The system, named TL-CDRNet, was developed by Lanxin Zhong of Hainan Vocational University in Haikou, China. Its starting point is a familiar one in modern language technology: a pretrained multilingual encoder-decoder Transformer, in this case the mBART-50 model with roughly 611 million backbone parameters, twelve encoder layers, twelve decoder layers, and a SentencePiece vocabulary of more than 250,000 tokens. But rather than simply fine-tuning this backbone on a rewriting task, the study treats cultural adaptation as a transfer learning problem in which the crucial question is what knowledge can safely move between cultures and what must stay behind. Four newly designed modules wrap around the pretrained model to answer that question at different levels of granularity.

The first module addresses a tension that has long shaped cross-lingual and cross-cultural modeling. Some knowledge is genuinely universal: factual semantics, thematic content, and basic narrative structures transfer well between cultural domains. Other knowledge is deeply local: metaphors, symbolic preferences, value judgments, and pragmatic conventions that can actively harm output quality if forced through a shared channel, a phenomenon known as negative transfer. TL-CDRNet handles this with a shared-specific transfer adapter module that inserts two lightweight bottleneck feed-forward networks into each of the twelve encoder layers. One adapter learns cross-cultural commonalities while the other, conditioned on the target culture, learns culture-specific representations. Learnable weighting parameters then balance the two contributions at every layer, allowing the model to discover dynamically where the boundary between the universal and the local actually lies.

Separating the two knowledge streams, however, is not enough, because the appropriate balance differs from one cultural pair to the next. When the source and target cultures are close, much of the shared representation is useful and should flow freely; when they are distant, aggressive sharing risks importing assumptions that do not belong. The second module therefore learns a 64-dimensional embedding for each cultural-linguistic domain, normalizes it, and computes the Euclidean distance between source and target embeddings as a scalar measure of cultural distance. A sigmoid gating function converts that distance into a transfer coefficient: culturally close pairs receive stronger injection of shared knowledge, while distant pairs shift weight toward target-culture-specific representations. Importantly, the study stresses that these embeddings model corpus-level differences in expression conventions and narrative preferences rather than fixed national or ethnic identities, and no external cultural-index database is used.

Distance at the domain level is still a coarse signal, so a third module works at the level of individual tokens. A transferability-aware knowledge selection mechanism scores every hidden representation, estimating how applicable each piece of content is to the target-culture task. Factual information, core thematic elements, and universal expressions earn high transferability scores and pass through largely intact; content tightly bound to the source cultural context is blended with a local adaptation representation instead. This token-level filtering proved essential for semantic preservation in the experiments: removing it dropped the semantic consistency score from 0.918 to 0.892, indicating that without it the model drifts further from the original meaning while chasing cultural fit.

The fourth and arguably most distinctive module makes cultural discount itself a trainable objective. A discriminator network mean-pools the decoder’s hidden states and passes them through a three-layer multilayer perceptron to output a cultural discount score between zero and one, where higher values signal greater risk of communication loss in the target culture. The training loss then demands that the generated text score lower than the original unadapted text by at least a safety margin, a hinge-style formulation that turns reduction of cultural discount into a differentiable goal rather than a hoped-for side effect. Combined with standard generation loss, a semantic consistency loss based on cosine similarity of sentence embeddings, and a distribution alignment loss, this feedback closes the loop between transfer, generation, and cultural evaluation.

To test the framework, the researchers assembled a dataset of 9,000 samples across three culturally demanding scenarios: film synopsis adaptation, tourism promotional text rewriting, and descriptions of cultural creative and intangible heritage products. Each sample pairs a source text with source and target cultural-linguistic labels, a human-written target-culture reference adaptation, and a cultural discount annotation. Bilingual and multilingual annotators produced the reference rewrites, and target-domain reviewers then assessed fluency, semantic preservation, and cultural appropriateness, with three annotators per sample and disagreements sent back for revision. Inter-annotator agreement measured by Krippendorff’s alpha reached 0.742, and the data were split 80/10/10 into training, validation, and test sets with texts from the same original item kept together to prevent leakage.

The results show consistent gains over a range of baselines including direct fine-tuning of mBART, LoRA adapters, and a culture-aware BART variant. Against the strongest competitor, TL-CDRNet raised BLEU from 29.6 to 31.4, ROUGE-L from 44.0 to 46.1, BERTScore from 88.9 to 89.8, and semantic consistency from 0.900 to 0.918. The task-oriented metrics improved more dramatically: the cultural discount score fell from 0.361 to 0.298, and the cultural discount reduction rate climbed from 33.5 percent to 43.7 percent, an improvement of 10.2 percentage points. Performance held steady across all three scenarios, with reduction rates of 43.5, 41.9, and 40.3 percent for film, tourism, and heritage content respectively, while semantic consistency stayed between 0.897 and 0.922 throughout.

Ablation and transfer-strategy comparisons reveal where the gains come from. Moving from training from scratch through pretrained initialization, direct fine-tuning, bridging domain transfer, multi-stage transfer, and finally the complete model produced a monotonic rise in performance, with the full system lifting the reduction rate from 24.6 percent under direct fine-tuning to 43.7 percent, a transfer gain of 77.6 percent. Removing the shared-specific adapters cut the reduction rate to 34.6 percent; removing the cultural distance gate lowered it to 36.1 percent, with the sharpest degradation under distant-culture pairs; and removing the discount feedback loop caused the most severe decline of all. The framework also proved resilient under scarcity: with only five target-domain examples it achieved a reduction rate of 16.8 percent versus 8.3 percent for direct fine-tuning, and with one hundred examples it reached 39.8 percent against 22.4 percent. In cross-domain tests, where the model trained on one scenario and was evaluated on another, it maintained semantic consistency above 0.898 and reduction rates between 36.9 and 40.4 percent. A further analysis showed that although transfer gains decline as cultural distance grows for all methods, the slope is markedly gentler for TL-CDRNet than for LoRA, suggesting the distance-aware mechanisms genuinely mitigate the risks of long-range cultural transfer.

