Reproduce “Hierarchical Level-Wise News Article Clustering via Multilingual Matryoshka Embeddings” on OpenResearch

Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.

Universal app for Apple silicon and Intel. Requires macOS 11 or later.

About this paper

Hans W. A. Hanley, Zakir Durumeric

A framework for hierarchical clustering of multilingual news articles was developed by Stanford University researchers, utilizing Matryoshka embeddings and a level-wise reciprocal agglomerative clustering algorithm. The method achieved a Pearson correlation of 0.817 on the SemEval 2022 Task 8 test set, improving upon the previous best bi-encoder model's 0.801, and outperformed BERTopic in F1 scores for hierarchical clustering.

Key insights

  • Multilingual Matryoshka embeddings can effectively encode hierarchical semantic granularity, with lower dimensions capturing broad themes and higher dimensions capturing specific events.
  • A modified Reciprocal Agglomerative Clustering algorithm, when combined with these hierarchical embeddings, can automatically identify distinct news stories, topics, and themes without requiring a predefined cluster count.
  • Extensive data augmentation, particularly with stylistic and cross-lingual variations, is vital for training robust multilingual models that perform well across a wide range of languages.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.