A psychological education model integrating artificial intelligence-based BERT in physical education

Abstract

To address the limitations of traditional physical education (PE) approaches that rely heavily on manual observation for psychological assessment, this study integrates Natural Language Processing techniques with sports psychology education. A psychological education model incorporating the pre-trained Bidirectional Encoder Representations from Transformers (BERT) model was developed to enable intelligent and precise evaluation of students’ psychological states. First, a text sentiment analysis model was constructed to extract emotional features. The BERT model was then fine-tuned for sports-specific contexts to capture deep semantic features of the text. Together, these components formed a four-dimensional feature extraction framework—semantic, emotional, thematic, and psychological. Experiments were conducted on two publicly available datasets, Emotional First Aid Dataset 1 and Psy-Insight 2. The proposed BERT-LDA-Psych model achieved average accuracy, precision, recall, and F1 scores of 91.5%, 91.0%, 90.3%, and 90.6%, respectively, on sentiment analysis tasks. It also outperformed mainstream ensemble models in psychological state recognition and feature fusion efficiency. The proposed model exhibits stable performance on general psychological counseling text corpora, establishing a reference computational framework for subsequent psychological state recognition tasks within PE contexts. The current results indicate functional feasibility in controlled textual environments; however, transferability to authentic PE classroom discourse remains uncertain.

Introduction

In contemporary education, the coordinated development of physical and mental health has become a central goal. Physical education (PE) is no longer limited to physical training and the transmission of motor skills; it also plays a crucial role in cultivating students’ psychological well-being. During PE learning, students experience various emotional responses and may face psychological challenges such as low self-confidence or poor teamwork. These issues directly affect their learning outcomes and can hinder their holistic physical and mental development^1–3. In response, educational authorities in China have emphasized the deep integration of physical and psychological education, with policy documents highlighting the need to improve social psychological service systems. However, in practice, integrating PE with psychological education remains challenging. Traditional PE has long focused on physical performance standards and movement accuracy, often overlooking opportunities for psychological guidance embedded in physical activities. Many PE teachers lack formal training in psychology, making it difficult to accurately detect students’ emotional changes or provide personalized interventions tailored to individual psychological profiles^4,5. The model designed in this study takes into full account the practical requirements of frontline teaching. Instead of producing complex feature vectors or technical algorithmic parameters, it presents results in a visualized and interpretable format, including the classification of students’ psychological states, dominant emotional tendencies, and corresponding simplified intervention suggestions. This design enables PE teachers to quickly understand students’ psychological conditions in PE classes without requiring specialized knowledge in psychology. Existing psychological education programs often follow rigid, standardized formats, relying on manual observation or paper-based questionnaires. These approaches are inefficient, prone to subjective bias, and unable to identify students’ latent psychological risks in a timely manner, thus failing to meet the growing demand for personalized psychological education in modern PE⁶. Psychological assessment in traditional PE predominantly depends on manual observation conducted by instructors, which introduces constraints in representational granularity. Such a mechanism complicates the quantitative characterization of student emotional fluctuations and limits the precise detection and intervention of psychological states specific to sports-related contexts. Within PE classrooms, students often experience reduced self-confidence associated with delayed acquisition of motor skills and suboptimal athletic performance. In parallel, cooperative learning tasks and team-based competitions may trigger interpersonal coordination barriers and communication conflicts, while endurance training and physical challenges frequently induce anxiety responses and avoidance tendencies toward difficulty-oriented tasks. These manifestations exhibit contextual structures that differ markedly from psychological distress patterns observed in conventional academic learning environments. Conventional analytical paradigms remain insufficient in capturing real-time psychological dynamics during physical activity, and personalized intervention strategies tailored to sports-specific contexts are similarly underdeveloped. This gap produces a structural disconnect between psychological education methodologies and PE practice scenarios. To mitigate these limitations, the proposed model is designed to address representative psychological issues in PE settings, including diminished self-confidence, impaired teamwork interaction, and sports-related anxiety. The framework enables computational recognition of emotional states and psychological conditions at the textual representation level, thereby compensating for the lack of context-sensitive assessment in traditional evaluation approaches.

With the rapid advancement of artificial intelligence (AI), its applications in education have expanded, providing new approaches and technical support for integrating PE with psychological education. AI can efficiently process large-scale educational data and uncover hidden insights, overcoming the limitations of traditional teaching and enabling precise, individualized psychological interventions⁷. In this context, integrating AI into psychological education within PE and constructing scientifically grounded, efficient models has become a critical path to addressing current challenges and promoting high-quality development in PE.

Natural Language Processing (NLP), a key branch of AI, enables deep analysis and understanding of textual data. In PE, students’ verbal expressions, written reflections, and learning feedback provide rich text-based evidence of their psychological states^8,9. This study selects four key technologies and leverages their respective strengths to construct a multidimensional framework for extracting and analyzing psychological features, thereby providing technical support for developing a PE-oriented psychological education model. The Bidirectional Encoder Representations from Transformers (BERT) pre-trained model is particularly advantageous because it captures bidirectional contextual information, allowing precise extraction of deep semantic features from text. Unlike traditional unidirectional language models, BERT encodes semantic meaning by considering both preceding and following context, similar to how humans interpret a sentence in light of its surrounding text^10,11.

This study aims to develop a psychological education model that integrates AI-based BERT into PE. Specifically, it first applies the BERT pre-trained model to extract deep semantic features from students’ textual feedback. TextMind is used to extract psychological features, while a labeled Latent Dirichlet Allocation (LDA) topic model identifies thematic features. Sentiment analysis captures emotional characteristics, forming a multidimensional framework for psychological feature analysis. Based on the practical PE context, the study establishes psychological assessment criteria and personalized intervention strategies, culminating in a comprehensive psychological education model for PE.

