A total of 264 graduate and doctoral students were recruited to voluntarily participate in the study (195 women and 69 men). After informed consent, the subjects were first asked to freely write a few paragraphs in Chinese (at least 300 words) about their recent situation or feelings, similar to the way they naturally express themselves on social media such as Weibo. Moreover, they were required to fill out a questionnaire with a design based on the SWLS scale and including gender information.

All phases of the study were subject to ethical approval by the Institutional Review Board of the Institute of Psychology, Chinese Academy of Sciences with the ethical approval number H16003.

After word segmentation, the three sentiment word dictionaries were used on subjects' writing texts to calculate the word frequency of each emotion word category. Each subject's text was processed separately, and finally, 29 emotional features were extracted.

To examine the relationship between each emotion word category and life satisfaction, we first conducted analysis using the Pearson correlation coefficient. It was expected that there existed some redundant features that would weaken the prediction performance. Hence, features that did not contribute to making predictions were removed. In our study, we used the method of forward feature selection as the feature selection technique.

With the selected features, we started to train the regression model for predicting life satisfaction. Due to the small sample size used in this study, the support vector regression (SVR) method that performs well on small datasets was conducted. At the same time, we also used linear regression (LR), multi-layer perceptron regressor (MLP), random forest (RF), and decision tree (DT) to train the prediction model to compare the performances of different algorithms and attempt to select a better model. All models were trained through 5-fold cross-validation.

To test the validity of the predictive models, we calculated the Pearson correlation coefficients, which were the correlation between the predicted life satisfaction results of the model and the scores of the questionnaire. Higher correlation coefficients indicate better effect of the predictive model.

To test the reliability of the models, we adopted the method of split-half reliability testing in psychometrics. Concretely, we made use of another self-statement dataset, the textual posts of 50 Sina Weibo active users, because the split-half reliability test required more text data for each subject. First, the Weibo posts within a year of each Weibo user were spliced together to generate a single piece of text. To ensure the homogeneity of the split texts, we conducted the odd–even split-half method. For each piece of the text, all the sentences were numbered in order, and then, we obtained the odd half text containing the odd-numbered sentences and the even half text correspondingly. However, we conducted our predictive models on the odd half text and the even half text separately. Finally, the model-predicted life satisfaction scores from the odd and the even half text were compared through Pearson correlation analysis, with the correlation coefficient as the indicator of split-half reliability.

In our study, feature selection and model training were performed through scikit-learn, and other statistical analyses were conducted through SPSS 26.0.

To find out the relationship between emotion word categories and life satisfaction, we conducted Pearson correlation analysis on the word frequencies and the life satisfaction score, and the results are shown in Table 2.

As shown in Table 2, the positive emotion words were positively correlated with life satisfaction, and the negative emotion words were negatively correlated with life satisfaction. Among the positive words, the four categories of happy, respect, favorite, and glad were significantly correlated with life satisfaction with the highest correlation coefficient up to 0.24; among the negative emotions, the eight categories of anger, sorrow, fear, boredom, reproach, aversion, sadness, and N (negation + positive) were significantly correlated with life satisfaction with the highest correlation coefficient up to −0.26. Compared to the positive emotions, the negative emotions significantly correlated with life satisfaction with more word categories and relatively higher correlation coefficients.

After feature selection, the remaining features are shown in Table 3.

As shown in Table 3, the seven categories of happy, sorrow, boredom, reproach, glad, aversion, and N (negation + positive) were selected. Compared with positive emotion words, more negative emotion word categories were present in this list. It seemed that, in the emotion words lexicon reflecting life satisfaction, the negative words were richer than the positive words.

Appropriate parameter settings are essential for training to generate a well-performing machine learning model (32). After the tuning process of model parameters, the kernel function of SVR was set to “rbf,” the regularization coefficient was set to 1, the kernel coefficient was set to 100, the dimension of the polynomial kernel function was set to 2, and the other parameters used the default values.

After the 5-fold cross-validation, the values of the Pearson correlation coefficients between the predicted scores of life satisfaction and the scores of the SWLS questionnaire are shown in Table 4.

As shown in Table 4, the correlation coefficients between the predicted scores and the SWLS scores most achieved medium level (>0.30), while the values were varied among different algorithms. The SVR model had relatively higher validity, with the correlation coefficient between the predicted value and the true value reaching 0.42. The correlation coefficients between the predicted scores of the MLP model and the questionnaire score could also reach 0.40. For the other two algorithms RF and DT, the indexes of model validity were relatively lower. The correlation coefficients between the predicted scores of the DT model and the questionnaire score were only 0.27.

