Selecting a source language for zero-shot cross-lingual transfer is typically done by intuition or by defaulting to English, despite large performance differences across language pairs. We study whether linguistic similarity can predict transfer performance and support principled source-language selection. We introduce quantified WALS (qWALS), a typology-based similarity metric derived from features in the World Atlas of Language Structures, and evaluate it against existing similarity baselines. Validation uses three complementary signals: computational similarity scores, zero-shot transfer performance of multilingual transformers (mBERT and XLM-R) on four NLP tasks (dependency parsing, named entity recognition, sentiment analysis, and abusive language identification) across eight languages, and an expert-linguist similarity survey. Across tasks and models, higher linguistic similarity is associated with better transfer, and the survey provides independent support for the computational metrics.

Advancing healthcare with large language models: A scoping review of applications and future directions.

Medical Entity Recognition (MedER) is an essential NLP task for extracting meaningful entities from the medical corpus. Nowadays, MedER-based research outcomes can remarkably contribute to the development of automated systems in the medical sector, ultimately enhancing patient care and outcomes. While extensive research has been conducted on MedER in English, low-resource languages like Bangla remain underexplored. Our work aims to bridge this gap. For Bangla medical entity recognition, this study first examined a number of transformer models, including BERT, DistilBERT, ELECTRA, and RoBERTa. We also propose a novel Multi-BERT Ensemble approach that outperformed all baseline models with the highest accuracy of 89.58%. Notably, it provides an 11.80% accuracy improvement over the single-layer BERT model, demonstrating its effectiveness for this task. A major challenge in MedER for low-resource languages is the lack of annotated datasets. To address this issue, we developed a high-quality dataset tailored for the Bangla MedER task. The dataset was used to evaluate the effectiveness of our model through multiple performance metrics, demonstrating its robustness and applicability. Our findings highlight the potential of Multi-BERT Ensemble models in improving MedER for Bangla and set the foundation for further advancements in low-resource medical NLP.

Medical large language models (LLMs) perform well on medical NLP tasks, but lack models tailored for cancer phenotyping and diagnosis. Moreover, having tens of billions of parameters increases the computational burden in healthcare settings. To this end, we present CancerLLM, a 7-billion-parameter Mistral-style model trained on 2.7 M clinical notes and 515 K pathology reports across 17 cancer types, followed by fine-tuning on cancer phenotype extraction and diagnosis generation tasks. Our evaluation demonstrated that CancerLLM achieved strong performance on internal benchmarks, with F1 score of 91.78% on phenotyping extraction and 86.81% on diagnosis generation. It outperformed existing LLMs, with an average F1 score improvement of 9.23%. Additionally, the CancerLLM demonstrated its efficiency on time and GPU usage, and robustness comparing with other LLMs. We demonstrated that CancerLLM can potentially provide an effective and robust solution to advance clinical research and practice in cancer domain.

The efficiency of machine learning algorithms applied to various NLP tasks, particularly the automatic extraction of argumentation structures from texts, heavily depends on datasets annotation consistency. As a rule, annotation of large corpora for algorithms training relies on a joint effort of several annotators, detrimental to its consistency. The present article describes methods of modifying argumentation annotations for improving their consistency. Annotations in the study take form of argumentation graphs constructed in accordance with the Argument Interchange Format standard, using ArgNetBank Studio tools. These graphs contain nodes with statements text (for premises, conclusions) aggregated into arguments with edges through scheme nodes (which correspond to argument types from Walton’s compendium). The study proposes methods for automatic modification of graphs in case of annotators’ disagreement in identifying argumentative statements and arguments of specific types. At the level of statements, consistency improvement relies on two procedures: 1) removal of leaf nodes present only in one of the graphs; 2) transformation of two-argument sequences into one-argument upon fulfillment of specific conditions. For argument types, agreement increases through replacing infrequent narrow-focused schemes with more common general ones, as well as by applying a hierarchical system of schemes substitution rules based on their functional classification (a functional group unites schemes similar in semantic and textual expression properties). A typical quantitative way of measuring annotation consistency is employing inter-annotator agreement coefficients. Calculation of the Krippendorff α agreement coefficient demonstrates a considerable increase of consistency upon modifying a corpus of 160 annotations for 80 short scientific articles in Russian: the increase equals 25 % for argumentative statements and 19 % for argument types. However, an experiment in identifying argumentative sentences with an MLP classifier shows only a 5 % increase of F-measure after modifying the corpus. The increase of F-measure for identifying four argumentation types under analysis (Cause to Effect, Verbal Classification, Example, Practical Reasoning) is even less: no more than 2 %. We arrive at a conclusion that an improvement of the inter-annotator agreement coefficient is by itself insufficient for a considerable increase in identification efficiency values in practice.