Literature review

Psychological assessment tools encompass a wide range of methods, devices, and systems used to evaluate individual psychological traits and behaviors. These tools play a critical role in clinical psychology, educational assessment, human resource management, and career development. Early studies on sentiment analysis primarily focused on detecting and classifying specific emotional types in sentences or documents. Zhang et al.¹² enhanced emotional semantics by mapping common emoji vectors and word embeddings into the same sentiment space, and then employed a multi-channel convolutional neural networks (CNNs) for sentiment analysis. Kodati & Tene¹³ proposed the Transformer architecture, which replaced the recurrent structures of RNNs and the convolutional operations of CNNs with self-attention as the fundamental mechanism. Compared with RNNs, Transformers overcame sequence dependency limitations, and compared with CNNs, they captured long-range semantic relationships more effectively. Yan et al.¹⁴ developed a maturity assessment model for data management to improve sentiment analysis on Weibo. Their approach dynamically encoded text at both feature and word levels using WOBERT Plus and ALBERT, extracted local key features via convolution, integrated cross-channel and multi-head self-attention pooling to capture global semantic information, and applied multi-granularity feature interaction to fuse word-level and feature-level semantic representations effectively.

Current sentiment analysis methods mainly focus on fine-tuning BERT to construct sentiment classifiers, with limited exploration of cross-lingual information. Recent studies have addressed this gap. Guo & Li¹⁵ noted that BERT captured contextual semantics via a bidirectional Transformer encoder. Pre-training tasks, including masked language modeling and next sentence prediction, generate high-quality word embeddings. For sentiment analysis, fine-tuning requires adding a classifier to BERT’s output layer and optimizing parameters using cross-entropy loss. Feng et al.¹⁶ combined BERT fine-tuning with a multi-layer collaborative CNN. Fine-tuning BERT captured fundamental word embeddings while preserving bidirectional semantic encoding, and the CNN extracted hierarchical emotional features, supplementing domain-specific knowledge for deep feature fusion and accurate classification. To address limited textual data in social sciences and the performance constraints of traditional models, Laurer et al.¹⁷ proposed the BERT-NLI model. BERT was first pre-fine-tuned on natural language inference tasks to accumulate task-related knowledge and then further fine-tuned on political text sentiment analysis, reducing dependency on large annotated datasets.

While existing research has made progress in technical applications and psychological feature analysis, current models inadequately address the dynamic nature of PE, making it difficult to track real-time changes in students’ psychological states. This study integrates AI technologies with sports psychology education, enriching the intersection of PE and psychological research and expanding the application of BERT models in this domain.

Research methodology

Text sentiment analysis model

Text sentiment analysis forms the core foundation of the proposed psychological education model. Its primary goal is to accurately identify emotional tendencies in students’ textual feedback during PE, quantitatively extracting positive, negative, and neutral sentiment features to provide essential data for subsequent psychological state evaluation. First, collected PE text data were preprocessed through cleaning, tokenization, and noise reduction to generate standardized text sequences. These sequences were then mapped to computational vector representations via an embedding layer, and feature enhancement strategies were applied to improve the discriminative power of emotional features. Finally, the enhanced features were fed into a classifier, which produced probability outputs for each sentiment category through activation functions, completing the sentiment analysis task¹⁸. The core workflow is illustrated in Eq. (1):

1

In this workflow, T represents the original PE text data; denotes the text preprocessing function; is the text embedding function that transforms text sequences into vector representations; refers to the feature enhancement function, which strengthens the discriminability of emotional features; represents the sentiment classification function that outputs the sentiment category; and Emotion denotes the final sentiment analysis result.

Dynamic text embedding is performed in two stages. The first stage involves training word embeddings for PE-specific vocabulary using the Word2Vec model to construct a domain-specific word vector dictionary. The Word2Vec model employs the Skip-gram architecture, as illustrated in Fig. 1. In the Skip-gram model, the current word is used as input to predict words within a defined context window. During training, the model learns the semantic relationships between each word and its surrounding context words, encoding these relationships into the word vectors^19,20.

Fig. 1.

Skip-gram architecture.

The second stage introduces context-aware embeddings. Leveraging the Transformer encoder structure, this stage captures semantic variations of words in different contextual environments, addressing the limitations of static embeddings generated by Word2Vec. The core of context-aware embedding lies in the self-attention mechanism, which computes the association weights between each word and all other words in the text, dynamically adjusting the word vector representations. The self-attention computation is formulated as follows:

2

3

4

X denotes the base word vector matrix output by Word2Vec; , , and represent the weight matrices for the query, key, and value vectors, respectively; is the dimensionality of the query vectors; represents the self-attention output; denotes the layer normalization operation; and is the final dynamic text embedding vector.

To enhance the discriminability of emotional features, this study introduces an attention mechanism combined with a feature concatenation strategy to construct an emotion feature enhancement module. First, an emotion attention mechanism is applied to compute the association weight between each word and the sentiment categories, thereby reinforcing the representation of emotion-related words and suppressing irrelevant words²¹. The calculation of emotion attention weights is formulated as follows:

5

denotes the emotion attention weight matrix; is the transpose of the emotion attention weight matrix; represents the bias term; and is the activation function used to map the weight values to the range [0,1].

Next, the dynamic embedding vectors are multiplied by the emotion attention weights to obtain emotion-enhanced feature vectors. These vectors are then concatenated with statistical text features to form the final emotion feature vector. The feature concatenation is expressed as follows:

6

where represents the final emotion feature vector; denotes the feature concatenation operation; and is the statistical feature vector of the text.

In the preprocessing stage, this study introduces a domain-adaptive noise filtering mechanism. The mechanism integrates a specialized terminology lexicon for sports education with an affective dictionary to intelligently identify and correct spelling errors, non-standard abbreviations, and irrelevant symbols in the raw text, thereby improving overall data quality. In addition, a dynamic sliding-window approach is employed for tokenization. This method accommodates flexible lexical combinations across different contextual settings and establishes a solid foundation for subsequent sentiment feature extraction.

