Numerous studies have demonstrated that sleep plays a crucial role in physical recovery, memory consolidation, and emotional regulation (1). Scholars consider sleep to be a multifaceted phenomenon, proposing various sleep-related characteristics such as sleep quality (2), sleep duration (2), sleep problems (3) and sleep disorders (e.g., insomnia) (4). When these sleep issues arise, they are associated with an increased risk of developing various physical health conditions, including diabetes (5), obesity (6), cardiovascular diseases (7), and Alzheimer’s disease (8), as well as a heightened risk of mortality (9). In addition, sleep disorders are strongly linked to mental health problems, such as the experience of excessive negative emotions and insufficient positive emotions (10). To better understand sleep, researchers have developed various tools to assess different aspects of sleep, which can generally be categorized into objective and subjective measures. Objective sleep characteristics are typically measured using laboratory equipment or specialized wearable devices, to record multiple physiological signals during sleep (e.g., electroencephalography and eye movements). These measures are considered the standard clinical procedure for diagnosing sleep-related disorders. Subjective sleep characteristics are commonly assessed through self-reported questionnaires [e.g., the Pittsburgh Sleep Quality Index (PSQI) (11)] and sleep diaries.

Personality is considered one of the key factors related to sleep (4, 12). According to Akram et al. (4), the inherent nature of personality serves as both a predisposing and a potential maintaining factor for insomnia. A closer examination of personality’s susceptibility and its long-term effects can provide a better understanding of the etiology of insomnia (12). For example, Grandner (13) noted that elevated arousal levels may explain the connection between neuroticism and sleep quality, as neuroticism is associated with stronger physiological responses to stress, which can delay an individual’s return to a calm state after experiencing acute stress. To better understand personality, scholars have proposed various models from different perspectives. These include Cloninger’s psychobiological model (14), Cattell’s 16Factor Personality model (16 PF) (15), Eysenck’s personality model (16), the Alternative Model of Personality Disorders (AMPD) (17), and the Five-Factor Model (FFM) (18). Among the commonly used models, Cloninger’s psychobiological model emphasizes the integration of genetic and developmental factors and is widely applied in mental disorder research (14); the DSM-5’s Alternative Model of Personality Disorders (AMPD) focuses on traits such as negative affectivity, detachment, antagonism, disinhibition, and psychoticism, which are typically used to assess maladaptive personality traits and disorders (17). While all these models provide comprehensive analyses of personality, the FFM, also known as the Big-5 (19), stands out due to its demonstrated moderate-to-high longitudinal stability, reliability, and cross-cultural applicability (2). The FFM includes five broad traits: extraversion, agreeableness, conscientiousness, neuroticism, and openness to new experiences (18, 20).

Previous studies have shown that personality traits negatively affect various sleep characteristics, particularly sleep quality and chronotype (2, 3). In research based on the Five-Factor Model (FFM), neuroticism has consistently been identified as a stable negative predictor of sleep quality (2, 10, 21–25). Individuals with high neuroticism are more likely to hold metacognitive beliefs about sleep difficulties (24), experience more negative and fewer positive emotions (25, 26), and tend to be more hyperaroused, all of which adversely affect sleep quality (10). Beyond FFM-based research, the Cloninger model also shows a correlation between traits like harm avoidance and sleep disorders such as insomnia (27). Zakiei et al. (28) emphasized that not only traditional personality traits but also pathological traits (e.g., psychoticism and negative affectivity) are stable predictors of various sleep problems. Akram et al. (4) also found that insomnia is linked to negative or maladaptive personality traits, including neuroticism, perfectionism, worry, social inhibition, and avoidance. Hisler et al. (29) suggested that trait stress, rather than neuroticism, modulates the relationship between sleep and self-control. However, several recent studies have reported findings that challenge these widely accepted relationships, particularly the link between neuroticism and sleep quality (25, 30). Saksvik-Lehouillier et al. (25) found that individuals with high neuroticism experience fewer negative emotions under partial sleep deprivation than during normal sleep, whereas those with medium and low levels of neuroticism reported the opposite pattern. A further study of college students using Actigraph to objectively measure sleep found no relationship between neuroticism and sleep quality (30).

