搜索结果：Multilingual

In this paper, we test the efficacy of multilingual transformer-based large language models for spam classification tasks on the low-resource Kazakh language. Due to the lack of sufficient labeled data for spam emails written in the Kazakh language, we evaluate the ability of modern multilingual models for generalized tasks, especially the gains obtainable through supervised learning. In this analysis, we test a range of popular models such as bert-base-multilingual-cased, distilbert-base-multilingual-cased, and xlm-roberta for spam detection that are either trained for zero-shot tasks or through parameter-efficient supervised learning using LoRA. The experimental findings show that the zero-shot accuracy of multilingual LLMs is inconsistent, especially for the less common spam class, where the models tend to lack the ability to generalize because of the underrepresented Kazakh corpus. Yet, the models achieve considerable accuracy after being trained on a labeled corpus, where all models achieve accuracy above 0.99, especially XLM-R, which achieves a macro-F1 value of 0.99 and substantially reduces false negatives. A closer examination of these misclassified instances shows that common linguistic patterns, such as marketing necessities, financial jargon, or the conversational tone, play a significant role in their misclassification. The experimental results indicate that a limited amount of annotated instances for the Kazakh language, leveraging the efficiency of LoRA-based fine-tuning, can be effectively combined to improve the multilingual models' performance even on security-related tasks.

Language bias arises in systematic reviews when non-English studies are excluded owing to resource constraints. Large language models (LLMs) can mitigate this problem through multilingual processing. To assess whether direct multilingual LLM processing reduces language-based disparities in systematic review screening performance compared to translation-mediated approaches. Six state-of-the-art LLMs were evaluated under three conditions: (1) an English benchmark dataset (n = 2,911), (2) direct screening of non-English abstracts (n = 483), and (3) screening of machine-translated non-English abstracts. Performance was measured using sensitivity, specificity, F1 score, balanced accuracy, and workload reduction. All models achieved high sensitivity on English data (≥0.938). Translation-mediated screening substantially reduced sensitivity in some models (range: 0.47-0.54), whereas direct multilingual processing maintained high sensitivity (range: 0.71-1.00). Considerable differences were observed among models. Direct multilingual LLM screening may reduce language-related sensitivity disparities; however, the effects on downstream meta-analytic bias require further investigation.

Extended Reality (XR) technologies offer transformative potential for language education, yet current platforms largely neglect the accessibility needs of deaf and hard-of-hearing individuals. Existing solutions typically operate in single-language environments and lack integrated support for sign language and multimodal communication. There is a critical need for inclusive platforms that serve both deaf and hearing learners through cross-modal AI services embedded in immersive environments. This study presents a modular platform integrating six AI services: speech-to-text transcription (OpenAI Whisper), multilingual translation (Meta NLLB), text-to-speech synthesis (AWS Polly), sentiment analysis (RoBERTa), session summarisation (flan-t5-base-samsum), and International Sign (IS) translation via Google MediaPipe. An IS dataset of 750 gesture videos was processed to extract hand landmark coordinates mapped to 3D avatar animations within a Unity-based VR environment on Meta Quest 3 headsets. The system was validated through technical benchmarking of AI service performance, including comparative evaluation of text-to-speech services and multilingual translation models (NLLB-200 and EuroLLM 1.7B variants), load testing to assess platform. scalability, and end-to-end pipeline latency measurement for both the hearing and the deaf user pathways. The educational scenario was additionally evaluated in a companion pilot study, 50 which shares the same underlying AI services and provides complementary user-perception evidence. Technical benchmarking confirmed the platform's viability for real-time XR deployment. TTS benchmarking confirmed AWS Polly's lowest latency (50-100 ms first byte) at competitive cost. The EuroLLM 1.7B Instruct model achieved a BLEU score of 84.34, outperforming NLLB's 79.25. Load testing with 1,000 simulated concurrent users demonstrated average response times under 800 milliseconds with no critical failures. Avatar animation latency for IS sign rendering remained consistently under 300 milliseconds. End-to-end pipeline latency averaged 2.05 ± 0.31 s for the hearing pathway and 2.32 ± 0.34 s for the deaf (IS) pathway, both within accepted thresholds for conversational educational use. The companion pilot (N = 10) reported a mean overall experience rating of 4.6/5.0, 92% user satisfaction and unanimous (100%) demand for expanded language and sign-language support. 50. The results presented in this study focus on the technical feasibility of integrating cross-modal AI services within XR environments for accessible, multilingual language learning. The modular architecture enables independent scaling and adaptation to diverse contexts, laying the groundwork for equitable educational solutions aligned with EU digital accessibility objectives. Learning a new language can be challenging, and it is even more difficult for deaf individuals who rely on sign language. This study addresses the challenge by creating a virtual reality (VR) learning environment where a digital 3D character (avatar) can speak, translate, and perform sign language in real time. The system uses several artificial intelligence tools working together: one that converts speech into text, another that translates text between multiple languages, a third that converts text back into spoken language, and a fourth that translates text into International Sign Language gestures performed by the avatar. Users wear a VR headset and interact with the avatar in a virtual classroom where they can select their preferred language and receive immediate translations in both spoken and signed forms. The system’s technical performance was validated through benchmarking of AI translation models, text-to-speech services, end-to-end pipeline latency measurement, and scalability testing, confirming its suitability for real-time educational applications. Complementary user feedback gathered in a companion pilot study, which uses the same underlying AI services, is cross-referenced where relevant. 50 This research is a step toward creating virtual learning environments where language barriers and hearing limitations no longer prevent people from accessing education.

