This work addresses natural language processing in Spanish-language social media. The goal is twofold: first, the early detection of users at high risk of having gambling disorders, based on a set of their messages; second, determining the type of gambling addiction (Betting, Online Gaming, Trading, and Lootboxes). We employed supervised classification approaches; however, these methods inherently depend on the availability of large annotated corpora. In an attempt to cope with this challenge, we explored the incorporation of artificial user message threads generated using Generative Large Language Models, creating synthetic counterparts by usersegment for high or low risk users. For the early detection of user's gambling disorder risk, we proposed a hybrid Bi-LSTM model trained with the GroupDRO loss and enhanced by dual attention mechanisms-capturing learned and lexicon-based risk-alongside data augmentation. To support high-risk alarm decisions based on the classifier's output, we investigated the use of dynamic thresholds. Dynamic thresholds were intended as a trade-of between earliness in decision making and unfair alarm raising for users with long-enough message sequences. The approach achieved a competitive Macro-F1 score of 0.475. For classifying the type of addiction, the classification model employs a hierarchical Bidirectional Long Short-Term Memory (Bi-LSTM) achieving a Macro-F1 score of 0.856. In addition, we designed our approaches to be lightweight and versatile, capable of running on modest hardware without requiring GPU acceleration.
Gambling activities have been interwoven with human society throughout history, but recent decades have witnessed an unprecedented acceleration in their proliferation and accessibility. Of particular concern is the emergence of novel gambling modalities such as online casinos and in-game lootbox mechanisms [1], which have expanded the gambling ecosystem beyond traditional boundaries.
At the same time, sophisticated marketing strategies have emerged, using digital influencers on platforms like Twitch, where streamers promote gambling activities in ways that reduce the perceived risks associated with gambling [2]. Perhaps more concerning is the early exposure of younger demographics to gambling adjacent mechanics through colourful, engaging lootbox systems in video games potentially boosting gambling behaviours that may develop into clinical disorders later in life.
The relationship between exposure to gambling mechanics and the development of problematic gambling behaviours [3] highlights the critical importance of early detection and intervention systems across all demographic groups. Identifying at-risk individuals becomes essential for implementing timely preventive measures and appropriate support resources.
Given the critical importance of early detection for problematic gambling behaviours, two tasks are distinguished with the following specifications: 1. Detection of risk: early detection of gambling disorder risk for each user consists of seizing the risk of developing a gambling-related disorder for each user employing as fewer messages as possible. This can be approached as a binary classification task with two risk levels: low and high. 2. Type of addiction: regardless of the risk-level (either low or high risk), the type of addiction has to be determined, one out of these ones: Betting, Online Gaming, Trading, and Lootboxes. This can be approached as a multi-class classification in a mono-label setting, as each user has assigned one and only one of the labels as if the labels were mutually exclusive. The specifications state that the the type of addiction is estimated in the last round.
While these task definitions establish our objectives, efective approaches require consideration of previous work in this domain. The following section examines related research that informs our methodology.
To address this work we focused on antecedents in similar activities, such as the latest MentalRiskES [4, 5] and the CLEF eRisk related editions [6, 7]. In eRisk 2023, pathological gambling was already explored, but the focus remained exclusively on early detection, without addressing the specific type of addiction. In contrast, the gambling domain is introduced for the first time in MentalRiskES 2025, expanding the thematic scope of the shared task. While the addiction-type classification task is novel in this context, the underlying formulation as a multiclass classification problem has been present in previous MentalRiskES editions, albeit applied to other mental health conditions such as depression, eating disorders, or suicidal ideation.
In the previous edition of MentalRiskES, most of the participant systems focused on Transformerbased models, frequently relying on Spanish pre-trained variants such as RoBERTa [8] and BETO [9]. Several teams combined these models with diverse strategies to enhance classification performance. For instance, Ixa-Med [10] introduced a heuristic message-level re-labelling based on embedding similarity, while ELiRF-VRAIN [11] explored both classical methods like Support Vector Machines and Transformer-based architectures, including RoBERTa and Longformer, to manage long texts. UNEDGELP [12] adopted a two-step approach using BETO and ANN models, and UnibucAI [13] integrated RoBERTuito with the use of the Long Short-Term Memory (LSTM) architecture and experimented with context-aware strategies. These approaches reflect a growing trend towards combining deep learning with task-specific preprocessing and representation techniques.
