These results apply to retrospective classification of social media text and do not establish clinical effectiveness, prospective utility or the ability to identify individuals at risk in real time. The larger corpus contained 189,087 posts developed by the University of La Rioja within the PrevenIA suicide-prevention initiative, with 44.55% labeled as expressing suicidal ideation and 55.45% labeled as non-suicidal. S2i2 achieved 90.30% precision, 90.75% recall, 89.92% accuracy and a 90.34% F1-score on this dataset.

The smaller corpus was a Spanish translation of the SuicideWatch dataset, comprising 84,116 posts drawn from X and Reddit; 51.57% were labeled suicidal and 48.43% non-suicidal. On that corpus, the model reached 93.46% precision, 93.87% recall, 93.15% accuracy and a 93.51% F1-score. The held-out test sets contained 42,356 posts for the University of La Rioja corpus and 18,842 posts for SuicideWatch.

The researchers reported approximate 95% Wilson score confidence intervals for accuracy, treating it as a binomial proportion estimated from the respective test partitions. Rather than introducing a new neural-network component, the study’s contribution lies in integrating and systematically evaluating complementary architectures for a relatively understudied language and domain. S2i2 begins with bert-base-spanish-wwm-uncased, a Transformer model pretrained for Spanish using whole-word masking.

Self-attention converts each post into contextual token representations in which a word is interpreted according to the surrounding language. Those vectors then pass through a bidirectional LSTM, which reads the sequence in forward and backward directions, and a bidirectional temporal convolutional network, which applies dilated convolutions in both directions. Because the task classifies complete texts rather than generating the next token, the researchers treated strict temporal causality as unnecessary and used bidirectional processing to incorporate information on both sides of each position.

The temporal convolutional component is designed to capture patterns across a broad receptive field without the linear expansion of that field produced by ordinary convolutions. Dilation allows each filter to inspect earlier positions at increasing intervals, while the backward branch incorporates later context. A SpatialDropout1D layer suppresses entire embedding channels during training, batch normalization stabilizes activations, and max and average pooling aggregate the resulting features before dense layers produce a probability that a post belongs to the suicidal-ideation class.

The model was trained with binary cross-entropy loss and the AdamW optimizer, using early stopping and learning-rate reduction. Experiments ran in Python with Pandas, NLTK, Keras and TensorFlow on a Google Colab A100 virtual machine. Before modeling, the team removed URLs, email addresses, hashtags and punctuation, converted text to lowercase and eliminated redundant whitespace.

It also examined the language of posts labeled as suicidal. Frequent terms included quiero, siento, vida, morir, familia and ayuda, linking suicidal discourse with self-reference, emotional states, relationships, death and support-seeking. Frequent two-word expressions included quiero morir, quiero suicidarme, terminar vida, pensamientos suicidas, necesito ayuda, quiero vivir, seguir adelante and obtener ayuda.

Latent Dirichlet Allocation topic modeling identified recurring themes involving emotional experiences, interpersonal relationships, everyday situations, requests for help and explicit references to suicidal thoughts. These descriptive analyses characterized patterns in the two corpora rather than testing whether particular words or topics caused suicidal ideation. Sentiment analysis found negative polarity in 65.48% of suicidal-labeled texts in the translated SuicideWatch corpus and 71.58% in the University of La Rioja corpus.

About 98% of instances exceeded the subjectivity threshold, indicating that the analyzed material was dominated by personal opinions, emotions and individual experiences. The authors did not generate synthetic examples or augment the data, reasoning that paraphrasing could alter the meaning or intensity of suicidal expression and compromise label validity in this sensitive setting. Both datasets were relatively balanced, so no class weighting, oversampling or undersampling was applied.

The researchers stated that this choice preserved the original distributions and supported comparability with earlier work using the same corpora. Classical baselines established how much the hybrid architecture added. Random Forest, Decision Tree and Logistic Regression were trained and tested with the same corpora, splits, preprocessing and evaluation criteria as S2i2.

Random Forest and Logistic Regression provided relatively balanced precision and recall but lower overall results, while Decision Tree showed a marked decline in recall and reduced generalization in the reported comparison. The deep model exceeded these baselines on all reported metrics, with especially large gains in recall and F1-score. The researchers attributed the advantage to the division of labor among its components: BERT supplied Spanish contextual representations, BiLSTM modeled sequential dependencies, and BiTCN captured local and long-range temporal patterns.

These comparisons concern performance within the evaluated datasets and do not establish that the same ranking will hold for other platforms, languages or populations. S2i2 also outperformed two neural-network approaches associated with earlier Spanish-language studies, both of which were reimplemented for direct testing. Against the Ramírez-Cifuentes configuration on the smaller corpus, the full model improved precision by about 4.6 percentage points and recall by roughly three points.

It also exceeded the Castillo-Zúñiga configuration in precision and F1-score, although the latter was described as competitive on accuracy. The study considers recall especially relevant because false negatives can mean that risk-related text is missed, while precision affects how many false alarms accompany those detections. The authors cautioned that sarcasm, irony, figurative language, emotional ambiguity and culturally dependent expressions can complicate both forms of error in social media data.

The reimplemented comparisons were conducted under the study’s common experimental setting rather than reproducing every condition of the original publications. An ablation analysis indicated that the sequential components improved the model incrementally. A configuration using only the pretrained Transformer and final prediction layers produced 88.01% accuracy and an 88.30% F1-score on the smaller corpus.

Adding BiLSTM raised those values to 92.61% and 92.42%, respectively, while replacing it with BiTCN produced a slightly higher 93.10% accuracy and 93.04% F1-score. The complete architecture reached 93.15% accuracy and 93.51% F1-score, gains of 5.14 and 5.21 percentage points over the Transformer-only configuration. On the larger corpus, full S2i2 improved accuracy from 86.90% to 89.92% and F1-score from 87.08% to 90.34%.

Because every variant was retrained and evaluated with the same data partitions, the analysis supports the contribution of the blocks within this experimental setting, but it is not evidence of robustness in unseen clinical settings. Several limitations constrain interpretation of the findings. Both corpora consist of social media content, so the evaluation is confined to that domain.

One corpus is a translation of an English benchmark, and the study did not directly quantify how well semantic content was preserved during translation. The experiments were retrospective and offline; they did not test cross-domain transfer, prospective screening, clinical utility, dialect variation or performance across demographic groups. The model also requires cautious governance because misclassification can carry serious consequences, privacy-sensitive posts are being analyzed, and the training resources may not represent all Spanish-speaking communities.

The authors frame S2i2 as a research and support tool that could help specialist teams prioritize textual signals for review, not as an autonomous diagnostic system. Future work includes testing independent datasets from other platforms, forums, clinical narratives and counseling records, expanding coverage across Spanish-speaking countries, quantifying translation effects and examining privacy-preserving training methods such as federated learning.

Subject of Research: Evaluation of a hybrid natural-language-processing model for detecting suicidal ideation in Spanish-language social media posts.

Article Title: A BERT BiLSTM and BiTCN model for suicidal ideation identification in Spanish texts

Article References: Morales Ríos, A., Blandón Andrade, J. C., Martínez González, B., Sidorov, G., & Gil Arias, B. S. (2026). A BERT BiLSTM and BiTCN model for suicidal ideation identification in Spanish texts. Discover Artificial Intelligence, 6(1), Article 1048. https://doi.org/10.1007/s44163-026-02125-y

Image Credits: AI Generated

DOI: 10.1007/s44163-026-02125-y

Keywords: suicidal ideation, Spanish language, social media, natural language processing, BERT, BiLSTM, BiTCN, text classification, mental health, suicide prevention, model, suicidal