This study addresses the lack of global, long-term analyses of media narratives on Russia and Ukraine by examining their evolution from 2013 to 2024. Based on 22.2 million article titles from 14,622 sources operating in 194 countries and covering up to 74 languages, this study employs a hybrid human-AI approach combining multilingual clustering, cluster linking, manual annotation, entity analysis, and analysis of the language distribution of articles by Russian sources. We show that media coverage increased sharply around key geopolitical events, including the Euromaidan protests (2013-2014), the annexation of Crimea (2014), and Russia's full-scale invasion of Ukraine (2022). Multilingual clustering, in combination with manual annotation, identifies recurring patterns in coverage, including geopolitical tensions between Russia and the West, energy security, and disinformation campaigns. Cluster linking reveals how these patterns evolved over time in response to major events: clusters in 2013-2014 focused on the Russia-Ukraine gas dispute, the Euromaidan protests, and Crimea, while those in 2021-2022 centred on the full-scale invasion and its global repercussions. Entity analysis shows that media attention consistently concentrated on a small set of political figures, including Vladimir Putin, Volodymyr Zelenskyy, and Donald Trump, whose prominence varied across different phases of the conflict and key political events. Analysis of Russian media further indicates sustained publication in foreign languages, with increases during key geopolitical moments and differences in how the conflict is contextualised across language groups. Despite uneven regional and linguistic coverage, this study provides a large-scale, long-term, and multilingual overview of media coverage. Future research could extend this work through cross-country comparisons, linking narrative shifts to external variables, such as public opinion data or electoral outcomes, and applying the methodology to other large-scale datasets.

Large Language Models (LLMs) are being increasingly incorporated into decision-support systems. Nonetheless, a lack of clarity remains with reference to their reasoning processes, particularly in multilingual contexts. This uncertainty extends to cognitive biases - systematic errors in judgment, similar to those documented in human cognition. Existing research on cognitive biases in LLMs has focused primarily on English-language settings and a limited range of model families, leaving open the question of whether bias manifestations differ across input languages and distinct architectures. The current study investigated the ability of three widely used LLMs (ChatGPT, Claude, and Gemini) to solve cognitive tasks targeting availability heuristics and confirmation bias, comparing their performance to a human control group. The tasks were administered in English, Hebrew, and Russian, representing Germanic, Semitic, and Slavic linguistic contexts. The analytical dataset comprised 2,028 observations: 507 human responses, collected via dedicated online questionnaires, and 1,521 LLM-generated responses, obtained through API interfaces. Statistical analyses implemented Pearson's Chi-Square tests, post hoc comparisons with Bonferroni correction, logistic regression, and Firth penalized logistic regression to compare correctness patterns across models, tasks, and languages, with human performance serving as a baseline. The results revealed a "cognitive gap": LLMs consistently outperformed human participants on the rule-based deductive task, yet exhibited bias-mimicking error patterns in heuristic reasoning-based tasks. The observed effects varied significantly across languages and models, challenging the expectation of uniform multilingual performance and suggesting that LLM architecture interacts with linguistic structures in unpredictable ways. Overall, the findings indicate that cognitive bias expression in LLM outputs is not merely a technical constraint but a language-dependent phenomenon with practical implications for deployment in multilingual environments. The study emphasizes the need for cross-linguistic evaluation when assessing the reliability of LLM-based decision-support systems, particularly in domains where biased reasoning may affect judgment and decision quality.

Speech-language pathologists have important opportunities to deliver culturally and linguistically appropriate services for children with complex communication needs. International research has explored services in relation to available augmentative and alternative communication systems and collaboration with families, yet little is known about Australian speech-language pathologists' practices. The present study explores how Australian speech-language pathologists carry out their work with children with complex communication needs in multilingual families. Twenty-three (n = 23) Australian speech-language pathologists participated in semi-structured interviews about their experiences working with children who have complex communication needs in multilingual families. Transcribed interview data were analysed with a thematic analysis approach. One of the four main themes is explored in this paper, specifically how speech-language pathologists work to address challenges in providing interventions. Participants described how they work to: (a) Align their own and family expectations, (b) find shared and relevant ways to communicate with families, and (c) tailor resources to child and community needs. Findings raise similar challenges to international studies in this area, including the complexity of creating multilingual augmentative and alternative communication systems or adapting existing systems. Participants' experiences demonstrate both positive achievements or gains towards meeting families' needs and identify challenges faced.