In a set of users, = {1, 2, . . . , }, each user publishes their messages in diferent time-stamps. In this task, the messages are released sequentially, one message per round by each user (regardless of the real time-stamp). Formally, in round for user ∈ , a message () is released. Let us denote by [1,] = {(1), (2), . . . , ()} the sequence of messages at hand by round for user , that is, the history or accumulated set of messages. Naturally, the total number of messages published difers from user to user. Let us denote by the total number of messages published by user ; in other words, represents the ’last round’ and is dependant on the user.
In the detection of user risk, the optimization criteria involve both Macro-F1 and ERDE in an attempt to keep balance between prediction reliability and speed, while in the type of addiction task the optimization criterion is just Macro-F1. Regarding the optimization criteria, while for the detection task earliness is a key factor in the specifications, for the determination of the type of addiction, oddly enough, the prediction is received in the last round. As a result, for the detection task, the input information available is [1,] with as small as possible; by contrast, to determine the type of addiction, the input information available is [1,], all the messages up to the last round for the user.
Each message has to be represented as a numeric feature-vector that serves as the input to the classifier. The generation of this vector is critical as it has to convey semantics into a multidimensional numeric space (R). That is, semantic relatedness has to be encoded and computed in a vector space.
In an attempt to get the text represented as numeric feature-vectors we turned to available encoders with the aim of not investing time in task-specific fine-tuning. With this, just a shallow preprocessing was applied to avoid encoding incompatibilities, as follows: atypical and infrequent characters were substituted by a space character; emojis within the UTF-16 encoding spectrum were also replaced for a space character.
With regard to the text encoding, our approach turned to Google’s embeddings, to be precise, to the text-embedding-004 model1 [14]. We opted for these embeddings for several reasons: easy out-of-the-box deployment, competitive in general domains, multilingual approach and, above all, a large-context window, allowing inputs up to 2048 tokens without truncation. The number of tokens in the input is, needless to say, a bottleneck in large text processing and the development is being rapid. For example, the default limit was set to 512 tokens in BERT [15] and in 8000 in Gemini Embeddings [16, 17]. In this task, each message, (), tends to be short with an average of Δ9.58 ± 12.85 tokens per message.
Nevertheless, since we need to guarantee the ability to tackle all messages, we need to ensure that the sequence [1,] can be vectorized. The number of tokens by user (with all the messages concatenated) is, on average, |[1,]| = 615.53 ± 477.74. Further details about the token distribution per user are shown in Figure 1. Assuming roughly one token corresponds to four characters [18], a text of 120 characters would correspond to about 30 tokens, for example.
In any case, the strategy implemented to deal with input token limitation inherent to each
transformer tool could be described as follows: the sequence of all messages [1,] was processed in
consecutive sub-sequences {[1,], [+1,], . . . , [+1,]}. Next, each sub-sequence is embedded as
in expression (
: Σ* →− [+1,] →−
R
([+1,])
(
Finally, the mean of the embedded representations is obtained, as in (
(
[1,]) =
︁(
(
[1,]
), (
[+1,]), . . . , (
[,]))︁
(
So far we mentioned the means in which the inputs were vectorized. However, each input did not merely bring the message, instead, each input comprises four elements: the id_message identifier, the message, the date, and the platform. In our approach we decided to include, the date concatenated to the text message due to the fact that in preliminary experiments the date seemed to bring relevant nuances for both classification tasks. To sum up, each input instance vectorization was described as a vector embedding with both the date and message together. In an attempt to refrain from repetitive computations, the embeddings where cached locally. Hence, for cases where the text already existed, it only needed to be loaded from disk, saving time.
Data scarcity is a common challenge in mental health text classification tasks. While the original dataset provides a foundation, it lacks suficient examples of varied linguistic expressions for robust model training. Our architecture benefits from augmented data through improved generalization.
Not only suficient but also representative supervised data are the key cornerstone to train robust classification models. The absence of corpora is a challenge for inferred approaches. To tackle this, we turned to Generative Large Language Models (LLMs). Eventually, this resulted in the development of one of the most productive strategies in this work, despite its simplicity.
