Task 2 of the eRisk lab at CLEF 2025 focuses on the contextualized early detection of depression using user posts from Reddit. The HIT-SCIR team participate in this task, submitting five runs based on diferent configurations of our proposed Learnable Screening Model and Risk Post Bufer-Based Framework. Our approach involves several key components: contextual data augmentation using Large Language Models (LLMs) to simulate social interactions and generate summaries for training data; a core end-to-end learnable risky post screening model guided by symptom descriptions from established psychiatric scales; and a depression risk detector utilizing MentalBERT for classification. The oficial results on the test data demonstrate that our framework ranked ifrst across several evaluation metrics, notably F1-score, ERDE50, Flatency, and various ranking-based measures. This note describes the architecture, experimental setup, and performance analysis of our system, highlighting the value of integrating psychiatric knowledge into a learnable, context-aware model.
According to the World Health Organization (WHO), depression afects approximately 3.8% of the
global population1. In the United States, nearly 15% of adults experience at least one major depressive
episode during their lifetime [
With the proliferation of the internet, social media platforms have become conventional avenues
for individuals to openly express their thoughts and emotions [
This study explores the efectiveness of a technique that combines an online detection algorithm
based on a dynamic queue of at-risk posts [
Inspired by the implementation of Zhang et al. [
The screener calculates a risk score for each post based on its cosine similarity with descriptions from psychological scales. Subsequently, posts exhibiting higher risk scores are filtered for depression risk detection. The detector utilizes MentalBERT to acquire embedding representations for individual posts and their associated interaction information. It then models inter-post interactions using a Transformer and attention mechanisms, ultimately generating user features. Furthermore, the Straight-Through Estimator (STE) technique is employed to jointly train the screener and the detector. The screener updates its results based on the detector’s results. For final early detection testing, we use a dynamic risky post queue in conjunction with diferent alerting strategies for early depression detection. We evaluate the model’s performance under various parameter settings through five-fold cross-validation. The top three performing models are selected for ensemble voting, and diferent early decision strategies are configured across five submission runs. The results indicate that our method outperforms other participating systems on most evaluation metrics.
The remainder of this paper is structured as follows: Section 2 details the technical framework, Section 3 reports and analyzes the experimental results, and Section 4 concludes the study and discusses future directions.
Our overall process begins with a Contextual Data Augmentation phase. The subsequent framework, as illustrated in Figure 1, comprises two core stages: Psychiatric Scale-guided Risky Post Screening and Dynamic User-Level Early Risk Assessment Strategy.
The training data provided in eRisk2025 Task 2 consists of isolated user posts that inherently lack interactive context (e.g., comments or replies). However, posts encountered in test scenarios typically include associated contextual information, which proves crucial for accurate understanding and prediction. To bridge this discrepancy between the context-scarce training data and potentially context-rich test environments, and to enable our model to efectively leverage contextual information during testing, we
Posts Comments ... ... ... ... ...
Psychiatric Scale-Guided Screening Module
Depression Scale I feel depressed.
I always cry.
I am treating my depression. …… I am too tired to do things. ...
Screening
Model ...
Comment Summarization ...
Dynamic Memory
Buffer Score
Risk Prediction
Depressed Non-Depressed employ LLMs to generate simulated contextual information for the original user posts in the training data. This stage involves two specific steps:
Generation of Simulated Social Interactions: For each original post in the training set, a
pretrained generative LLM, denoted as LLMgen (e.g., we use the Phi-4 model [
{, }=1 = LLMgen(, Promptcomment) This step aims to supplement the original post with potential social reactions and discussion points, thereby providing a richer semantic context.
Summarization of Comment Content: Considering that multiple generated comments might contain redundant information or introduce unnecessary noise, we further summarize the comment set {, } generated in the previous step to extract its core semantics. This step also employs an LLM, LLMsum (in this study, we use the Phi-4 model), guided by a specific summarization instruction prompt, Promptsummary. The generated summary is denoted as .
= LLMsum({, }=1, Promptsummary) Through this contextual augmentation process, each original post is equipped with a highly condensed comment summary generated by the LLM. Both and serve as inputs to the subsequent risk prediction model. (1) (2)
We focus on how to extract key depression-related information from users’ post histories. Our approach involves using psychological scales to screen for useful posts and then detect depression. To prevent cascading errors, we have implemented an end-to-end joint training approach for both the screening and detection stages. This model facilitates a seamless flow between identifying relevant posts and ultimately detecting depression.
The model receives the user’s original post and its corresponding augmented context (comment
summary ) as input. We employ a pre-trained text encoder, EncoderFE, which is MentalBERT [
This module aims to assess the importance of each post based on psychiatric scales and to screen for representative posts.
Construction Symptom Templates and Embeddings: For the screening process, we first define a
scale item template set = {1, 2, . . . , }. These entries (symptom templates) are derived from
widely used and validated psychological depression assessment scales [
The prompt used is:
Subsequently, these two embedding vectors are concatenated to form the comprehensive feature The collection of importance scores for all posts is denoted as r = {1, 2, . . . , }.
Given a description of depression symptoms, retrieve the posts from user submissions that exhibit those symptoms. each template as: This process for obtaining the scale item template embeddings E = {e1 , . . . , e } is represented for e = EncoderE5(Prompt + ) content for this embedding step: where ‘Prompt‘ denotes the aforementioned task instruction, and + signifies concatenation.
Post Risk Score Calculation: For each user post , its textual content is directly encoded using the same EncoderE5 to obtain the post embedding e′ . Crucially, no prompt is prepended to the post Then, we calculate the cosine similarity between this post embedding e′ (from Eq. 6) and all scale item embeddings e (from Eq. 5) in the knowledge base. The maximum similarity value is taken as the importance score of post :
e′ = EncoderE5() =
max 1≤ ≤ ︃(
e′ · e ‖e′ ‖‖e ‖ )︃
Diferentiable Dynamic Post Screening Mask Generation : Our goal is to select the top ′ posts with the highest importance scores (where ′ can be a proportion or a fixed number). Traditional Top-′ operations are non-diferentiable. To enable end-to-end learning, we employ a Straight-Through Estimator (STE) method that allows for gradient propagation. First, based on the sequence of importance scores [1, . . . , ], a threshold is calculated (e.g., determined via quantile to identify the cut-of for the top ′ scores). A preliminary hard selection mask m′ is generated: where I is the indicator function. To achieve diferentiability, we construct the final mask m as follows: m′ = I{≥ } m = r + m′ − detach(r) where the detach(· ) operation prevents gradients from flowing back through its arguments. This construction ensures that during the forward pass, the mask behaves similarly to the hard selection ′, while during the backward pass, gradients can flow to the importance score r calculation, allowing parameters of the screening process to be optimized. This results in the mask sequence m = [1, 2, . . . , ].
The mask mentioned above helps us filter out irrelevant information. We will now integrate this mask into subsequent operations in a diferentiable form.
Mask-Guided Sequence Encoding: The user’s sequence of comprehensive post features [x1, . . . , x] and its corresponding diferentiable dynamic feature selection mask sequence [1, . . . , ] are fed into one or more Transformer encoder layers (TransformerEncoderLayer). Within the selfattention mechanism, we modify the masking mechanism for all attention layers as follows: S =
QKT
^ √ , , =
exp(, ) ∑︀=1 exp(,) In this modification, where represents the mask value for the -th post in the sequence being attended to, the self-attention mechanism ensures that during training, posts where = 0 (referring to the mask of the -th post in the overall user sequence) do not participate in subsequent computations; only posts with = 1 are considered. The detection model requires access to all posts during training to ensure updates to the screening process. However, during inference, we can discard the posts where = 0, which does not increase the model’s inference time. The output of the Transformer layers is: [h1, h2, . . . , h] = TransformerEncoderLayer([x1, . . . , x], attention_mask = [1, . . . , ]) (11) Here, h is the context-aware representation of the post after Transformer encoding.
User-Level Representation Aggregation: To obtain a single vector representing the user’s overall state, we aggregate the output sequence from the Transformer encoder [h1, . . . , h]. huser is obtained through avg-pooling.
Risk Prediction: Finally, the aggregated user-level representation huser is fed into a multi-layer perceptron (MLP) classifier, which outputs the probability ^ of the user having depression risk: ^ = (MLP(huser)) (12) where is the Sigmoid activation function. The entire model is trained end-to-end by minimizing the binary cross-entropy loss (BCEWithLogitsLoss) between the predicted probabilities and the true labels. We jointly optimize the filtering of important posts and the final classification. (8) (9) (10)
To more authentically simulate the application scenarios of early risk detection, we adopt the following dynamic assessment process:
Dynamic Memory Bufer : For each user, we maintain a dynamic memory bufer of capacity . This bufer stores the posts from the user’s recent history that are considered most relevant to depressive manifestations, based on their symptom relevance scores . When a new post is published by the user, its value is compared with those of the posts currently in the bufer. If the new post’s relevance is higher and the bufer is full, the post with the lowest relevance is removed, and the new post is added. This ensures that the model always makes judgments based on the user’s most relevant recent posts.
Sequential Prediction and Alert Mechanism: User posts are treated as a time series. At each time step (i.e., after a user publishes a new post and the memory bufer is potentially updated), the model (Section 2.2) performs a depression risk prediction based on the current set of posts in the memory bufer. If the model’s predicted probability exceeds a pre-set decision threshold for consecutive times (e.g., = 2), the system identifies the user as being at risk of depression and triggers an alert.
The organizers provide task-specific corpora constructed from user posts on Reddit during designated time periods. The data is distributed in XML format, containing user IDs, timestamps, post titles, and textual content. For Task 2, the training corpus features binary classification labels distinguishing between depression cases and control groups. Notably, all classifiers for Task 2 were trained exclusively on this dataset without incorporating any external data sources. This approach ensures the evaluation reflects the models’ performance under controlled conditions.
We submit the results of five runs, each based on a distinct decision-making approach for early detection.
The core of all approaches is a risk prediction model implemented as a voting ensemble. This ensemble
was constructed by selecting the three best-performing individual detection models from a candidate
pool, a selection process guided by a five-fold cross-validation method on the training corpus. The five
submitted runs (i.e., decision-making approaches) utilize this same pre-selected voting ensemble but
difer in their operational parameters ( , ) as follows:
1. Run 0: =0.5, =1
2. Run 1: =0.6, =1
3. Run 2: =0.7, =1
4. Run 3: =0.5, =2
5. Run 4: =0.5, =3
The performance of the proposed frameworks on the training set is quantitatively assessed using
precision ( ), recall (), and F1-score ( 1) [
Table 1 presents the performance of the five HIT-SCIR test runs on decision metrics (Precision / Recall / 1-score) and temporal metrics (5 / 50 / / speed / ). Among all participating teams, HIT-SCIR-4 achieves the first rank in 1-score (0.85) and 50 (0.03), and second rank in 5 (0.06). Furthermore, HIT-SCIR-4 and HIT-SCIR-2 are tied for first place in the 1.00 1.00 0.58 1.00 1.00 0.84 metric (both 0.82). Regarding other metrics within the team: HIT-SCIR-0 exhibits the highest Recall () (0.96). HIT-SCIR-0, HIT-SCIR-1, and HIT-SCIR-2 perform identically and better than the other two submissions on (4.00) and speed (0.99); these three runs also outperform HIT-SCIR-3 (0.08) and HIT-SCIR-4 (0.09) on the 5 metric (0.06).
Ranking-based evaluation employs standard information retrieval metrics such as Precision at 10 ( @10) and Normalized Discounted Cumulative Gain ( ) to assess user risk levels sorted in descending order. Table 2 shows that the HIT-SCIR team’s runs achieve first place in the vast majority of ranking metrics. Specifically: First, in the "1 writing" evaluation, all team runs rank first in @10 (1.00) and @10 (1.00); however, the team does not achieve the top rank for the @100 metric (0.58). Second, when evaluating with one hundred posts ("100 writings"), all team runs rank first in @10 (1.00) and @10 (1.00). For the @100 metric, the results of HIT-SCIR-0, HITSCIR-1, HIT-SCIR-2, and HIT-SCIR-3 (0.84) rank first, while HIT-SCIR-4 (0.83) performs slightly lower and does not achieve first place in this specific instance. Third, when evaluating with five hundred posts ("500 writings"), all team runs achieve first place in all three metrics: @10 (1.00), @10 (1.00), and @100 (0.89). Finally, when evaluating with one thousand posts ("1000 writings"), all team runs also achieve first place in all three metrics: @10 (1.00), @10 (1.00), and @100 (0.90).
The eRisk 2025 Task 2 underscores the complexities of early, contextualized depression detection from social media data. Our HIT-SCIR team proposes and evaluates a multi-stage framework centered around a learnable, psychiatric scale-guided screening model. This model is augmented by Large Language Models (LLMs), which are employed for two key purposes: first, to perform contextual augmentation on the training data by simulating interactions, and second, to generate concise summaries of contextual information (e.g., comment threads) for both training and testing data. This summarization process ofers the advantage of distilling relevant contextual signals while filtering out irrelevant information, thereby facilitating more focused and eficient processing by subsequent model components. Empirical analysis from our five submitted runs reveals the significant benefits of this integrated approach. The direct incorporation of knowledge derived from psychiatric scales into a diferentiable screening mechanism, combined with the LLM-refined contextual information and the specialized representations from MentalBERT, allows our models to efectively identify and prioritize risk-indicative posts. This results in leading performance on key metrics such as 1-score, 50, and . While traditional methods might struggle with the nuanced and evolving nature of online discourse, our end-to-end learnable system demonstrates robust adaptability. Future work will focus on refining the contextual augmentation techniques, exploring more sophisticated modeling of multi-party conversational dynamics within threads, and further enhancing the screening module’s sensitivity to subtle or emerging signs of depression. We also plan to investigate the framework’s generalizability to other mental health conditions.
During the preparation of this work, the authors use Gemini 2.5 Pro for grammar and spelling checks. After using this tool, the authors carefully review and edit the content as needed and take full responsibility for the publication’s content.
Here we provide the detailed templates in Table 3. Following prior work [