The study is candid about its limits. The cultural discount score and reduction rate are produced by a learned discriminator and cannot fully replace human judgment of appropriateness, emotional resonance, and audience acceptance; the author notes that future work should incorporate evaluations from target-culture participants. The baselines also exclude the newest instruction-tuned multilingual large language models, and the framework currently handles text only, leaving multimodal formats such as short video scripts and image-text generation for later exploration. There are ethical considerations too: aggressive adaptation to a target culture can erode the source culture’s distinctiveness, and biased training data risks producing stereotyped cultural portrayals, which is why the paper recommends review of culturally sensitive outputs by experts from both sides. Even so, the core demonstration stands: by teaching a model not just what cultures say but how far apart they sit, transfer learning can turn cultural adaptation from an afterthought of translation into an explicit, optimizable goal, offering a practical technical pathway for film distribution, tourism promotion, and the international reach of cultural products.

Subject of Research: Transfer learning for reducing cultural discount in cross-cultural AI text adaptation

Article Title: Transfer learning guided by cultural distance enables effective cross cultural content adaptation

Article References: Zhong, L. (2026). Transfer learning guided by cultural distance enables effective cross cultural content adaptation. Discover Artificial Intelligence, 6(1), Article 1225. https://doi.org/10.1007/s44163-026-02267-z

Image Credits: AI Generated

DOI: 10.1007/s44163-026-02267-z

Keywords: transfer learning, cultural discount, cross-cultural adaptation, multilingual NLP, Transformer, mBART, machine translation, negative transfer, cultural distance, semantic consistency, low-resource learning, intangible cultural heritage

The scale of that mismatch is well documented. The World Economic Forum estimates that a large share of the global workforce will need reskilling within just a few years as artificial intelligence reshapes skill profiles across industries. National labour-market reporting in Kazakhstan echoes the same tension, noting that university programmes are updated slowly and do not always reflect current demand. Massive open online courses, or MOOCs, seem like an obvious remedy, since platforms such as Coursera offer thousands of courses that universities can embed into credit-bearing programmes. But the sheer abundance of courses creates its own problem: with so many overlapping offerings, learners and curriculum designers cannot easily identify which courses develop the specific competencies that are simultaneously missing from a programme and demanded by employers.

The research team, led by scientists at L.N. Gumilyov Eurasian National University and partner institutions, argues that this is not a generic recommendation problem but a cross-domain alignment task. Existing systems typically integrate only one or two domains, such as vacancies and skills or courses and skills, and almost always operate in a single language, usually English. The new framework instead couples three domains at once: university curricula, IT vacancies scraped from the hh.kz jobs platform, and Coursera MOOCs, all in English, Russian, and Kazakh. According to the authors, this is the first educational knowledge graph to semantically connect all three domains in a multilingual setting.

At the heart of the system is a heterogeneous knowledge graph, a database structure in which different types of entities, from competency statements to job postings to course titles, are stored as nodes connected by typed relationships. The researchers formalised a property-graph ontology covering all three domains and instantiated it automatically. The deployed graph contains 39,810 nodes and 52,661 edges, including 2,226 unique skills, 2,404 vacancies, and 336 courses. Because the graph is modular, new courses, vacancies, or skills can be added without disturbing the existing structure, allowing the market picture to be refreshed as demand evolves.

The trickiest technical challenge was aligning skill phrases across three languages. Vacancy text is notoriously noisy, full of near-duplicates, misspellings, and inconsistent formatting, with phrases like

That multilingual challenge deserves closer attention, because it is what separates this framework from most prior educational recommenders. The researchers fine-tuned a multilingual Sentence Transformer, specifically the paraphrase-multilingual-mpnet-base-v2 model, on a purpose-built trilingual paraphrase corpus using a multiple-negatives-ranking objective. This training approach teaches the model to pull semantically equivalent skill phrases closer together in embedding space while pushing unrelated phrases apart, even when those phrases appear in different languages or in the abbreviated, inconsistent style typical of job postings. The payoff was measurable: fine-tuning improved Russian and Kazakh alignment without degrading performance on English, and the team also released an English skill-pair evaluation set so that other groups can benchmark against their results. For countries with mixed-language labour markets, this matters enormously, because a vacancy written in Russian describing a database administration role must be matched to a Kazakh-language curriculum outcome and an English-language Coursera course description before any recommendation can be trusted.

Once skills were aligned across languages, the researchers faced the problem of graph sparsity. A knowledge graph linking competencies to courses directly would contain relatively few edges, which starves graph neural networks of the structural signal they need to learn meaningful representations. The team’s solution was densification through skill communities. They compared four clustering algorithms and selected agglomerative hierarchical clustering, which groups similar skills into communities that then serve as additional pathways through the graph. Fine-tuning the embeddings improved this step as well: the silhouette coefficient, a standard measure of clustering quality, rose from 0.271 to 0.285, and the proportion of clusters judged erroneous by human experts fell from 19.7 percent to 14.0 percent. These communities act as semantic bridges, letting information flow between a competency mentioned in a curriculum and a related skill demanded in a vacancy even when no direct edge connects them.

