Simeng Han, Hailey Schoelkopf, Yilun Zhao, Zhenting Qi, Martin Riddell, Wenfei Zhou, James Coady, David Peng, Yujie Qiao, Luke Benson, Lucy Sun, Alexander Wardle-Solano, Hannah Szabó, Ekaterina Zubova, Matthew Burtell, Jonathan Fan, Yixin Liu, Brian Wong, Malcolm Sailor, Ansong Ni, Linyong Nan, Jungo Kasai, Tao Yu, Rui Zhang, Alexander Fabbri, Wojciech Maciej Kryscinski, Semih Yavuz, Ye Liu, Xi Victoria Lin, Shafiq Joty, Yingbo Zhou, Caiming Xiong, Rex Ying, Arman Cohan, Dragomir Radev. Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing. 2024.

As the impacts of climate change become evident worldwide, organisations across all sector of the economy are actively aiming at reducing their carbon footprints. Video streaming has been in the spotlight since the beginning of the pandemic. Previous work has identified opportunities to reduce electricity consumption, yet methods to reliably estimate the carbon reduction potential from interventions on video streaming systems are currently lacking. In particular, not enough consideration is given to the consistent methodological approach to impact assessment required to adequately account for the complex interactions between changes to a service and operational and structural effects at internet-scale systems at varying time scales. In this text, we review the state of knowledge and propose a consistent short-term marginal approach for the assessment of the short-term decarbonisation potential of interventions. We illustrate this with a simplified example intervention and contrast it to previous methodologically inconsistent approaches, in which we evaluate the temporary reduction of the video spatial resolution of user-generated content video on demand from 1080p to 720p in a fixed ladder scenario over one month in the UK. We find that the carbon reductions mainly come from savings at user devices, but are overall negligible at 0.5 gCO2e per day per typical laptop-viewer.

Unfaithful text generation is a common problem for text generation systems.In the case of Data-to-Text (D2T) systems, the factuality of the generated text is particularly crucial for any real-world applications.We introduce R2D2, a training framework that addresses unfaithful Data-to-Text generation by training a system both as a generator and a faithfulness discriminator with additional replacement detection and unlikelihood learning tasks.To facilitate such training, we propose two methods for sampling unfaithful sentences.We argue that the poor entity retrieval capability of D2T systems is one of the primary sources of unfaithfulness, so in addition to the existing metrics, we further propose named entity based metrics to evaluate the fidelity of D2T generations.Our experimental results show that R2D2 systems could effectively mitigate the unfaithful text generation, and they achieve new state-of-the-art results on FeTaQA, LogicNLG, and ToTTo, all with significant improvements.