A standardized personality lexicon for enhancing personalized human-machine interaction

Abstract

Personality, as a stable and coherent set of behavioral and cognitive patterns, significantly influences linguistic expression, emotional regulation, and cognitive functioning. The Big Five personality traits—neuroticism, extraversion, openness, agreeableness, and conscientiousness— are especially relevant for understanding to language use and social interaction, making them foundational for developing of personality-informed natural language processing (NLP) systems. Despite this, existing personality lexicons often lacks rigorous validation, show weak alignment with linguistic features and personality traits, and fail to adapt to dynamic language environments such as social media. This study presents the construction and empirical validation of a personality lexicon derived from established psychological scales, dictionaries, and literature. Validation using real-world participant data yielded high hit rates across all Big Five dimensions (all > 0.70; mean = 0.787) and their 30 corresponding facets (all > 0.60; mean = 0.768). This lexicon provides a robust foundation for advancing computational personality assessment and supports applications in personalized NLP, large language models, and mental health prediction.

Background & Summary

Personality is a relatively stable and coherent set of patterns of thoughts and behaviors shaped by the interaction of genetics and environment factors¹. Over the past century, it has been extensively studied, leading to the development of numerous theories and models. Among these, the Big Five Personality traits model, developed from the lexical hypothesis, has emerged as one of the most robust and widely accepted frameworks. This model has five core traits: neuroticism, extraversion, openness, agreeableness, and conscientiousness. These traits have been linked to linguistic abilities, emotional regulation, social behavior, and, notably, cognitive functions^1–5, particularly information processing and logical reasoning^6,7.

Therefore, understanding personality is critical to advancing automatic personality recognition technologies, and enhancing the accuracy of interpreting individual differences in language use, behavior, cognition, and emotion in natural language processing⁸. For example, embedding personality-based corpora into chatbot systems enables the dynamic analysis of users’ inputs—including their thoughts, emotions, and preferences—allowing the system to adapt its responses accordingly⁹. This development demonstrates the importance of a personality-informed model in intelligent human-robot interaction. Notably, the lexical hypothesis posits that personality differences are encoded in language¹⁰, establishing a theoretical bridge between linguistic patterns and trait psychology. However, even when individuals from diverse cultural backgrounds achieve similar scores on standardized personality tests, their personality traits may still exhibit significant variations in actual behavior and linguistic expression¹¹. Existing research indicates that the newly added dimension of the HEXACO model—Honesty-Humility—better captures personality traits relevant to specific cultural contexts, reflecting the cross-cultural diversity of personality theories¹². As the primary medium for personality expression, language reveals how identical traits are encoded and presented across cultures¹³. Therefore, it is necessary to construct a cross-cultural personality lexicon, particularly within the Chinese context. In this way, building personality lexicons with varying levels of granularity is valuable for enhancing the performance of large language models and advancing personalized human-computer interaction.

While prior studies confirm correlations between LIWC categories and personality traits¹⁴, their explanatory power remains limited (5.1–38.5% variance explained). Current Chinese lexicons like Wang et al.‘s dataset¹⁵ suffer from unvalidated annotations (F1 = 0.33 for facets), and social-media-adaptive tools like TextMind¹⁶ lack scalability. To address these gaps, this study: integrates multi-source seed words from standardized scales (e.g., IPIP-NEO), lexical databases (e.g., LIWC), and literature related to personality-related word; implements two-stage validation with real-world participant ratings; delivers a granularly annotated lexicon covering 5 dimensions and 30 facets.

In our research, we developed a reliable large, high-quality personality lexicon that covers both coarse- and fine-grained personality categories and can be used for comprehensive personality perception tasks. We also detailed a systematic construction and validation process, demonstrating its robustness across the five dimensions and 30 facets using real participant data.

This lexicon enables three novel applications: (1) Dynamic personality adaptation in chatbots using valence-facet mapping¹⁷; (2) Mental health screening via social media text mining¹⁸ (e.g., detecting high Neuroticism markers); (3) LLM personalization through trait-aligned prompt engineering¹⁹.