The core methodological innovation, however, lies in how the researchers framed the recommendation task itself. Most graph-based educational recommenders cast the problem as binary link prediction, essentially asking whether a connection between a learner’s competency and a course exists or does not. This study instead treats recommendation as continuous edge-weight regression, predicting a graded relevance score on the competency-to-course relation. That formulation is better suited to education, where the fit between a competency and a course is a matter of degree rather than a yes-or-no question. A course might partially cover a skill, or cover it deeply, and the ranking of candidate courses should reflect those gradations.

To solve this regression task, the team compared three graph neural network encoders with distinct architectural philosophies. GraphSAGE is a classical homogeneous architecture that samples and aggregates neighbourhood features and has long been a workhorse of graph machine learning. HAN, the heterogeneous attention network, extends this idea to graphs with multiple node and edge types by learning attention over different meta-paths, which are predefined sequences of relations connecting nodes of the same type. HGT, the heterogeneous graph transformer, goes further by parameterising attention separately for each node and edge type, allowing the model to weigh messages differently depending on what kinds of entities are involved. On a held-out 20 percent test graph, HGT achieved the best ranking quality with an NDCG@10 of 0.714, outperforming both GraphSAGE and HAN. When short meta-paths were used, performance climbed dramatically to an NDCG@10 of 0.987 with a root-mean-square error of just 0.108, suggesting that concise, well-chosen relation sequences help the transformer generalise far better than longer ones.

NDCG, or normalised discounted cumulative gain, is worth unpacking for readers outside information retrieval. It measures how well a ranked list places the most relevant items near the top, rewarding systems that front-load their best recommendations. The strong NDCG figures indicate that the model does not merely identify good courses somewhere in a long list; it places the most market-relevant courses in the first few positions, which is precisely what a curriculum designer or a student with limited time needs. The authors describe the resulting recommendations as front-loaded and actionable, a property that follows directly from this ranking behaviour.

The practical demonstration came when the system was applied to the Information Systems programme of a Kazakhstani university. With just five top-ranked courses, the framework expanded the programme’s market-aligned competency profile by 36.2 percent. Allowing the full catalogue raised that figure to 170 percent. The steep difference between these two numbers carries an important message for educational institutions: a small, carefully chosen set of online courses captures a substantial share of the achievable alignment gain, and diminishing returns set in quickly after that. This front-loaded structure means that even modest investments in MOOC integration, perhaps a handful of courses embedded into an existing programme under an integrated assessment model, can meaningfully narrow the gap between certified and demanded competencies.

The choice of Kazakhstan as a testing ground gives the work particular policy relevance. The country’s national strategy explicitly frames digital transformation and human-capital development as intertwined priorities and recognises that artificial intelligence can analyse and recommend promising development paths. Kazakhstani universities have already begun embedding MOOC modules into credit-bearing courses, but decisions about which courses to embed have historically relied on manual organisational effort. By linking internal artefacts such as programmes, syllabi, and competency maps to external resources such as vacancies, MOOCs, and skill classifiers through semantic correspondence rather than administrative paperwork, the framework offers a template for what the authors call intellectualising an institution’s information resources.

Reproducibility is another notable strength. The researchers released all datasets, models, graph snapshots, and code, which is still uncommon in educational data mining, where proprietary institutional data often prevents independent verification. The modular design of the property graph also means the approach is not locked to its initial data sources. New vacancies can be scraped from the labour market as demand evolves, additional courses can be ingested from other platforms beyond Coursera, and the ontology can be extended to domains beyond information technology. Because the graph is instantiated automatically rather than hand-curated, keeping the market picture current becomes a pipeline maintenance task rather than a recurring research project.

Several caveats temper the enthusiasm. The evaluation focused on a single programme at a single university, so the reported gains, while striking, may not generalise uniformly across disciplines with different competency structures. The vacancy data came from one platform, hh.kz, and labour-market coverage varies across sectors and regions. The edge-weight regression targets, like any supervised signal, inherit whatever biases exist in the training data, including the possibility that certain skills are overrepresented in vacancies relative to their long-term importance. And while the trilingual corpus was sufficient to demonstrate improved cross-lingual alignment, Kazakh-language training data remains scarcer than English or Russian, which may cap how well the embedding model captures the nuances of Kazakh skill terminology.

Even so, the study points toward a broader shift in how educational institutions might manage curriculum relevance. Rather than waiting for slow, standards-driven programme revisions, universities could treat alignment as a continuous, data-driven process: monitor live vacancies, map demanded skills against certified competencies, and recommend targeted online interventions that close specific gaps. The finding that heterogeneous graph transformers outperform homogeneous architectures on this task, and that framing recommendation as regression rather than binary prediction improves ranking quality, offers methodological guidance for the growing field of knowledge-graph-based educational systems. As artificial intelligence continues to reshape which skills matter, frameworks of this kind, grounded in multilingual semantic alignment and cross-domain graph reasoning, may become standard infrastructure for keeping higher education tethered to the realities of the labour market it serves.

Subject of Research: A multilingual heterogeneous knowledge graph framework that aligns university curricula, labour-market vacancies, and MOOCs using edge-weight regression with heterogeneous graph transformers for educational recommendation.

