[Paper Review] Multilingual Clustering of Streaming News

Sebastião Miranda, Artūrs Znotiņš+2 more|arXiv (Cornell University)|Sep 3, 2018

Complex Network Analysis Techniques17 references4 citations

TL;DR

This paper proposes an online, multilingual clustering system that groups streaming news articles into monolingual and crosslingual story clusters using dynamic centroid updates and crosslingual embedding alignment. It achieves state-of-the-art F1 scores of 84.0 on multilingual news streams in English, Spanish, and German by leveraging a pivot-language strategy with English as the reference for crosslingual similarity.

ABSTRACT

Clustering news across languages enables efficient media monitoring by aggregating articles from multilingual sources into coherent stories. Doing so in an online setting allows scalable processing of massive news streams. To this end, we describe a novel method for clustering an incoming stream of multilingual documents into monolingual and crosslingual story clusters. Unlike typical clustering approaches that consider a small and known number of labels, we tackle the problem of discovering an ever growing number of cluster labels in an online fashion, using real news datasets in multiple languages. Our method is simple to implement, computationally efficient and produces state-of-the-art results on datasets in German, English and Spanish.

Motivation & Objective

• To enable real-time, scalable clustering of multilingual news streams into coherent story clusters.
• To address the challenge of discovering an ever-growing number of story clusters in an online fashion across multiple languages.
• To interlace monolingual and crosslingual clustering by maintaining separate monolingual centroids and linking them via crosslingual similarity.
• To improve clustering performance by using English as a pivot language for crosslingual alignment.
• To develop a system that supports media monitoring by enabling efficient search, visualization, and exploration of multilingual news stories.

Proposed method

• The system maintains monolingual cluster centroids in a language-specific bag-of-words space and crosslingual centroids in a shared multilingual embedding space of dimensionality $k_2$.
• Each incoming document is embedded into both monolingual and crosslingual vector spaces using pre-trained multilingual sentence embeddings.
• Monolingual clustering is performed online using a centroid-based method with dynamic thresholding to merge or create new clusters.
• Crosslingual clustering is achieved by computing similarity scores between monolingual clusters across languages using the crosslingual embedding space.
• A pivot-language strategy is applied, where all crosslingual cluster decisions are made relative to English clusters, improving alignment quality.
• The system supports online correction of past clustering decisions by re-evaluating similarity scores as new documents arrive.

Experimental results

Research questions

• RQ1 Can an online clustering system effectively group multilingual news articles into coherent story clusters without prior knowledge of the number of clusters?
• RQ2 How does the performance of online multilingual clustering compare to monolingual baselines when using shared crosslingual embeddings?
• RQ3 Does using English as a pivot language improve crosslingual clustering accuracy compared to global thresholding across all language pairs?
• RQ4 To what extent can dynamic centroid updates and similarity-based merging maintain cluster coherence in a continuous stream?
• RQ5 How does the integration of monolingual and crosslingual clustering stages affect overall clustering quality in a streaming setting?

Key findings

• The proposed system achieves a state-of-the-art F1 score of 84.0 on crosslingual clustering for English, Spanish, and German news streams.
• Using English as a pivot language for crosslingual similarity yields significantly better results than global thresholding across all language pairs.
• Global thresholding for crosslingual distances fails to find a universally effective threshold, indicating language-pair-specific differences in embedding similarity.
• The system maintains high monolingual clustering performance while enabling effective crosslingual linking, with minimal degradation in F1 when comparing monolingual to crosslingual results.
• The use of a two-stage clustering approach—monolingual first, then crosslingual linking—reduces the complexity of crosslingual matching and improves scalability.
• The method supports online correction of past clustering decisions, enabling robust adaptation to evolving news stories in real time.