Researchers Build Multilingual Corpus of Multi-Word Expressions for Translation Systems

Linguists and machine translation researchers have developed AlphaMWE, a large annotated dataset designed to improve how translation systems handle multi-word expressions—phrases whose meanings cannot simply be derived from their individual words, such as "kick the bucket" or "spill the beans."

Building the Dataset#

The team constructed the corpus by starting with English source material and translating it into Arabic, Chinese, German, Italian, and Polish using four major translation services: Microsoft Bing, Google Translate, Baidu Fanyi, and DeepL. Rather than relying solely on automatic output, researchers manually reviewed and corrected every translation, then carefully annotated instances of verbal multi-word expressions (phrases where a verb serves as the core element). The Arabic corpus includes both standard and dialectal versions from Egypt and Tunisia. Each sentence underwent multiple rounds of human review to ensure consistency and accuracy.

Key Challenge: Translation Complexity#

One major finding from the project is that state-of-the-art machine translation systems struggle consistently with multi-word expressions. The researchers documented specific error patterns across the systems tested, revealing that translating these semantically opaque combinations remains a genuine technical hurdle. This difficulty affects the quality of translated output and underscores why human intervention remains essential in professional translation workflows.

Intended Uses#

The resulting dataset aims to serve the research community across several domains. Linguists can use it to study how multi-word terms vary across languages; translation engineers can leverage it to identify and improve weak points in their systems; and information extraction specialists can build better tools for recognizing these expressions in text. The dataset's multi-layered human curation—combining post-editing and annotation—makes it a substantially cleaner resource than raw machine output alone.