Given the constraint that large-scale publicly annotated corpora remain scarce in PE scenarios, a two-stage learning strategy is adopted. The initial stage performs base model training on general psychological counseling datasets (EFA-1 and PI-2), establishing foundational semantic representations, while the subsequent stage transfers the learned parameters into the PE domain through text-level domain adaptation mechanisms in sports-related contexts. Within the model construction process, domain-specific adaptation is explicitly introduced. A specialized lexicon for the PE field is first constructed, covering high-frequency semantic units such as motor skill acquisition, physical training processes, team-based sports activities, and classroom instructional feedback. During preprocessing, domain noise filtering is applied alongside standardization of terminology to reduce lexical variability and improve consistency across textual inputs. Following this, the BERT model is fine-tuned under sports-oriented contextual constraints, where PE classroom corpora are incorporated into adaptive training to strengthen semantic encoding of sport-related emotional and psychological expressions. In parallel, within the Labeled Latent Dirichlet Allocation (LDA) topic modeling stage, domain-specific psychological categories such as sports confidence, teamwork interaction, and performance-related anxiety are introduced to align topic distributions with instructional realities in PE environments. Finally, given the colloquial structure and fragmented expression patterns commonly observed in sports-related textual data, syntactic parsing strategies and feature fusion mechanisms are reconfigured. This adjustment enhances structural robustness and improves the model’s capacity to generalize toward authentic PE classroom discourse.

BERT pretrained model and fine-tuning

The BERT model employs a Transformer encoder as its core architecture, consisting of multiple stacked Transformer encoder layers, as illustrated in Fig. 2. In its pretraining phase, BERT leverages two tasks: the Masked Language Model (MLM) and Next Sentence Prediction (NSP), which capture semantic representations at the word and sentence levels, respectively^22–24. Since the core architecture of BERT is composed of stacked Transformers, and Transformers encode text using an attention mechanism, the model can assign different weights to each word based on its relationships with other words and sentences in the text.

Fig. 2.

BERT model architecture.

The BERT model outputs a contextualized representation for each token, which can be used in downstream tasks such as sentiment analysis. For sentiment analysis, the representation of the [CLS] token is typically used as the overall sentence embedding:

Output = Transformer(Input) 7

The [CLS] embedding can then be passed through a fully connected layer and a softmax layer for classification:

8

Probabilities = softmax(Logits) 9

where W and b denote the weights and bias of the fully connected layer, respectively.

During the Masked Language Model (MLM) stage, BERT randomly selects a subset of tokens from the input and replaces them with the special [MASK] token²⁵. The model’s objective is to predict these masked tokens. This process allows BERT to learn contextual relationships between words, improving its understanding and generation of text. In the Next Sentence Prediction (NSP) task, BERT receives two sentences as input and predicts whether the second sentence follows the first in the original text. This task enables the model to capture sentence-level relationships and better understand the semantics and context of text sequences.

During the fine-tuning stage, a layer-wise learning rate strategy is adopted, in which different learning rates are assigned to different layers of the BERT model. Specifically, layers closer to the input are assigned lower learning rates to preserve the stable capture of fundamental linguistic features. In contrast, layers closer to the output are assigned higher learning rates to accelerate adaptation to sports education–specific tasks. This strategy balances generalization capability and task specificity, thereby improving fine-tuning efficiency and effectiveness. In addition, adversarial training is introduced to enhance model robustness. Small perturbations are added to the input data during training to improve resistance to noise and variation.

Given that the available training corpus consists primarily of general psychological counseling texts rather than domain-specific PE data, domain adaptation in this study is not realized through direct retraining on sports corpora. Instead, a knowledge-infused adaptation paradigm is introduced, where external PE domain knowledge is injected into the semantic representation process. Within this design, a domain dictionary comprising approximately 2,500 PE-specific lexical units is constructed and utilized as an extension to the BERT embedding vocabulary. The construction protocol of this lexicon is summarized in Table 1. Annotation and validation are independently performed by two PE professors and one computational linguistics expert, ensuring cross-disciplinary semantic consistency. Inter-annotator reliability reaches 0.87 under the Cohen’s Kappa metric, indicating substantial agreement across labeling decisions. After annotation completion, newly incorporated domain terms are aligned with the pre-existing semantic embedding space through the WordPiece tokenizer of BERT. For compound expressions that cannot be directly mapped to single lexical units, a subword decomposition strategy is applied to preserve semantic compositionality while maintaining compatibility with the original embedding structure.

Table 1.

Construction process of the sports terminology dictionary.

Term category	Quantity	Examples	Sources
Sports event names	380	Vaulting horse, three-step layup, relay race, forward roll	Compulsory Education Physical Education and Health Curriculum Standards
Skill evaluation vocabulary	620	Stiff movements, poor sense of rhythm, uncoordinated exertion, incorrect body posture	PE academic papers and textbooks
Classroom organization language	450	Rotation drills, group competitions, protection and assistance, free activity time	Primary and secondary school PE classroom observation records
Physical sensation descriptions	520	Weak legs, short of breath, lack of wrist strength, unstable center of gravity	Collection of students’ post-class PE reflection texts
Sports emotional expression patterns	530	"Afraid to perform in front of everyone", "Always feeling like dragging the team down"	PE teacher interviews and psychological counseling cases

Sports emotional expression patterns refer to the linguistic configurations through which students articulate psychological states in PE scenarios. Through cross-domain analysis between PE knowledge structures and psychological counseling corpora, a set of core expression patterns is formalized, as summarized in Table 2. Within general psychological counseling texts, these patterns appear as heterogeneous variants with surface-level lexical differences but shared underlying structural regularities. The proposed model performs domain transfer by capturing these abstract syntactic and semantic structures, rather than relying on direct memorization of domain-specific lexical content. This mechanism enables adaptation from counseling discourse to sports-related contexts through structural pattern alignment at the representation level.

Table 2.

Sports emotional expression patterns.