Using the original training set as a basis, we generated artificial variants of the original messages focusing, primarily, on these aspects: verb tenses, register (formal or informal) and, orthographic and grammatical errors. The motivation was to get semantically similar sentences keeping the label (both risk level on the first task or addiction type, on the second). We discarded generating negative or speculative variants not to risk the reference label. Shallowly speaking, we are generating not a complete digital twin but somehow a synthetic counterpart of given users.
Formally, given the original supervised set of sequences of messages and user label, as in (
{(̃︁ [1,], )}=1
̃︀ =
{( ℳ̃︁
[1,], ) ∈ ̃︀ : = }
with ∈ {0, 1}
(
With regard to the specific generative model selected, we opted for Gemini Flash 2, to be precise, gemini-2.0-flash-thinking-exp [19]. We configured it for data augmentation by instructing it to generate Spanish text variations that preserve the original meaning and key information while utilizing diferent sentence structures and synonyms. The system varies formality in register and tone without altering the core message and retains the same emotional content and sentiment. It ensures that the 2Gemini Flash, developed by Google available at https://ai.google.dev/gemini-api/docs/thinking output remains natural and fluent to enhance friendliness, which may include emojis, colloquial slang, and intentionally varied orthographic and grammatical mistakes (though it avoids repeating identical errors). The system preserves any clinical significance and risk indicators in mental-health-related texts. Furthermore, it avoids reusing original words, phrases, punctuation, or sentence patterns, as it prioritizes data augmentation as its primary objective. The output format is presented as a JSON array that mirrors the input structure and maintains the class labels. For each original message, we generated one variant; We could have augmented artificially the corpus with more variants per user-message, but decided not to, in an attempt to avoid overfitting.
At this point the question arising was whether augment the texts of both types of users (high and low risk) or augment only the segment on which the classifiers were more error-prone. Preliminary experiments were conducted augmenting, in turns, each segment and jointly augmenting both of them. These results were helpful and shed light to make the decision.
In this section we provide details of the approaches involved to cope with Task 1 i.e. early detection of user gambling risk (either low or high): the message classifier, the strategy to optimize the classifier and the strategy to make a decision to generate the user label with the information provided by the message classifier.
The estimated risk level of gambling disorders was addressed as supervised binary classification. The core of the approach rests in a Bidirectional Long Short-Term Memory (Bi-LSTM) [20] network (made available in Pytorch) augmented with attention mechanisms following the idea by Zhou et al. [21] but improved with a more modern transformer based self-attention [22] together with a lexicon based attention [23]. The bidirectional nature of the LSTM allows the model to capture dependencies in both forward and backward directions, a thing that is essential when tackling complex tasks like determining the risk of developing gambling disorders.
Let us provide some technical details to favour reproducibility. The architecture starts with an input normalization and regularization stage, applying moderate dropout (20%) and layer normalization to stabilize training. Followed by two stacked Bi-LSTM layers with a pretty aggressive dropout rate of 50% between layers. Each layer processes the input bidirectionally, capturing sequential dependencies in both forward and backward directions. This bidirectional design enables the model to incorporate contextual information from the entire sequence, crucial for identifying complex gambling risk patterns that may manifest throughout the text. What distinguishes this model is its specialized attention mechanism, implemented through a custom module where the manually-curated lexicon attention and the inferred attention are combined: • Inferred Attention: A standard single head attention, as described by Vaswani et al. [22], automatically discovers which parts of the input sequence are most relevant for a successful detection. • Lexicon Attention: A predefined set of risk-related terms (e.g., “debt”, “chasing losses”, “addiction”) is embedded into vectors using the Google’s embedding model as describe in section 3.1. These vectors guide the attention mechanism by computing a risk-relevance score for each hidden state, boosting the influence of gambling-specific features. Approximately, the risk-related termvocabulary includes above 200 terms generated via LLMs and subsequently refined through manual review. The terms fall broadly into the following categories: Clinical and diagnostic terms; Problematic gambling behaviours; Financial distress indicators; Psychological and emotional indicators; Social and family impact; Treatment and self-help terms; Terms for betting, trading, lootboxes, online gambling and crypto; Rationalization and denial phrases; Colloquial high-risk expressions; Extreme consequence indicators; and Severe addiction indicators.