Abstract Subjective text classification tasks, such as abuse detection and stance analysis, often suffer from high levels of annotator disagreement. Conventional approaches typically collapse these disagreements into a single ground truth, thereby discarding valuable supervision signals. We propose MO-WEL (Multi-Objective Weighted Ensemble Learning), a novel framework that explicitly leverages annotator disagreement by jointly optimising ensemble weights and size under multiple objectives. Candidate predictors are trained on diverse label projections obtained through random sampling or annotator-specific selection, and ensemble weights are optimised with respect to three complementary losses: F1 score, cross-entropy and Manhattan distance, alongside a regularisation term. Experiments on four benchmark datasets (ConvAbuse, HS-Brexit, MD-Agreement and ArMIS) show that MO-WEL consistently outperforms strong baselines in accuracy, calibration, and distributional alignment. A case study further demonstrates that MO-WEL produces predictions that balance majority correctness with minority annotator perspectives, yielding interpretable and reliable outputs. Our findings highlight the importance of modelling annotator diversity and suggest ensemble optimisation as a principled means of incorporating disagreement into subjective NLP tasks.

Fine-tuning large language models in federated learning with fairness-aware prompt selection.

Arabic, an inflectional language with a rich morphology and complex syntactic structures, demands robust approaches for effective normalization and lemmatization. While, most existing NLP models focus on English, Modern Standard Arabic remains understudied, particularly in lemmatization tasks. In this paper, we introduce AraBART, the first Arabic model to feature an end-to-end pre-trained encoder-decoder, leveraging the BART architecture. We used the Arabic-PADT UD Treebank and Farasa corpus. Performance was assessed with accuracy, precision, recall, and F1 score. The results show that AraBART surpasses strong baselines, including transformer-based models such as AraT5, mT5, and BERT, achieving a 4.72% improvement in lemmatization accuracy. More importantly, AraBART has achieved an accuracy of 94.71%, approaching the performance of Farasa (97.32%). Furthermore, incorporating parts of the Farasa corpus into our training process showed a clear improvement in accuracy, revealing AraBART's effectiveness for broad applications in Arabic NLP tasks, including text summarization and machine translation.

