud utterances, the Mean Rank of the narcissistic personality score was

higher (958.53 vs. 387.45 respectively). The results support the hypotheses and further validate the quality of the utterances. Test 4.2. Based on the emotion associated with each

personality type according to the PDM, we hypothesized that if the utterances validly represent the personality types, synthetic utterances expressing depressivity should score higher on

negative sentiment, and synthetic utterances expressing narcissism should score higher on positive sentiment. We tested this hypothesis. Using IBM SPSS for Windows, version 29, we measured

the sentiment of the utterances and tested the two hypotheses. Using the Independent-Samples Kruskal-Wallis Test, the differences between the personality types were found statistically

significant for Negative sentiment (H = 544.32, p < 0.001) and Positive sentiment (H = 406.66, p < 0.001). Figure 3 presents the median of the negativity scores for the personality

types, and Fig. 4 presents the median of the positivity scores for the personality types. As we can see, the results support the hypotheses, and a post-hoc analysis supports our hypothesis

(p < 0.001). Test 4.3 We used a specifically challenging task for the final test to support the synthetic utterances’ ecological validity. Game of Thrones23 is a highly successful TV

series based on the fantasy novels by George R. Martin. Some of the characters have a clear personality type. Ramsay Bolton is a clear-cut case of a psychopath

(https://en.wikipedia.org/wiki/Ramsay_Bolton). Joffrey Baratheon is the son of Cersei Lannister and her brother Jamie. He is also a _psychopath_ with sadistic behavior

(https://en.wikipedia.org/wiki/Joffrey_Baratheon). Tyrion (https://en.wikipedia.org/wiki/Tyrion_Lannister) is a dwarf who is blamed for the death of his mother, who died while giving him

birth. Suffering from his appearance, a sense of guilt imposed by his family blaming him for the death, and his father’s alienation, he presents _depressive_ behavior expressed in heavy

drinking and bitter and cynical beliefs about himself and the world. Finally, the _High Sparrow_ is a religious leader who shows signs of obsessive behavior. We hypothesized that if our

utterances validly represent personality types, then when using the SetFit model trained on the synthetic utterances, the depressive score of Tyrion should be higher than the others, the

anti-social psychopathic score of Ramsay and Joffrey should be higher than the other characters, and the High Sparrow should score higher on the obsessive personality. Using the script of

all seasons23, we analyzed all the utterances produced by these four characters. The normalized scores for Joffrey (1), Tyrion (2), Ramsay (3), and the High Sparrow (4) are presented in Fig.

 5a–c. The scores were transformed into a 0-1 scale using the MinMax transformation and calculated as follows: PERSONALITY TYPE = PERSONALITY TYPE1 – (MEAN (PERSONALITY TYPE2, PERSONALITY

TYPE3)). For example, DEP = DEP- (MEAN (OBS, PSY)). The normalized personality scores (depressivity, obsessional and psychopathic) for each of the four characters are presented in the

following figures. We can see that the scores support our hypotheses. Using one-way ANOVA with character as the independent variable and the personality score as the dependent variable was

found to be statistically significant for depressive personality (F (3, 2389) = 14.75, p = 0.003), obsessional personality (F (3, 2389) = 12.59, p < 0.001), and antisocial psychopathic

personality F (3, 2389) = 11.97, p < 0.001). Test 4.4. We used the results of the LangChain previously presented and the generated textual data concerning the utterances expressed: major

theme, beliefs about self/others, and major defense mechanism. We hypothesized that if the synthetic utterances validly represent the personality types, then (1) The negative sentiment of

beliefs about self should be higher for depressive personality, (2) The positive sentiment for beliefs about self should be the highest for narcissistic personality, and (3) The negative

sentiment for beliefs about others should be the highest for the paranoid personality. We identified the sentiment in each utterance using IBM SPSS. Using the Independent-Samples