Article Title: A multilingual heterogeneous knowledge graph framework for curriculum–vacancy–MOOC alignment: edge-weight regression with heterogeneous graph transformers for educational recommendation

Article References: Serikbayeva, S., Sambetbayeva, M., Ramazanova, V., Lamasheva, Z., Yerimbetova, A., Makhatova, V., & Sadirmekova, Z. (2026). A multilingual heterogeneous knowledge graph framework for curriculum–vacancy–MOOC alignment: edge-weight regression with heterogeneous graph transformers for educational recommendation. Frontiers in Artificial Intelligence, 9, Article 1906883. https://doi.org/10.3389/frai.2026.1906883

Image Credits: AI Generated

DOI: 10.3389/frai.2026.1906883

Keywords: knowledge graph, educational recommendation, heterogeneous graph transformer, MOOCs, skills gap, curriculum alignment, graph neural networks, multilingual NLP, sentence transformers, edge-weight regression, labour market, Kazakhstan

The work, led by Abubakar Sulaiman Gezawa of Sanming University together with colleagues at Hanshan Normal University and Bayero University, tackles three persistent gaps in existing multimodal translation systems. The first is architectural: because captioning and translation are usually trained as independent modules, gradients cannot flow between them, so errors propagate irreversibly and visual context is lost after the first stage. The second is scaling: many multilingual systems require a separate decoder or substantial parameter growth for every additional language. The third, and arguably most subtle, is grounding: attention mechanisms in conventional systems may interact with visual features superficially, but they routinely fail to preserve fine-grained details such as gender, count, spatial relations and object attributes throughout the translation process.

The team’s answer is a unified architecture built around a Transformer decoder that supports two operating modes with shared weights. In captioning mode, the model looks at an image and generates an English description. In translation mode, a small learnable vector called a language token is prepended to the embedded English caption, steering the same decoder toward the desired target language. Because the captioning and translation objectives are optimized jointly through a combined loss, the visual encoder receives complementary gradients from both tasks, encouraging representations that are simultaneously good at recognizing objects and at preserving the attributes that translators need—exactly the information that a cascaded system loses the moment it hands its text to a second model.

The first core technical innovation is what the authors call cross-modal representation alignment. A ResNet-50 convolutional encoder, pretrained on ImageNet, converts each input image into a seven-by-seven grid of 49 visual tokens, projected into the same 512-dimensional space used by the language model. These tokens are then made available to every one of the six decoder layers through a tri-attention scheme: causal self-attention over previously generated tokens, cross-attention to the English caption’s encoder outputs, and cross-attention to the visual features, each with eight attention heads and independent query, key and value projections. The ordering is deliberate. Self-attention first establishes linguistic context; caption cross-attention grounds the representation in the semantic content of the source description; and visual cross-attention acts last, refining the representation with direct evidence from the image so that details such as plurality or gender can re-enter the computation even if the English caption under-specified them.

The second innovation is the language-token conditioning strategy. Instead of training a dedicated decoder per language, the model learns a single 512-dimensional embedding for each of five typologically diverse targets—Chinese, French, Arabic, Spanish and Hindi—spanning Sino-Tibetan, Romance, Semitic and Indo-Aryan families and four different scripts. Placed at position zero of the input sequence, the language token participates in self-attention, so every caption and generated token can attend to it, and it also joins the visual cross-attention stream, allowing language-specific visual interpretation. When generating Arabic, for instance, where gender agreement is grammatically mandatory, the token can help route visual gender information to the right morphological decisions. The authors report that learned tokens specialize: the French token attends strongly to determiner positions where gender must be resolved, while the Arabic token focuses on verb morphology. A shared vocabulary of 60,000 subword units, learned with SentencePiece across all languages, enables cross-lingual transfer of common subwords, particularly between French and Spanish.

The empirical results, obtained on MS COCO for captioning and the OPUS-100 parallel corpus for translation, show consistent gains over cascaded baselines. The unified model achieves a BLEU-4 score of 44.8 for English captioning on the MS COCO Karpathy test split, alongside a CIDEr of 153.2, and an average BLEU of 41.1 across the five target languages—improvements of 2.2 and 2.0 points respectively over cascaded systems. COMET scores, a neural metric that correlates better with human judgment of semantic fidelity, average 0.732 for the proposed model compared with roughly 0.698 for text-only baselines. Perhaps most striking is the efficiency: the language-token approach requires 218 million parameters in total, a 43 percent reduction compared with the 312 million needed when separate decoders are trained for each language, while actually producing slightly better average translation quality.

The gains are not distributed evenly, and that pattern is instructive. The largest improvements over cascaded baselines appear for Arabic and Hindi, languages with rich inflectional systems in which gender, number and case must be marked inside the word itself. In a cascaded pipeline, an English caption saying “two children playing” can lose the count information needed to generate the correct Hindi plural; the unified model preserves that cue through per-layer visual attention. French and Spanish, which mark gender largely through articles and function words that can sometimes be recovered from context alone, show more modest gains of about 1.9 BLEU. Manual inspection of 500 errors confirmed the mechanism: cascaded baselines produced 2.3 times more attribute errors and 1.8 times more spatial relationship errors than the unified model, which cut overall error rates by 58 percent relative to the cascaded approach.

An ablation study disentangled the contribution of each component. Removing cross-modal alignment caused the largest degradation, dropping CIDEr by 8.6 points and average BLEU by 1.6, while adding only 8 million parameters—evidence that the benefit comes from architectural integration rather than sheer capacity. Replacing joint training with a conventional two-stage scheme cost 1.8 BLEU, confirming that end-to-end optimization acts as a useful regularizer: the visual encoder and caption decoder develop representations that balance fluent English generation with preservable visual semantics. A late-fusion variant that injected visual features only at the decoder’s final layer fared worst among the ablations, underscoring the paper’s central claim that visual grounding must operate throughout decoding, not as an afterthought, to influence agreement, word order and inflection.