Purpose The purpose of the proposed work is to enhance Arabic-English translation and other Arabic NLP tasks by addressing the unique linguistic challenges of the Arabic language. The Arabic Language Attention Transformer (ALAT) integrates morphological awareness, diacritic sensitivity, and right-to-left positional encoding to improve semantic understanding, syntactic accuracy, and contextual representation. By overcoming limitations in existing models like AraBERT and mBERT, ALAT aims to deliver superior performance in sentiment analysis, machine translation, and named entity recognition, setting a new benchmark for Arabic NLP applications. Design/methodology/approach The proposed ALAT employs an advanced attention-based architecture tailored for Arabic-English translation. It integrates a Morphological Awareness Module (MAM) to capture root-based patterns, a Diacritic-Aware Embedding Layer for semantic disambiguation, and a custom Right-to-Left Positional Encoding aligned with Arabic script. The model enhances the standard Transformer by incorporating morphological and diacritic features into its self-attention mechanism, improving contextual representation. ALAT is trained on diverse Arabic-English parallel corpora using cross-entropy loss with optimization techniques like the Adam optimizer, achieving superior performance in translation accuracy, as demonstrated by improved BLEU scores. Findings The proposed ALAT model significantly outperforms existing Arabic NLP models in key tasks. It achieved an F1-score of 92.4% in sentiment analysis, 86.1% in named entity recognition, and a BLEU score of 47.8 in machine translation, surpassing models like AraBERT and mBERT. ALAT's Morphological Awareness Module and Diacritic-Aware Embedding Layer improved semantic accuracy, while right-to-left positional encoding enhanced syntactic coherence. The model also demonstrated superior performance in morphological disambiguation (94.5% accuracy) and diacritic restoration (97.2% accuracy), establishing its effectiveness in addressing Arabic's linguistic complexities. Originality/value The originality of the proposed work lies in the development of the ALAT, a model specifically designed to address Arabic's unique linguistic challenges. Unlike existing models, ALAT integrates a Morphological Awareness Module, Diacritic-Aware Embedding Layer, and Right-to-Left Positional Encoding, enhancing semantic understanding and syntactic accuracy. Its innovative attention mechanism captures both morphological and diacritic features, significantly improving performance across NLP tasks. This work adds substantial value by setting a new standard for Arabic NLP, offering a robust, language-specific solution that outperforms general multilingual models in translation, sentiment analysis, and named entity recognition.

We introduce Bielik 7B v0.1, a 7-billion-parameter generative text model for Polish language processing. Trained on curated Polish corpora, this model addresses key challenges in language model development through innovative techniques. These include Weighted Instruction Cross-Entropy Loss, which balances the learning of different instruction types, and Adaptive Learning Rate, which dynamically adjusts the learning rate based on training progress. To evaluate performance, we created the Open PL LLM Leaderboard and Polish MT-Bench, novel frameworks assessing various NLP tasks and conversational abilities. Bielik 7B v0.1 demonstrates significant improvements, achieving a 9 percentage point increase in average score compared to Mistral-7B-v0.1 on the RAG Reader task. It also excels in the Polish MT-Bench, particularly in Reasoning (6.15/10) and Role-playing (7.83/10) categories. This model represents a substantial advancement in Polish language AI, offering a powerful tool for diverse linguistic applications and setting new benchmarks in the field.

Implicit content, such as implicatures and presuppositions, is a key feature of political discourse, allowing speakers to convey meaning indirectly and influence audience interpretation. While Large Language Models (LLMs) have demonstrated impressive capabilities in natural language understanding, their ability to process implicit meaning in real-world contexts remains an open question. This study investigates whether state-of-the-art LLMs can detect and classify implicit content in Italian political speech. Using a subset of the IMPAQTS corpus we assess nine multilingual models, both open-weight and proprietary. The study comprises two tasks: a binary detection task, where models determine whether a given sentence contains implicit content, and a binary classification task, in which models identify whether the implicit content is conveyed through implicature or presupposition. To enhance model performance, we employ six different prompting techniques. Results reveal that while some proprietary models exhibit moderate success in detecting implicit content, none surpass chance-level performance in classification. Open-weight models consistently underperform, with accuracy scores hovering near random guessing. Among prompting strategies, more structured techniques achieve marginal improvements in detection but fail to enhance classification accuracy. These findings highlight the persistent challenges LLMs face in pragmatic reasoning, defining implicit content detection and classification as unresolved tasks in NLP.