Methods

Seed word extraction

Scale-derived words: selected from validated instruments (e.g., NEO-PI-R, BFI-2) based on Cronbach’s α > 0.85. The seed words in this study primarily came from three areas: personality scales, word dictionaries, and research literature, all selected through a comprehensive literature review. For personality scales, relevant studies were identified through searches of PsycINFO, PubMed, Web of Science, Scopus, and JSTOR, using the following keywords: (“Personality Traits” OR “Big Five” OR “NEO” OR “Five Factor Model” OR “neuroticism” OR “conscientiousness” OR “openness” OR “extraversion” OR “agreeableness”) AND (“questionnaire” OR “personality inventory” OR “scale” OR “personality test”). Scales included in the final selection were well-established, widely used, and demonstrated validity and reliability.

Lexicon sources: for word dictionaries or lexicon, searches were conducted using the keywords: (“Personality Traits” OR “Big Five” OR “NEO” OR “Five Factor Model” OR “neuroticism” OR “conscientiousness” OR “openness” OR “extraversion” OR “agreeableness”) AND (“word dictionary” OR “word lexicon” OR “word library” OR “linguistic inquiry” OR “LIWC”). Only dictionaries that were highly cited and commonly used in personality research were included.

Literature sources: for research literature, search terms included “Big Five Personality Traits”, “Big Five”, “NEO”, “Five Factor Model”, “neuroticism”, “conscientiousness”, “openness”, “extraversion”, “agreeableness”, “behavior*“, “language*“, and “textual*“. Studies were screened based on the following inclusion criteria: (1) published within the past 5–10 years; (2) to ensure the quality and credibility of the research literature, English-language publications with at least 50 citations at the time when search was performed (2024/3-2024/6) or published in high-quality, influential, and peer-reviewed journals (a list of eligible journals is provided in Supplementary Materials 1); (3) Chinese-language studies published in core Chinese academic journals. Exclusion criteria were: (1) inaccessible the full text, such as databases that are not open to the public; (2) studies that did not report words or linguistic features related to personality traits. Full texts of the selected literature were reviewed by trained psychology researchers, and personality-related terms with statistically significant associations (P < 0.05) were extracted.

Seed word extraction results

Scales

Seed words were first extracted from widely used personality assessment instruments. These included the NEO series (Neuroticism-Extraversion-Openness Personality Inventory)²⁰ and the BFI (Big Five Inventory) series, which are standard tools for assessing the five major personality traits: Neuroticism, Extraversion, Openness, Agreeableness, and Conscientiousness.

The following scales were utilized in this study to extract relevant words:

1. IPIP-NEO-120: This is a shortened version of the NEO scale, containing 120 items, making it suitable for quick assessments. Each personality dimension is measured by 24 items, effectively capturing an individual’s personality traits^21,22.

2. IPIP-NEO-300: This version includes 300 items, providing a more detailed and comprehensive personality assessment. Each dimension is measured by 60 items, making it ideal for in-depth psychological research and personality analysis²³.

3. NEO-PI-R: The most comprehensive version of the NEO scale, with 240 items, evaluates the five major personality traits and their six sub-dimensions. This detailed breakdown aids in a deeper understanding of individual differences^24–27.

4. The BFI (Big Five Inventory) series also evaluates the five major personality dimensions but offers a more concise set of options compared to the NEO series. This study primarily used the BFI-44 scale, which contains 44 items, with approximately 9 items assessing each personality trait. It is the most widely used version of the BFI series²⁸.

5. FFMRF (Five-Factor Model Rating Form): Developed by Thomas A. Widiger and other psychologists in the late 1990s, this tool is based on the Five-Factor Model of personality. It uses simple adjectives or phrases to assess the five major personality traits. Due to its high reliability and validity, it has been widely applied across various cultures and languages, especially for quick screening and preliminary personality assessments in large-scale surveys or online research^29–31.

6. CBF-PI (California Brief Personality Inventory): Developed by a psychology research team from the University of California in the early 2000s, this inventory provides a concise and rapid assessment of the five major personality traits. It has demonstrated good internal consistency and construct validity in multiple studies, simplifying traditional long-form assessments and making personality evaluations more efficient³².