South Africa's speech-language therapy (SLT) and audiology professions face challenges in achieving linguistic and cultural integration (i.e., the meaningful incorporation of diverse linguistic and cultural perspectives into training and practice), a critical aspect for effectively serving the country's diverse population. Limited curriculum content on indigenous languages and cultural competence, along with low diversity of academic and clinical training staff (staff), may hinder students' preparedness for multilingual and multicultural clinical practice. This study explored undergraduate students' views and self-reported practices regarding linguistic and cultural integration during their professional training. To explore the views and practices of South African SLT and audiology undergraduate students concerning linguistic and cultural integration in their training. A cross-sectional convergent mixed-methods survey design was used to gather quantitative and qualitative data from a purposive sample of 48 third- and fourth-year SLT and audiology students across four South African universities. Data were collected using a structured online questionnaire, including Likert-scale items and open-ended questions. Quantitative data were analysed using descriptive and inferential statistics, while thematic analysis was applied to qualitative responses. Within this sample, participants recognised the importance of linguistic and cultural competence but report feeling inadequately prepared to implement these skills in clinical settings. Key barriers included insufficient curriculum coverage of cultural topics, reliance on untrained interpreters, and a lack of bilingual resources. Students from indigenous language backgrounds reported higher ratings of the importance of linguistic integration than their English-speaking peers (p < .05). Qualitative themes suggested a perceived need for greater curriculum responsiveness, increased staff diversity, and enhanced institutional support to facilitate culturally competent practice. These exploratory findings suggest that curriculum responsiveness, institutional support, and staff diversity may warrant further consideration within ongoing efforts aimed at improving student preparedness for practice in South Africa's multilingual and multicultural healthcare environment. These findings point to the need for ongoing educational transformation efforts to better serve the diverse needs of the South African population.

This study developed and validated monolingual and bilingual sentence-bidirectional encoder representations from transformers (SBERT) models for detecting cancer recurrence within Thai-English electronic medical records (EMRs) from Thai cancer hospitals. A multicentre dataset of 32&#x2009;436 documents from 1250 patients was used for model development. External validation involved an independent dataset of 9244 documents from 384 patients across two Thai cancer hospitals. Performance was benchmarked against a fine-tuned PubMedBERT (MetBERT). The development dataset included breast (43.9%), colorectal (12.1%), cervical (28.0%) and head and neck (16.0%) cancers. MetBERT achieved the highest area under the precision-recall curve (AUPRC) for locoregional versus no recurrence (11.1%) and locoregional versus distant recurrence (91.7%), while monolingual-SBERT excelled at distant versus no recurrence (32.0%). External validation demonstrated MetBERT superiority for locoregional versus no recurrence (9.30%-21.50%). For distant versus no recurrence, bilingual-SBERT performed best with AUPRC 17.55%-24.39%. While MetBERT led in distinguishing locoregional versus distant recurrence (88.30%-94.70%), bilingual-SBERT demonstrated robust external validation performance (AUPRC 85.25%-91.80%). Low AUPRC values (9%-32%) reflect the extreme class imbalance in real-world data (~1% recurrence prevalence). Despite this, fine-tuned MetBERT achieved highest performance, while bilingual-SBERT demonstrated superior robustness during external validation. This validates sentence embedding models for handling mixed Thai-English medical records in multilingual clinical environments. Sentence embedding frameworks provide a practical, generalisable solution for detecting cancer recurrence within multilingual EMRs. Despite text-length constraints, these models are suitable for clinical integration as a screening tool for cancer registry workflows.

