In Neural Machine Translation (NMT) with Large Language Models (LLMs), prompting has become the predominant approach for adapting to a new translation task without requiring extensive fine-tuning data. However, when translating low-resource language pairs, conventional prompts, built as simple linear text, struggle to represent richer dependency or constituency syntax, making it difficult for LLMs to grasp the source language’s syntactic patterns and semantic nuances and thus impairing translation quality. To address this challenge, this paper proposes Syntax-Aware Structured Prompting (SASP). Since word-level embeddings are insufficient for capturing the overall semantics of a sentence and are susceptible to interference from sentence length and word frequency, we encode source and candidate sentences with sentence-level embeddings and retrieve several semantically similar sentences from the target-language monolingual corpus. Subsequently, each retrieved sentence undergoes fine-grained dependency parsing to extract clause-level subject-verb-object structures as well as part-of-speech information. These syntactic patterns are then organized into clause-level structural templates and integrated with the retrieved example sentences to form a structured prompt, enhancing translation quality. We evaluate SASP on language pairs between Mongolian-Chinese (Mo-Zh), Uyghur-Chinese (Ug-Zh), and Tibetan-Chinese (Ti-Zh) using the CCMT2019 corpus. Experimental results show that SASP consistently improves translation quality across all tasks, achieving up to a 13.4

Large language models often face challenges in low-resource agglutinative language text summarization tasks due to poorly designed prompts, leading to core information dilution, reduced fidelity, and critical information loss caused by the complex grammatical structures of agglutinative languages. For traditional Mongolian, a typical low-resource agglutinative language, this paper proposes a Nucleus-to-Periphery Retrieval-Iterative Prompting (NPRIP). This method first guides the model to extract highly condensed semantic core information (events, persons, time, etc.) from the original text. Subsequently, through multiple rounds of self-refinement iteration, it progressively expands peripheral details (background, causes, consequences, secondary facts, etc.). The model performs fact consistency checks, redundancy removal, and fidelity correction on the current draft, achieving gradual improvements in information completeness and fidelity. To enhance Mongolian language representation, we perform parameter-efficient fine-tuning on LLaMA3-8B using the CCMT2019 Mongolian-Chinese parallel corpus, and construct a larger abstractive Mongolian news summarization dataset MoSum along with its augmented version. Experiments on traditional Mongolian text summarization tasks demonstrate that our proposed method significantly outperforms multiple baseline models on automatic evaluation metrics including ROUGE-1, ROUGE-2, and ROUGE-L. This validates the effectiveness of core-priority structured iterative prompting in low-resource agglutinative language summarization scenarios.

Representation Learning in Neural Machine Translation (NMT) aims to transform language into continuous vector, thereby capturing rich semantic information and enabling efficient translation. However, it typically relies on large-scale data to adequately acquire linguistic semantic and syntactic relationships, which fail to be fully captured in low-resource scenarios, resulting in vector representations with limited generalization capability. Recently, the emergence of pre-trained models has introduced a novel paradigm for low-resource NMT tasks, nevertheless, model’s inherent anisotropy still limits the effect of representation learning. Therefore, this paper proposes a semantic feature optimization method based on contrastive learning, which adds the model’s self-predictions to the negative sample set and leverages BLEU and similarity to rank all samples, further calibrating and optimizing the feature space of the low-resource model from the subtle differences between diverse samples. We conduct extensive experiments on three low-resource translation tasks, which demonstrates the significant advantages of the contrastive training strategy in enhancing representation learning.

