Scientific texts in the biomedical domain are often complex and inaccessible to non-experts, posing challenges for comprehension and decision support. The CLEF 2025 SimpleText track aims to evaluate systems that can simplify scientific content and identify hallucinations in generated text. In this paper, we describe the participation of the RECAIDSTechTitans team in four subtasks: Task 1.1 (Sentence Simplification), Task 1.2 (Document Simplification), Task 2.1 (Spurious Sentence Detection - Post-hoc), and Task 2.2 (Spurious Sentence Detection - Sourced). We developed prompt-based fine-tuning pipelines using T5 transformer models, tailored for each subtask using lightweight preprocessing and structured input prompts. Our best performance was observed on Task 1.2 with a SARI score of 33.89, demonstrating the potential of compact, domain-adapted models in biomedical NLP. We also explored hallucination detection challenges, highlighting the need for future work in error-aware generation and model grounding strategies.
The CLEF 2025 SimpleText Track focuses on enhancing the accessibility of scientific texts through
simplification and controlled content generation [
Our primary goal was to explore how transformer-based architectures, particularly T5, could be adapted for domain-specific classification problems in scientific and biomedical contexts. These tasks are essential for improving structured access to knowledge and play a key role in the simplification pipeline—enabling better content indexing, retrieval, and interpretation. We developed text-to-text classification pipelines using HuggingFace’s T5 models, which were trained on paragraph-level inputs with carefully designed prompt formats. Training and inference were conducted using Google Colab and relied on curated CSV datasets provided by the organizers.
The research on text simplification, particularly within scientific and technical domains, highlights
various approaches and challenges in making complex texts more accessible. He et. al., emphasize the
importance of alternative representations beyond traditional text simplification, such as graph-based
visualizations, to enhance consumer comprehension of dietary supplement information [
Spring et. al., explore multi-level text simplification in German, utilizing source labels and pretraining
to adapt standard language to specific CEFR levels [
Fatima et. al., extend the scope to cross-lingual science journalism, aiming to generate summaries
in local languages for non-expert audiences [
The identification of hallucination in text generated by multimodal models has garnered significant
research attention, with various approaches addressing the underlying causes and detection methods.
Cui et. al., highlight that while models like GPT-4V(ision) can process visual and textual data
simultaneously, their hallucination behaviors remain inadequately understood, prompting the development of
benchmarks such as Bingo to systematically assess bias and interference challenges associated with
hallucinations [
Mitigation strategies are also a focal point. Chen et. al., emphasized that enhancing vision annotations
and employing more discriminative vision models can improve the accuracy of responses, thereby
reducing hallucinations [
Further insights into the internal mechanisms of hallucination are provided by Yang et. al., who find
that hallucinations tend to concentrate in deeper layers of LVLMs, with a strong attention bias toward
text tokens [
In specialized domains such as medical imaging, Khanal et. al., introduce hallucination-aware
ifnetuning, which not only detects but also corrects hallucinations, addressing the critical implications
of hallucination in sensitive applications [
The datasets used in the CLEF 2025 SimpleText Track were designed to evaluate the performance
of NLP models in classifying scientific text at both the source and topic levels. Four subtasks were
addressed: Task 1.1 (Sentence Simplification), Task 1.2 (Document Simplification), Task 2.1 (Spurious
Sentence Detection (Post-hoc)), and Task 2.2 (Spurious Sentence Detection (Sourced)). Each task
involved predicting labels for biomedical paragraphs that originated from diverse sources such as
Cochrane Reviews and PubMed abstracts. For all tasks, the data was structured in CSV format and
included clearly labeled training, validation, and test splits. Each row in the dataset consisted of a single
paragraph of text and an associated label, either indicating its source (e.g., "Cochrane", "PubMed") or its
primary topic (e.g., "Treatment", "Diagnosis", "Epidemiology"). The sourced versions of the tasks (2.1
and 2.2) also included metadata linking the paragraph back to its originating document, which is useful
for evaluating faithfulness and alignment in simplification scenarios [
Example (Document Simplification - Task 1.2): "This study analyzed 450 patients undergoing chemotherapy to compare the eficacy of drug A versus drug B in reducing tumor size over a six-month period. The results suggest drug A had a higher response rate."