The authors are candid about the system’s remaining weaknesses. Visually similar objects such as cups and bowls are occasionally confused, a limitation partly traceable to the coarse seven-by-seven feature grid, where adjacent objects can blend within the same cell. Attribute hallucination persists in lower-resource languages; in one documented case, an image of two children with a red ball was rendered in Hindi with the wrong count, and a woman holding an umbrella became a man in Arabic. Spatial prepositions such as “behind” versus “next to” are sometimes flattened, because encoding three-dimensional depth from a two-dimensional projection remains genuinely hard. These failures occurred in only 4.7 percent of inspected test examples, however, and appeared two to three times more frequently in the cascaded systems the new model is designed to replace.

Beyond the benchmark numbers, the researchers frame the work in terms of biologically inspired computation, arguing that per-layer cross-modal modulation echoes the layered multisensory integration observed in human cortex, while the language tokens act as a sparse gating mechanism akin to experience-dependent linguistic plasticity. The practical implications reach further than the analogy: accessibility services for visually impaired users in diverse linguistic communities, multilingual educational content, and cultural heritage documentation all depend on descriptions that stay faithful to a visual scene across languages. With training completed in about 18 hours on four NVIDIA A100 GPUs and inference running beam search at modest width, the architecture is deliberately positioned for deployment rather than spectacle. Future directions the team outlines include depth-aware visual encoders for spatial reasoning, language-aware optimization for low-resource languages, and extension of the token-conditioning scheme toward zero-shot translation of unseen languages—steps that could turn a strong multilingual captioner into a genuinely universal visual translator.

Subject of Research: A unified multimodal neural model that integrates image captioning and multilingual machine translation with cross-modal alignment and language-token conditioning

Article Title: Cross-modal alignment with language tokens for unified multilingual image-to-text generation

Article References: Gezawa, A. S., Junejo, N. U. R., Abdulkadir, S. A., Liu, C., & Lin, G. (2026). Cross-modal alignment with language tokens for unified multilingual image-to-text generation. Frontiers in Computational Neuroscience, 20, Article 1826455. https://doi.org/10.3389/fncom.2026.1826455

Image Credits: AI Generated

DOI: 10.3389/fncom.2026.1826455

Keywords: multimodal machine translation, image captioning, cross-modal alignment, language tokens, neural machine translation, Transformer decoder, ResNet-50, MS COCO, OPUS-100, BLEU score, multilingual NLP, vision-language models

The stakes are particularly high in India, which has the second-largest population of internet users worldwide. According to figures cited in the study, India had more than 375 million social media users in 2019, rising to 518 million in 2020, with projections suggesting the number could reach 1.5 billion by 2040. Affordable mobile data has driven this explosive growth, and platforms such as Facebook, Twitter, and WhatsApp now carry content in dozens of languages. Misinformation that circulates in regional languages can provoke local sentiments and influence health decisions, as demonstrated during the COVID-19 pandemic, when false claims about remedies and vaccines spread faster than official corrections could keep up.

The researchers distinguish carefully between related concepts that are often conflated. Disinformation refers to deliberately created and spread falsehoods, rumors are unverified pieces of information circulated with intent to mislead, and misinformation is false content shared without necessarily intending to deceive. The motivations behind such content range from spreading religious hatred to gaining political advantage or disseminating false health information. Because fact-checking websites themselves can carry biases, and because human verification cannot scale to billions of posts, automated detection systems have become an essential line of defense, and the study argues that these systems must work across the languages people actually use.

At the heart of the new work is a two-level attention architecture that mirrors how humans read documents. The model first processes each sentence word by word using a Bidirectional Long Short-Term Memory network, which reads text in both forward and backward directions to capture context from either side of every word. A custom attention layer then assigns dynamic weights to individual words, learning during training which terms matter most for distinguishing fake from genuine content. The weighted word representations are combined into sentence vectors, and a second Bi-LSTM layer with its own attention mechanism models relationships between sentences, producing a document-level representation that feeds into a final sigmoid classifier.

This hierarchical design is not merely an engineering convenience. Fake news articles often contain misleading lexical patterns, emotionally polarized expressions, and contextual inconsistencies distributed across multiple sentences rather than concentrated in any single phrase. By attending to both words and sentences, the model can capture local semantic cues and document-level structure simultaneously. The attention weights also improve interpretability, allowing researchers to see which parts of an article drove a classification decision, a significant advantage over opaque black-box approaches.

A crucial contribution of the study is linguistic rather than architectural. Because no publicly available fake news dataset existed for Marathi, the team built one from scratch, collecting 1,957 news articles from Marathi outlets including Loksatta, Lokmat, Maharashtra Times, and Zee News, of which 707 were labeled fake and 1,250 true. Labels were assigned based on source credibility, fact-checking reports, and consistency across multiple platforms, with manual review of every article. For Hindi, the researchers combined two existing datasets from Kaggle and GitHub, while English experiments used the widely adopted ISOT dataset. To address severe class imbalance in the low-resource languages, the team applied translation-based augmentation using the IndicTrans neural translation framework, generating additional training samples while keeping the test set untouched to prevent data leakage. The final augmented corpus contained 45,386 English, 10,481 Hindi, and 5,000 Marathi samples.

The choice of word embeddings proved important. FastText, developed by Facebook’s AI Research lab, represents each word as a bag of character n-grams, allowing it to capture morphological information and generate vectors for out-of-vocabulary words from their sub-word components. This property is especially valuable for morphologically rich languages like Hindi and Marathi, where word forms vary extensively. The researchers compared FastText against GloVe and random initialization under identical hyperparameters, finding that pretrained embeddings substantially improved performance, with FastText also converging fastest at 383.06 seconds of training time compared with 501.22 seconds for GloVe.