Several studies suggest that conscientiousness and extraversion are positively correlated with sleep quality (2, 21, 23), as individuals higher in these traits tend to be both psychologically and physiologically healthier and experience fewer negative emotions in stressful situations (26). However, some studies have contradicted these findings, suggesting that the relationship between personality traits and sleep quality may not be consistent across different contexts (22, 30). For example, Križan and Hisler (22) found no significant relationship between extraversion and objectively measured sleep quality, and Mead et al. (30) reported that extraversion may even negatively impact sleep quality.

Although research generally shows that agreeableness and openness are unrelated to sleep quality (21, 23, 30), Cellini et al. (10) found that when cognitive reappraisal, inhibition, emotions, and hyperarousal were included in the regression model, agreeableness was the only personality trait that predicted sleep quality. Spears et al. (9) also suggested that agreeableness may indirectly predict mortality risk by affecting daytime impairments and sleep. Furthermore, Leger et al. (26) found evidence suggesting that openness might be associated with better sleep quality, as individuals with high openness tend to experience fewer negative emotions and may, therefore, have better sleep quality.

Due to the limitations of traditional methods in large-scale data collection, real-time analysis, and cost efficiency, researchers have increasingly turned to social media in recent years to assess mental health indicators, leveraging its natural user expressions and traceable characteristics as complementary tools (31–35). These studies focused on three characteristics: behavioral features [e.g., number of posts, timing of posts, frequency of likes (31, 36), and comments on other posts (37)]; multimedia features [e.g., videos (38) and images (39)]; and text features. For text features, three semantic representation methods were mainly used to understand the content: discrete representations [e.g., term frequency-inverse document frequency (40)], closed vocabulary methods [e.g., Linguistic Inquiry and Word Count (41) and self-defined dictionaries (42)], and open vocabulary methods [e.g., Word2Vec (43) and BERT (44)]. Among these, open vocabulary methods using deep learning-based word embeddings provide the most effective representation of social media text (45).

Regarding personality, most research on personality assessment has focused on English text (46), with few studies examining the Chinese context using distributed representation methods for data training models. For example, Mahajan et al. (47) found that online-revealed personalities align with users’ true personalities. Cutler and Condon (48) conducted factor analysis on word embeddings of English text using BERT and compared the results with earlier lexical studies. They found that agreeableness, extraversion, and conscientiousness traits were well-replicated, but neuroticism and openness were not. Considerable research attention has focused on examining the relationship between sleep characteristics and their associated variables on social media platforms such as Twitter, Weibo, and Reddit (49). For example, Liu et al. (50) examined sleep-related user attributes like region and education levels, and Yao et al. (51) investigated how sleep quality as a symptom accompanies other mental health issues within a depression community, identifying co-occurrences with fear, negative expectations, and suicidal intentions. Relatively few studies have evaluated sleep problems using natural language processing techniques. For instance, Tian et al. (52) employed a support vector machine algorithm to detect sleep-related complaints in social media posts and identify key topics associated with insomnia.

Previous studies have established connections between personality traits and sleep characteristics, but inconsistent methodologies have produced mixed results (4). Two persistent issues complicate research in this area. First, data availability remains limited. Most existing datasets that include both sleep and personality measures are small and not publicly accessible, which restricts the generalizability of findings. Second, research approaches are often disconnected. Although social media platforms offer rich natural language data, most studies focus exclusively on either personality assessment or insomnia detection, with little attention to the interaction between the two (53).

This study aims to investigate the relationship between personality traits and both model-assessed sleep problems and self-reported sleep quality. Specifically, we focus on four main objectives:

As shown in Figure 1 , we followed a systematic process for project design, data collection, and implementation. (1) First, we collected microblogs from Sina Weibo, China’s largest social media platform, and constructed two datasets. The first dataset was a large-scale collection, consisting of 4,860,000 posts from 15,251 users. The second dataset combined microblogs with survey responses from 923 Sina Weibo users (336 active users). For data collection and management, we used PyCharm 2021.3.1 (Community Edition, JetBrains) to process the microblog text. (2) We then built two sleep assessment models: Model 1 was designed to determine whether a post was sleep-related, and Model 2 assessed whether a post indicated a sleep problem. For personality assessment, we applied BERT-based word embeddings in combination with long short-term memory (LSTM) (55) regression models incorporating an attention mechanism (Model 3). (3) To analyze the relationship between personality traits (FFM) and sleep characteristics, we used both self-reported data and model-generated assessments. As part of an exploratory analysis, we applied the three semantic models, which independently assess personality traits and sleep problems, to the large-scale microblog dataset to further investigate their relationship in this context.