Nonmedical opioid use is an urgent public health challenge, with far-reaching clinical and social consequences that are often underreported in traditional healthcare settings. Social media platforms, where individuals candidly share first-person experiences, offer a valuable yet underutilized source of insight into these impacts. In this study, we present a named entity recognition (NER) framework to extract two categories of self-reported consequences from social media narratives related to opioid use: ClinicalImpacts (e.g., withdrawal, depression) and SocialImpacts (e.g., job loss). To support this task, we introduce RedditImpacts 2.0, a high-quality dataset with refined annotation guidelines and a focus on first-person disclosures, addressing key limitations of prior work. We evaluate both fine-tuned encoderbased models and state-of-the-art large language models (LLMs) under zero- and few-shot in-context learning settings. Our fine-tuned DeBERTa-large model achieves a relaxed tokenlevel F1 of 0.61 [95% CI: 0.43-0.62], consistently outperforming LLMs in precision, span accuracy, and adherence to task-specific guidelines. Furthermore, we show that strong NER performance can be achieved with substantially less labeled data, emphasizing the feasibility of deploying robust models in resource-limited settings. Our findings underscore the value of domain-specific fine-tuning for clinical NLP tasks and contribute to the responsible development of AI tools that may enhance addiction surveillance, improve interpretability, and support real-world healthcare decision-making. The best performing model, however, still significantly underperforms compared to inter-expert agreement (Cohen's kappa: 0.81), demonstrating that a gap persists between expert intelligence and current state-of-the-art NER/AI capabilities for tasks requiring deep domain knowledge. The dataset, annotation guidelines, appendix, and training scripts are publicly available to support future research.**https://github.com/SumonKantiDey/Reddit_Impacts_NER.

This study presents a max-margin–based approach for sentence boundary segmentation in Indonesian paragraphs, addressing a persistent challenge in Natural Language Processing applications. Conventional rule-based or sequential methods often struggle to distinguish ambiguous punctuation marks, particularly in contexts involving abbreviations, numerical expressions, hierarchical sentence structures, and direct quotations. To overcome these limitations, this research formulates sentence segmentation as a paragraph parsing task, enabling the model to capture both local boundary cues and global structural patterns within a paragraph. A manually annotated corpus of 12,000 paragraphs from news articles, public documents, and academic texts was developed to provide diverse linguistic structures and punctuation variations. The proposed model integrates local punctuation features, structural constraints from the Indonesian EYD standard, and global paragraph coherence through a max-margin discriminative parsing framework. Experimental results show that the model achieves strong performance on the test set, with a precision of 0.93, recall of 0.91, and F1-score of 0.92, significantly outperforming a rule-based baseline. Error analysis further highlights improvements in handling ambiguous cases such as abbreviations, numerical formatting, and direct quotations with nested punctuation. The findings demonstrate that a structured max-margin approach delivers more reliable sentence boundary segmentation and can enhance downstream NLP tasks requiring accurate sentence-level text processing.

The proliferation of online gambling promotions in YouTube comment sections poses significant challenges for content moderation on Indonesian digital platforms. Although transformer models have proven effective for various Indonesian-language NLP tasks, no systematic comparative evaluation exists for detecting online gambling promotions on YouTube, nor has research explored model sensitivity to hyperparameters in this context. This research identifies the optimal transformer model and configuration for detecting Indonesian-language online gambling promotion comments on YouTube. A total of 26,455 YouTube comments were collected from February to July 2025 and stratified into balanced training (18,926 comments) and validation sets (3,340 comments), plus an imbalanced testing set (4,189 comments with 28.05% promotions and 71.95% non-promotions) reflecting realistic platform conditions. Nine fine-tuning experiments were conducted with three transformer models (IndoBERT, IndoBERTweet, XLM-RoBERTa) using three learning rates (1e-5, 2e-5, 3e-5). Evaluation employed accuracy, precision, recall, F1-score, and AUC-ROC metrics. Results show IndoBERT with learning rate 1e-5 achieved best performance (F1-score 99.57%, recall 99.49%), outperforming IndoBERTweet (F1-score 98.58%) and XLM-RoBERTa (F1-score 99.28%). Interestingly, the formal corpus-trained model (IndoBERT) proved more effective than the social media model (IndoBERTweet), indicating that gambling promotion language patterns tend to be structured despite appearing in informal contexts. IndoBERT demonstrated greatest stability to learning rate variations (standard deviation 0.0011), while XLM-RoBERTa offered fastest inference time (2.48 ms) with optimal performance-efficiency balance. These findings provide practical recommendations for automated content moderation systems on Indonesian social media platforms, with IndoBERT for maximum accuracy scenarios and XLM-RoBERTa for large-scale real-time deployment.