The experimental results were striking. The HAN model achieved 99.42 percent accuracy on English news, 99.04 percent on Hindi, and 80.90 percent on Marathi, outperforming traditional machine learning baselines including Logistic Regression, Linear Support Vector Machines, and Random Forest, as well as deep learning models such as CNN and standalone Bi-LSTM. Against multilingual transformers, the picture was more nuanced: mBERT reached 98.72 percent overall accuracy and DeBERTa-v3-base 98.66 percent, slightly below the proposed model’s 98.90 percent, but DeBERTa’s performance collapsed to 74.42 percent on Marathi, exposing the sensitivity of large pretrained transformers to data imbalance. Notably, the instruction-tuned LLaMA 3.1 8B model, evaluated through prompting rather than fine-tuning, managed only 67.98 percent accuracy in zero-shot settings and 84.76 percent with six examples in context, with 1,638 failed predictions out of 10,965 test instances.

Ablation experiments confirmed that both attention levels contribute meaningfully. A Bi-LSTM-only model achieved an F1-score of 97.43 percent, rising to 97.82 percent with word-level attention alone and 98.17 percent with sentence-level attention alone, while the complete hierarchical architecture reached 98.72 percent. Sentence-level attention proved more influential than word-level attention in isolation, suggesting that contextual dependencies across sentences play a decisive role in identifying deceptive content. Paired t-tests across multiple random seeds showed all improvements over baselines were statistically significant, with p-values below 0.01.

The study is candid about its limitations. Marathi’s error rate of 19.10 percent, compared with just 0.58 percent for English and 0.96 percent for Hindi, reflects the challenge of low-resource detection, and qualitative analysis showed that the worst failures involved fake articles written in a style so similar to legitimate news that the model assigned incorrect predictions with high confidence. The authors note that generalization to other Indic languages remains unexplored and that multimodal signals such as images and metadata are not yet incorporated. Future work will expand the dataset to additional Indian regional languages and integrate textual and visual information using transformer-based Indic language models. For now, the research demonstrates that carefully designed attention mechanisms, paired with sub-word embeddings and thoughtful data augmentation, can bring state-of-the-art misinformation detection to languages that have long been left behind.

Subject of Research: Multilingual fake news detection using hierarchical attention networks for English, Hindi, and Marathi news articles

Article Title: Hierarchical attention network for multilingual fake news detection

Article References: Nandgaonkar, S., & Mane, S. (2026). Hierarchical attention network for multilingual fake news detection. Discover Artificial Intelligence, 6(1), Article 1417. https://doi.org/10.1007/s44163-026-02425-3

Image Credits: AI Generated

DOI: 10.1007/s44163-026-02425-3

Keywords: fake news detection, hierarchical attention network, multilingual NLP, Bi-LSTM, FastText embeddings, Marathi language, Hindi language, misinformation, deep learning, low-resource languages, data augmentation, natural language processing

The authors trace how social media became the default data source for crisis informatics. Following the massive diffusion of platforms during the first decade of the 2000s, it became clear that they could act as crowdsourced newsfeeds on virtually any topic. Among the options, Twitter quickly emerged as the most suitable for research, thanks to its micro-blogging format of frequent short posts and, crucially, the availability of user-friendly application programming interfaces. Even though full access required a subscription fee, the free tier of the APIs allowed enough data to be collected to support work across many domains. Rival platforms never achieved the same standing in the scientific community: Facebook was difficult to access, while YouTube, Instagram and Flickr posed challenges for processing multi-modal content with the methodologies available at the time.

The analytical toolkit built around Twitter data relied heavily on network analysis to explore community structures and information diffusion, alongside natural language processing to identify crisis events and assess public perception. Yet researchers have long documented the limitations of these data, including demographic and other selection biases among users, technical difficulties posed by the volume, unstructured nature and multilingualism of the content, and the overflow of unverified material that facilitates the spread of misinformation. More recently, the integration of transformer-based architectures and large language models has marked a significant shift in crisis monitoring. BERT-based models have dominated event detection and classification tasks, while decoder-only models such as GPT-4 and LLaMA-2 have gained traction for their reasoning and few-shot capabilities. Even so, the authors note, these models still struggle with the informal language and domain-specific jargon characteristic of crisis communication.

The turning point came with Elon Musk’s acquisition of Twitter in 2022 and the operational changes that followed: the rebranding to X, altered content moderation policies, and, most consequentially for scientists, the end of the free API tier in February 2023. These developments prompted many researchers and practitioners to reconsider their reliance on the platform for crisis communication, marking what the authors describe as a potential end to the era when Twitter served as a primary conduit for scientific discourse during public emergencies. In the wake of these events, alternative micro-blogging platforms such as Bluesky, Mastodon and Threads have emerged and attracted large migrations of users. However, considerable uncertainty remains about when, or even whether, some of these networks will reach the critical mass needed to make them useful for crisis management.

A bibliometric analysis conducted by the team underscores the scale of the disruption. Counting Scopus-indexed articles whose titles mention a specific online social network alongside a crisis management keyword, restricted to four data-oriented subject areas, the authors chart the rise and subsequent decline of platform-specific crisis research. The pattern reflects a broader cooling of enthusiasm for social media data in the field, driven not only by access costs but also by a second, quieter constraint: regulation.