The combination of the two attention mechanisms is encompassed as in (
Combined Attention = Inferred attention ⊙ (1 + Lexicon attention)
(
– Group formation: In our implementation, 5 groups are formed dynamically based on the error quantiles of the predictions inspired by [29]. Instances are ranked by decreasing error and split by error percentiles (leading to 5 groups). During training stage, the model would progressively allocate more attention to groups with higher errors. – Weight update mechanism: As a contribution in our work, in favour of gradual weight updates based on persistent performance patterns (rather than strong weight fluctuations due to temporary or sporadic diferences), instead of employing the original GroupDRO, in our case the group weights are updated using an exponential moving average approach [30]. • Training corpus: Each training instance in the training corpus was built as the concatenation of all the messages (together with the date) by user with its corresponding user-risk label (), namely, = {([1,], )}=1. Thus, the size of the training corpus was bound to the number of users (| | = = | |), a small set of = 357 users (350 in the supervised train set and 7 in the trial set), this is why we enhanced the corpus with artificially generated inputs as mentioned in section 3.2. In diferent runs, we explored augmenting only users with = 0 or both = 0 and = 1, as stated in the settings of each task (Tables 2 and 6), accordingly, thereby doubling the size of the supervised dataset. Needless to say, we could have generated more than a single synthetic counterpart for each user and generate more instances to assess the impact of the size of artificially generated data in the learning stage, but this point remains open for further research.
The decision strategy determines whether a user is classified as either high or low risk. In our case,
this decision is based on the cumulative risk assessment across rounds. For each round , and each user
, the classifier provides the likelihood of high risk evidences in the sequence of messages [1,], (),
as in (
:
Σ* →−
[1,] →−
[0, 1]
( ()
[1,]) =
(
At round the system makes the decision to label each user as high risk if () exceeds a dynamic
threshold (). The threshold increases over time (or more specifically, over rounds) as defined in
(
Note that, while this dynamic threshold is a step ahead towards handling realistic situations, it is also true that it promotes conservative decisions detrimental to detection speed or earliness. We consider the proposal of the threshold an open research question worth of further research as it is a cornerstone in the task.
In this section we provide details of the methodology followed to cope with Task 2 that is, to determine the type of addiction. The methods proposed are bound to the diference in criteria from Task 1 to Task 2 in what earliness and available information regards (mentioned in page 2).
The classification model employs a hierarchical Bidirectional Long Short-Term Memory (Bi-LSTM) [20] architecture (with the help of the Pytorch library) for predicting gambling disorder risk across four categories. The approach begins with a LockedDropout (20%) mechanism [31](by contrast to basis. high risk and 0 indicates low risk.
The motivation to use a lineally increasing threshold is to reward the gain of a stronger evidence as
more data become available so that user decision label is not benefitted for the mere fact of having a
longer chat history (). In other words, if the threshold had been constant (e.g. 0) then
have always surpassed it with suficient enough messages (with
large enough) and, thus, end up
always raising the alarm of high risk. In order to prevent this and put the focus on more realistic
situations we opted for a dynamically increasing threshold. Our approach mitigates raising alarms by
mere exposure to further messages and becomes slightly more realistic than with a constant threshold
setting. In an attempt to avoid raising the alarm by some , we could have also addressed the user
accumulated risk score (
0 otherwise the approach described in section 3.3.1). This mechanism applies consistent masking across sequence dimensions, preserving temporal coherence during training. It was used because this approach typically works best with RNNs.
The core of the architecture consists of two stacked Bi-LSTM layers [20]. Each layer processes the sequence bidirectionally, capturing sequential dependencies in the text in both forward and backward directions, while a more aggressive dropout of 40% was employed between the layers, this was possible as having Bi-LSTM layers practically duplicate the network’s dimension, e.g. a 64 dimension input will transform into a 128 dimension output after passing a Bi-LSTM layer. With a unidirectional LSTM this would not have been an issue. At last a final dropout layer of 30% precedes the classification head. The terminal hidden state from the second Bi-LSTM layer serves as the sequence representation, which is then projected through a fully connected layer to produce a multi-dimensional output (as many as class-labels in this mono-label and multi-class classification task i.e. 4).