As the SVR model performed best in predicting life satisfaction, we further tested the split-half reliability of this model. With 50 Sina Weibo users' posts, we divided each user's posts into two halves, according to the parity of the chronological order. The trained SVR model as described earlier was applied to the two halves of textual data, achieving two sets of predicted life satisfaction scores, and then, the Pearson correlation coefficient was calculated as the index of split-half reliability. The test result of the SVR model was quite high (r = 0.939, p < 0.001), showing that the prediction scores calculated by this model were highly correlated on different self-statement texts of the same individual.

The results of modeling (Table 4) in this study showed that it is possible to recognize individuals' life satisfaction based on the emotion words used in their self-statement texts. In this study, we built the prediction models with SVR, and the correlation coefficient between the predicted and real-life satisfaction scores was 0.42. This validity was considered to be acceptable and meaningful in the field of personality and social psychology (33). When tested on social media texts, the model had also shown high split-half reliability. Specifically, the present study used only the emotion word frequencies as the extracted features for prediction models rather than any other linguistic or behavioral features. Compared with the intricate combinations of diverse features in previous modeling studies, our findings demonstrated the effect of the emotion elements in recognizing life satisfaction and provided a more concise and comprehensible predictive model.

The different performances of models in our results indicated which method was more suitable to build the life satisfaction prediction model based on the self-statement text. In terms of the validity of the model, the SVR-based models performed relatively better, and the correlation coefficient between predicted and real scores reached the top of our study. Compared with the other algorithms in this study, previous researchers have found that the SVR algorithm was relatively more stable and robust and performed well, especially in small datasets (34). It also reminded us that different machine learning algorithms have different characteristics; therefore, researchers should choose the appropriate machine learning algorithm according to their research tasks and the nature of the research data.

The results of this study (Table 2) indicated that the positive emotion words were positively correlated with life satisfaction, and the negative emotion words were negatively correlated with life satisfaction. Previous research found that individuals' mood can influence their life satisfaction (23, 24), that is, the emotion words in the self-statement texts in our study reflected the individuals' true emotional experience, and the valence of these emotion words reflected the direction that the life satisfaction changed.

According to the results of correlation analysis (Table 2), compared with positive emotion words, negative emotion words had a higher correlation coefficient with life satisfaction. From the results of the feature selection (Table 3), there were more negative emotion words than positive emotion words among the selected features. It seemed to indicate that in self-statement texts, negative emotion words conveyed more information about life satisfaction and made a greater contribution to the prediction of life satisfaction. The result was consistent with Baumeister's finding, which showed that negative events, experiences, relationships, and psychological states affected our feelings more than positive factors (35). Furthermore, compared with positive emotions, negative emotions were more differentiated, and could be described with more words (26). This asymmetry was also reflected in the dictionary used in this study, where the negative emotions contained more categories compared with the positive emotions.

In this study, the seven emotional features of happy (PA), sorrow (NB), boredom (NE), reproach (NN), glad (MH), aversion (ME), and N (negation + positive) were finally selected for modeling (Table 3). As shown by the correlation results (Table 2), these seven features were also those with the highest correlation coefficients with life satisfaction. These results suggested that these emotion categories played an important role in reflecting life satisfaction, which may bring meaningful inspiration to our understanding of the relationship between specific emotions and life satisfaction.

Our study built a predictive model with good validity and reliability which could identify individuals' life satisfaction only through the emotion words in their self-statement. While self-statements are quite common on today's social media and easy to obtain, this prediction model could be a choice of a measurement method to detect social media users' life satisfaction. Compared with previous studies, our limited emotion word features as the input of the model are much easier to extract and understand. It would be particularly suitable for a fast, large-scale, cost-effective investigation of life satisfaction in online scenarios, serving the public mental health and social wellbeing promotion. In addition, through the prediction study, we revealed the relationship between specific emotion word categories and life satisfaction and obtained new knowledge of the linguistic characteristics reflecting life satisfaction. The emotion categories provide specific clues for further exploring the factors and mechanisms in real life determining individuals' life satisfaction. The general importance of emotion expressions in our study also suggested that it may be possible to develop life satisfaction detection methods through other modalities of emotion expression, such as facial expressions.