Since May 2018, the European Union’s General Data Protection Regulation has set the terms for how personal data can be collected, stored and reused, and the authors argue it remains the only example of a unified privacy law at continental scale. While the GDPR is now a fundamental component of EU privacy and human rights law, it also represents a significant limitation on how social media data can be used in research projects, particularly EU-funded ones such as Horizon and ERC grants. Reproducibility suffers because researchers are often unable to archive and share datasets containing information that could directly or indirectly identify individuals. Moreover, when users withdraw their consent from a platform, all corresponding records must be removed from datasets, which implies monitoring systems capable of detecting consent withdrawal events such as account deletions and updating datasets accordingly. Over time, as more users delete accounts or revoke consent, datasets shrink, further complicating longitudinal analyses. The authors also raise concerns that Twitter/X’s new policies may no longer fully satisfy key GDPR obligations, and they point to limited but suggestive evidence that people may nonetheless agree to share personal data during emergencies.

What, then, can replace the Twitter firehose? The authors examine the leading alternatives and find that none can fully substitute for the platform’s former role in disaster management. TikTok lacks a real-time public stream. Reddit and Telegram offer only fragmented access, with content dispersed across independent subreddits and channels that must each be individually identified and monitored rather than aggregated into a single feed. Bluesky and Mastodon suffer from extremely limited adoption. These constraints position the alternative platforms as valuable complements rather than standalone replacements, and the authors stress that each platform must be matched to its audience and user behaviors, and to the regional contexts and analytical objectives of a given crisis.

The core of the article’s argument is a call for multi-source crisis monitoring frameworks that combine online news media and radio broadcasts with whatever social media data remain accessible. Beyond improving coverage, such diversification reduces the selection biases inherent to any single source by capturing heterogeneous communities and communication practices. Media-based sources carry their own editorial and political biases, but they offer complementary advantages. Local radio stations in particular remain key information channels in low-connectivity contexts and provide access to under-represented community-level perceptions that are largely invisible on global online platforms; recent work has used machine learning to analyze radio broadcasts about the Ebola outbreak in Uganda, for example. When combined with national and international press, these sources partially compensate for selection biases and contribute to more geographically equitable situational awareness. Compared with user-generated content, media sources also allow better control of misinformation through source filtering and credibility assessment, and they raise fewer legal and ethical challenges under data protection regimes such as the GDPR. The authors argue that collective efforts are needed to establish governance frameworks enabling controlled access to social media data during crises, so that such data can be integrated into operational decision-making.

Moving to multi-source infrastructures introduces new technical problems of its own. The same events are often reported across multiple channels, requiring robust deduplication and information fusion methods to avoid redundancy and contradictory signals. Multilingual processing remains a major bottleneck, especially for under-resourced languages prevalent in many crisis-prone regions; Africa alone has more than 2,000 indigenous languages, and dedicated multilingual models will be needed to handle them. Without focused work on low-resource natural language processing, the authors warn, multi-source systems risk reproducing existing geographical and linguistic inequities in crisis response. Their conclusion is that future crisis management systems must integrate a broad panel of data, including multimodal and multilingual content, while tackling two central challenges: the fusion and ingestion of heterogeneous data, and the mitigation of biases, which generative AI can reproduce and amplify, potentially entrenching long-standing inequalities. The era of one platform, one feed and one methodology is ending, and what replaces it will determine how quickly the world can see a disaster unfolding.

Subject of Research: The impact of restricted social media access and data protection regulation on crisis management research and the shift toward multi-source crisis monitoring

Article Title: Rethinking data sources for crisis management: How restricted social media access is reshaping current practices

Article References: Interdonato, R., Decoupes, R., Roche, M., Syed, M. A., Teisseire, M., & Valentin, S. (2026). Rethinking data sources for crisis management: How restricted social media access is reshaping current practices. PLOS Complex Systems, 3(6), e0000112. https://doi.org/10.1371/journal.pcsy.0000112

Image Credits: AI Generated

DOI: 10.1371/journal.pcsy.0000112

Keywords: crisis management, social media data, Twitter/X API, GDPR, disaster monitoring, natural language processing, large language models, misinformation, multilingual NLP, online news media, radio broadcasts, data fusion

The review’s central finding is a gap between what these models are optimized to do and what global teams actually need. LLMs are trained, evaluated, and celebrated largely on semantic performance: whether a translation preserves meaning, whether a summary is factually consistent, whether a benchmark score improves. But project communication depends not only on what is said, on how, when, and to whom it is said. A semantically perfect translation of an indirect refusal may still fail catastrophically if the recipient expects directness, and a flawlessly worded status update can undermine trust if it ignores local norms about hierarchy and disagreement. The authors call this dimension cultural-pragmatic performance, and they find that current evaluation practices barely touch it.

The methodology behind this conclusion is itself notable for its discipline. The team followed established scoping review guidelines, including the PRISMA-ScR reporting standard and the five-stage framework of Arksey and O’Malley, searching four major databases: Web of Science, Scopus, ABI/INFORM, and EBSCO Business Source, supplemented by targeted screening of ACL and NeurIPS papers and arXiv preprints where methodological detail was sufficient. From 1,452 identified records, the screening funnel narrowed to 228 full-text assessments and finally to 69 included studies. The corpus was organized deductively into four thematic domains: human-AI collaboration, cross-cultural project communication, LLM cultural and multilingual optimization, and the challenge of measuring cultural-pragmatic performance. A CiteSpace keyword co-occurrence network of 312 nodes confirmed that the field is organized around linked but largely separate subareas spanning AI methods, multilingual language technologies, intercultural communication, and project application contexts.