The aim of this paper is to deeply analyze the unexpected side effects of people during the Covid-19 pandemic using the RNN based NLP sentiment analysis model. The normalized correlation values that is obtained by computing the cases values between the people behavior extracted and covid-19 reported case also has values close to 1 million by the end of June 2020 provided in dataset. In this research work, with more time, we would like to continue from the results we have achieved by training the RNN with NLP based sentiment analysis model for more extended periods of time for predicting the behavior of people during Covid-19 pandemic with 76.71% of accuracy which is high as compared to the CNN, such as days or weeks, in order to see how results can improve. The advancement in this field created an urge in me to research more on the techniques and methodologies developed for covid-19 extraction. During the outbreak of an epidemic, it is of immense interest to monitor the effects of containment measures and forecast of outbreak including epidemic peak affecting the behavior of people. To confront the change in behavior, a simple RNN based NLP sentiment analysis model is used to simulate the number of affected patients of Coronavirus disease. Our initial problem formulation involved investigating the ideal conditions and preprocessing for working with a specific NLP task: predicting the behavior during the specific time of May 20 – June 20 in 2020 for all four traits of common people during the Covid-19 pandemic.

In recent years, the dominance of Large Language Models (LLMs) such as BERT and RoBERTa has led to remarkable improvements in NLP tasks, including mental illness detection from social media text. However, these models are often computationally intensive, requiring significant processing time and resources, which limits their applicability in real-time or resource-constrained environments. This paper proposes a lightweight yet effective hybrid model that integrates a 6-layer RoBERTa encoder with a capsule network architecture to balance performance, interpretability, and computational efficiency. The contextual embeddings generated by RoBERTa are transformed into primary capsules, and dynamic routing is employed to generate class capsule outputs that capture high-level abstractions. To validate performance and explainability, we employ LIME (Local Interpretable Model-Agnostic Explanations) to provide insights into feature contributions and model decisions. Experimental results on benchmark mental health datasets demonstrate that our approach achieves high accuracy while significantly reducing inference time, making it suitable for deployment in real-world mental health monitoring systems.1.To design a computationally efficient architecture for mental illness detection using a lightweight RoBERTa encoder integrated with capsule networks.2.To perform a detailed time complexity analysis highlighting the trade-offs between performance and efficiency.3.To enhance model interpretability through LIME-based feature attribution, supporting transparent and ex- plainable predictions.

Bangladesh is a socio-culturally diverse country where perspectives on women's freedom vary significantly. Social media sites are now important places for sharing feminist ideas through public comments and posts. This study offers a detailed collection of 6,830 comments in Bangla about feminism to help analyze public opinion. Most of this data comes from Facebook, with some also from Instagram and Twitter. Data collection involved systematic extraction from public groups and targeted hashtag searches, including (women's rights). Native Bangla speakers meticulously annotated each comment by hand to ensure that it was topical and to identify any abusive language. This manual validation procedure guarantees a high-quality dataset appropriate for the study of online gender-based violence in the Bangla language context, sentiment analysis, and abusive language analysis, among other machine learning and NLP tasks. In addition, comments were divided into three sentiment classes: neutral, negative, and positive. This allowed for thorough analysis of feminist discourse on Bangladeshi social media and supervised learning.

