Teenager depression tendency detection system based on large language model

Abstract

The invention relates to a teenager depression tendency detection system based on a large language model, and belongs to the technical field of natural language processing and mental health detection. According to the system, a Chinese social media corpus is constructed through multi-dimensional data preprocessing, a GraphRAG architecture in which a knowledge graph and a vector database are coordinated is innovatively adopted, and an Llama3.1 model is optimized in combination with an LoRA fine tuning technology, so that the recognition capability of teenager metaphor expression and complex semantics is remarkably improved. The method has good performance in the aspect of detecting the risk level of depression tendency, and a Chinese social media corpus with about 200,000 pieces of data is constructed. Compared with the prior art, the method has the advantages that the metaphor expression recognition rate is increased, the problems of overfitting, insufficient field adaptability and the like caused by data scarcity are solved, data safety is ensured through privatization deployment, dynamic evaluation and psychological intervention suggestions are supported, and a new solution is provided for psychological health monitoring of teenagers.

Description

Technical Field

[0001] The present invention relates to the field of natural language processing and mental health detection technology, and in particular to a system for detecting adolescent depression tendencies based on a large language model, which aims to effectively detect and evaluate adolescent depression tendencies using advanced technical means. Background Art

[0002] In recent years, mental health problems among teenagers have become increasingly serious. Social media has provided teenagers with a channel to vent their emotions, but the negative related information is increasing, and most of it is expressed in a metaphorical way (for example, "want to sleep forever", "eat candy", etc.). The boundary between this metaphorical expression and everyday language is blurred, posing a huge challenge to early identification and timely intervention.

[0003] Traditional mental health testing methods have numerous drawbacks. Traditional screening scales (such as the PHQ-9) are frequency-restricted, resulting in omission rates as high as 56%. There is a severe shortage of professional psychological counselors. Automatic monitoring based on keyword matching has a high false positive rate. Questionnaires and clinical interviews are subject to significant subjective interference, limiting testing frequency and data collection scope, making it difficult to achieve dynamic and comprehensive tracking of individual mental health.

[0004] While natural language processing (NLP) technology has brought new directions to mental health testing, current research faces technical bottlenecks. Deep learning models have poor interpretability, and Transformer-based models are complex, making it difficult to explain how to extract mental health features from text. Data scarcity leads to widespread overfitting, and public datasets are limited in size. For example, the commonly used depression detection dataset "DAIC-WOZ" contains only 189 samples. Natural language semantics are complex, and existing NLP systems have low accuracy in recognizing metaphorical expressions, making it difficult to capture psychological state clues implicit in text. Although large language models (LLMs) have powerful language understanding and generation capabilities, when applied to the vertical field of mental health, they lack adaptability to domain knowledge, making it difficult to accurately grasp the nuances of mental health expertise and prone to making inaccurate judgments and suggestions when faced with complex psychological cases.

[0005] In response to the above problems, the present invention proposes a system for detecting adolescent depressive tendencies based on a large language model. This method constructs a high-quality Chinese social media corpus to filter valuable information from messages posted by adolescents on social media; constructs a knowledge graph based on the corpus data, identifies entities such as risk factors and establishes associations to form a structured knowledge network that fits the actual expressions of adolescents; utilizes the efficient retrieval capabilities of the vector database, uses LLM to encode text fragments into high-dimensional vectors, and quickly matches information similar to adolescent messages. At the same time, using the GraphRAG architecture and LoRA fine-tuning technology, we can deeply understand the special expressions in adolescent messages, accurately identify implicit information such as metaphors and hints, and improve the accuracy of detecting depressive tendencies; adopts private deployment to ensure data security and privacy, ensure that adolescent information is not leaked, provide reliable technical support for adolescent mental health monitoring, and effectively make up for the shortcomings of existing technologies. Summary of the Invention

[0006] The present invention aims to provide a system for detecting adolescent depression tendencies based on a large language model, to solve the problems of low accuracy, poor adaptability, and insufficient interpretability in the existing technology for detecting adolescent depression tendencies, and to provide reliable technical support for adolescent mental health monitoring.

[0007] The detection system of the present invention includes a data processing unit, an architecture construction unit, a model optimization and deployment unit, and a detection evaluation unit.