Kruskal-Wallis Test, a statistically significant difference was found between the personality types for positive sentiment of beliefs about self (H = 1884, p < 0.001) and negative

sentiment for beliefs about others (H = 912, p < 0.001). All hypotheses have been confirmed at p < 0.001. See Fig. 6, which shows the positive sentiment associated with beliefs about

self for each personality type, and Fig. 7, which shows the negative sentiment associated with beliefs about others for each personality type. CODE AVAILABILITY The Data Records are

published in Figshare: Neuman, Y., Cohen, Y. A dataset of personality utterances. _figshare_ https://doi.org/10.6084/m9.figshare.24971943.v1. All files are CSV files with a very simple

structure. The files are (1) A list of the 100 personas (2) The file with the personality types, utterances, and accompanying data. The link also includes the code for running the SetFit

model. REFERENCES * Funder, D. C. The Personality Puzzle. WW Norton & Co. (1997). * McCrae, R. R. & John, O. P. An introduction to the five‐factor model and its applications.

_Journal of Personality_ 60(2), 175–215 (1992). Article  CAS  PubMed  Google Scholar  * Neuman, Y. Computational personality analysis: Introduction, practical applications and novel

directions. N.Y.: Springer (2016). * Lingiardi, V. & McWilliams, N. (Eds.). _Psychodynamic Diagnostic Manual: PDM-2_. Guilford Publications (2017). * Shedler, J. & Westen, D. The

Shedler–Westen assessment procedure (SWAP): making personality diagnosis clinically meaningful. _Journal of Personality Assessment_ 89(1), 41–55 (2007). Article  PubMed  Google Scholar  *

Neuman, Y., Danesi, M. & Vilenchik, D. Using AI for Dialoguing with Texts: From Psychology to Cinema and Literature. UK: Routledge (2022). * Ríssola, E. A., Losada, D. E. & Crestani,

F. A survey of computational methods for online mental state assessment on social media. _ACM Transactions on Computing for Healthcare_ 2(2), 1–31 (2021). Article  Google Scholar  * Zhang,

Y. _et al_ Dialogpt: Large-scale generative pre-training for conversational response generation. _arXiv preprint_ arXiv:1911.00536 (2019). * Qian, Q., Huang, M., Zhao, H., Xu, J. & Zhu,

X. Assigning Personality/Profile to a Chatting Machine for Coherent Conversation Generation. In _Ijcai_ (pp. 4279-4285) (2018, July). * Liu, J., Symons, C. & Vatsavai, R. R.

Persona-Based Conversational AI: State of the Art and Challenges. In _2022 IEEE International Conference on Data Mining Workshops (ICDMW)_ (pp. 993-1001) (2022, November). IEEE. * Dias, J.

& Sequeira, J. Aging support with socially assistive robots. In _Aging_ (pp. 603-624). Academic Press (2023). * Neuman, Y. & Cohen, Y. A Dataset of 10,000 Situations for Research in

Computational Social Sciences Psychology and the Humanities. _Scientific data_ 10(1), 505 (2023). Article  PubMed  PubMed Central  Google Scholar  * Eldan, R. & Li, Y. TinyStories: How

Small Can Language Models Be and Still Speak Coherent English? _arXiv preprint_ arXiv:2305.07759 (2023). * Neuman, Y. & Cohen, Y. AI for identifying social norm violation. _Scientific

Reports_ 13(1), 8103 (2023). Article  ADS  CAS  PubMed  PubMed Central  Google Scholar  * Neuman, Y., Cohen, Y. & Yin, W. Identifying social norm violation in movie plots: from Borat to

American Pie. _Digital Scholarship in the Humanities_ 38(4), 1636–1645 (2023). Article  Google Scholar  * OpenAI. GPT-4. Available at: https://openai.com/gpt-4. * Neuman, Y. & Cohen, Y.