Pattern ID	Pattern name	Structural template
S1	Skill failure-Self-attribution	"Practiced [skill] for [duration] but [failure outcome], feeling [self-evaluation]"
S2	Social comparison-Avoidance motivation	"[Peer] is more [advantage dimension] than me, and I do not want to [participate behavior]"
S3	Physical sensation-Anxiety coupling	"Every time I [action], my [body part] undergoes [physiological response], feeling [emotion]"
S4	Public evaluation-Competence exposure anxiety	"Do not want to [action] in [occasion], fearing [reaction of others]"
S5	Team role-Diffusion of responsibility	"During [team task], I [role behavior], feeling [team experience]"
S6	Authority interaction-Learned avoidance	"After [authority figure] applied [correction method], I [behavioral change]"

Each input vector in BERT is a sum of three embeddings: (1) Token embeddings: Represent the word or subword itself^26,27. BERT uses WordPiece tokenization, which splits words into a limited set of subword units to balance vocabulary coverage and flexibility²⁸. (2) Position embeddings: Encode the position of each token in the sequence. (3) Segment embeddings: Distinguish between two sentences, which is especially useful in tasks involving sentence pairs, such as question-answering.

Multi-dimensional psychological feature extraction

The 19-category labeling scheme adopted in this study is derived from general psychological counseling classification standards rather than being designed specifically for PE teaching contexts. Within this taxonomy, a subset of categories exhibits direct relevance to sports scenarios, including academic pressure, interpersonal relationship dynamics, self-esteem levels, stress-related emotions, adolescent psychological characteristics, and teamwork-related behaviors. Other categories, such as divorce-related issues, sexual orientation, and substance abuse, are retained to preserve the structural completeness of the original labeling system. However, these categories are excluded or masked during downstream analysis focused on sports psychological interpretation. The introduction of this general-purpose labeling framework serves as an extensible scaffolding for psychological classification tasks. It provides a transitional representation space that supports subsequent refinement, wherein domain-specific categories tailored to PE contexts—such as sports confidence, competitive anxiety, team coordination, and athlete burnout—can replace or refine the existing label structure in future iterations.

In traditional LDA, topics are often generated randomly, which may result in irrelevant or ambiguous topics. Labeled LDA, by integrating topic information with label information, can better capture label-relevant topic features. This enables the model to generate topics more closely aligned with labels, improving interpretability. Labeled LDA extends classic LDA by incorporating supervised learning concepts. In the context of psychological counseling, labeled LDA can extract topic features from consultation dialogues by leveraging datasets annotated with counseling question labels, providing a more accurate reflection of clients’ concerns. Table 3 lists example consultation question categories²⁹. Based on this standard, the 19 categories of labels employed in this study are selected, merged, and reconstructed by three psychological counselors in conjunction with PE teaching scenarios. The specific mapping relationships are defined as follows: (1) academic pressure is mapped to PE learning pressure; (2) interpersonal relationships are mapped to peer interaction in the PE classroom; (3) self-esteem is mapped to athletic self-efficacy; (4) stress is mapped to competitive stress or training stress; (5) adolescent issues are mapped to adolescent sports participation psychology. The remaining general categories, such as divorce and substance abuse, are excluded from the psychological modeling for PE scenarios.

Table 3.

Types of counseling questions.

No	Type	No	Type	No	Type
1	Academic concerns	8	Interpersonal relationships	15	Obsessive–compulsive disorder
2	Work-related stress	9	Romantic relationship issues	16	Sexual orientation
3	Family conflicts	10	Divorce	17	Sexual issues
4	Substance abuse	11	Breakups	18	Parent–child relationships
5	Grief	12	Self-exploration	19	Other
6	Insomnia	13	Low self-esteem	–	–
7	Stress	14	Adolescence issues	–	–

The psychological counseling category labels in Table 3 were manually predefined based on established professional standards. Specifically, three licensed psychological counselors with more than ten years of clinical experience were invited to jointly construct a 19-category framework in accordance with the Psychological Counseling Classification Standard (Third Edition) and the contextual characteristics of PE settings. These categories include key dimensions such as academic stress, interpersonal relationships, self-perception, and sports performance anxiety. In the LDA-based topic modeling process, these predefined labels were used as supervisory signals to guide topic extraction. A Labeled LDA approach was adopted, in which each document was associated with relevant labels, and topic distributions were learned by maximizing the joint probability of documents and labels. This semi-supervised strategy ensured that the extracted topics remained consistent with the professional taxonomy of psychological counseling while capturing domain-specific psychological patterns in PE contexts. The labeling process was further validated through expert consensus meetings. The inter-rater agreement, measured using Cohen’s Kappa coefficient, reached 0.85, indicating a high level of consistency among experts.

Essentially, the LDA model defines a joint probability distribution over observed and latent variables. It infers latent variables by computing conditional probability distributions across the observed data. The general framework of LDA is illustrated in Fig. 3. In this framework, the hyperparameter α\alphaα controls the sparsity of topic distribution within documents, i.e., how many topics a document contains, while β\betaβ controls the sparsity of word distribution within topics, i.e., how many unique words a topic contains.

Fig. 3.

LDA model framework.

This study employed a labeled LDA topic model to extract topic features from psychological counseling dialogue texts. For a given counseling text and its corresponding consultation label , a topic distribution is selected from the prior over topic distributions. Here, is a KKK-dimensional vector representing the proportion of text XXX across KKK topics. The selection of is influenced by the label , ensuring that documents with the same label have similar topic distributions. Formally, this can be expressed as:

10

where α is the Dirichlet hyperparameter controlling the sparsity of topic distribution within a document, and represents the conditional distribution of topics given the label, capturing the influence of on the document’s topic allocation.

In the context of user psychological crisis recognition from counseling texts, semantic features, topic features, psychological features, and sentiment features are first concatenated and then fed into a fully connected layer. This layer allows for further transformation to extract more expressive combined features:

11

The output vector HHH from the fully connected layer is then passed through a softmax layer to classify the user’s psychological state. The model is optimized using the cross-entropy loss function to enhance prediction accuracy and stability:

12

13

where and denote the weights and bias of the softmax layer, is the true label of sample i for class c, is the predicted probability of class c, and C is the total number of classes.

Syntactic parsing was first performed to extract dependency relations such as the core predicate, subject, object, negation markers, and modification structures. A graph convolutional network (GCN) was then applied to encode the syntactic dependency graph, producing a syntactic structural feature vector S. This representation effectively captures complex sentence patterns, including negation, contrast, and modification, thereby reducing misinterpretation of sentiment in structurally complex expressions.