7. HEXACO: Created by Michael C. Ashton and Kibeom Lee in 2001, the HEXACO personality model offers a six-factor framework, which includes Honesty-Humility, Emotionality, Extraversion, Agreeableness, Conscientiousness, and Openness to Experience. This model extends the traditional Five-Factor Model by adding the Honesty-Humility dimension and has been validated in numerous cross-cultural studies, showing high reliability and validity^33–35.

8. Chinese Adjective Big Five Personality Inventory: Developed in the late 1990s by a team of Chinese psychologists, this inventory is tailored to the Chinese linguistic and cultural context. It uses a series of adjectives to assess the Big Five personality traits and has demonstrated good psychometric properties, making it suitable for personality research within the Chinese cultural setting³⁶.

9. Cattell’s 16 Personality Factors (16PF): Developed by psychologist Raymond Cattell, the 16PF scale assesses 16 distinct personality factors, offering a comprehensive yet intricate personality description. This tool is widely used in career planning and psychological evaluations^37,38.

10. Eysenck Personality Questionnaire (EPQ): Created by Hans Eysenck in 1970, the EPQ focuses on assessing three major dimensions: Neuroticism, Extraversion, and Psychoticism. It provides insight into how individuals interact with the external world and their emotional stability. The EPQ has been widely applied and validated globally, demonstrating high reliability and validity^39–41.

Representative seed words extracted from these instruments are listed in Table 1.

Table 1.

Examples of scale seed word extraction.

Source Scale	Words	Original Item Description
IPIP-NEO-120	Anxious	Worries about many things
IPIP-NEO-300	Friendly	Easily makes friends
BFI-44	Energetic	Is full of energy
FFMRF Five-Factor Word List	Angry, distressed	Angry hostility
CBF-PI	Persistent, determined, persevering	Once I set a goal, I work hard to achieve it
HEXACO	Uninterested in aesthetics, indifferent	I find visiting art galleries boring.
Chinese Adjective Big Five Personality Scale	Silent	Silent
Cattell 16 Personality Factors (16PF)	Boring, humorous	I am not good at telling jokes or funny stories
Eysenck Personality Questionnaire (EPQ)	Vulnerable	Are your feelings easily hurt?

Lexicon databases

Seed words were also drawn from two well-established lexical resources: the LIWC lexicon and the 2818 Personality Trait Descriptor Lexicon. The LIWC (Linguistic Inquiry and Word Count) lexicon, a widely recognized tool in psycholinguistic research, categorizes words into psychologically meaningful domains such as emotion, cognition, and social interaction^14,42,43. From this lexicon, words with strong personality relevance were identified with the input of professional psychologists. The 2818 Personality Trait Descriptor Lexicon includes adjectives commonly used to describe personality traits, each evaluated for familiarity, frequency, and relevance in prior psychological studies. Its structured, open-access format provides a robust foundation for personality-related lexical analysis⁴⁴. Examples from these databases are shown in Table 2.

Table 2.

Example of dictionary seed word extraction.

Words	Source Dictionary/Corpus	Dimensions
Unbelievable	LIWC	Function; Negative emotion; Affect
Unpopular	LIWC	Cognitive process; Discrepancy
Indecisive	LIWC	Affect; Negative emotion; Anxiety
Trust	LIWC	Affect; Positive emotion
Adventurous	LIWC	Affect; Positive emotion
Smart and Capable	Personality trait descriptor	/
Abnormal	Personality trait descriptor	/
Absent-minded	Personality trait descriptor	/
Restrained	Personality trait descriptor	/
Abstract	Personality trait descriptor	/

Research literature

Seed words were extracted from published research literature^14,45–60. Examples of these extracted seed words are shown in Table 3.

Table 3.

Example of seed word extraction from research literature.

