LTI Colloquium: COMET: a Neural Framework for State-of-the-art MT Evaluation скачать в хорошем качестве

LTI Colloquium: COMET: a Neural Framework for State-of-the-art MT Evaluation

Speaker: Alon Lavie @ Unbabel / Carnegie Mellon University Website: https://unbabel.com/research/people/a... Title: COMET: a Neural Framework for State-of-the-art MT Evaluation Abstract: Delivery of high-quality Machine Translation (MT) is only possible with reliable evaluation metrics to inform modelling and system development. The translation workflows we develop at Unbabel require highly-adapted MT systems which are regularly retrained to continuously deliver customer-specific, accurate translations. Unfortunately, with current state-of-the-art neural MT systems, traditional metrics such as BLEU and METEOR have been shown to no longer correlate well with human judgments, and in particular, they poorly distinguish between fine-grained accuracy distinctions of top performing MT models. This can result in misinformed MT development decisions that affect the quality of translations for our customers. To address this challenge, we recently developed COMET - a new neural-based framework for training automated MT evaluation models that are demonstrated to exhibit new state-of-the-art levels of correlation with human judgments. Our framework leverages recent breakthroughs in cross-lingual pretrained language modeling resulting in highly multilingual and adaptable MT evaluation models that exploit information from both the source input and a target-language reference translation in order to more accurately predict MT quality. We showcase our framework by training and evaluating COMET models for three different types of human judgments: Direct Assessments, Human mediated Translation Edit Rate (HTER) and Multidimensional Quality Metrics (MQM). Our models achieve new state-of-the-art performance on the WMT 2019 and 2020 Metrics shared tasks and are sensitive to fine distinctions typical of high-performing MT systems. The COMET framework and our top-performing pretrained evaluation models are freely available open-source. In this presentation we present an overview of the COMET framework and highlight its capabilities through assessments of the COMET models we have trained, their correlation with human judgments of translation quality, and their utility in practice for evaluating and contrasting MT models developed at Unbabel. #NLProc #MachineTranslation