[0008] (1) Data processing unit: The data processing unit performs multi-dimensional cleaning, screening, and preprocessing on the original social media data, and constructs a Chinese social media corpus. The specific steps are as follows:

[0009] a. Data crawling: Collect comments from the "Zoufan" Weibo comment section to obtain comments related to people with depression;

[0010] b. Data storage: The acquired data is stored as a txt file.

[0011] c. Data cleaning: Remove special characters, advertising links, handle missing values ​​and invalid emoticons, and filter out text containing clues to depression, while retaining some data without depression tendencies;

[0012] d. Data annotation: Combine the cleaned data with core assessment dimensions, develop annotation guidelines that are consistent with the characteristics of social media texts, annotate the texts according to their level of depression, and construct a Chinese social media corpus.

[0013] (2) Architecture construction unit: GraphRAG architecture that uses knowledge graph and vector database to collaborate.

[0014] The specific steps are as follows:

[0015] a. Knowledge Graph Construction: Analyze the Chinese social media corpus processed by the data processing unit. Identify entities related to adolescent depression and determine the relationships between them. Map entities and relationships into a low-dimensional vector space, numerically representing the semantic connections between them for easier system calculation and understanding. This constructs a structured knowledge network, providing rich background knowledge support for subsequent processing.

[0016] b. Vector Database Integration: Each text fragment is encoded as a high-dimensional vector, quickly matching information similar to the input text to improve retrieval efficiency. Hierarchical clustering is also performed on the text in the corpus. Based on the clustering, a summary is generated for each category, summarizing the core information of the text in that category, facilitating subsequent rapid retrieval and understanding.

[0017] c. Retrieval and Information Integration: When the detection system receives a text query containing information about young people, the GraphRAG architecture searches both the vector database and the knowledge graph. The vector database calculates the similarity between the input text and the vectors of text fragments in the database, quickly selecting the most relevant text fragments as search results. The knowledge graph then locates relevant concepts and relationships based on the keywords or entities in the query, providing structured knowledge information.

[0018] (3) Model optimization and deployment unit: Select LoRA (Low-Rank Adaptation) fine-tuning technology to fine-tune the Llama3.1 model and perform integrated deployment, and conduct ablation experiments and evaluation index calculations on the detection system. The specific steps are as follows:

[0019] a. Model fine-tuning: Llama 3.1-8B was selected as the base model and fine-tuned on the LLaMA-Factory platform. Chinese conversation samples were used for command fine-tuning to improve the model's ability to understand and process text related to adolescent depression in Chinese conversation scenarios. The fine-tuning results were evaluated using metrics.

[0020] b. Model Deployment: The fine-tuned Llama 3.1 model is integrated into the Ollama framework, working in conjunction with the GraphRAG architecture. During system operation, when a text query containing information about adolescents is received, the vector database in the GraphRAG architecture first calculates vector similarity based on the text query and retrieves relevant text fragments from the social media corpus. Simultaneously, the knowledge graph provides structured knowledge based on the entities and relationships in the query. The information retrieved from the vector database and the knowledge graph is fused to generate richer contextual information. This fused contextual information is then fed into the deployed fine-tuned Llama 3.1 model, which analyzes and determines this information, outputting an assessment of adolescent depressive tendencies.

[0021] c. Ablation Experiment: An ablation experiment was designed to omit the knowledge base, comparing the performance of the detection system of the present invention with that of a model lacking key components. A randomized validation strategy was used, splitting the data into training and test sets to ensure the generalizability of the evaluation results. The test set data was selected for experiments using both the detection system of the present invention and the model lacking key components.

[0022] d. Evaluation metric calculation: For the detection system of the present invention, an evaluation framework including depression severity is constructed. The model's prediction results are compared with the manually annotated true labels on a sample-by-sample basis.

[0023] (4) Detection and Evaluation Unit: Receives text containing information about adolescents, determines their depression tendency level, and provides recommendations. The specific steps are as follows:

[0024] a. The model first performs deep semantic analysis on the input text from adolescents to identify key emotional signals, behavioral expressions, or physiological states;

[0025] b. Retrieve pre-built knowledge graphs and vector databases based on the extracted risk features;

[0026] c. Determine the level of depression tendency and provide suggestions.

[0027] The advantages of the present invention compared with the prior art are:

[0028] 1. Sophisticated and Efficient Data Processing: Existing technologies lack sufficient depth in processing social media data, making it difficult to extract valuable information. The data processing unit of this invention constructs a high-quality Chinese social media corpus through multi-dimensional cleaning, screening, and preprocessing. Natural language processing techniques are used to remove special characters, advertising links, and invalid emoticons. Keyword extraction algorithms are employed to filter text, and annotation is performed using annotation guidelines developed in conjunction with authoritative evaluation scales. This ensures data accuracy and usability, providing a solid foundation for subsequent model training and analysis.

[0029] 2. Scientific and rational architecture design: Traditional methods lack a framework for effectively integrating knowledge and retrieval information. This paper adopts the GraphRAG architecture, integrating knowledge graphs with vector databases, and employs LLM as an aid. This can better handle the complex language phenomena found in adolescent online expressions, enhance the model's ability to capture clues to depressive tendencies, and strengthen the model's ability to understand and process complex contexts.

[0030] 3. Model optimization and deployment offer significant advantages: Existing models suffer from poor adaptability and data security issues when handling adolescent mental health testing tasks. The model optimization and deployment unit of this invention uses Chinese conversation samples to fine-tune the base model's instructions. Leveraging LoRa technology, it reduces training parameters and computational costs, improving the model's performance in Chinese conversation scenarios. This private deployment safeguards data security and privacy, ensuring stable model operation and providing strong support for testing.

[0031] 4. Strong ability to understand complex language: Teenagers often use internet buzzwords, emoticons, and metaphors to express their emotions on social media, and existing technologies struggle to accurately understand these specialized expressions. This invention, leveraging the GraphRAG architecture and the LoRA-tuned Llama 3.1 model, can more accurately parse these specialized expressions, avoiding detection errors caused by misunderstandings and more precisely capturing the true psychological state of teenagers.

[0032] 5. Provide comprehensive psychological analysis and intervention recommendations: Existing technologies often only provide a simple determination of the presence of psychological problems, but struggle to deeply analyze psychological trends and provide targeted intervention recommendations. The detection system of this invention not only determines the level of depression but also captures the user's psychological trends by analyzing emotional vocabulary, sentence logic, and daily life information in conversations. This provides comprehensive and accurate information to psychological intervention volunteers, enabling them to take timely and effective measures to better protect the mental health of young people.

[0033] This invention focuses on the detection of depressive tendencies in adolescents. Through a series of innovative technologies and methods, it demonstrates significant advantages in data processing, model architecture, and privacy protection, providing a more reliable solution for adolescent mental health monitoring. BRIEF DESCRIPTION OF THE DRAWINGS

[0034] The present invention will be better understood from the following detailed description of the embodiments of the present invention in conjunction with the accompanying drawings.

[0035] Figure 1 It is a flow chart of the overall steps of the present invention.

[0036] Figure 2 This is a diagram of the detection and evaluation unit of the present invention. DETAILED DESCRIPTION

[0037] The implementation of this invention involves several key steps, each carefully designed and debugged to ensure that the adolescent depression detection system based on a large language model can operate efficiently and accurately. The details are as follows:

[0038] (1) Data processing: Responsible for multi-dimensional cleaning, screening and preprocessing of social media raw data to build a Chinese social media corpus. The specific steps are as follows:

[0039] a. Data crawling: Using web crawler strategies, we browsed and crawled the comment section of the Weibo post "Zoufan" (https: / / weibo.com / u / 1648007681) to obtain comments related to people with depression, including: time, commenter ID, commenter, comment (including emoticons), person being replied to, and the person's link to the reply.

[0040] b. Data storage: Use " / t" as a delimiter to separate each piece of data, and store the acquired data in a txt file.

[0041] c. Data cleaning: Use the re.sub() function and regular expression matching to remove special characters and advertising links, use the Unicode range to filter invalid emoticons, delete missing text data directly, and use a keyword extraction algorithm to filter out text containing clues to depression, while retaining some data without depression tendencies;

[0042] d. Data Annotation: The cleaned data was combined with depression assessment scales (such as the Beck Depression Inventory (BDI) and the Centre for Epidemiological Studies (CES-D)) and the core assessment dimensions of the Diagnostic and Statistical Manual of Mental Disorders (DSM-5). An annotation guide was developed that conforms to the characteristics of social media text. The texts were annotated with the degree of depression (four severity levels: no depressive tendency, mild depressive tendency, moderate depressive tendency, and severe depressive tendency). A Chinese social media corpus of approximately 200,000 data items was constructed.

[0043] (2) Architecture construction: Adopt the GraphRAG architecture that combines knowledge graph and vector database. The specific steps are as follows:

[0044] a. Knowledge Graph Construction: Using natural language processing technology and a large language model-based extraction method, the data processing unit analyzes the Chinese social media corpus. Using word segmentation techniques, the text is broken down into word units. Using the large language model, entities associated with adolescent depression are accurately identified. After identifying the entities, the LLM uses semantic relationship extraction techniques to determine the relationships between them. Using the Knowledge Graph Embedding (KGEM) algorithm, entities and relationships are mapped into a low-dimensional vector space, numerically representing the semantic connections between them for easier model calculation and understanding. This results in the construction of a structured knowledge network, providing rich background knowledge support for subsequent model processing.

[0045] b. Vector Database Integration: LLM is used to encode each text segment into a high-dimensional vector, enabling rapid matching of information similar to the input text and improving retrieval efficiency. The Leiden algorithm, based on graph theory, divides node communities by optimizing modularity, effectively clustering similar texts. Based on clustering, a summary is generated for each category, summarizing the core information of that category of text, facilitating subsequent rapid retrieval and understanding.

[0046] c. Retrieval and Information Integration: When the detection system receives a text query containing information about young people, the GraphRAG architecture searches both the vector database and the knowledge graph. The vector database calculates the similarity between the input text and the vectors of text fragments in the database, quickly selecting the most relevant text fragments as search results. The knowledge graph then locates relevant concepts and relationships based on the keywords or entities in the query, providing structured knowledge information.

[0047] (3) Model optimization and deployment: Select LoRA fine-tuning technology to fine-tune the Llama3.1 model and integrate and deploy it, and perform ablation experiments and evaluation index calculations on the detection system. The specific steps are as follows:

[0048] a. Model Fine-tuning: Llama 3.1-8B was selected as the base model and fine-tuned on the LLaMA-Factory platform. Nearly 81,000 Chinese conversation samples were carefully selected from Chinese social media platforms, popular online forums, and professional conversation corpora for command fine-tuning. During fine-tuning, key parameters such as the learning rate of 5e-5 and the cutoff length of 512 were adjusted. After 50 rounds of iterative training, the model weights were continuously optimized to improve the model's ability to understand and process text related to adolescent depression in Chinese conversation scenarios. The fine-tuning results were evaluated using the predict bleu-4, predict rouge-1, predict rouge-2, predict rouge-1, and predict runtime metrics, as shown in Table 1.

[0049] Table 1 Comparison of fine-tuning results

[0050]

[0051] b. Model Deployment: The fine-tuned Llama 3.1 model is integrated into the Ollama framework, working in conjunction with the GraphRAG architecture. During system operation, when a text query containing information about adolescents is received, the vector database in the GraphRAG architecture first calculates vector similarity based on the text query and retrieves relevant text snippets from the social media corpus. Simultaneously, the knowledge graph provides structured knowledge information based on the entities and relationships in the query. The information retrieved from the vector database and the knowledge graph is fused to generate richer contextual information. This fused contextual information is then fed into the deployed fine-tuned Llama 3.1 model, which analyzes and determines this information, outputting an assessment of the adolescent's depressive tendencies. For example, if the query text mentions "I've been feeling sad lately and don't want to go to school," the vector database finds similar text snippets, and the knowledge graph provides the relationship between "sadness" and "depressive tendencies," as well as possible interventions. The fine-tuned Llama 3.1 model integrates this information to determine the adolescent's potential level of depressive tendencies and provides appropriate recommendations.

[0052] c. Ablation Experiment: An ablation experiment was designed to omit the knowledge base, comparing the performance of the detection system of the present invention with that of a model lacking this key component. Taking the user statement "My mom knows I'm depressed, but I can't tell her why because even if I do, she can't help me, and neither can the world. She wants me to live a healthy and happy life, but I never can," as an example, the model lacking the GraphRAG component primarily provides a structured plan for improving mental health, such as getting enough sleep, maintaining a balanced diet, and engaging in physical activity, but it fails to assess depression. The detection system of the present invention, on the other hand, not only assesses depression as moderate to severe, but also analyzes its impact and suggests next steps for seeking help, such as sharing feelings with family and friends or contacting local mental health services or hotlines. This demonstrates the crucial role of the GraphRAG and LoRA fine-tuning mechanisms in improving the model's professionalism, accuracy, and practicality.

[0053] d. Calculation of evaluation indicators: For the detection system of the present invention, a two-dimensional evaluation framework including depression degree (4 severity levels: no depression tendency, mild depression tendency, moderate depression tendency, severe depression tendency) is constructed. A random partitioning verification strategy is adopted, and the data set is divided into a training set and a test set in a ratio of 7:3 to ensure the generalization of the evaluation results. By comparing the model prediction results with the manually annotated true labels, the Precision (precision), Recall rate (recall rate) and F1 score (F1 value) evaluation indicators are calculated. The calculation formula is as follows:

[0054]

[0055] TP represents the number of samples whose actual category is a specific risk category (e.g., medium risk) and whose model predictions also apply to that risk category; FP refers to the number of samples that do not actually belong to a risk category but are incorrectly predicted by the model to belong to that risk category; and FN is the number of samples that actually belong to a risk category but the model is unable to correctly identify and predict other risk categories. The model's predictions were compared sample by sample with the manually annotated true labels. The highest F1 score for depression tendency detection reached 0.8333, and the performance fluctuations between each depression tendency were small, indicating that the system has a stronger ability to identify depressive tendency characteristics and more stable prediction performance, which can provide more reliable technical support for mental health assessment.

[0056] (4) Detection and Evaluation Unit: Receives text containing information about adolescents, determines their depression tendency level, and provides recommendations. The specific steps are as follows:

[0057] a. Deep Semantic Parsing: After receiving text containing information about adolescents, the detection and evaluation unit applies deep semantic parsing technology from natural language processing to conduct a detailed, sentence-by-sentence and paragraph-by-paragraph analysis of the text. Through lexical analysis, syntactic analysis, and semantic understanding algorithms, key emotional signals within the text are identified, such as frequently occurring negative vocabulary ("despair," "helplessness," "pain," etc.); descriptions of emotional fluctuations; behavioral descriptions, such as "no longer participating in social activities" and "frequent insomnia"; and physiological state-related descriptions, such as "long-term loss of appetite" and "persistent physical pain." This process aims to accurately capture potential information related to depressive tendencies within the text.

[0058] b. Risk Signature Retrieval: After completing deep semantic analysis and extracting the aforementioned risk signatures, the detection and evaluation unit searches the pre-built knowledge graph and vector database based on these signatures. By matching the extracted risk signatures with knowledge nodes in the knowledge graph and vectors in the vector database and calculating similarities, the detection and evaluation unit uncovers relevant knowledge and similar cases related to the risk signatures of the current youth text.

[0059] c. Determination of Depression Propensity and Recommendations: Based on the results retrieved from the knowledge graph and vector database, the detection and assessment unit uses pre-set assessment models and algorithms to quantitatively determine the adolescent's depression propensity level, categorizing it as low-risk, medium-risk, or high-risk. Furthermore, personalized professional recommendations are generated for each propensity level, such as psychological counseling recommendations, medical advice, and family care guidance. Ultimately, the judgment results and recommendations are provided as feedback.

[0060] This paper is based on an Ubuntu 22.04 system, using an NVIDIA A10 GPU, PyTorch 2.3.1, Python 3.11, and other deep learning libraries. Llama 3.1-8B was selected as the base model, and fine-tuned using the LoRA fine-tuning technique and the LLaMA-Factory platform. The learning rate was set to 5e-5, the number of training epochs was 50, and the truncation length was 512. The model was deployed based on the Ollama framework and visualized using Streamlit.

[0061] The embodiments of the present invention are described in detail above. Specific implementation methods are used herein to illustrate the present invention. The description of the above embodiments is only used to help understand the method of the present invention. At the same time, for those skilled in the art, according to the ideas of the present invention, there may be changes in the specific implementation methods and application scopes. In summary, the contents of this specification should not be understood as limiting the present invention.

Claims

1. A system for detecting adolescent depression tendency based on a large language model, characterized by: include: (1) Data processing unit, used to perform multi-dimensional cleaning, screening and preprocessing of social media raw data to build a Chinese social media corpus; (2) Architecture construction unit, using the GraphRAG architecture that combines knowledge graph and vector database; (3) Model optimization and deployment unit, which uses fine-tuning technology to fine-tune the base model and conducts private deployment, designs ablation experiments that omit the knowledge base, and introduces multiple evaluation indicators; (4) Detection and evaluation unit, which receives texts containing information about adolescents, determines their depression tendency level and gives suggestions.