Words	Source Literature
Gambling	Brunborg, G. S., Hanss, D., Mentzoni, R. A., Molde, H. & Pallesen, S. Problem gambling and the five-factor model of personality: a large population-based study. Addiction 111, 1428–1435, 10.1111/add.13388 (2016).
Depression	Kern, M. L., Eichstaedt, J. C., Schwartz, H. A., Dziurzynski, L., Ungar, L. H., Stillwell, D. J., Kosinski, M., Ramones, S. M. & Seligman, M. E. The online social self: an open vocabulary approach to personality. Assessment 21, 158–169, 10.1177/1073191113514104 (2014).
Perfectionism	Burcaş, S. & Creţu, R. Z. Perfectionism and neuroticism: Evidence for a common genetic and environmental etiology. J. Pers. 89, 819–830, 10.1111/jopy.12617 (2021).
Low Pressure	Kern, M. L., Eichstaedt, J. C., Schwartz, H. A., Dziurzynski, L., Ungar, L. H., Stillwell, D. J., Kosinski, M., Ramones, S. M. & Seligman, M. E. The online social self: an open vocabulary approach to personality. Assessment 21, 158–169, 10.1177/1073191113514104 (2014).
Contemplative	Chapman, B. P. & Goldberg, L. R. Act-frequency signatures of the Big Five. Pers. Individ. Differ. 116, 201–205, 10.1016/j.paid.2017.04.049 (2017).
Excellent	Eichstaedt, J. C., Kern, M. L., Yaden, D. B., Schwartz, H. A., Giorgi, S., Park, G. & Ungar, L. H. Closed- and open-vocabulary approaches to text analysis: a review, quantitative comparison, and recommendations. Psychol. Methods 26, 677–698, 10.1037/met0000349 (2021).
Health	Giorgi, S., Le Nguyen, K., Eichstaedt, J. C., Kern, M. L., Yaden, D. B., Kosinski, M., Seligman, M. E. P., Ungar, L. H., Schwartz, H. A. & Park, G. Regional personality assessment through social media language. J. Pers. 90, 405–425, 10.1111/jopy.12674 (2022).
Loneliness	Kern, M. L., Eichstaedt, J. C., Schwartz, H. A., Dziurzynski, L., Ungar, L. H., Stillwell, D. J., Kosinski, M., Ramones, S. M. & Seligman, M. E. The online social self: an open vocabulary approach to personality. Assessment 21, 158–169, 10.1177/1073191113514104 (2014).
Stifling	Kern, M. L., Eichstaedt, J. C., Schwartz, H. A., Dziurzynski, L., Ungar, L. H., Stillwell, D. J., Kosinski, M., Ramones, S. M. & Seligman, M. E. The online social self: an open vocabulary approach to personality. Assessment 21, 158–169, 10.1177/1073191113514104 (2014).
Prayer	Koutsoumpis, A., Oostrom, J. K., Holtrop, D., van Breda, W., Ghassemi, S. & de Vries, R. E. The kernel of truth in text-based personality assessment: A meta-analysis of the relations between the Big Five and the Linguistic Inquiry and Word Count (LIWC). Psychol. Bull. 148, 843–868, 10.1037/bul0000381 (2022).

Lexicon annotation

Seed words from Chinese sources were retained in their original form and English terms were translated according to the Oxford Tenth Edition Dictionary. Labeling was performed in two stages.

Trait mapping: Two psychological researchers independently mapped words to dimensions/facets using IPIP-NEO-120 facet definitions (Table S1 (see Supplementary Information document)); five dimensions, each comprising six facets)^61,62, as specified by the original scales or, for lexicon and literature sources, based on those reported in the corresponding studies.

Valence coding: each term was annotated for its emotional valence—classified as positive, negative, or neutral. Emotional valence was defined as the affective sentiment associated with a given word. A positive valence indicated that the word conveyed a favorable emotional sentiment, a negative valence reflected an unfavorable sentiment, and a neutral valence indicated the absence of strong emotional sentiment.

Lexicon annotation results

After consolidating the collected words and removing duplicates, a total of 6084 seed words were identified. Figure 1 shows a word cloud illustrating representative descriptors for each of the five Big Five personality dimensions.

Fig. 1.

Examples of personality word in the five dimensions.

The detailed counts for each dimension and facet are provided in Table 4. Examples of lexicon annotations are shown in Table 5.

Table 4.

Summary of counts of word annotation results.