In the inference stage, the settings were as follows: • Initialization: The model employs Xavier Normal initialization [24] (as it was the case of the approach described in section 3.3.1 for Task 1) for all weights, while biases are initialized to zero. This initialization strategy helps to maintain appropriate activation magnitudes throughout the network during forward propagation. As for the optimizer and scheduler are the same ones used in (3.3.1) AdamW [25] and ReduceLROnPlateau [26]. • Optimization criteria: This classifier employed the well-known CrossEntropyLoss [ 32] function for the optimization stage.
• Training corpus: The training corpus is identical to the one defined in 3.3.1.
Given the sequence of messages available from the 1st round to round by user , i.e. [1,], the
classifier provides the confidence of each addiction type (with ∈ {Betting, Online Gaming, Trading,
Lootboxes}) as ( [1,], ) = 1 in (
([1,]) = arg max ([1,], )
The second means, (), exploits the first impressions mostly (denoted as FIM), as in (
With the limitation that the total number of messages (last round, ) is user dependant, in the real scenario our system submitted the decision made thus far, ^(), every round (1 ≤ ≤ ) while the specifications stated that the approaches would be ranking just with the predictions sent on the final round ( = ). This might have an undesirable impact in carbon emissions.
Registered as SoloResearch team, denoted after the family name of the first author, Sologuestoa, we attained a rank of 11 over 38 in Task 1 and 5 over 32 in Task 2. Below we analyze the performance of the methods presented.
Our methods were assessed employing the data made available in MentalRiskES 2025 [33, 34, 35]. Some details about the corpus are shown in Table 1.
Three runs were submitted each of which with the specifications stated in Table 2. The settings difer
on the dynamic threshold set to make the decision, () described in (
On the one hand, classification accuracy metrics achieved in this task together with the final rank are shown in Table 3 and, on the other hand, their corresponding speed and latency-weighted accuracy are given in Table 4.
Note that, in Table 3 Run 2 achieved better Macro-F1 than the other two runs, with a score of 0.475 (Rank 11 over a total of 38 runs taking part in this task) in comparison to Run 0 and 1 (ranked in 15th and 13th positions respectively). Nevertheless, detection speed is a key factor in this task and, as shown in Table 4, Run 0 resulted in the fastest approach, leading to the best Latency-weighted F1.
The methods presented were assessed by means of the corpus employed in MentalRiskES 2025 [33, 34, 35]. Some details are summarized in Table 5.
Table 7 presents the performance metrics for Task 2, focused only on accurate detection capability. In Task 2, we demonstrated strong detection capabilities. Our models achieved high Macro-F1 scores (0.856) , with run 0 and run 1 achieving identical performance metrics, meaning that the voting system had little relevance since run 0 and run 1 were identical expecting the voting system. Notably, our run 2 showed a slight decrease in Macro-F1 marking the importance of our data augmentation since the runs 0 and 1 were made with the augmented data from both risk classes 0 and 1 and the last run used only augmented data from the low risk class. Overall, we can appreciate our eforts, which placed us as the 3rd team in the competition rankings for this task.
In connection to the SDGs, the carbon emissions and computational eficiency of each run were also assessed. Table 8 presents our hardware configuration and emissions data. The results demonstrate exceptional eficiency together with competitive performance across both tasks.
Having presented the experimental results and in an attempt to provide further discussion, we delved
into evaluating both performance and resource consumption, considering the straightforward eficiency
ratios in (
Macro F1 Ef RAM = RAM Size ,
Macro F1 Macro F1 Ef Energy = Energy Used , Ef Emissions = Emissions
Macro F1
Ef GPU = GPU Energy ,
(
In this work we describe some approaches to deal with the early detection of gambling disorder risks and classification of addiction types proposed in MentalRiskES 2025. We restricted ourselves to approaches that could run in a regular laptop without GPUs or intensive computation resources.
