CroCoSum: A Benchmark Dataset for Cross-Lingual Code-Switched Summarization

TL;DR

CroCoSum introduces the first benchmark for cross-lingual code-switched summarization, pairing English technology news with Chinese–English code-switched summaries and comprising over 24k English sources and 18k summaries, with more than 92% containing code-switching. The authors evaluate pipeline, end-to-end, and zero-shot baselines, finding that prior CLS pretraining provides little to no benefit on CroCoSum and that end-to-end mBART-50 often yields the best automatic metrics. However, code-switching metrics reveal misalignment between model outputs and human switching, highlighting evaluation challenges for code-switched generation and the need for more diverse CLS resources. They also perform qualitative error analyses of switching behavior, underscoring the practical importance of robust evaluation and data design for realistic code-switching in CLS applications.

Abstract

Cross-lingual summarization (CLS) has attracted increasing interest in recent years due to the availability of large-scale web-mined datasets and the advancements of multilingual language models. However, given the rareness of naturally occurring CLS resources, the majority of datasets are forced to rely on translation which can contain overly literal artifacts. This restricts our ability to observe naturally occurring CLS pairs that capture organic diction, including instances of code-switching. This alteration between languages in mid-message is a common phenomenon in multilingual settings yet has been largely overlooked in cross-lingual contexts due to data scarcity. To address this gap, we introduce CroCoSum, a dataset of cross-lingual code-switched summarization of technology news. It consists of over 24,000 English source articles and 18,000 human-written Chinese news summaries, with more than 92% of the summaries containing code-switched phrases. For reference, we evaluate the performance of existing approaches including pipeline, end-to-end, and zero-shot methods. We show that leveraging existing CLS resources as a pretraining step does not improve performance on CroCoSum, indicating the limited generalizability of current datasets. Finally, we discuss the challenges of evaluating cross-lingual summarizers on code-switched generation through qualitative error analyses.

Figures (6)

• Figure 1: (a) A Data Example of Source Article and Target Summary Pair. (b) Baseline Approaches.
• Figure 2: Illustration of Error Types.
• Figure :
• Figure :
• Figure :