Dimension	Dimensions Word Count	Facet	Facets Word Count
Neuroticism	1158	Anxiety	198
		Anger	232
		Depression	155
		Self-conscientiousness	254
		Immoderation	123
		Vulnerability	196
Extraversion	1674	Friendliness	463
		Gregariousness	198
		Assertiveness	311
		Activity Level	246
		Excitement-seeking	181
		Cheerfulness	274
Openness	1187	Imagination	129
		Artistic interests	108
		Emotionality	148
		Adventurousness	193
		Intellect	281
		Liberalism	329
Agreeableness	1764	Trust	117
		Morality	259
		Altruism	299
		Cooperation	437
		Modesty	216
		Sympathy	436
Conscientiousness	1332	Self-efficacy	164
		Orderliness	200
		Dutifulness	175
		Achievement striving	215
		Self-discipline	264
		Cautiousness	314

Table 5.

Example of seed word annotation.

Words	Source	Dimensions	Facets	Emotional valence
Creative	BFI-44	Openness	Ideas	Positive valence
Steady	FFMRF	Neuroticism	Anger and Hostility	Positive valence
Pessimistic	FFMRF	Neuroticism	Depression	Negative valence
Vulgar	HEXACO	Openness	Art	Negative valence
Meddlesome	LIWC	Conscientiousness	Self-discipline	Negative valence
Clumsy	EPQ	Agreeableness	Compliance	Negative valence

Data Curation

The data curation process consisted of two phases. The first phase involved a pilot study to test the functionality of the curation tool and to preliminarily assess the feasibility of using the Big Five Personality Questionnaire and the personality lexicon to evaluate annotation reliability. Based on the validation results from this pilot phase, formal data collection was then conducted.

Sample source

Participants were recruited using a convenience sampling approach through online advertisements, and the questionnaire was distributed via the Wenjuanxing platform (a Chinese online survey distribution tool). Inclusion criteria were: 1) adults aged 18 to 65 years; and 2) individuals who provided informed consent and were able to independently complete the questionnaire. Exclusion criteria were: 1) a self-reported history of a professionally diagnosed severe psychological or neurological disorder; and 2) individuals unable to comprehend or follow the study instructions; and 3) minors.

The study protocol was approved centrally by the Ethics Review Committee of West China Hospital of Sichuan University (process number 20232426). Informed consent was obtained from all participants prior to their participation in the study.

Data curation tools

Demographic Variables

Demographic variables included gender (male, female), age (18–24 years, 25–35 years, 36–50 years, or over 50 years), employment status (employed or unemployed), education level (junior high school, high school, vocational school, college, bachelor’s degree, or graduate degree).

Big five personality questionnaire

The Chinese version of the International Personality Item Pool–NEO 120-item (IPIP-NEO-120) questionnaire was used to assess participants’ Big Five personality traits, with Cronbach’s α coefficients indicated good internal consistency at both the domain and facet levels ($\alpha$ = 0.63–0.88)^61,63. Based on their scores, participants were categorized as “High,” “Average,” or “Low” for each personality dimension and facet. “High” for scores in the top 30th percentile, “Low” for scores in the bottom 30th percentile, and “Average” for scores in the middle 40% (31st–69th percentile)⁶¹.

Lexicon rating task

Two versions of the lexicon rating task were developed for the pilot and formal phases of the study. In the pilot phase, 241 seed words labeled across different personality dimensions were randomly selected from validated personality questionnaires with clearly defined traits. These words were divided into six sets of 40, and each participant was randomly assigned one set. Participants rated each word using a five-point Likert scale, where higher scores indicated greater personal relevance (0 = Very inaccurate to 4 = Very accurate). In the formal phase, 3,600 words without clearly defined dimensions or facets in existing dictionaries or literature were selected from the seed lexicon. Words from validated personality questionnaires were excluded. These words were divided into 60 sets of 60, based on response time data from the pilot phase. Each participant was randomly assigned one set and rated the words using the same five-point Likert scale.

Data curation results

Pilot study

A total of 50 participants were included in the pilot study, the majority of whom were female. Approximately half of the sample were young adults, with more than 50% having jobs and having completed undergraduate or higher education. Detailed demographic characteristics are shown in Table 6.

Table 6.

Characterization of the general demographic information in the pre-curation (n = 50).

