Domain-specific NLP models critically depend on labelled data that is expensive and scarce in expert fields. While large language models can generate synthetic data, they lack mechanisms to enforce domain-specific factual and logical constraints. This paper introduces KASTLE, a neuro-symbolic framework that integrates domain knowledge graphs directly into the synthetic text generation process through a novel constraint-guided attention mechanism. Unlike existing approaches that apply symbolic reasoning as post-hoc filters, KASTLE embeds domain constraints at every decoding step, enabling context-aware knowledge injection under few-shot conditions. The framework features three components: a knowledge-augmented generator with real-time constraint satisfaction, a multi-layer quality filter assessing syntactic, semantic, domain, and factual dimensions, and a meta-learning engine that dynamically optimises the mixing of synthetic and real data. Extensive evaluation across medical, legal, financial, technical, and scientific domains demonstrates that KASTLE achieves an average F1 Score of 89.5, outperforming the strongest baseline (InstructAug) by 9.1 percentage points and the strongest retrieval-augmented baseline (RAG-Enhanced) by 11.2 percentage points while reducing labelled data requirements by at least 80%.
Named Entity Recognition (NER) is a major task in Natural Language Processing (NLP), and it plays an important role in information retrieval, relation extraction, machine translation, and knowledge graph construction. In recent years, various approaches have been applied to develop NER models, including rule-based methods, machine learning techniques, and deep learning models. In particular, deep learning models have demonstrated state-of-the-art performance in resource-rich languages. However, the scarcity of annotated datasets has constrained research on low-resource languages such as Assamese. Furthermore, deep learning models for text processing in Assamese have not been extensively explored for downstream NLP tasks when compared to resource-rich languages. In this article, we propose a stacked ensemble approach by combining transformer-based models with a Bi-LSTM meta-learner to develop an NER system for the Assamese tourism domain. The fundamental principle of the stacked ensemble NER is to leverage the outputs of base models through a secondary meta-classifier to improve the performance. We use three multilingual transformer models: mBERT, XLM-RoBERTa, and MuRIL as base learners. Additionally, to accomplish this task, we construct a domain-specific dataset comprising 7166 sentences and 94,604 words, annotated into eight NE classes. Experimental results yield micro-average, macro-average, and weighted-average F1 scores of 82.80%, 81.56%, and 83.01%, respectively, for the proposed ensemble model.
Relation extraction for Assamese, a morphologically rich Indo-Aryan language spoken by roughly 15 million people, has been held back by the absence of an annotated corpus, a trained model, and any standardised evaluation protocol. We address all three gaps in this work. We first introduce AREX, the first manually annotated Assamese relation extraction dataset, comprising 1,927 triples across ten semantic relation categories drawn from biographical Wikipedia text, with inter-annotator agreement of κ=0.84. We benchmark four fine-tuned multilingual transformer encoders together with one instruction tuned large language model under a controlled five seed protocol, and we find that task specific fine tuning of the smaller multilingual encoders outperforms zero shot prompting of the substantially larger LLM, with mBERT reaching a macro F1 of 0.772 ± 0.012 against Llama-3.1-8B at 0.719 ± 0.016. To exploit the complementary strengths of the four fine tuned models, we then propose a Relation Guided Adaptive Ensemble that computes per relation weights from validation F1-scores and aggregates the corresponding probability distributions at inference. The ensemble achieves a macro F1 of 0.782 ± 0.007, significantly improving on mBERT under a paired t-test at p=0.049 and reducing across seed variance by 40%. An ablation study confirms that both the F1-guided weighting and the per relation granularity are individually necessary, with uniform weight averaging performing worse than the strongest single model. The extracted triples are serialised into an RDF knowledge graph of 2,184 edges and 1,467 entities, queryable through SPARQL and enriched with temporal attributes for time anchored relations. A manually verified stratified sample of 50 triples yields a triple level precision of 84.0% with a 95% Wilson confidence interval from 71.7 to 91.2 percent and verification agreement κ=0.86. AREX, the trained models, and the constructed knowledge graph collectively provide the first end-to-end reproducible baseline for Assamese relation extraction and demonstrate a practical pipeline for extending information extraction to other morphologically rich low resource languages.
Melanoma is one of the most life-threatening forms of skin cancer, where early detection improves survival outcomes. Despite recent advances in Deep Learning (DL) for dermoscopic image analysis, existing approaches face critical limitations, including reliance on image-only inputs, inability to incorporate clinical metadata, and unstable performance on imbalanced datasets observed in real-world screening. Furthermore, many architectures are computationally heavy, limiting their suitability for clinical deployment. To address these gaps, in this paper, we propose a lightweight multi-modal diagnostic framework (BNDrop-Xnet) that integrates dermoscopic image features with clinical metadata (age, sex, and lesion location) through a weighted fusion strategy. Our framework consists of four main stages: (1) data preprocessing, including normalization, stratified sampling, and SMOTE-based class balancing; (2) image model creation using a four-block Convolutional Neural Network (CNN) architecture with Conv2D, Batch Normalization, MaxPooling, and Dropout layers; (3) metadata model construction with impact-factor computation and Multi-Layer Perceptron (MLP)-based learning of diagnostic feature contributions; and (4) ensemble integration through a weighted fusion and decision threshold optimization. It was evaluated on the ISIC 2020 dataset and achieved 90% accuracy, 80% precision, and a 73% F1-Score, outperforming several state-of-the-art models. These results demonstrate the effectiveness of combining visual and clinical cues for robust melanoma screening. BNDrop-XNet offers an efficient, interpretable, and meaningful solution for real-world deployment in resource-limited environments.
Zero-shot and few-shot learning methods offer effective solutions to data scarcity in Natural Language Processing (NLP), especially for under-resourced languages like Arabic and for mixed-language (code-switched) text. However, their capabilities in such challenging settings remain underexplored. In this work, we present the first systematic comparison of two prompt-based few-shot learning approaches; Pattern-Exploiting Training (PET) and LM-BFF (Better Few-Shot Fine-Tuning) on Arabic and Arabic-English code-switched classification tasks (Named Entity Recognition and Sentiment Analysis). Our experiments show that LM-BFF achieves a new state-of-the-art result on an Arabic-English sentiment dataset (ArEnSA), outperforming the previous best F1-score by approximately 1.5% points, and consistently surpasses PET in sentiment classification, particularly under code-switched conditions. Meanwhile, PET remains a strong baseline for structured tasks such as Arabic NER. These findings highlight the efficacy of prompt-based fine-tuning in low-resource, multilingual NLP, and underscore the potential of such techniques for underrepresented languages and code-switched text.