Dysarthria is a frequent motor speech disorder following a stroke, affecting up to 42% of survivors and resulting in reduced speech intelligibility and diminished quality of life. Clinical assessments, such as the Frenchay Dysarthria Assessment, Second Edition (FDA-2), rely heavily on the subjective judgment of speech-language pathologists (SLPs), which limits comparability and scalability. Telepractice solutions have the potential to extend access to care, but validated digital tools that combine automatic analysis with clinically usable interfaces remain scarce. This study aimed to develop and evaluate a web-based application that integrates automatic speech recognition (ASR) and acoustic analysis into a user-centered dashboard for SLPs. Specifically, we investigated: (1) whether ASR can provide intelligibility scores comparable to those of human listeners; (2) the usability of the system in 2 iterative cycles with SLPs; and (3) the feasibility of presenting clinically relevant acoustic features to support telerehabilitation. A user-centered design process was followed, involving contextual inquiry, requirements gathering, prototype development, and iterative testing with SLPs. The analytical core of the prototype included an ASR module (Whisper Large-v3) to compute intelligibility scores, combining word error rate-based accuracy with sentence-level and word-level alignment. Phoneme-level error highlighting was implemented to identify frequent substitution or deletion patterns. In parallel, an acoustic module extracted clinically relevant measures, including fundamental frequency (mean and range), intensity (mean and variability), and vowel formants (F1-F2 space), supplemented by sustained phonation duration. A pilot validation compared ASR-based intelligibility scores with transcriptions from 8 lay listeners for 3 patients with dysarthria performing the Frenchay Dysarthria Assessment-2 word and sentence tasks. Usability was evaluated in 2 cycles with 8 and 4 SLPs, respectively, using the System Usability Scale and structured questionnaires. In the pilot validation, ASR performance was comparable to, and in some cases better than, untrained human listeners for individuals with mild and moderate dysarthria, though performance declined with severe cases. Both usability cycles yielded excellent System Usability Scale scores (cycle 1: mean 88.4, SD 4.6; cycle 2: mean 91.7, SD 4.1). Core workflow elements, including navigation, session upload, and intelligibility score presentation, were consistently rated highly. Feedback evolved from bug reports and requests for clearer terminology in cycle 1 to suggestions for advanced analytic features in cycle 2, such as additional voice-quality indices and integrated note-taking. The prototype demonstrates that automatic intelligibility scoring and acoustic analysis can be integrated into a clinically usable, web-based dashboard. While current limitations include reliance on English-only phoneme analysis, limited advanced acoustic features, and lack of regulatory compliance, the application achieved excellent usability and shows promise for scalable telerehabilitation. Future work should expand multilingual support, incorporate additional acoustic measures, and validate the tool in larger clinical cohorts.

The aim of this quality improvement initiative was to increase colorectal cancer (CRC) screening rates from 17.75% and 28.45% to 40% in patients aged 45-49 years and 50-75 years, respectively, at a resource-limited primary care clinic within 1 year. We identified a major gap in CRC screening in both age groups (45-49&#x2009;and 50-75 years). We implemented multifaceted, patient-centred interventions in a community clinic serving a diverse population. The intervention incorporated the use of stool-based testing (faecal immunochemical test and multi-targeted stool DNA test), colonoscopy referrals, patient navigation, multilingual education, electronic health record alerts and a real-time tracking dashboard. The primary outcome measure was CRC screening completion rates. Colonoscopy and stool test completion rates were the process measures. In 45-49 years, we observed a sustainable, steady increase in CRC screening rates from 17.75% (n=400) to 41.25% during the study period and 59.50% 6&#x2009;months poststudy, with a median rate of 52.38% in a run chart and a mean of 37.00% in the statistical process control (SPC) chart. In 50-75 years, we observed a sustainable increase in CRC screening rate from 28.45% (n=2879) to 46.16% during the study period and to 61.06% 6&#x2009;months poststudy, with a median rate of 55.47% in a run chart and mean of 43.00% in a SPC chart. In 45-49 years, colonoscopy completion rates increased from the baseline rate of 25.93% (n=108) to 37.04% during the study period and to 43.52% 6&#x2009;months poststudy. In 50-75 years, colonoscopy completion rates increased from the baseline rate of 74.01% (n=654) to 92.05% during the study period and to 97.55% 6&#x2009;months poststudy. We exceeded our goals to increase CRC screening in both age groups. Patient-centred care and system-integrated interventions may substantially increase CRC screening among underserved individuals.

Session dialogue assessment based on machine learning is gradually becoming an effective solution for therapeutic alliance measurement which is an important factor for successful psychotherapy. However, most existing models assume clean and pre-structured dialogue transcripts, whereas real-world counseling documentation often contains heterogeneous case reports. This gap limits the applicability of current automated assessment models in realistic documentation scenarios. In this work, we propose a framework for automated working alliance assessment from complex, multilingual reports. First, language-specific BERT models are fine-tuned to process case reports across different languages, enabling accurate speaker role delineation and dialogue structuring. Second, Gemini-2.5-Flash is leveraged to annotate the dialogues with working alliance ratings. Third, a hybrid feature representation strategy is then developed to jointly capture linguistic style and semantic content from the counseling dialogues. Furthermore, an entropy-based mutual information analysis is conducted to identify the most informative linguistic features. Finally, the extracted hybrid features serve as inputs to XGBoost for alliance assessment. In experiments, the proposed framework shows better performance in the comparison with SOTA methods and generalization ability.