Variable Name	Frequency	Percentage
Gender
Female	32	64%
Male	18	36%
Age Group
18–35 years	35	70%
36–50 years	4	8%
Over 50 years	11	22%
Employment Status
Employed	31	62%
Unemployed	19	38%
Educational Level
Junior High School	2	4%
High School	1	2%
Vocational School	4	8%
College	5	10%
Bachelor’s Degree	21	42%
Graduate Degree	17	34%

Most participants scored within the moderate range on the Big Five personality traits. However, Neuroticism showed greater variability, with more participants with either high or low levels of Neuroticism compared to the other dimensions. Distributions of personality traits are detailed in Table 7.

Table 7.

Big Five personality traits in the pre-curation (n = 50).

Big Five Personality Dimension level	Frequency	Percentage
Agreeableness	93.90 ± 11.40a
Low	1	2%
Medium	33	66%
High	16	32%
Extraversion	82.80 ± 11.50a
Low	2	4%
Medium	42	84%
High	6	12%
Openness	88.8 ± 11.15a
Low	3	6%
Medium	41	82%
High	6	12%
Conscientiousness	89.52 ± 12.53a
Low	5	10%
Medium	37	74%
High	8	16%
Neuroticism	65.70 ± 13.66a
Low	5	10%
Medium	40	80%
High	5	10%

α: mean (std).

Formal data curation

The formal data curation included a total of 329 participants, the majority of whom were female. Most participants were young adults, and a large portion of them were unemployed. Nearly all participants had completed undergraduate or higher education. Detailed demographic characteristics can be found in Table 8.

Table 8.

Characterization of the general demographic information in formal data curation (n = 329).

Variable Name	Frequency	Percentage
Gender
Female	226	68.69%
Male	103	31.31%
Age Group
18–24 years	215	65.35%
25–35 years	95	28.87%
36–50 years	12	3.65%
Over 50 years	7	2.13%
Employment Status
Employed	245	74.47%
Unemployed	84	25.53%
Educational Level
Vocational School	4	1.22%
Junior High School	8	2.43%
High School	8	2.43%
College	16	4.86%
Bachelor’s Degree	139	42.25%
Graduate Degree	154	46.81%

In terms of personality traits, the highest average score was observed in the Agreeableness dimension (91.86 ± 10.82), while the lowest average score was in the Neuroticism dimension (65.61 ± 16.38). The majority of participants fell within the moderate range for the Big Five personality traits. Trait Distribution details are provided in Table 9.

Table 9.

Big Five Traits (n = 329).

Dimension	Low(%)	Medium(%)	High(%)	M ± SD
Agreeableness	2.7	70.8	26.5	91.86 ± 10.82
Extraversion	2.6	79.9	17.5	85.33 ± 11.12
Openness	4.0	80.0	16.0	91.42 ± 9.72
Conscientiousness	7.0	75.0	18.0	90.10 ± 12.02
Neuroticism	16.4	67.6	16.0	65.61 ± 16.38

Data Records

Datasets. The full dataset, as well as all raw material used for this study is publicly available at: 10.6084/m9.figshare.29596547.v3⁶⁴.

The main dataset is named personality_lexicon_Chinese_version.xlsx (identical to personality_lexicon_English_version.xlsx). This file contains all annotated personality adjectives, including the source, dimension, facet, and emotional polarity of each personality adjective. The column “id” contains a consecutive number identifying each individual lexical word. The dimension and facet indicate to which dimension and facet of personality each adjective belongs. In addition, the source and emotional valence (positive, negative, neutral) of the adjectives are provided.

Other raw material includes the data collected from both the pilot study and the formal study. The pilot data are stored in pilot_study_Chinese_version.xlsx (identical to pilot_study_English_version.xlsx), and the formal data are stored in formal_data_revised_Chinese_version.xlsx. (identical to formal_data _revised_English_version.xlsx) Notably, the file formal_data_Chinese_version.xlsx contains two sheets: the first sheet includes the 241 personality adjectives used in the pilot study, and the second sheet includes the 3600 personality adjectives used in the formal study. In consideration of the differences between Chinese and English, we have provided two versions of the data, allowing future researchers to choose the language version that best meets their research needs.