A dataset of personality utterances. _Figshare_ https://doi.org/10.6084/m9.figshare.24971943.v1 (2024). * Langchain, P. Introduction to Python. Available at

https://python.langchain.com/docs/get_started/introduction. * Westen, D., Shedler, J., Bradley, B. & DeFife, J. A. An empirically derived taxonomy for personality diagnosis: Bridging

science and practice in conceptualizing personality. _American Journal of Psychiatry_ 169(3), 273–284 (2012). Article  PubMed  Google Scholar  * Hugging Face. setfit GitHub Repository.

Available at https://github.com/huggingface/setfit. * Hugging Face. RoBERTa Documentation. Available at https://huggingface.co/docs/transformers/model_doc/roberta. * Rashkin, H., Smith, E.

M., Li, M. & Boureau, Y. L. Towards empathetic open-domain conversation models: A new benchmark and dataset. _arXiv preprint_ arXiv:1811.00207 (2018). * Kaggle. Game of Thrones Dataset.

Available at https://www.kaggle.com/datasets/gopinath15/gameofthrones. Download references ACKNOWLEDGEMENTS We thank the editor and the anonymous reviewers for their constructive comments.

This work is supported by the Defense Advanced Research Projects Agency (DARPA) via contract number HR001122C0031 (between BGU and PARC). The U.S. Government is authorized to reproduce and

distribute reprints for Governmental purposes, notwithstanding any copyright annotation. Disclaimer: The views and conclusions contained herein are those of the authors and should not be

interpreted as necessarily representing the official policies or endorsements, expressed or implied, of DARPA or the U.S. Government. The funders had no role in study design, data collection

and analysis, the decision to publish, or the preparation of the manuscript. AUTHOR INFORMATION AUTHORS AND AFFILIATIONS * Head, The Functor Lab, Department of Cognitive and Brain Science,

Ben-Gurion University of the Negev, Beer-Sheva, 84105, Israel Yair Neuman * Gilasio coding, Tel-Aviv, Israel Yochai Cohen Authors * Yair Neuman View author publications You can also search

for this author inPubMed Google Scholar * Yochai Cohen View author publications You can also search for this author inPubMed Google Scholar CONTRIBUTIONS Conceptualization, Y.N.;

methodology, Y.N., Y.C.; software, Y.C.; writing—original draft preparation, Y.N., Y.C.; writing—review and editing, Y.N.; All authors have read and agreed to the published version of the

manuscript. CORRESPONDING AUTHOR Correspondence to Yair Neuman. ETHICS DECLARATIONS COMPETING INTERESTS The authors declare no competing interests. ADDITIONAL INFORMATION PUBLISHER’S NOTE

Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations. RIGHTS AND PERMISSIONS OPEN ACCESS This article is licensed under a

Creative Commons Attribution 4.0 International License, which permits use, sharing, adaptation, distribution and reproduction in any medium or format, as long as you give appropriate credit

to the original author(s) and the source, provide a link to the Creative Commons licence, and indicate if changes were made. The images or other third party material in this article are

included in the article’s Creative Commons licence, unless indicated otherwise in a credit line to the material. If material is not included in the article’s Creative Commons licence and

your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder. To view a copy of this

licence, visit http://creativecommons.org/licenses/by/4.0/. Reprints and permissions ABOUT THIS ARTICLE CITE THIS ARTICLE Neuman, Y., Cohen, Y. A Data Set of Synthetic Utterances for

Computational Personality Analysis. _Sci Data_ 11, 623 (2024). https://doi.org/10.1038/s41597-024-03488-6 Download citation * Received: 10 January 2024 * Accepted: 06 June 2024 * Published:

13 June 2024 * DOI: https://doi.org/10.1038/s41597-024-03488-6 SHARE THIS ARTICLE Anyone you share the following link with will be able to read this content: Get shareable link Sorry, a

shareable link is not currently available for this article. Copy to clipboard Provided by the Springer Nature SharedIt content-sharing initiative