Enhancing Cross-Lingual Embedding Alignment with Additive Keywords for International Trade Product Classification

ICLR 2026 Conference Submission16994 Authors

Keywords: Bilingual classification, Cross-lingual embedding alignment, Low-resource languages, International trade classification, Harmonized System codes

Abstract: Cross-lingual embedding alignment plays an important role in enabling effective multilingual classification tasks. Although multilingual pretrained language models and fine-tuning techniques are increasingly adopted, current approaches inadequately address specialised domains, where domain-specific terminology and mixed-language content present unique challenges that hinder classification accuracy. This work considers the problem of automatically classifying text-based descriptions of international trade transactions with respect to an international standard Harmonized System (HS) code taxonomy. We propose a novel method that incorporates mixed-language keyword embeddings to improve cross-lingual alignment, focusing on bilingual models, and subsequently leverages this alignment for downstream classification tasks, with particular applicability to low-resource domains. Using a supervised learning framework implemented through neural network architectures, the model is trained on pairs of product descriptions and their corresponding extracted keywords. Experimental results on benchmark bilingual datasets demonstrate significant and consistent improvements in classification performance over baseline models, including in low-resource target language scenarios. The findings demonstrate the effectiveness of incorporating additive keywords as a strategy for cross-lingual embedding alignment, thereby enhancing representation quality and improving classification accuracy.

Primary Area: unsupervised, self-supervised, semi-supervised, and supervised representation learning

Submission Number: 16994