Jul 2026· International Journal of Scientific Research in Computer Science Engineering and Information Technology· 0 citations
TL;DR
A hybrid detection framework which combines semantically deep embeddings from the RoBERTa transformer with a set of carefully designed language statistics and linguistic statistics and shows excellent resistance to the surface-level adversarial paraphrasing strategy.
Abstract
The influx of large language models (LLMs), like GPT-4, Claude, and Llama, has made distinguishing between natural and artificial content more difficult. The existing detection algorithms have three inherent drawback: they have limited ability to detect on short sentence-length texts containing less than 60 words, they are domain-specific, and they are fragile to adversarial attacks that involve synonym-replacement and paraphrasing. To overcome all three limitations, this paper proposes a hybrid detection framework which combines semantically deep embeddings from the RoBERTa transformer with a set of carefully designed language statistics (vocabulary richness, burstiness, and information entropy) and linguistic statistics (part-of-speech distributions, Flesch Reading Ease scores). The feature vector of the resulting embedding a 778-dimensional vector is handed to an ensemble of gradient-boosting trees, specifically XGBoost. The experiments are carried out on a processed database of 27,333 essays from professional, technical and social media writing, that is written by students and generated by AI tools. The proposed model is able to classify 99.00%, with an F1-score of 0.9932 and an area under the ROC curve (AUC) of 1.0. The framework achieves a 99.67% accuracy rate for perfect precision and recall for AI-generated content in the presence of synonym-based adversarial paraphrasing, showing excellent resistance to the surface-level adversarial paraphrasing strategy. Near-perfect cross-domain generalization is demonstrated and inference latency is 0.351ms per sample on average, which makes them suitable for realtime applications. These findings pave the way for scalable, accurate, adversarial-resilient AI-generated text detection using hybrid feature fusion, a paradigm that is encouraged by the results presented herein.
As natural language processing (NLP) systems increasingly operate on heterogeneous data, faithful social bias evaluation requires benchmarks that are both domain-sensitive and scalable. However, existing resources still face a core trade-off: template-based benchmarks scale easily but often lack linguistic authenticity, whereas naturally occurring examples (NOEs) reflect real usage but are expensive to curate and often provide limited coverage across domains and social groups. To bridge this gap, we investigate whether large language models (LLMs) can automatically adapt template-based bias datasets to specific domains using zero-shot prompting. We evaluate the effectiveness of LLMs along three dimensions: bias consistency, which measures agreement between bias estimates derived from adapted templates and NOEs; content preservation, which assesses whether the original semantic intent is maintained; and domain adherence, which evaluates how well the generated text reflects the linguistic characteristics of the target domain. Across the Equity Evaluation Corpus (EEC) and the Identity Phrase Templates Test Set (IPTTS), and across three domains—IMDb, Twitter, and Wikipedia Talk Pages—using both automatic and human evaluations, we show that domain-adapted templates capture real-world bias patterns more faithfully than standard templates. Overall, our results highlight the importance of domain variation in fairness research and position LLM-based adaptation as a scalable and reproducible framework for robust social bias assessment in NLP.
Tamara Quiroga, Felipe Bravo-Marquez, Valentin Barrière· ACM Transactions on Intellig...· 0 citations
Text classification using traditional machine learning techniques has been used in natural language processing (NLP) tasks to distinguish translated from non-translated languages, with high accuracy scores indicating the distinctive style of translated languages. While deep learning (DL) has demonstrated impressive performance in terms of representation learning and capturing nuanced patterns in natural language data, DL models act as black boxes, making their results difficult to interpret. This study addresses this issue by demonstrating an explainable AI (XAI) DL framework in a case study of United Nations (UN) meetings. The framework consists of three stages: i) train a variational autoencoder (VAE) combined with BERT embeddings converted from translated and non-translated texts; ii) utilize the majority vote from three classifiers selected from a stacked ensemble to classify the VAE’s latent representations; iii) implement a perturbation-based XAI method to interpret the DL model’s decisions. The results indicate that the VAE-based model effectively distinguishes the two text types, with accuracy scores above 0.8. The XAI analysis reveals that interpreting the VAE-based model’s decision uncovers stylistic differences between the two text types beyond superficial lexical and syntactic features. This proof-of-concept study demonstrates the potential of the XAI DL framework in other NLP studies that aim to analyze style.
The rapid proliferation of digitally distributed news has made large-scale automated verification an urgent research priority, as deceptive content spreads across social platforms faster than manual reviewers can evaluate it. This work introduces a two-stage deep learning pipeline in which a pretrained BERT encoder generates rich contextual token embeddings that are subsequently processed by a Bidirectional Long Short-Term Memory (BiLSTM) network, enabling the classifier to exploit both sentence-level semantics and document-level narrative flow simultaneously. Experiments on the WELFake benchmark [1] following a standardised cleaning pipeline—lowercasing, URL stripping, tokenisation, stop-word removal, and WordNet lemmatisation— yielded 98.7 % accuracy and an F1-score of 0.986. This surpasses a BERT-only baseline by 3.1 percentage points and exceeds several previously published state-of-the-art results. The gains confirm that coupling transformer-based semantic representations with recurrent sequential modelling produces a measurable and reproducible improvement in misinformation detection. Planned extensions include multilingual evaluation and knowledge-distilled encoder replacements targeting real-time throughput constraints.
Ardra P Namboodiri, Archa P S, Honey Mol O· 2026 6th International Confe...· 0 citations
This work proposes TinyAttack, a novel adversarial attack framework designed to exploit vulnerabilities in LLMs through Unicode-based stylistic transformations, which utilises five Unicode variants to modify the visual rendering of text without altering its underlying semantic or syntactic structure.
M. Mamta, B. Grecu, O. Cocarascu· Annual Meeting of the Associ...· 0 citations