This study presents a comparative evaluation of transformer-based models and traditional machine learning approaches for automated resume classification—a key task in optimizing recruitment workflows. While traditional approaches like Support Vector Machines (SVM) with TF-IDF demonstrated the highest performance (93.26% accuracy and 95% F1-score), transformer models such as DistilBERT and RoBERTa showed competitive results with 93.27% and 91.34% accuracy, respectively, and fine-tuned BERT achieved 84.35% accuracy and an F1-score of 81.54%, indicating strong semantic understanding. In contrast, Word2Vec + LSTM performed poorly across all metrics, highlighting limitations in sequential modelling for resume data. The models were evaluated on a curated resume dataset available in both text and PDF formats using accuracy, precision, recall, and F1-score, with preprocessing steps including tokenization, stop-word removal, and lemmatization. To address class imbalance, we applied stratified sampling, macro-averaged evaluation metrics, early stopping, and simple data augmentation for underrepresented categories. Model training was conducted in a PyTorch environment using Hugging Face’s Transformers library. These findings highlight the continued relevance of traditional models in specific NLP tasks and underscore the importance of model selection based on task complexity and data characteristics

Retrieval-augmented generation (RAG) enriches prompts with external knowledge, but it often relies on additional infrastructure that may be impractical in resource-constrained or offline settings. In addition, updating the internal knowledge of a language model through retraining is costly and inflexible. To address these limitations, we propose an explainable and structured prompt augmentation pipeline that enhances inputs using pre-trained models and rule-based extractors, without requiring external sources. We describe this approach as an orchestrated LLM workflow: a structured sequence in which lightweight LLM modules assume specialized roles. Specifically, (1) an extractor module identifies factual triples from input prompts by combining dependency parsing with a rule-based extraction algorithm; (2) a scorer module, based on a generic lightweight LLM, evaluates the importance of each triple via its self-attention patterns, leveraging internal beliefs to promote explainability and trustworthy cooperation with the downstream model; (3) a performer module processes the augmented prompt for downstream tasks in supervised fine-tuning or zero-shot settings. Much like in a theater staging, each module operates transparently behind the scenes to support and elevate the performer’s final output. We evaluate this approach across multiple performer architectures (encoder-only, encoder-decoder, and decoder-only) and NLP tasks (multiple-choice QA, open-book QA, and summarization). Our results show that this structured augmentation with scored facts yields consistent improvements compared to baseline prompting: up to a 28.78% accuracy improvement for multiple-choice QA, up to a 9.42% BLEURT improvement for open-book QA, and up to a 18.14% ROUGE-L improvement for summarization. By decoupling knowledge scoring from task execution, our method provides a practical, interpretable, and low-cost alternative to RAG in static or knowledge-limited environments.

IntroductionWe propose a hybrid methodology to evaluate the alignment between structural communities inferred from interaction networks and the linguistic coherence of users' textual production in online social networks. Understanding whether community structure reflects language use allows for a more nuanced validation of Community Detection Algorithms (CDAs) beyond assuming their outputs as ground truth.MethodsUsing Twitter data on climate change discussions, we compare several CDAs by training Natural Language Processing Classification Algorithms (NLPCA), such as BERTweet-based models, on the communities they generate. Classification accuracy serves as a proxy for the semantic coherence of CDA-induced groups. This comparative scoring approach offers a self-consistent framework for evaluating CDA performance without requiring manually annotated labels. We also introduce a coverage–precision trade-off metric to assess community-level performance.ResultsOur results show that the best CDA/NLPCA combinations predict a user's community with over 85% accuracy using only three short sentences. This demonstrates a strong alignment between structural and linguistic patterns in online discourse.DiscussionOur framework enables scoring CDAs based on semantic predictability and allows prediction of community membership from minimal textual input. It offers practical benefits, such as providing proxy labels for low-supervision NLP tasks, and is adaptable to other social platforms. Limitations include potential noise in CDA-generated labels but the approach offers a generalizable method for evaluating CDA performance and the coherence of online social groups.