A hierarchical fusion strategy was subsequently designed to integrate the four-dimensional feature space with syntactic information. First, the semantic feature vector extracted by BERT was fused with the syntactic feature vector S through a weighted attention mechanism, which adaptively assigns weights to different components and strengthens the interaction between semantic content and syntactic structure:

14

Next, the fused semantic–syntactic representation was sequentially concatenated with the sentiment feature , topic feature , and psychological feature , followed by a nonlinear transformation, replacing the conventional simple concatenation approach:

15

Finally, the integrated feature representation was fed into a fully connected layer to generate the final classification feature vector H:

16

This fusion strategy embeds syntactic dependency structures into a multi-dimensional feature framework, effectively addressing misclassification issues caused by negation and modification in complex sentence constructions. It further enhances the model’s ability to accurately interpret colloquial and fragmented text commonly found in PE teaching scenarios.

For semantic feature extraction, in addition to using the fine-tuned BERT model, this study introduces syntactic parsing techniques to extract more fine-grained semantic relationships. By analyzing sentence structure, syntactic parsing enables the model to capture dependency relations among sentence components, thereby improving its understanding of core meanings and affective tendencies. For topic feature extraction, the parameter settings of the Labeled LDA model are further optimized. The hyperparameters α and β are adjusted to better control topic sparsity and word distribution sparsity. A topic relevance evaluation mechanism is also introduced to periodically assess the quality and coherence of generated topics. Irrelevant or low-quality topics are removed in a timely manner to ensure that the extracted topic features remain closely aligned with the requirements of psychological assessment in sports education.

For psychological feature extraction, the application scope of the TextMind tool is further expanded. In addition to basic quantitative analysis of psychological traits, domain-specific knowledge in sports education is incorporated to define customized indicators related to learning motivation, self-confidence, and teamwork ability.

To systematically validate the contribution of each component in the four-dimensional feature fusion framework, a hierarchical feature ablation strategy was designed. The model was incrementally built upon the base BERT semantic representation by progressively introducing sentiment features, LDA-based topic features, and psychological features, resulting in four experimental configurations: Configuration A (BERT semantic features only), Configuration B (BERT + sentiment features), Configuration C (BERT + sentiment + LDA topic features), and Configuration D (full four-dimensional model). This design allows for the observation of the marginal contribution of each additional feature dimension, thereby providing empirical evidence for the necessity of multi-dimensional feature integration. Feature fusion was implemented using a gated attention mechanism, which dynamically adjusts the weights of different feature dimensions. This enables the model to adaptively emphasize the most relevant features according to the characteristics of the input text.

Model architecture overview

In this study, multiple NLP components are integrated to construct the BERT-LDA-Psych model. The structural relationships among the model components and the corresponding data flow are detailed in Table 4. The BERT encoder and the syntactic GCN operate as parallel feature extraction pathways. Specifically, BERT is responsible for contextual semantic encoding, whereas the GCN executes syntactic structure encoding, and the outputs from both pathways are combined at the gated fusion layer. This design prevents gradient vanishing while preserving text representations at diverse granularities. The first 10 layers of the BERT model are frozen during the initial phase, whereas the final 2 layers participated in the fine-tuning process. The GCN, gated network, and emotional attention mechanism are trained from scratch. In contrast, LDA and TextMind served as pre-trained tools and did not participate in the end-to-end backpropagation.

Table 4.

Structural components and data flow of the BERT-LDA-Psych model.

Module level	Component name	Input dimension	Output dimension	Relationship with other modules
Input layer	Raw text input	Variable-length text sequence	Raw string	→ Preprocessing Layer
Preprocessing layer	Domain-adaptive noise filtering + Dynamic sliding-window tokenization	Cleaned text sequence	Standardized token sequence	→ Embedding Layer
Embedding layer	Word2Vec Skip-gram + BERT Token/Position/Segment embeddings	Vocabulary size V × embedding dimension	Sequence length L × 768	→ Feature Extraction Layer (Word2Vec → BERT, BERT → GCN)
Feature extraction layer	BERT + Syntactic GCN	d = 768 (BERT-base) | Dependency relation dimension = 30	L × 768 (BERT) + L × 300 (GCN)	BERT → Emotional enhancement | GCN → Gated fusion
Feature enhancement layer	Emotional attention mechanism + Statistical feature concatenation + Adversarial training perturbation	768 + statistical feature dimension | Perturbation dimension = 76	L × (768 + 50) = L × 818	→ Gated Fusion Layer
Feature fusion layer	Gated attention fusion + LDA topic model + TextMind psychological features	4 × 768 | Number of topics K = 1 | Number of psychological indicators =	1 × 512	Receives four-channel features → Outputs fused representation → Classification Layer
Classification output layer	Fully connected layer + Softmax classifier	Fused feature dimension = 512	Number of categories C = 3 (Emotion) or C = 9 (Psychology)	Receives fused features → Outputs probability distribution

Experimental design and performance evaluation

Experimental materials

Experiments were conducted on two publicly available datasets: Emotional First Aid Dataset 1 (EFA-1) and Psy-Insight 2 (PI-2). EFA-1 was collected by mental health institutions in collaboration with university psychology laboratories and covers common psychological scenarios such as anxiety, depression, and stress. Only dialogue texts with at least three to six conversation turns were selected, and each user utterance was annotated with sentiment labels. PI-2 contains 431 Chinese psychological counseling dialogues, each comprising at least five turns, with sentiment labels assigned to every user utterance. Both datasets were split into training, validation, and test sets, as summarized in Table 5.

Table 5.

Experimental dataset statistics.

Dataset	EFA-1		PI-2
	#Samples	#Turns	#Samples	#Turns
Training	7462	28,345	298	2030
Test	1060	4504	45	294
Validation	2138	9034	88	571

The EFA-1 and PI-2 datasets used in this study originate from psychological counseling scenarios. Although they provide high-quality annotations of psychological states and textual features, they differ from real-world PE classroom environments. In particular, these datasets lack domain-specific contextual characteristics of PE, such as real-time feedback during motor skill acquisition, interactive expressions in team-based activities, and descriptions of physical sensations unique to sports participation. Therefore, the experimental results primarily reflect the model’s baseline performance in psychological state recognition rather than its direct applicability in PE settings.

