Jul 2026· International Conference on Control, Decision and Information Technologies· pp. 1185-1190· 0 citations· 20 references
Abstract
The challenge of cross-lingual semantic similarity detection is a significant problem in the context of multilingual educational software tools. This paper proposes a novel approach using ensemble learning and knowledge distillation for the development of an efficient and interpretable cross-lingual semantic similarity detection model for the English-French language pair. The methodology is based on the fusion of knowledge from the MiniLM encoder representation using a lightweight attention mechanism, LaBSE encoder with support for language-independent semantic representations, and the BERT encoder with the ability to produce dense contextual vector representations. The knowledge is then distilled using a Multi-Layer Perceptron (MLP) architecture for the development of the semantic similarity detection model. The experimental results show that the ensemble architecture attains a validation F1-score of 0.930, while the knowledge distillation student model retains a robust F1-score of 0.918 with a low computational footprint (3.17M parameters, 42MB memory, 5.77ms inference). This demonstrates the viability of knowledge distillation for the transfer of ensemble-level semantic knowledge into a compact architecture for the context of resource-constrained educational tools.
Semantic text similarity is an especially hard task that can be performed in the Uzbek language because of its rich morphological and the absence of annotated data. The given paper introduces a very effective Bidirectional Encoder design using the monolingual model named “Bidirectional Encoder Representation from Transformers for Uzbek language” and optimized to provide scalable semantic search. A combination of knowledge distillation and metric learning are applied in a semi-supervised approach that is employed by the model. Triplet Loss with Hard Negative Mining is used to enhance the discriminative ability of the vector space. One such innovation is supporting the Matryoshka Representation Learning, which allows the model to produce dynamically truncated dimensions of embeddings. The architecture proposed wound up with a Spearman correlation of 0.835 within the Uzbek test set, topping the state-of-the-art results. Furthermore, Matryoshka Representation Learning achieves 6-fold compression of vectors with a negligible accuracy drop (ρ=0.816), having high-computation efficiency to implement Natural Language Processing systems in low-resource systems.
B. Muminov, N. Allaberganova, Olimjon Mamadiyorov et al.· International Conference on...· 0 citations
This paper proposes SeSyCo, a Semantic-Symbolic Knowledge Consensus framework, which leverages the semantic space to diverge monolingual queries into broad multilingual evidence, and subsequently utilize the symbolic space to eliminate language discrepancies, converging the gathered information into a robust consensus for precise SPARQL generation.
Yu Zhang, Ran Song, Xiaofei Gao et al.· Proceedings of the 32nd ACM...· 0 citations
The production of misinformation on digital platforms has been a concern in recent years, especially in interlingual contexts – people who create content in different languages and language elements that are integrated into written content.the existing fake news detectors use mostly only one language, and they are not adequately cross-linguistically generalizable in using the valuable detectors especially with low-resource languages. To overcome these weaknesses, the present paper suggests an LLM-based multilingual fake news detector which combines cross-lingual semantic alignment and contextual reasoning. The proposed approach involves the fine-tuned multilingual language model, which involves a transformer to derive contextual embeddings of multilingual text. The semantic alignment process (contrastive) is used to project the representation of the various languages to a common embedding space, and hence the transfer of knowledge. Moreover, semantic inconsistencies and misguiding tendency contribute to the reasoning of the model deducing abilities applied in the presence of attention processes. Experiments with the mixed English, Hindi, and Tamil data reveal that the proposed model leads to the accuracy of 94.3 and the F1-score of 93.6, which is better than a baseline model, including SVM, LSTM, and mBERT. The results enable mentioning the great overallization possibility and low-resource, multilingual applicability of the model. The paper gives a scaled and efficient approach to tackling the real-life aspect of multilingual fake news detection.
T. Divya, A. Meenakshi· International Conference Com...· 0 citations
This paper proposes a cross-lingual semantic modeling algorithm design for language transfer learning, aiming to improve the quality and efficiency of multilingual semantic representation by using better semantic alignment and semantic fusion strategies. This method employs a deep neural network semantic representation model and combines statistical alignment methods for cross-lingual semantic alignment, thereby expanding the common semantic domains between languages. To address the issue of low-resource languages, a transfer learning strategy of task adaptation and feature adaptation is used, which significantly enhances the cross-lingual learning effect. Additionally, the integration of multimodal information is utilized for improved semantic restoration, thereby enhancing the stability of semantic expression. Experimental results show that this method outperforms traditional methods in terms of sentence meaning accuracy, convergence speed, and computational efficiency. The system can be applied in cross-lingual language environments and has good portability.
Zhenhua Lin· The 2026 International Confe...· 0 citations
This study proposes a bilingual academic chatbot based on a semantic retrieval approach using the Multilingual BERT (mBERT) transformer architecture to support academic information services in higher education. The dataset was constructed from official academic information at Garut Institute of Technology, including new student admissions, academic calendars, institutional profiles, and lecturer and staff data. The data were organized in a bilingual question–and–answer format in Indonesian and English. The mBERT model was fine-tuned using a Sentence-BERT framework to generate sentence embeddings for semantic retrieval tasks, with MultipleNegativesRankingLoss applied during training. Model performance was evaluated using BERTScore to measure semantic similarity between chatbot responses and human reference answers. Experimental results show that the fine-tuned model outperformed the base model, achieving an average F1-score improvement from 0.7638 to 0.8152 for Indonesian and from 0.7556 to 0.8005 for English. The results also demonstrate more stable score distributions, indicating consistent semantic performance. The optimized model was subsequently integrated into a web-based prototype to enable real-time bilingual academic question answering. These findings confirm that combining mBERT with semantic retrieval effectively enhances the relevance and contextual accuracy of chatbot responses, thereby supporting digital transformation and improving the efficiency of academic services in higher education.
This paper proposes a general optimization framework that combines a vocabulary pruning method with a targeted fine-tuning protocol for MNMT models, and reduces the vocabulary size from over 128,000 to approximately 10,000 tokens, enabling a 60% memory saving without any loss in performance.
Ahmed Amine Aliane, N. Semmar, H. Aliane· 0 citations