The first thematic domain, human-AI collaboration, reveals a field still deciding what AI even is in this context. Some studies treat AI systems as active communication partners that shape information flow, task coordination, and interaction structure; others treat them as coordination machinery for task execution, resource allocation, and workflow management. Empirical work on generative chatbots shows project managers appropriating them for communication and decision support, with reported gains in documentation and repetitive tasks, but also documented concerns about judgement displacement and communication trust. A third strand examines trust itself as a relational construct, finding that transparency mechanisms, explainable outputs, and clear decision logic matter more to calibrated reliance than raw task performance, and that opaque system behavior or misaligned expectations can quietly corrode it.

The second domain grounds the problem in decades of intercultural communication research. Cultural values demonstrably shape how information is interpreted, how feedback is given, and how conflict emerges in multilingual teams, with cultural intelligence, the capacity to interpret unfamiliar cultural cues and adapt behavior accordingly, consistently associated with better communication outcomes. Studies of high-context and low-context communication styles, drawing on Edward Hall’s classic framework, document how indirect requests, implicit disagreement, and context-dependent emotional regulation create coordination friction when team members operate with different inferential defaults. Perhaps most striking is the evidence on multilingual coordination: language proficiency asymmetries systematically shape participation, influence, and trust formation, meaning that non-native speakers often have less voice in project discussions even when their task expertise is comparable. Language, in other words, is not just a skill but a social cue that distributes power.

Against this backdrop, the technical literature on LLM optimization looks simultaneously impressive and incomplete. Eleven studies in the review focus on cross-lingual semantic accuracy through multilingual pre-training, transfer experiments, and benchmark evaluations, reporting genuine improvements, particularly for high-resource languages, but also persistent disparities for low-resource languages and diminishing returns as language coverage expands. Six studies explore prompting strategies, showing that culturally conditioned prompts, grounded in nation- or community-specific contexts, can shift model outputs in directness, softening strategies, and stance expression. Seven more examine model-level adaptation through culturally conditioned fine-tuning and data augmentation, with measurable but uneven gains that favor well-represented cultures and leave underrepresented ones behind. The pattern is consistent: technical capability is real, but it clusters where data and attention already are.

The evaluation problem is where the review delivers its sharpest critique. The studies that measure LLM performance across languages rely overwhelmingly on semantic or task-oriented metrics. One study found LLM-based translation systems outperformed baselines in roughly 41 percent of evaluated language pairs by BLEU score, a benchmark-specific result rather than proof of general superiority. Another showed that factual consistency varies substantially across languages even when input semantics are stable. Studies of computational politeness can detect features like hedging and honorifics with reasonable accuracy, but treat each feature independently rather than as part of a unified cultural construct. Most damning is the finding on cross-lingual consistency: a model may generate semantically similar answers across languages while reproducing culturally specific expectations about hierarchy, gender roles, or civic responsibility as if they were universal. Consistency, the review argues, is not appropriateness.

There is a constructive thread here too. Research from intercultural communication itself shows that cultural effectiveness can be quantified through multidimensional competence scales covering cognitive, affective, and behavioral dimensions, distinguishing communicative appropriateness and mutual understanding from mere semantic consistency. By analogy, the authors suggest, LLM evaluation frameworks could incorporate an explicit cultural alignment score, assessing whether outputs demonstrate culturally situated interpretation rather than linguistic stability alone. Such a shift would move the field beyond the current situation where technically fluent outputs may still reproduce exclusion, obscure power asymmetries, or weaken trust in exactly the global project environments where the stakes are highest.

The review also situates LLMs within an older and often forgotten ecology of language mediation. Translation and interpreting scholarship has long treated cross-language communication as purposeful, situated professional practice embedded in organizational language policies, workflows, and asymmetrical power relations, not as a lexical mapping problem. When LLMs generate or evaluate multilingual project communication at scale, they enter this ecology without the situated judgement, accountability, and explicit role boundaries that professional mediators carry. The authors are careful not to anthropomorphize: describing LLMs as socio-technical actors does not attribute human understanding to them, but it does identify their effects as outcomes of interactions among model design, organizational practices, professional mediators, users, and governance arrangements. Automation, as translation-technology research has shown, reshapes agency and labor rather than merely adding efficiency.

The practical agenda that emerges is demanding but concrete. Future evaluation should combine semantic performance with human judgment, scenario-based interaction, pragmatic appropriateness, translation-purpose criteria, and governance measures, in line with frameworks such as the NIST AI Risk Management Framework and UNESCO’s ethics recommendation. Organizations should specify when professional translators, interpreters, or bilingual domain experts retain review and decision authority over AI-mediated communication. And the research community itself must confront its blind spots: the review relied on English-language publications, a choice that may underrepresent locally published scholarship and culturally specific practices, shaping both the thematic map and the apparent gaps. The bounded conclusion the authors offer is worth internalizing by anyone deploying AI in a multinational team: LLMs in global project environments warrant analysis not only as productivity tools or language models, but as components of mediated communication operating under conditions of linguistic diversity, cultural difference, and organizational asymmetry. Fluency, it turns out, is the easy part.

Subject of Research: Culturally adaptive large language models in cross-cultural project management

Article Title: Literature review on large language models (LLMs) for cross-cultural project management

Article References: Shen, Y., Cetindamar, D., & Zhang, Y. (2026). Literature review on large language models (LLMs) for cross-cultural project management. AI & SOCIETY. https://doi.org/10.1007/s00146-026-03357-3

Image Credits: AI Generated

DOI: 10.1007/s00146-026-03357-3

Keywords: large language models, cross-cultural communication, project management, cultural intelligence, human-AI collaboration, multilingual NLP, pragmatics, machine translation, AI evaluation, cultural bias, AI governance, scoping review