To partially alleviate the data mismatch across scenarios, a small validation set simulating PE scenarios, designated as PE-Sim 100, is constructed in this study. Dialogue samples highly relevant to PE contexts are screened from the EFA-1 and PI-2 datasets, and are subsequently rewritten contextually by three psychological counselors with backgrounds in PE. The rewriting principles require replacing general psychological scenarios with PE classroom contexts, and transforming abstract emotional descriptions into expressions of sports experiences, while preserving the original emotional labels. Consequently, a total of 100 simulated PE scenario texts are obtained to evaluate the performance degradation of the proposed model under a context that closely approximated authentic PE teaching. Although this validation set does not consist of genuine PE classroom data, a comparison with the original datasets allows for a preliminary assessment of the robustness of the model regarding scenario migration.

To prevent data leakage, this study adopts a dialogue-level segmentation strategy instead of turn-level segmentation. Specifically, each student’s complete dialogue history is treated as an independent unit and randomly split into training, validation, and test sets in a 7:1.5:1.5 ratio. This ensures that no dialogue turns from the same student appear simultaneously in both the training and test sets, thereby improving the reliability of the evaluation results. The choice of dialogue-level segmentation is motivated by the continuity and individual specificity of psychological states. Different turns within the same student’s dialogue are often inherently correlated. Although this strategy slightly reduces the amount of training data, it more realistically reflects the model’s generalization ability to unseen students. To further examine the impact of the segmentation strategy, an additional comparative experiment using turn-level splitting is conducted. The results show that turn-level segmentation yields approximately 1.3% higher test performance, indicating a potential overestimation caused by data leakage. Therefore, all results reported in the main text are based on the more rigorous dialogue-level segmentation strategy.

Experimental environment

All experiments were conducted on a Windows 10 platform with a “software + hardware” co-optimized environment to ensure efficiency and stability for training deep learning models such as BERT fine-tuning and LDA topic modeling. This setup avoided computational bottlenecks and excessive training time. Hardware and software configurations are listed in Table 6. Mixed-precision training was employed to accelerate convergence and reduce GPU memory usage. The software environment was based on Python.

Table 6.

Experimental hardware and software configuration.

Category	Configuration
Hardware	CPU: Intel Core i9-12900 K
	GPU: NVIDIA RTX 3090
	Memory: 64 GB DDR5
	Storage: 1 TB SSD + 4 TB HDD
Software	Programming language: Python 3.9
	Deep learning framework: PyTorch 1.12.0
	Text processing libraries: jieba 0.42.1, TextMind 2.0
	Data analysis libraries: Pandas 1.5.3, NumPy 1.24.3
	Visualization libraries: Matplotlib 3.7.1, Seaborn 0.12.2

Parameters setting

The proposed BERT-LDA-Psych model was trained using the Adam optimizer with a learning rate of 0.001, a decay coefficient of 0.9, and a weight decay of 1e-5 to prevent overfitting. The batch size was set to 32, with 50 training epochs and an early stopping strategy. Training stopped if validation accuracy did not improve over five consecutive epochs, and the best model parameters were saved. A Dropout layer with a probability of 0.3 was introduced to randomly deactivate neurons, enhancing model generalization.

Performance evaluation

Figures 4, 5, 6 present the performance of the proposed model and several mainstream baselines on the EFA-1 and PI-2 datasets, along with average metrics. The BERT-LDA-Psych model outperformed all comparison models in both datasets. Specifically, it achieved an average accuracy of 91.5%, precision of 91.0%, recall of 90.3%, and F1 score of 90.6%, representing improvements of 4.6%, 4.6%, 4.7%, and 4.6% over the strongest baseline, the fine-tuned BERT model. While classical deep learning models can extract sequential text features, they lack bidirectional contextual semantic understanding, limiting their performance. The standard BERT fine-tuning model, without adaptation to the characteristics of sports teaching texts, showed insufficient domain-specific applicability. In contrast, the proposed model incorporates sports-specific BERT fine-tuning, labeled LDA topic features, and TextMind-derived psychological features, effectively enhancing sentiment feature discriminability and classification accuracy.

Fig. 4.

Performance comparison of models on EFA-1.

Fig. 5.

Performance comparison of models on PI-2.

Fig. 6.

Average performance comparison of all models.

The ultimate goal of this study was to achieve accurate recognition of students’ psychological states in PE. To evaluate this, the proposed model was tested on a comprehensive psychological state recognition task. The baseline models included current mainstream ensemble approaches in the field, namely TextMind + LSTM and BERT + TextMind. Model performance was assessed using psychological state recognition accuracy and feature fusion efficiency as evaluation metrics. The experimental results are presented in Fig. 7.

Fig. 7.

Comparative performance of different models for comprehensive psychological state recognition in PE.

Table 7 reports the detailed results of the ablation experiments. Configuration A, which uses only BERT-based semantic features, achieves F1 scores of 84.2% on the EFA-1 dataset and 81.7% on the PI-2 dataset. This confirms the effectiveness of pretrained language models in capturing baseline semantic understanding. When sentiment features are added in Configuration B, performance increases significantly to 87.5% and 85.3%, indicating that the sentiment dimension plays an important role in psychological state recognition. Further incorporation of LDA-based topic features (Configuration C) yields additional improvements of 2.1% and 1.9%, suggesting that topic-level information helps distinguish contextual variations across different psychological states. Finally, the full four-dimensional model (Configuration D) achieves the best performance, with F1 scores of 92.1% and 90.9%, representing gains of 7.9% and 9.2% over the baseline Configuration A. When the syntactic GCN is removed from the complete model (Configuration E), the F1 score decreased by 2.3% on the EFA-1 dataset and by 2.7% on the PI-2 dataset. These reductions demonstrate that syntactic structural information plays a significant role in enhancing emotional comprehension within complex colloquial texts. This progressive improvement clearly demonstrates the distinct contribution of each feature dimension. In particular, the psychological feature dimension provides the most substantial gain, which is consistent with the inherently multidimensional nature of psychological states in PE contexts. Overall, the ablation results strongly support the design validity of the proposed multi-dimensional feature fusion framework.