Both files contain demographic information, Big five inventory (IPIP-NEO-120) data, and rating scores for each study participant. Within the demographic information, the column “id” contains a consecutive number identifying each individual study participant; other details are described in the Methods. The big five inventory data include responses to 120 items as well as scores across the five dimensions of personality. All remaining columns represent the rating scores assigned to each lexical word by participants. To facilitate participants’ understanding of each adjective, each lexical word was embedded into a complete sentence during the rating process.

Technical Validation

Validation process

Data from both the pilot and formal phases underwent the same two-step validation process.

Big five personality questionnaire (IPIP-NEO-120) validation

Reliability (the extent to which the questionnaire provides consistent or reliable measurements of a particular construct) was assessed using Cronbach’s alpha to measure internal consistency. Validity (the extent to which the questionnaire reflects the construct it is intended to measure) was evaluated using the Kaiser-Meyer-Olkin (KMO) measure of sampling adequacy and Bartlett’s test of sphericity. Following standard psychometric practice⁶⁵, these two metrics pertain to the validity and reliability assessment of the personality assessment instrument, aiming to confirm that trait scores derived from the IPIP-NEO-120 possess psychometric reliability, thereby serving as benchmark labels for subsequent dictionary construction and evaluation.

Lexicon validation

To evaluate the accuracy of personality-related word annotations, a two-level validation procedure was conducted at both the Big Five dimension and facet levels. Only words labeled with either positive or negative emotional valence were included; neutral terms were excluded from analysis. Participants were stratified into high/medium and low trait groups based on their scores on the IPIP-NEO-120 questionnaire. For each word, average ratings were computed separately within each trait group, requiring a minimum of three participant ratings per group to ensure reliability.

Validation was based on whether participants’ ratings aligned with their corresponding trait levels. For all dimensions and facets except Neuroticism, a word was considered a hit for high/medium trait participants if positively valenced words received high ratings (≥3) or negatively valenced words received low ratings (<3), excluding neutrals to enhance discriminative power. Conversely, for low trait participants, a hit occurred when positively valenced words received low ratings (<3) or negatively valenced words received high ratings (≥3). For example, if a participant had high agreeableness and also rated positively valenced words highly (≥3), it was considered a hit; conversely, if a participant had low agreeableness and rated negatively valenced words highly (≥3), it was considered a hit. For Neuroticism—conceptually associated with negative valence—this logic was reversed: participants with high/medium Neuroticism were expected to rate negatively valenced words highly (≥3) or positively valenced words low (<3), while those with low Neuroticism were expected to do the opposite. For example, if a participant had high neuroticism and also rated negatively valenced words highly (≥3), it was considered a hit; conversely, if a participant had low neuroticism and rated positively valenced words highly (≥3), it was considered a hit.

Each word thus contributed two classification instances: one from the high/medium group (yielding either a true positive [TP] or false negative [FN]) and one from the low group (yielding either a true negative [TN] or false positive [FP])⁶⁶. Based on previous studies^67,68, our study was more concerned with overall accuracy and therefore hit rate was used as an assessment metric. This allowed for the construction of confusion matrices and the calculation of hit rates for each dimension and facet. Hit rate calculation:

$$\mathrm{Hit\; rate}=\frac{\mathit{TN}+\mathit{TP}}{\mathrm{TP}+\mathrm{FN}+\mathrm{FP}+\mathrm{TN}}$$

*Where:

TP: High or medium-trait user rates positive word ≥3

TN: Low-trait user rates positive word <3

Reverse logic for Neuroticism*

The confusion matrix for word validation is shown in Table 10.

Table 10.

Confusion matrix for word validation.

		Big Five Personality Test Levels
		Dimension/Facet level is medium or high	Dimension/Facet level is low
Word Score	Positive valence word score ≥ 3 Negative valence word score < 3	True Positive (TP)	False Positive (FP)
	Positive valence word score < 3 Negative valence word score ≥ 3	False Negative (FN)	True Negative (TN)

The full process of constructing the personality lexicon is shown in Fig. 2.

Fig. 2.

Personality lexicon construction process.

Validation results

Pilot study

In the pilot study, IPIP-NEO-120 questionnaire demonstrated good psychometric properties, with high reliability (Cronbach’s alpha = 0.876) and satisfactory validity (KMO = 0.794), confirming its suitability as a gold-standard measure.