Low-resource neural machine translation (NMT) is often hindered by the scarcity of parallel corpora, which significantly restricts the performance improvement of translation models. Traditional data augmentation methods, such as back-translation and lexical substitution, are primarily confined to surface-level transformations at the lexical and syntactic levels, making it challenging to generate training samples that exhibit rich semantic diversity. To address this limitation, this paper proposes a semantic-level parallel corpus augmentation method based on Abstract Meaning Representation (AMR). The proposed method first parses sentences into AMR graphs in order to capture their core semantic structures. Then, multiple augmentation strategies targeting AMR graph structures are designed to generate diverse semantic variants at the semantic level. These variants are subsequently decoded into bilingual parallel corpora using a multilingual AMR-to-text generation model for Mongolian and Chinese. Experimental results on Mongolian-Chinese NMT tasks show that the proposed method improves the baseline by 2.0 BLEU points and significantly outperforms traditional augmentation methods such as back-translation and lexical substitution. This paper offers a novel semantic-level approach to alleviating the corpus bottleneck in low-resource NMT.

To address traditional Chinese Grammatical Error Correction (CGEC) limitations—weak semantic understanding, inflexible error classification, unstable revision—this paper proposes the Ant Colony-Inspired Multi-Agent CGEC Method (ACMA-CGEC), supported by Large Language Models (LLMs). ACMA-CGEC adopts an ant colony-inspired three-tier architecture (global control, information memory, task execution): Queen Ant Agent (global control core) formulates a Disassembling-Analyzing-Processing task chain and enables max 3-iteration backtracking for adaptive scheduling; Global Process Memory Pool (GPMP, sole memory carrier) unifies structured data storage to ensure multi-agent data consistency. In task execution: Disassembling Ant Agent provides standardized linguistic units via semantic-aware segmentation and part-of-speech tagging; Analyzing Ant Agent realizes decoupled syntactic-semantic detection for precise error recognition; Processing Ant Agent uses dual-model collaboration for optimal revision (adhering to minimal modification and original meaning preservation). This end-to-end CGEC method innovates in collaborative architecture, decoupled detection, and adaptive revision, demonstrating advantages in detection accuracy and revision flexibility, and offering a modular, adaptive paradigm for CGEC research.

A key challenge faced by deep neural networks is the gradual loss of input data features during the layer-by-layer feature extraction process, which becomes particularly serious in the task of small object detection. Although the existing theories attribute this performance degradation to the phenomenon of information bottleneck, we propose AGRC-ViMamba, which is a new framework aimed at reducing information attenuation and improving the recognition accuracy of small targets. Our method integrates two key components: (1) Multi-level attention gated reversible column (ML-AGRC), which serves as a reversible auxiliary branch to maintain a reliable gradient flow; (2) The improved feature extractor SPD-Conv-SS2D module based on the VSS architecture. The backbone network adopts this enhanced VSS module and the auxiliary branches inspired by YOLOv9, effectively retaining fine-grained target features during forward propagation while minimizing parameter overhead to the greatest extent. A large number of experiments have proved the effectiveness of this model. Compared with the state-of-the-art (SOTA) method, without reducing the recall rate, the accuracy of this model on the VisDrone and AI-TOD datasets has increased by 1.5 https://github.com/wizardly1985/AGRC-ViMamba .

This paper proposes an innovative deep learning model based on a multimodal framework for efficient keyword spotting in Mongolian lead-type newspaper images. The model integrates visual and textual data to enhance keyword localization accuracy and robustness, supporting both Query-by-Example (QbE) and Query-by-String (QbS) tasks. To address homoglyphic heterogeneity in Mongolian graphemes (i.e., different characters sharing identical glyphs), an intermediate code mapping mechanism based on Unicode standardization is applied, which unifies morphologically identical characters into a normalized intermediate representation layer, achieving a QbS accuracy of 95.57

The scarcity of parallel corpora for Mongolian and Chinese constrains the performance of Mongolian-Chinese neural machine translation (NMT), particularly manifesting in inadequate accuracy in translating specialized terminology. To address this limitation, this study adopts a lexically constrained augmentation strategy that constructs pseudo-source sentences by appending Chinese constraint words to Mongolian source texts, while enforcing the inclusion of these constraints in the output to improve translation accuracy. However, this approach presents two inherent drawbacks: processing pseudo-sentences with a single encoder tends to induce semantic interference, while the introduced constraint words may exacerbate alignment errors during decoding. To overcome these limitations, this paper propose a Constraint-Augmented Mongolian-Chinese NMT method (CANMT) based on dynamic feedback alignment. The method employs a dual-encoder architecture to isolate bilingual representations, coupled with a dynamic feedback alignment module that progressively reduces alignment errors through iterative reffnement, thereby enhancing overall translation performance.