We developed two sleep assessment models. The first model assessed each post on “whether it is related to sleep” (Model 1). For posts that are assessed as being “related to sleep”, the second model evaluated “whether there are sleep problems” (Model 2). Models 1 and 2 shared a similar structure, and both utilize a BERT fine-tuning approach with a downstream fully connected neural network to build a classifier. To ensure computational efficiency, the text was first segmented into sentences. Due to BERT’s input limitation of 512 characteristics, sentences exceeding this length were truncated. After segmentation, each sentence was first passed through a pre-trained BERT Chinese model to convert each word into word embeddings, which were subsequently input into a Transformer Encoder (Trm). The Trm established semantic relationships between words according to the context. Then the output from the Trm was then fed into four types of classification models: a linear fully connected layer (BERT + Linear), a convolutional neural network (BERT + CNN), a recurrent neural network (BERT + LSTM), and a recurrent neural network with an attention mechanism (BERT + LSTM + Attention). These models were chosen to assess different types of contextual and sequential relationships within the text, allowing us to test various architectures for better performance.

To further evaluate the performance and robustness of our models, we also converted each sentence into word embeddings using a pre-trained ERNIE model [Enhanced Representation through Knowledge Integration; (56)], which was then input into a linear classifier (ERNIE + Linear). This additional model served as a comparative baseline, allowing us to assess the impact of using a different pre-trained language model for the task. The results from these models were compared to determine the most effective architecture for assessing sleep-related content and sleep problems in microblog posts, providing insights into which approach best captured the nuances of sleep-related language.

We split the labeled dataset for Model 1 (N = 2,454) and the labeled dataset for Model 2 (N = 1,123) into training, validation, and testing sets using a ratio of 8:1:1. Specifically, 80% of the data was used for training the model, 10% for model validation, and the remaining 10% for testing. This approach ensured that the models were trained on a sufficient amount of data while retaining a separate dataset for testing performance and tuning. To evaluate the performance of the models, we used Precision (P), Recall (R), and F1 score as key metrics. These metrics were chosen because they offer a balanced view of model performance. In our evaluation, the F1 score was prioritized because it combines both Precision and Recall into a single metric, offering a more comprehensive understanding of the model’s performance.

The personality assessment model based on microblog text comprised five components: the BERT embedding layer, sentence fusion layer, LSTM layer, attention layer, and regression layer ( Figure 2 ). This model evaluated personality as a trait variable, assuming it remains stable over time. Therefore, we aggregated multiple posts from each user to form a sequence of text data, which was crucial for capturing the consistency of the user’s personality expression over time. Initially, word embeddings for each post were generated using the Chinese BERT pre-trained model. Since BERT had an input limitation of 512 tokens, posts longer than this threshold were truncated, and a “split then merge” approach was applied to ensure that the word embeddings accurately represented longer posts. In the sentence fusion layer, the word embeddings from the most recent 100 posts were selected for subsequent training. For users with fewer than 100 posts, zero vectors were used for padding. This resulted in each user’s word embeddings being represented as a 100 × 768 matrix. The sequence of word embedding matrices for each user was then input into the LSTM layer. LSTM, a specialized type of recurrent neural network, was particularly suited for handling sequence data. Here, each user’s posts were treated as a sequence, allowing the LSTM to capture the temporal dynamics and semantic relationships across the multiple posts of each user. By processing these sequences, the model could learn how a user’s personality traits were reflected and expressed over time in the context of their social media posts.

To enhance the model’s focus on relevant personality-related features, we incorporated an attention mechanism. This mechanism allocated attention to different parts of the sentence vectors, enabling the model to better capture key personality-related content from each post, thereby improving classification accuracy. Finally, a linear regression model was applied to the output of the attention layer to generate the personality scores.