Table 7.

Ablation study results (F1 Score, %).

Configuration	Feature composition	EFA-1	PI-2	Performance gain (vs A)
Configuration A	BERT semantic features only	84.2	81.7	Baseline
Configuration B	BERT + sentiment features	87.5	85.3	+ 3.3 / + 3.6
Configuration C	BERT + sentiment + LDA topic features	89.6	87.2	+ 5.4 / + 5.5
Configuration D	BERT + sentiment + LDA topic + psychological features	92.1	90.9	+ 7.9 / + 9.2
Configuration E	Full model w/o syntactic GCN	89.8	88.2	+ 5.6 / + 6.5

To ensure the robustness and statistical significance of the results, all models were trained and evaluated over 10 repeated runs with different random seeds. The proposed BERT-LDA-Psych model shows statistically significant improvements over the strongest baseline (fine-tuned BERT) across accuracy, precision, recall, and F1 score (p < 0.01). Specifically, on the EFA-1 dataset, the mean accuracy of the BERT-LDA-Psych model is 92.1% (95% CI 91.8–92.4%), significantly higher than that of the fine-tuned BERT model at 87.5% (95% CI 87.1–87.9%). On the PI-2 dataset, the corresponding values are 90.9% (95% CI 90.5–91.3%) and 86.3% (95% CI 85.8–86.8%), respectively. These statistical analyses confirm the reliability of the performance gains and the superiority of the proposed model. Overall, the results demonstrate that the BERT-LDA-Psych model consistently outperforms single-dimensional feature baselines on psychological counseling text datasets.

To evaluate the effectiveness of model migration to PE scenarios, the performance differences between the base BERT model and the domain-adapted BERT-LDA-Psych model were tested on the PE-Sim 100 validation set. As shown in Table 8, the F1 score of the unadapted base BERT model decreased to 78.3% on the PE-Sim dataset, representing a decline of 9.6% compared to that on the EFA-1 dataset. This reduction indicated that the general model suffered from significant semantic comprehension deviations when processing sports-specific contextual expressions. Conversely, the F1 score of the BERT-LDA-Psych model after sports domain adaptation recovered to 85.7%, which constrained the performance degradation within 6.4%, thereby validating the effectiveness of the domain adaptation strategy. However, this result remained lower than the performance of the model on the original psychological counseling datasets, indicating a semantic gap between the simulated scenario and the authentic PE classroom. This finding demonstrated that the current study only completed the feasibility validation of the technical path; the actual performance of the model in authentic PE classrooms still requires subsequent empirical research to be verified.

Table 8.

Performance comparison of scenario migration in simulated PE contexts (F1 score, %).

Model configuration	EFA-1	PI-2	PE-Sim 100	Scenario migration degradation (%)
Base BERT (Unadapted)	87.9	85.2	78.3	9.6 (EFA-1 → PE-Sim)
BERT-LDA-Psych (Domain-adapted)	92.1	90.9	85.7	6.4 (EFA-1 → PE-Sim)
Performance degradation margin	–	–	–	3.2

To evaluate the auxiliary value of the model in practical applications, a comparative experiment on PE teacher-model collaboration was designed. Six frontline PE teachers with more than five years of teaching experience and three school psychological counselors were invited to identify the psychological states within 50 student post-class reflection texts. The identification tasks included: (1) sentiment inclination judgment (positive/negative/neutral); (2) primary psychological problem type identification (self-confidence, teamwork, academic pressure, interpersonal relationships, and others); (3) intervention advice generation. The evaluation indicators comprised identification accuracy, annotation consistency, and response time.

The experimental results are detailed in Table 9. In the sentiment inclination judgment task, the average accuracy of the PE teachers was 72.4%, that of the psychological counselors was 85.1%, and that of the BERT-LDA-Psych model reached 91.5%. In the psychological problem type identification task, the accuracy of the PE teachers was only 58.3%, which was significantly lower than the 82.7% achieved by the psychological counselors and the 89.2% achieved by the model. This difference indicated that non-professional teachers faced obvious capability bottlenecks when identifying the deep psychological problems of students. Regarding the generation of intervention advice, the advice generated by the model was based on a preset strategy repository, outperforming the teachers in terms of standardization and consistency, but underperforming compared to the psychological counselors in terms of specificity and flexibility. In terms of response time, the model required an average of 0.8 s to process a single text sequence, whereas the PE teachers took an average of 4.5 min, and the psychological counselors spent an average of 6.2 min.

Table 9.

Results of the human-AI collaborative comparative experiment (n = 50).

Evaluation task	PE teachers (n = 6)	Psychological counselors (n = 3)	BERT-LDA-Psych	Improvement margin of the model relative to teachers	Gap/advantage between the model and experts
Sentiment inclination judgment	72.4%	85.1%	91.5%	+ 19.1%	+ 6.4%
Psychological problem type identification	58.3%	82.7%	89.2%	+ 30.9%	+ 6.5%
Intervention advice generation	0.52 (relative to expert reference standard)	0.89 (expert consensus standard)	0.76 (relative to expert reference standard)	+ 46.2% similarity	− 14.6% similarity
Processing time per text sequence	4.5 min	6.2 min	0.8 s	337.5-fold efficiency improvement	465-fold efficiency improvement