The key challenge in Fine-Grained Visual Classification (FGVC) lies in capturing local subtle differences. Although Transformers excel at modeling long-range dependencies, their patch partitioning scheme in image processing tends to weaken the association between local and global features. While sliding windows alleviate neighborhood fragmentation, they still struggle to avoid the block-like discretization of features. To address this issue, we propose the Mamba Concept-Guided Learning (M-CGL) framework, which consists of two novel components: the Mamba Semantic Concept Modeling (M-SCM) module and the Mamba Semantic Concept Fusion (M-SCF) module. The M-SCM module enhances the inter-relationships among fine-grained features to extract more discriminative representations. The M-SCF module fuses discriminative features and feature maps across multiple stages, enabling hierarchical concept alignment and preserving spatial-semantic consistency throughout the network. While M-SCF ensures intra-sample semantic coherence, it neglects inter-sample structural relations. Thus we adopt SoftTriple loss to explicitly enforce intra-class compactness and inter-class separability, enhancing discrimination among visually similar categories. Experiments show that our method achieves accuracy gains of 0.54% on CUB-200, 1.56% on Aircraft and 5.5% on Fiber, respectively.

Currently, data augmentation is a primary technique for improving the performance of Neural Machine Translation (NMT) in low-resource settings. However, traditional data augmentation methods, while alleviating issues related to data sparsity, are often influenced by noise, leading to syntactic errors and semantic discrepancies, which in turn degrade the quality of NMT outputs. To address this challenge, this paper proposes a self-supervised data augmentation approach that integrates contrastive learning to pull similar features closer together, thereby reducing the noise introduced by conventional augmentation techniques. Moreover, to resolve the issue of excessive low-frequency words in traditional low-resource agglutinative language NMT, commonly used solutions such as stem-and-affix segmentation can preserve basic semantic information but depend on manually curated dictionaries, which lack flexibility. Although Byte Pair Encoding (BPE) is a statistical method, it fails to capture word-level semantic features. In light of these challenges, this paper introduces a morphological recombination approach to further enhance translation quality. Specifically, we propose a convolutional gated morphological attention mechanism in the encoder to capture and amplify morphological features within the input sequence, while a morphological cross-attention mechanism in the decoder ensures that these features are effectively leveraged to guide the translation process. Experiments conducted on the Mn-Zh task using various gated attention mechanisms demonstrate an average BLEU score improvement of 2.62

Back-translation has been proven effective in enhancing the performance of Neural Machine Translation (NMT), with its core mechanism relying on synthesizing parallel corpora to strengthen model training. However, while traditional back-translation methods alleviate the data scarcity in low-resource machine translation, their dependence on random sampling strategies ignores the semantic quality of monolingual data. This results in the contamination of model training through the inclusion of substantial low-quality samples in the generated corpora. To mitigate noise interference, additional training iterations or model scaling are required, significantly increasing computational costs. To address this challenge, this study proposes a Semantic Uncertainty Sampling strategy, which prioritizes sentences with higher semantic uncertainty as training samples by computationally evaluating the complexity of unannotated monolingual data. Experiments were conducted on three typical low-resource agglutinative language pairs: Mongolian-Chinese, Uyghur-Chinese, and Korean-Chinese. Results demonstrate an average BLEU score improvement of +1.7 on test sets across all three translation tasks, confirming the method’s effectiveness in enhancing translation accuracy and fluency. This approach provides a novel pathway for the efficient utilization of unannotated data in low-resource language scenarios.

Sequence-to-sequence neural machine translation (NMT) has achieved great success with many language pairs. However, its performance remains constrained in low-resource settings such as Mongolian–Chinese translation due to its strong reliance on large-scale parallel corpora. To address this issue, we propose ILFDN-Transformer, a Mongolian–Chinese NMT model that integrates implicit language features and a deliberation network to improve translation quality under limited-resource conditions. Specifically, we leverage the BART pre-trained language model to capture deep semantic representations of source sentences and apply knowledge distillation to integrate the resulting implicit linguistic features into the Transformer encoder to provide enhanced semantic support. During decoding, we introduce a deliberation mechanism that guides the generation process by referencing linguistic knowledge encoded in a multilingual pre-trained model, therefore improving the fluency and coherence of target translations. Furthermore, considering the flexible word order characteristics of the Mongolian language, we propose a Mixed Positional Encoding (MPE) method that combines absolute positional encoding with LSTM-based dynamic encoding, enabling the model to better adapt to complex syntactic variations. Experimental results show that ILFDN-Transformer achieves a BLEU score improvement of 3.53 compared to the baseline Transformer model, fully demonstrating the effectiveness of our proposed method.

Globally, with over 7, 000 languages, only about 20 are well-resourced. Low-resource languages face bottlenecks in traditional supervised learning and fine-tuning of pre-trained models due to data scarcity and informal text usage. Research on sentence representation for low-resource languages is crucial for overcoming data scarcity constraints, enhancing semantic modeling capabilities, and promoting the equitable development of natural language processing technology across the global linguistic landscape, as well as advancing various downstream task applications. This paper surveys the evolution of research in lowresource sentence representation, focusing on analyzing the core concepts, effectiveness, and limitations of traditional methods, cross-lingual knowledge transfer, and contrastive representation learning methods. It also discusses prospects for the development of low-resource sentence representation learning.

This paper presents the Mongolian Meta-Adapter (MMA), An innovative methodological system developed to simultaneously mitigate lexical formation intricacies and annotation scarcity in Mongolian affective computing. MMA integrates three key components: a rule-based morphological tokenizer, a parameter-efficient meta-learning scheme, and cross-lingual adversarial training. To preserve the semantic consistency of Mongolian's agglutinative structure, we develop a rule-guided morphological parser with auto-annotation, achieving 92.7

To address the challenges of slow synthesis speed, unstable quality, limited emotional expressiveness, and the lack of controllable speaking rate in Mongolian emotional speech synthesis, this paper proposes a speech Rate-Controllable Mongolian emotional speech synthesis model based on improved Tacotron2 (SRC-IT2). First, an end-to-end Mongolian speech synthesis module is constructed based on an improved Tacotron2 framework, incorporating the unique linguistic characteristics of the Mongolian script. The front-end processing is optimized accordingly, and a G2P-Seq2Seq model is employed to achieve accurate grapheme-to-phoneme conversion for Mongolian characters. Next, on top of the end-to-end synthesis framework, a joint text-audio emotion analysis module is integrated to effectively learn and represent emotional style features specific to Mongolian speech. Finally, a style encoder and speaking rate control variable are embedded into the acoustic modeling process, further enhancing Tacotron2’s ability to dynamically adjust the speaking rate during emotional speech generation. Experimental results demonstrate that the proposed model produces more natural-sounding speech with improved emotional expressiveness and enables effective real-time control over speaking rate in Mongolian emotional speech synthesis.

In Mongolian Text Sentiment Analysis (MTSA), the scarcity of annotated sentiment datasets and the insufficient consideration of syntactic dependency and topological structural information pose significant challenges to accurately capturing semantics and effectively extracting emotional features. To address these issues, this paper proposes a Mongolian Text Sentiment Analysis model based on ConvBERT and Graph Attention Network (MTSA-CG). Firstly, the ConvBERT pre-trained model is employed to extract textual features under limited data conditions, aiming to mitigate the shortcomings caused by data scarcity. Concurrently, textual data are transformed into graph-structured data, integrating co-occurrence, dependency, and similarity information into a Graph Attention Network (GAT) to capture syntactic and structural cues, enabling a deeper understanding of semantic and emotional connotations for more precise sentiment classification. The proposed multi-graph fusion strategy employs a hierarchical attention mechanism that dynamically weights different graph types based on their semantic relevance, distinguishing it from conventional graph aggregation methods. Experimental results demonstrate that, in comparison with various advanced baseline models, the proposed method significantly enhances the accuracy of MTSA.

Chinese Spelling Correction is a critical task in natural language processing, yet traditional methods often suffer from weak global semantic coherence, poor domain generalization, and uncontrollable corrections. This paper proposes C2F-MAC, a Coarse-To-Fine framework integrating Large Language Model (LLM)-based domain adaptation with Multi-Agent Collaboration. In the coarse-grained phase, lightweight fine-tuning via Low-Rank Adaptation (LoRA) significantly reduces computational overhead while enhancing cross-domain terminology recognition. The fine-grained phase introduces a multi-agent mechanism driven by state machine protocols, forming a closed-loop "detect-correct-validate" workflow: agents collaboratively perform error localization, minimal-edit corrections, and necessity verification. Experiments demonstrate that C2F-MAC outperforms baseline models in cross-domain scenarios, achieving a balanced optimization of correction fidelity and resource efficiency. This work provides a robust solution for Chinese Spelling Correction with enhanced domain adaptability and controllability.

Improving neural machine translation (NMT) for the Mongolian-Chinese language pair is challenging due to the lack of high-quality parallel data. This study explores various noise enhancement techniques to enhance the Mongolian-Chinese Neural Machine Translation (MNMT) model’s translation quality. Techniques such as swap, token, delete, and source. Experimental results show that these methods significantly improve translation quality, with the source method yielding the most substantial enhancement. These findings indicate that noise enhancement effectively addresses data scarcity and quality issues, providing a robust strategy for improving MNMT performance.

To address the inadequacies in feature extraction and fusion in Mongolian Speech Emotion Recognition (SER), this paper proposes a Mongolian SER model based on the Whisper (MSER-Whisper). Firstly, a method is introduced to process the layer-wise feature representations extracted from the encoder of the Whisper model. Log-Mel spectrograms are fed into Whisper to obtain comprehensive multi-layer encoder features. After dimensionality reduction, an attention mechanism is constructed over these features to identify the layer-wise weight distributions corresponding to different emotion categories. This approach not only leverages the shared representations learned from large-scale training data by the open-source Whisper model—thus reducing the learning difficulty under low-resource Mongolian scenarios—but also fully exploits the acoustic and speech information captured at different layers of the pre-trained model. In addition, the model employs a cross-attention mechanism, where the final Prosodic Features (PFs) guide the final Spectral Features (SFs) to facilitate correlated emotion mining, enabling dynamic adjustment of weight information for effective feature fusion. Experimental results demonstrate that, compared to various advanced baseline models, the proposed method significantly improves the accuracy of Mongolian SER.

To address the limitations of monolithic analysis and uncontrollable revisions in traditional Chinese grammatical error correction (CGEC) methods, this paper proposes a multi - agent collaborative CGEC framework with three core capabilities—memory capability, execution capability, and planning capability. Driven by large language models (LLMs), the framework functions via: (1) the Tokenization Agent generating structured part-of-speech sequences as semantic primitives; (2) the Syntax Validation Agent and Semantic Verification Agent conducting parallel formal rule detection and logical-semantic verification to identify conflicts and inconsistencies; (3) the Error Classification Agent integrating multi-source features for dynamic error classification and precise localization with linguistically logical labels; and (4) the Error Correction Agent adhering to the minimal intervention principle to generate and select optimal revision plans via LLMs and a three-dimensional evaluation model. Experiments show the framework significantly boosts error correction performance, with MACGEC achieving an F0.5 of 44.13 on NaCGEC, outperforming DeepSeek-R1 32B COT (29.96) by 47.3