We trained five separate regression models for the Big Five personality traits, using normalized selfreported CBF-PI scores as labeled inputs. Normalizing personality scores improved training efficiency, reduced the impact of data sparsity and outliers, and led to a more stable model. In addition to the primary model, we compared the BERT-based model’s performance with simpler models such as a linear fully connected layer (BERT + Linear) and a recurrent neural network (BERT + LSTM). The effectiveness of these models was evaluated by comparing the predicted personality scores with the labeled values using Pearson’s correlation coefficient and RMSE. This comparison allowed us to assess the relative performance of different models in capturing personality traits from user-generated microblog content.

The results of the sleep assessment models (i.e., Models 1 and 2) and comparisons with alternate models are shown in Table 1 . When determining whether a post is related to sleep (Model 1), the best performance was achieved when using only the BERT fine-tuning approach (accuracy = 95.09%, precision = 97.27%, F1 score = 0.95). When determining whether a sleep-related post indicated a sleep problem (Model 2), the best performance was achieved when using only the BERT fine-tuning approach, which had an accuracy of 91.30%, a precision of 91.65%, and an F1 score of 0.91.

It is noteworthy that the assessment models for sleep-related and sleep problem posts, which used only BERT or ERNIE fine-tuning followed by a fully connected neural network output, performed better than those that incorporated additional deep learning algorithms. This indicates that, for our tasks, fine-tuning pre-trained models is sufficient to achieve good classification results. In fact, the addition of more complex models tended to degrade performance. This aligns with the findings of Mohammadi and Chapon (57), who demonstrated that when fine-tuning BERT pre-trained models, fully connected neural networks perform better than more complex classifiers for text classification tasks. Simpler classifiers can maximize the use of BERT’s text representation capabilities, resulting in better performance, whereas complex classifiers are more prone to issues like overfitting, which can degrade performance.

We applied the sleep assessment models (Model 1 and Model 2) to a total of 73,735 posts from 336 valid participants. Among these, 2,709 posts were identified as related to sleep, and 1,578 of them were further identified as expressing sleep problems. For each participant, we calculated: (1) the total number of posts (TN, i.e., posting frequency); (2) the number of sleep-related posts (NSR); (3) the number of posts indicating sleep problems (NSP); (4) the proportion of sleep-related posts (PSR); and (5) the proportion of posts with sleep problems (PSP). Among the 336 users, 192 posted sleep-related content (with a maximum of 168 such posts and a highest proportion of 53.30%), and 172 users posted content expressing sleep problems (with a maximum of 81 such posts and a highest proportion of 53.33%).

We initially used self-reported sleep quality (i.e., PSQI scores) to evaluate the validity of sleep characteristics assessed by Model 1 and Model 2. However, PSQI scores were not significantly correlated with any of the four model-assessed sleep characteristics. To further explore potential influencing factors, we examined the correlation between the number of posts indicating sleep problems (NSP) and the total number of posts (TN). A strong positive correlation was observed (Pearson’s r = 0.64, p <.001), indicating possible collinearity. To address this, we conducted a moderation analysis with PSQI score as the independent variable, the number of posts indicating sleep problems (NSP) as the dependent variable, and the total number of posts (TN) as a moderating variable. Age and gender were included as covariates. The results showed that poorer self-reported sleep quality (i.e., higher PSQI scores) significantly predicted a greater number of sleep problem posts (NSP) (β = 0.68, p = .015). In addition, the interaction between sleep quality and the total number of posts (TN) was significant (β = 0.12, p = .006), suggesting a moderating effect of user activity. Simple slope analysis ( Figure 3 ) indicated that when the posting frequency was low, the PSQI score did not significantly predict the number of sleep problem posts (NSP) (β = 0.06, p = .817). However, at a high level of posting frequency, higher PSQI score significantly predicted more posts expressing sleep problems (β = 1.49, p <.001). A similar pattern was observed for the number of sleep-related posts (NSR). From these findings, we conclude that:

Conclusion 1: Self-reported sleep quality significantly predicts model-assessed sleep problems, and this relationship is moderated by users’ total posting frequency.