Discussion

This study implemented an intelligent assessment of students’ psychological states by integrating text sentiment analysis, multi-dimensional feature extraction, and psychological state recognition. Its key advantage lies in the deep adaptation of textual features and the collaborative integration of multiple techniques, providing a practical reference for applying NLP in the context of PE and psychological instruction. Rather than relying on a single semantic or sentiment feature, the study constructed a four-dimensional feature framework comprising semantic, sentiment, topic, and psychological features. This design overcomes the limitation of single-dimensional feature representation commonly seen in traditional text analysis. The incorporation of label-guided LDA effectively addressed the issues of ambiguous topics and weak relevance to the teaching context, ensuring that feature extraction aligned closely with the psychological assessment needs of real educational scenarios. Previous studies, such as Zhu et al.³⁰, demonstrated that integrating multimodal data—combining text and images—enhances sentiment analysis performance, highlighting the benefits of multimodal fusion. In contrast, this study achieved nearly equivalent representational power using only textual data, by deeply mining internal features and fusing multiple dimensions. Furthermore, the domain-adapted fine-tuning of BERT for PE texts mitigated the shortcomings of general pre-trained models in handling domain-specific vocabulary and sentiment expressions. This aligns with the approach proposed by Widmann & Wich³¹, who demonstrated that domain-specific model fine-tuning could substantially improve the recognition of textual sentiment features. This study extended that idea to a concrete educational scenario, confirming the value of domain adaptation in educational text analysis. Compared with Wang and Wang³², who employed multimodal physiological signals for emotion recognition and emphasized the limitations of single-modal data in capturing psychological states, the present study focused on textual data. However, it incorporated multi-dimensional feature fusion to achieve a more comprehensive extraction of psychological information. This strategy mitigated the inherent constraints of single-modal approaches while maintaining methodological consistency. The core innovation of this work lies in integrating general NLP techniques with context-specific characteristics of PE teaching. The model demonstrates that only by adapting and fusing technology around the specific textual features and task requirements of the educational scenario can AI genuinely serve the practical needs of teaching and learning.

The current model does not explicitly account for the differentiated psychological needs between individual and team sports, nor does it incorporate the psychological characteristics of students across different educational stages. As a result, there are performance limitations in fine-grained PE scenarios. Future work will explore modular architectures and zero-shot learning to enable context-aware adaptability. In particular, heuristic prompt-based psychological sequence algorithms proposed by Tan³³ can be leveraged to embed psychological priors—such as personality profiling and contextual adaptation—into prompt engineering, thereby improving cross-scenario generalization. In terms of data modality limitations, future research will focus on developing a multimodal fusion framework. Specifically, physiological signals and visual cues will be integrated to construct a unified multimodal psychological state recognition model. This integration will adopt a hierarchical fusion strategy, including low-level fusion for raw signal processing, mid-level fusion for cross-modal feature extraction, and high-level fusion for decision integration. To further enhance the model’s international applicability, multilingual adaptation mechanisms can also be introduced. Through cross-lingual transfer learning, the model would be capable of processing PE feedback in different languages. Related technical directions may draw on the work of Tan et al.³⁴, which demonstrates that multimodal fusion improves emotion recognition accuracy by an average of 15.8% compared to unimodal approaches. Such multimodal and multilingual extensions would make the proposed framework more aligned with the complexity of real-world PE environments, providing a more comprehensive psychological assessment tool for global PE practitioners. While the BERT-LDA-Psych model demonstrates clear advantages over single-dimensional feature models on psychological counseling datasets, it should be emphasized that these performance gains were observed under a specific data distribution. Its generalizability to PE contexts still requires further empirical validation.

Conclusion

Research contribution

This study proposed a psychological education model tailored for PE settings by integrating a pretrained BERT language model with adaptive optimizations specific to the characteristics of PE contexts. Rather than introducing a fundamentally new methodology, this study systematically integrated existing NLP techniques and adapted them to a specific application scenario, constructing a four-dimensional feature extraction framework. The aim was to address the limitations of traditional PE psychological assessment, which largely relied on manual observation and questionnaire-based evaluation. Compared with existing studies, the key contribution of this model lies in its domain-specific adaptation for PE text data. This includes the integration of a sport-related terminology lexicon, targeted tuning of sports-related emotional expression patterns, and a coordinated multi-dimensional feature fusion mechanism. Experimental results on publicly available psychological counseling datasets demonstrate that this integrated approach significantly improves psychological state recognition accuracy compared to single-feature models, providing a feasible technical pathway for applying NLP techniques in PE contexts. Experimental results indicated that the proposed model achieved high recognition accuracy on publicly available psychological counseling datasets, demonstrating its potential application value for psychological assessment in PE teaching. However, because authentic PE classroom data were not utilized, the complete effectiveness of the model within PE teaching scenarios remains to be directly validated.

Future works and research limitations

Despite these contributions, the study has several limitations. Although the model is adapted for PE scenarios, it does not yet account for differences in psychological characteristics across educational levels or different sports activities, leaving room to improve scene-specific adaptability. Future research could focus on personalized optimization of the model for diverse student groups and sport types, enhancing its applicability across varied educational contexts. This study has an important limitation in that the data used are derived from general psychological counseling scenarios rather than real PE classroom settings. Although domain adaptation techniques are employed to optimize the model, its actual performance in PE environments still requires validation through in-situ data collection and empirical testing. This limitation prevents the study from making a definitive claim regarding the model’s effectiveness in PE practice. Future work will focus on collecting authentic textual data from PE classrooms, constructing a dedicated psychological state annotation dataset for PE, and conducting model validation and performance evaluation in real teaching environments. In addition, the current model has not been specifically adapted to account for differences across various sports disciplines and educational stages, leaving room for further improvement in adaptability. Based on these limitations, future research should develop a structured dataset that captures psychological characteristics across different sports and grade levels, enabling the model to learn scenario-specific psychological representations. Meanwhile, modular adaptation and lightweight fine-tuning strategies can be introduced to allow the model to rapidly adjust to the practical requirements of different sports types and educational stages.

Author contributions

Yuan Wei: Conceptualization, methodology, software, validation, formal analysis, investigation, resources, data curation, writing—original draft preparation Noor Hamzani Farizan: methodology, software, validation, formal analysis, investigation Cheng Ji: writing—review and editing, visualization, supervision, project administration, funding acquisition.

Funding

This research received no external funding.

Data availability

The datasets used and/or analyzed during the current study are available from the corresponding author Cheng Ji on reasonable request via e-mail t15033305901@foxmail.com.

Declarations

Competing interests

The authors declare no competing interests.

Ethics approval

This article does not contain any studies with human participants or animals performed by any of the authors. All methods were performed in accordance with relevant guidelines and regulations.