In the lexicon validation, all Big Five dimensions achieved hit rates above 0.7, indicating strong annotation accuracy. However, due to the limited sample size, rank-order consistency could not be reliably assessed for each dimension and facet. Detailed validation results are shown in Table 11.

Table 11.

Validation results of personality lexicon dimensions in the pre-curation (n = 50).

Questionnaire/Sub-questionnaire	Hit Rate
Overall	0.817
Openness	0.765
Conscientiousness	0.860
Extraversion	0.885
Agreeableness	0.718
Neuroticism	0.857

Formal data curation

In the formal data curation, IPIP-NEO-120 questionnaire consistently demonstrated good psychometric properties, with both reliability and validity scores exceeding 0.80 (Table 12) and the hit rates for each of the Big Five dimensions were all above 0.70, with an overall average hit rate of 0.787. At the facet level, all hit rates exceeded0.60, with an overall average hit rate of 0.768 (Table 13). These results indicate that the questionnaire reliably captured participants’ personality traits and that the lexicon annotations were accurate.

Table 12.

Reliability and validity of the Big Five Personality Inventory in formal data curation (n = 329).

Questionnaire/Sub-questionnaire	Reliability	Validity
Overall	0.820	0.843
Openness	0.794	0.775
Conscientiousness	0.883	0.881
Extraversion	0.825	0.839
Agreeableness	0.845	0.826
Neuroticism	0.923	0.911

Table 13.

Lexicon Validation Performance.

Dimension	Hit Rate	Strongest Facet (HR)	Weakest Facet (HR)
Agreeableness	0.823	Morality (0.874)	Trust (0.656)
Extraversion	0.776	Cheerfulness (0.860)	Gregariousness (0.700)
Openness	0.734	Emotionality (0.825)	Artistic interests (0.652)
Conscientiousness	0.770	Dutifulness (0.814)	Self-efficacy (0.660)
Neuroticism	0.832	Immoderation (0.881)	Vulnerability (0.711)
Overall	0.787	—	—

As shown in Table 14, we compared the Chinese personality lexicon proposed in this study with existing cross-lingual personality lexicon.

Table 14.

Comparison of the proposed Chinese personality lexicon with existing personality lexicon across different languages.

Resource name	Language(s)	Personality model	Number of entries	Granularity	Validation approach	Public
This work	Chinese; English	Big Five	6,084 words	facet-level, dimension-level, word-level	329 real-sample ratings	Yes
LIWC 201514	English; Chinese	/	～6,400 entries	word-level, dimension-level	expert ratings, partial real-sample ratings	Commercial
BigFive: a Chinese textual dataset15	Chinese	Big Five	13,478 phrases	facet-level, dimension-level, phrase level	3 expert ratings	No
Personality trait descriptor44	English	/	2,818 words	word-level	1572 real-sample ratings	Yes

Compared to existing resources, our personality lexicon is the first publicly available, facet-level Big Five personality lexicon for Chinese—grounded in psychometric theory and validated through large-scale empirical data. In contrast, most existing large-scale personality lexicons across languages are limited to simple word lists, often lacking fine-grained facet-level categorization and robust empirical validation.

Usage Notes

Researchers who want to conduct further analyses with the Personality Lexicon can download the full dataset “personality_lexicon.xlsx”, which is described in Data Records and is available at 10.6084/m9.figshare.29596547.v3⁶⁴. In addition to the processed dataset, we also provide all raw datasets.

Author contributions

T.J., J.J. and W.Z. conceptualized the study and originated the dataset creation. T.J., X.K., Y.Y., H.Z., X.H., J.J. and Y.L. participated in the collection, annotation, and validation of the dataset. The manuscript was drafted by T.J. and edited by X.S., H.C. and Y.L. All authors have read and approved the final manuscript.

Data availability

The data supporting the findings of this study are openly available in figshare at 10.6084/m9.figshare.29596547.v3⁶⁴.

Code availability

No code was generated or used during the study.

Competing interests

The authors declare no competing interests.

Supplementary information

The online version contains supplementary material available at 10.1038/s41597-026-06783-6.