The predictive performance of the personality (Big Five) models (Model 3) is shown in Table 2 . When the input consisted of sentence vectors extracted using BERT features, the LSTM regression model based on the attention mechanism performed the best (average RMSE = 0.186), with predicted values for the test set significantly positively correlated with the questionnaire scores. Using the linear regression model or the LSTM model separately resulted in poor performance, with low, non-significant correlation coefficients between the predicted and questionnaire scores and higher RMSE. It is noteworthy that the performance of the personality assessment model (Model 3) varied across the five traits. Openness (r = 0.50, p = .003) and conscientiousness (r = 0.51, p = .003) performed well, with correlations around 0.5, whereas extraversion (r = 0.27, p = .021), agreeableness (r = 0.33, p = .013), and neuroticism (r = 0.24, p = .047) showed poorer performance. The self-reported questionnaires indicated generally high scores, with uneven training data for agreeableness. Results for agreeableness and extraversion may also have been affected by social desirability factors, such as acceptance, pro-social behavior, and sociability (58). Furthermore, agreeableness, which is a highly evaluative trait, may lead to inaccuracies in self-assessment and, in turn, inconsistent results (59).

We examined the correlations between self-reported sleep quality (i.e., PSQI score), the number of sleep-related posts (NSR), the number of posts indicating a sleep problem (NSP), and the proportion of posts with sleep problems (PSP) and personality traits (measured by CBF-PI scores) ( Supplementary Table S1 in Supplemental Materials ). The results showed significant associations between all five personality traits (openness, conscientiousness, extraversion, agreeableness, and neuroticism) and PSQI scores. Then we used these four model-assessed sleep characteristics as dependent variables and other variables (i.e., personality traits, gender and age) that showed significant correlations with these characteristics as independent variables in the regression model. Results in Table 3 show that agreeableness significantly predicted better self-reported sleep quality (β = 0.14, p <.01), while neuroticism significantly predicted poorer self-reported sleep quality (β = 0.4, p <.001). Therefore, we concluded that:

Conclusion 2: Openness, conscientiousness, extraversion, and agreeableness were significantly associated with better self-reported sleep quality, while neuroticism was linked to poorer self-reported sleep quality in the correlation analysis. Agreeableness and neuroticism remained significant predictors in the multivariate regression.

We applied the sleep assessment models (i.e., Models 1 and 2) to the dataset comprising 73,735 posts from 336 surveyed users. We then explored the relationships between sleep characteristics and personality traits at both the post and user levels to provide a comprehensive view.

At the post level, “whether a post is related to sleep (SR)” and “whether a post indicates a sleep problem (SP)” were used as independent variables, and self-reported personality traits (CBF-PI scores) were used as dependent variables in independent samples t-tests. Due to the small proportion of sleep-related posts for all personality traits except agreeableness, we applied corrections for unequal variances using IBM SPSS Statistics for Mac (Version 26.0.0.2). The results are shown in Figure 4 . We found two key results: (1) Posts related to sleep (SR) had significantly lower scores in openness (t = 20.88, p <.001), conscientiousness (t = 19.23, p <.001), extraversion (t = 17.16, p <.001), and agreeableness (t = 11.67, p <.001), and had significantly higher scores in neuroticism (t = 17.07, p <.001); (2) Compared to posts without sleep problems, posts expressing sleep problems (SP) had significantly lower scores in openness (t = 17.52, p <.001), conscientiousness (t = 13.99, p <.001), extraversion (t = 12.47, p <.001), and agreeableness (t = 7.70, p <.001), and significantly higher scores in neuroticism (t = −12.93, p <.001). Therefore, we drew the following conclusion:

Conclusion 3: Openness, conscientiousness, extraversion and agreeableness measured by CBF-PI scores are significantly associated with fewer model-assessed sleep problems, whereas neuroticism is significantly was associated with more model-assessed sleep problems at the post level.

At the user level, we first examined the correlations between personality traits (measured by CBF-PI scores) and the four model-assessed sleep characteristics (the number of sleep-related posts (NSR), the proportion of sleep-related posts (PSR), the number of posts indicating a sleep problem (NSP) and the proportion of posts with sleep problems (PSP)) ( Supplementary Table S1 in Supplemental Materials). The number of sleep-related posts (NSR) was significantly negatively correlated with openness (r = −0.12, p = .031), conscientiousness (r = −0.16, p = .006), and extraversion (r = −0.11, p = .046). The number of posts indicating a sleep problem (NSP) was significantly negatively correlated with openness (r = −0.14, p = .013) and conscientiousness (r = −0.16, p = .003). The proportion of posts with sleep problems (PSP) was significantly negatively correlated with openness (r = −0.14, p = .013; Supplementary Table S1 ). Further regression analysis revealed that conscientiousness was a significant negative predictor of the number of sleep-related posts (NSR) (β = −0.13, p = .032).

Further, we explored the relationship between model-assessed sleep characteristics and model-assessed personality traits from a big data-driven perspective. We randomly selected 15,251 users who had posted sleep-related content and crawled all of their original posts (N = 4,864,600) made between January 2020 to January 2023 using the user IDs. We then selected users who had posted more than five original posts (N = 13,753; 11,035 women, 2,627 men, and 91 with undisclosed gender), resulting in a total of 3,960,000 posts. We applied the sleep assessment models (i.e., Models 1 and 2) and the personality assessment model (Model 3). Results showed that 12,623 users had written sleep-related posts, which totaling 447,600 posts (11.3% of all posts), and 12,274 users had written about sleep problems, which totaled 332,600 posts (8.4% of all posts). Based on the analysis results from the surveyed dataset, we used two model-assessed sleep characteristics: the number of sleep-related posts (NSR) and the number of posts indicating a sleep problem (NSP), and conducted a correlation analysis between these sleep characteristics and personality traits. As shown in Table 4 , the number of sleep-related posts (NSR) was significantly correlated with lower scores in conscientiousness (r =−0.07, p <.001), extraversion (r =−0.16, p <.001), and agreeableness (r =−0.14, p <.001), and significantly correlated with higher scores in neuroticism (r =0.23, p <.001). For the number of posts indicating a sleep problem (NSP), the results were similar to those for the number of sleep-related posts (NSR), with the addition of a significant positive correlation with openness (r = 0.02, p = .027).

We analyzed the model-assessed personality scores in the large dataset and found two key issues: (1) The standard deviation (SD) of model-assessed personality trait scores was lower than that of self-reported scores in the surveyed dataset (N = 336; Supplementary Tables S2 , S3 ). Since the model’s training goal was to minimize the loss function, predictions tended to cluster around the mean value. (2) There was collinearity among the model-assessed personality traits, likely due to the use of the same dataset for training the models of the five dimensions. Despite these issues, the model-assessed personality traits was significantly correlated with the self-reported personality traits ( Table 2 ), indicating that the relative magnitude of each user’s personality scores reflected their position within the entire group.

To address these two issues, we ranked and grouped the model-assessed personality traits (N = 13,753) and selected the top and bottom 27% of the data to form low-score and high-score groups, respectively, for each dimension. The 27% selection was chosen to ensure a clear distinction between the low-score and high-score groups, capturing significant extremes while minimizing overlap and potential collinearity from the central range. Figure 5 shows that: (1) users in the high-score group for conscientiousness (t = 14.96, p <.001), extraversion (t = 23.36, p <.001), and agreeableness (t = 19.65, p <.001) had significantly fewer sleep-related posts than those in the low-score group for these traits, while users in the high neuroticism group had significantly more sleep-related posts than those in the low neuroticism group (t = −21.02, p <.001); (2) users in the high-score group for openness (t =−2.18, p = .029) and neuroticism (t =−17.54, p <.001) had significantly more posts indicating a sleep problem than those in the low-score group for these traits, while users in the high-score group for conscientiousness (t = 11.42, p <.001), extraversion (t = 22.68, p <.001), and agreeableness (t = 21.58, p <.001) had significantly fewer posts indicating a sleep problem than those in the low-score group for these traits. Thus, we concluded that:

Conclusion 4: From a big data-driven perspective, higher model-assessed sleep characteristics were significantly associated with lower scores in model-assessed conscientiousness, extraversion, and agreeableness, and with higher scores in neuroticism. These findings largely align with the results from self-reported data.