Multilingual Neural Machine Translation (MNMT) enhances low-resource language translation by leveraging related languages, yet its application to Creole languages like Chavacano, a Spanish-based Creole spoken in the Philippines, remains underexplored. This study investigates the influence of Austronesian (Cebuano, Hiligaynon) and Indo-European (Spanish, English) parent languages on Chavacano MNMT performance. We fine-tuned an mT5-base model on the ChavacanoMT dataset, which comprises 1,044,185 samples spanning 20 language pairs involving Chavacano and related languages. Ablation experiments were conducted by excluding individual languages from training to evaluate their impact on model performance. Performance was assessed using BLEU, chrF++, and BERTScore metrics, with Paired Bootstrap Resampling applied to determine statistical significance. This study reports that Austronesian languages significantly influence Chavacano translation quality more than Indo-European languages (5 BLEU points, p =0.001). Cebuano and Hiligaynon enhance translation quality ( p ⪋0.025), while Spanish contributes less due to structural divergence (VSO vs. SVO) despite Chavacano’s 77% lexical similarity with Spanish. Our model achieves a 17 BLEU point improvement in Chavacano-English translation over prior benchmarks. These findings highlight the role of structural similarity in MNMT and inform auxiliary language selection for low-resource Creole translation, contributing new benchmarks and insights for Chavacano NLP.
Contrastive learning demonstrates superior performance in semantic embedding by capturing relationships between data samples to enhance model representation. However, its effectiveness heavily depends on positive sample construction and the choice of objective functions. Positive samples must be carefully designed to ensure the model can identify meaningful similarities while mitigating the impact of noisy samples. To address this, we propose JCCSE, a novel method that constructs positive samples by splitting, encoding, aggregating, and projecting text. Text is decomposed into segments, encoded to extract semantic content, aggregated to highlight relationships, and finally projected into a semantic space suitable for learning. Additionally, we introduce two supervised loss functions that synergize with the standard contrastive loss, enhancing the discriminative power of the semantic space and improving model performance. Experimental results on MRPC, Quora, and Contrast Clause datasets demonstrate that JCCSE achieves accuracy scores of 89.5%, 92.5%, and 92.1%, with corresponding F1-scores of 91.7%, 91.8%, and 93.3%, respectively. These results indicate that our method enhances the representational quality of semantic embeddings, effectively addresses core challenges in contrastive learning, and lays a solid foundation for further applications in natural language processing.
Medical response generation in low-resource and linguistically complex languages such as Arabic requires precise, context-aware advice, making Prompt Engineering a critical mechanism for controlling large language model behavior and improving output quality. The demand for scalable and reliable Arabic medical support systems has increased interest in prompt-based controllability, particularly where model fine-tuning and large annotated datasets are impractical. However, existing LLMs often show limited clinical relevance and semantic alignment when prompted with unstructured, patient-authored Arabic medical queries. In this work, we propose ST ructur E d st A tic- D ynamic PE (STEAD-PE), a prompt-only framework for Arabic medical response generation. STEAD-PE operates without model fine-tuning or parameter updates and leverages expert-curated Arabic medical posts paired with clinically validated treatment plans. The framework introduces case-specific structured metadata Domain, Severity, and Diagnosis—dynamically injected at inference time to enable clinically grounded and context-aware responses. We evaluate STEAD-PE across thirteen state-of-the-art LLMs, including Arabic-specialized, multilingual, and general-purpose models, using BERTScore for semantic similarity. Results show consistent and monotonic improvements in semantic and clinical relevance across all models, with STEAD-PE (Dynamic) achieving the highest BERTScore-measured semantic alignment in every case. Among evaluated models, LFM2-350M attains the best overall BERTScore (86.50%), while Qwen-32B improves from 80.89% (Zero-Shot) to 84.90% (Static) and 85.76% (Dynamic). These findings establish STEAD-PE as a robust, model-agnostic PE framework and highlight the importance of dynamically injected clinical context in reducing linguistic ambiguity and improving semantic fidelity in low-resource medical NLP settings.
In multilingual communication, dialectal differences, code-switching, and culturally influenced persuasive approaches make phishing and social engineering attacks more widespread, testing existing detection technologies. Traditional monolingual and multilingual transformer-based methods concentrate on lexical and semantic similarities and are less resistant to paraphrase or translation-induced semantic changes. TRIDENT-SE (Translation-Divergence and Intent Graph-based Detection of Social Engineering) is a hybrid cross-lingual cyber threat detection system that uses massive language models and graph-based behavioral reasoning. An instruction-tuned Large Language Model contextualizes English, Modern Standard Arabic, regional Arabic dialects, and romanized Arabic phrases. TDE modules assess semantic differences between original and translated messages, enhancing translation noise and language change resilience. An Intent–Action Graph Reasoning (IAGR) module simulates attacker strategies by correlating urgency, impersonation, credential solicitation, authority exploitation, and money demands. TRIDENT-SE increases F1-score by 9.4% over monolingual baselines and 5.1% over state-of-the-art multilingual transformer models in code-switching, dialect-shift, and adversarial paraphrase situations on an annotated Arabic–English dataset. Arabic-speaking human-centered evaluation participants had 18.7% more correct risk assessments and 22.3% less harmful user responses. Multilingual phishing and social engineering detection is robust and interpretable using translation-aware semantic analysis and intent-driven graph learning.
With the proliferation of digital platforms and smartphones, users increasingly express opinions in their native languages through online reviews. However, fine-grained sentiment analysis for such low-resource languages remains underexplored, as most Aspect-Based Sentiment Analysis (ABSA) methods target resource-rich languages such as English. To address this gap, we present Odia-TriGCN, a hybrid framework for Aspect Sentiment Triplet Extraction in the low-resource Odia language. It integrates tree-distance relations (dependency and constituent trees), syntactic dependencies, and biaffine semantic attention to capture rich structural and semantic cues. The model further leverages auxiliary features from an Odia news corpus and cross-lingual supervision using an Odia–English parallel corpus. We also incorporate a customized dependency parser trained on manually annotated data following the Universal Dependencies (UD) scheme, to obtain consistent head-dependent relations. Additionally, we extend the existing Odia ABSA dataset and introduce Odia-Triplet , a manually annotated dataset comprising aspect-opinion-sentiment triplets. Experiments across five domains show that Odia-TriGCN outperforms the strongest compared non-LLM baseline in each domain by 5.6 to 7.1 absolute F1 points. The framework is further evaluated in cross-lingual (Odia→Hindi, Odia→Bengali) and code-mixed (Odia–English, Odia–Hindi–English, Odia–Bengali–English) settings, providing further evidence of its applicability in multilingual contexts.
