2016/12/14 by Rico Sennrich, Sennrich, Rico · 1 citation
Computer Science · #Computation and Language (cs.CL) #FOS: Computer and information sciences #Natural Language Processing Techniques #Text Readability and Simplification #Topic Modeling
paper · pdf · doi:10.48550/arxiv.1612.04629
openalex publication_date 2016/12/14 · openalex created_date 2022/10/03 · openalex updated_date 2026/07/28
Analysing translation quality in regards to specific linguistic phenomena has\nhistorically been difficult and time-consuming. Neural machine translation has\nthe attractive property that it can produce scores for arbitrary translations,\nand we propose a novel method to assess how well NMT systems model specific\nlinguistic phenomena such as agreement over long distances, the production of\nnovel words, and the faithful translation of polarity. The core idea is that we\nmeasure whether a reference translation is more probable under a NMT model than\na contrastive translation which introduces a specific type of error. We present\nLingEval97, a large-scale data set of 97000 contrastive translation pairs based\non the WMT English->German translation task, with errors automatically created\nwith simple rules. We report results for a number of systems, and find that\nrecently introduced character-level NMT systems perform better at\ntransliteration than models with byte-pair encoding (BPE) segmentation, but\nperform more poorly at morphosyntactic agreement, and translating discontiguous\nunits of meaning.\n