Label (Topic): Treatment
Our approach to the CLEF 2025 SimpleText Track was centered around the use of the T5 transformer model, known for its flexibility in handling text-to-text tasks. We preprocessed the input paragraphs by formatting them into prompts such as "classify: <paragraph>" to suit the requirements of source and topic classification. This prompt-driven setup allowed us to train a unified model that could generalize across both scientific and biomedical content. We used HuggingFace’s implementation of T5 and performed training and validation on Google Colab with eficient memory and compute scaling. To handle all four subtasks—Task 1.1, Task 1.2, Task 2.1, and Task 2.2 we maintained a consistent architecture and adjusted only the input encoding and dataset splits. The model training employed a learning rate of 3e-4, batch size of 16, and ran for 4 to 10 epochs depending on task complexity. For sourced tasks (2.1 and 2.2), metadata was included in evaluation to assess model grounding and consistency with the source document. We ensured that all outputs adhered strictly to CLEF’s JSONbased submission format with uniquely defined run_ids for tracking. Beyond standard classification, we also conducted internal experiments using CLEF’s distortion error taxonomy to test model robustness against simplification inconsistencies [ 17]. Although formal submission for hallucination detection tasks wasn’t part of our main track, our system was evaluated informally for issues like overgeneralization and topic shift. Overall, the pipeline demonstrated that prompt-based T5 models ofer a scalable and efective solution for low-resource scientific text classification. The workflow is shown in Figure 1.
Our best result was achieved on Task 1.2, with a score of 33.8926. While T5 performed consistently across subtasks, some submissions showed negative scores due to format errors or model limitations. We observed that diferent submission configurations impacted results significantly, pointing to the need for thorough tuning.
In this task, the primary evaluation focuses on 37 abstracts comprising 587 aligned sentences, following the same alignment format as Cochrane-auto and related datasets. As participants, we note that for this track overview paper, the organizers chose to evaluate all submissions for both Task 1.1 and Task 1.2 at the document level. This unified approach ensures consistency in ground truth and allows for comparable scoring across the two tasks. Similarly additional evaluation on the larger set of 217 abstracts with 4,293 source sentences paired with 217 plain language summaries with 3,641 sentences were done. Results for Task 1.1 is presented in Table 1. Similarly, the results for Task 1.2 is presented in Table 2. To better illustrate our system’s behavior, we include a few examples of sentence-level simplification below: • Original: “The treatment led to a statistically significant improvement in survival rate.”
Simplified: “The treatment helped patients live longer.” • Original: “Participants were excluded if they had prior exposure to the drug in the last 6 months.”
Simplified: “People who used the drug in the last 6 months were not included.” • Original: “This systematic review included 25 studies evaluating the impact of physical activity on blood pressure. Most studies were randomized controlled trials involving adults over 50. The authors concluded that moderate exercise significantly reduced systolic and diastolic blood pressure levels.” Simplified: “This review looked at 25 studies on how exercise afects blood pressure. Most involved people over 50. The authors found that moderate exercise lowered blood pressure.”
Task 2 is tested with a corpus comprising 2,659 manually annotated sentence–simplification pairs. Each simplified sentence could exhibit multiple error types, framing the task as a multi-label classification problem. The provided error taxonomy was hierarchically structured into four main categories (A–D), each encompassing several fine-grained error types. Evaluation metrics included both micro and macro F1 scores computed at the group level. Additionally, a "No Error" class was used to indicate cases where no simplification errors were detected. The results for Task 2.1 and 2.2 are shown Table 3 and Table 4 respectively.
This work demonstrates the potential of transformer-based architectures like T5 for text simplification and classification in scientific domains. Our models showed notable performance, especially in topic classification tasks. Future work will explore model ensembling, grounding methods, and evaluation on hallucination-related subtasks to improve system robustness. The results of our T5-based system for CLEF 2025 SimpleText Track Task 1 demonstrate that careful model selection and fine-tuning on domain-specific data can outperform larger, general-purpose models. While some teams used massive models such as Zephyr-7B or Flan-T5, our more modest configuration achieved competitive or superior performance in both primary (SARI) and secondary metrics. This outcome aligns with findings from previous ImageCLEF tracks, including MedCLIP, where resource-eficient transformer models showed strong potential when adapted correctly to biomedical and scientific text. Moreover, our results underscore the importance of data preprocessing and submission formatting, as initial failures were caused not by modeling limitations but by submission protocol violations.
We thank the organizers of the CLEF 2025 SimpleText Track for curating the datasets and organizing the shared task. We also thank our faculty mentors and the Department of Artificial Intelligence and Data Science at Rajalakshmi Engineering College for their support and motivation throughout this project.
During the preparation of this manuscript, the authors used ChatGPT and Grammarly for grammar correction, spelling improvement, and sentence rephrasing. These tools were only used to enhance the readability and language quality of the paper. All technical content, including model design, experiments, and analysis, was produced entirely by the authors. The authors reviewed all AI-assisted edits and take full responsibility for the final content. [16] Y. Shu, H. Lin, Y. Liu, Y. Zhang, G. Zeng, Y. Li, Y. Zhou, S.-N. Lim, H. Yang, N. Sebe, When semantics mislead vision: Mitigating large multimodal models hallucinations in scene text spotting and understanding, arXiv preprint arXiv:2506.05551 (2025). [17] M. Shardlow, R. Nawaz, Neural text simplification of clinical notes with domain-specific pretraining, in: Proceedings of the 20th Workshop on Biomedical Language Processing (BioNLP 2021), 2021, pp. 91–100.