We resorted to available multi-lingual approaches and employed data augmentation by segment via generative LLMs. For Task 1, the most challenging early detection scenario, we developed a hybrid approach incorporating Bi-LSTM adjusted with GroupDRO loss function and dual attention mechanisms (with both learned-risk and lexicon-risk awareness) and augmenting training data. We opted for a conservative risk detection strategy even at the expense of increasing latency with the aim to get a realistic decision strategy rather than challenge-tailored. The approach delivered competitive results with a Macro-F1 score of 0.475. For Task 2, our hierarchical Bi-LSTM architecture proved highly efective, achieving a Macro-F1 score of 0.856 and positioning us the third-best team in the competition.
As for the lessons learned, in what the impact of the artificial data obtained with generative LLMs is regarded, we observed an interesting phenomenon, that is, augmenting the low-risk class segment of data yielded the most substantial improvements. This suggests that enhancing information of what "normal" (or low-risk) behaviour is, paradoxically, improves high-risk identification, the conjecture is that clearer boundaries between classes are learned this way.
It is worth highlighting that our approach demonstrated fairly good computational eficiency, operating without GPU resources and with minimal RAM requirements (16GB) while maintaining competitive performance. This eficiency-performance balance is particularly relevant for potential real world deployment in mental health monitoring systems where computational resources may be limited.
For future work, we find several core aspects worth exploring e.g. (i) learnable decision making strategies (i.e. learnable thresholds) together with a thorough experimental framework; (ii) the impact of each attention mechanism (inferred and lexicon) and, particularly, terms that should be involved in the lexicon-attention; (iii) alternative loss functions for both tasks with systematic comparisons. The task itself, could be re-considered as multi-label given that the addiction types are not mutually-exclusive and could co-appear in real settings.
To sum up, integrating well-established LSTM architectures with current techniques as strategic data augmentation and specialized attention mechanisms has demonstrated powerful for mental health risk and addiction detection, particularly in scenarios with limited labelled data.
This work was partially funded by LOTU grant (TED2021-130398B-C22 funded by MICIU/AEI/ 10.13039/501100011033 and by the European Union NextGenerationEU/PRTR) and also by the Spanish Ministry of Science and Innovation (EDHIA PID2022-136522OB-C22) with the second author being awarded a grant from the Spanish Ministry of Education within Formación de Profesorado Universitario program (FPU23/01068); it has been also funded by the Basque Government (IXA IT1570-22).
During the preparation of this work, the authors used ChatGPT in order to improve the writing style and perform grammar and spelling checks. After using this tool, the authors reviewed and edited the content as needed and take full responsibility for the publication’s content. in: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022, pp. 16742–16751. [31] S. Merity, N. S. Keskar, R. Socher, Regularizing and optimizing lstm language models, 2017. URL: https://arxiv.org/abs/1708.02182. arXiv:1708.02182. [32] A. Mao, M. Mohri, Y. Zhong, Cross-entropy loss functions: Theoretical analysis and applications, 2023. URL: https://arxiv.org/abs/2304.07288. arXiv:2304.07288. [33] P. Álvarez-Ojeda, M. V. Cantero-Romero, A. Semikozova, A. Montejo-Ráez, The PRECOM-SM Corpus: Gambling in Spanish Social Media, in: Proceedings of the 31st International Conference on Computational Linguistics, 2025, pp. 17–28. [34] J. Á. González-Barba, L. Chiruzzo, S. M. Jiménez-Zafra, Overview of IberLEF 2025: Natural Language Processing Challenges for Spanish and other Iberian Languages, in: Proceedings of the Iberian Languages Evaluation Forum (IberLEF 2025), co-located with the 41st Conference of the Spanish Society for Natural Language Processing (SEPLN 2025), CEUR-WS. org, 2025. [35] A. M. Mármol-Romero, P. Álvarez-Ojeda, A. Moreno-Muñoz, F. M. P. del Arco, M. D. MolinaGonzález, M.-T. Martín-Valdivia, L. A. Ureña-López, A. Montejo-Ráez, Overview of MentalRiskES at IberLEF 2025: Early Detection of Mental Disorders Risk in Spanish, Procesamiento del Lenguaje Natural 75 (2025).
Act as a data augmentation expert. Create a different variation of the following text while: This text is classified as {classification} (where 1 indicates HIGH risk content and 0 indicates LOW risk).
For mental health-related text, ensure that clinical significance and risk indicators are preserved.
Format your response as a JSON array with the same structure as your input.
The original text that you shall modify: