Enhancing NLP for low-resource languages (LRLs) is vital for supporting linguistic diversity, yet a fundamental challenge lies in the cross-lingual semantic gap. To address this, a common approach is to initialize LRL word embeddings with those from high-resource languages, thereby enriching their semantic space. However, subsequent embedding adaptation is typically performed through uniform updates across the vocabulary, which fails to allocate larger corrections to semantically difficult or infrequent tokens (e.g., culturally specific terms in Amharic or Somali) that need them most. To address this, we propose Selective Embedding Correction (SEC), a parameter-efficient framework for dynamic semantic correction. SEC models embedding adjustments as sparse low-rank residuals regularized by a group-lasso objective. This enables data-driven selectivity that refines poorly aligned tokens while preserving well-aligned representations. Experiments on named entity recognition, part-of-speech tagging, and natural language inference across multiple low-resource languages show that SEC achieves strong improvements over baselines with only about 1% additional trainable parameters.

Aspect category detection (ACD) in sentiment analysis aims to identify the aspect categories mentioned in a sentence. In this paper, we formulate ACD in the few-shot learning scenario. However, existing few-shot learning approaches mainly focus on single-label predictions. These methods can not work well for the ACD task since a sentence may contain multiple aspect categories. Therefore, we propose a multi-label few-shot learning method based on the prototypical network. To alleviate the noise, we design two effective attention mechanisms. The support-set attention aims to extract better prototypes by removing irrelevant aspects. The query-set attention computes multiple prototype-specific representations for each query instance, which are then used to compute accurate distances with the corresponding prototypes. To achieve multi-label inference, we further learn a dynamic threshold per instance by a policy network. Extensive experimental results on three datasets demonstrate that the proposed method significantly outperforms strong baselines.