2. The adolescent depression tendency detection system based on a large language model according to claim 1 is characterized in that: The multi-dimensional cleaning of the data processing unit includes removing special characters, invalid emoticons, advertising links, correcting grammatical errors, and processing missing data values.

3. The adolescent depression tendency detection system based on a large language model according to claim 1 is characterized in that: In the architecture construction unit, a knowledge graph is constructed based on the processed data to identify and associate risk factor entities; a vector database constructs a similarity graph of the semantic vectors of text fragments through the Leiden algorithm and performs community detection to form a hierarchical clustering structure, and then combines LLM to generate a summary for each cluster.

4. The adolescent depression tendency detection system based on a large language model according to claim 1 is characterized in that: In the model optimization and deployment unit, Chinese dialogue samples are used to fine-tune the base model, and the fine-tuned model is integrated. A two-dimensional evaluation framework for depression level is adopted, and a random partitioning verification strategy is used to divide the training set and test set. The model prediction results are compared with the manually labeled real labels to calculate the evaluation indicators and conduct ablation experiments.

5. The adolescent depression tendency detection system based on a large language model according to claim 1 is characterized in that: After receiving the text containing information about teenagers, the detection and evaluation unit first inputs the text into the fine-tuned and deployed system. The model analyzes the text, identifies the clues and emotional tendencies related to depressive tendencies contained therein, and gives relevant suggestions.

6. A method for detecting depression tendency in adolescents based on the detection system according to any one of claims 1 to 5, characterized in that: The following steps are involved: (1) Use the data processing unit to process the raw social media data and construct a Chinese social media corpus; (2) Build the GraphRAG architecture through the architecture building unit to complete the knowledge graph construction and vector database integration; (3) Use the model optimization and deployment unit to fine-tune the base model, conduct ablation experiments and calculate evaluation indicators after private deployment, and optimize the model based on the evaluation results; (4) Use the detection and evaluation unit to receive texts containing information about adolescents, determine their level of depression tendency and give suggestions.

7. The method for detecting depression tendency in adolescents according to claim 6, characterized in that: When building a Chinese social media corpus, the specific operations for multi-dimensional cleaning, screening, and preprocessing of raw data are as follows: (1) Use natural language processing technology to remove special characters, advertising links, and invalid emoticons; (2) Using keyword extraction algorithms to screen out texts containing clues to depressive tendencies; (3) Combining authoritative assessment scales at home and abroad, depression assessment scales, and the core assessment dimensions of the Diagnostic and Statistical Manual of Mental Disorders, formulate annotation guidelines that conform to the characteristics of social media texts and annotate the texts with depression levels.

8. The method for detecting depression tendency in adolescents according to claim 6, characterized in that: When building the GraphRAG architecture: (1) In terms of knowledge graph construction, based on the processed data, risk factor entities are identified and a structured knowledge network is constructed based on semantic relationships; (2) When integrating the vector database, LLM is used to convert text fragments into high-dimensional vectors. The Leiden algorithm is used to construct a similarity graph for the semantic vectors of the text fragments and perform community detection to form a hierarchical clustering structure. LLM is then used to generate a summary for each cluster. (3) During the retrieval and generation process, when the model receives a query, it searches both the vector database and the knowledge graph, and integrates the retrieved information as the input context for the fine-tuned model.

9. The method for detecting depression tendency in adolescents according to claim 6, characterized in that: When fine-tuning the base model and evaluating it for private deployment: (1) Fine-tune the base model using selected Chinese dialogue samples, adjusting key parameters such as learning rate and batch size; (2) Integrate the fine-tuned models to achieve rapid deployment and inference across environments; (3) Design an ablation experiment with the knowledge base omitted to compare the performance of the complete model and the model lacking key components. For the fully configured model, build a two-dimensional evaluation framework and use a random partitioning verification strategy to divide the dataset. (4) By comparing the model prediction results with the manually labeled true labels, various evaluation indicators are calculated and the model is optimized and improved based on the evaluation results.

10. The method for detecting depression tendency in adolescents according to claim 6, characterized in that: When conducting a test assessment: (1) First, perform deep semantic analysis on the input adolescent text to identify key emotional signals, behavioral expressions, or physiological states; (2) Retrieve pre-built knowledge graphs and vector databases based on the extracted risk features; (3) Determine the level of depression tendency and give suggestions.

Images