Weihua Zheng, Zhengyuan Liu, Tanmoy Chakraborty, Weiwen Xu, Xiaoxue Gao, Bryan Chen Zhengyu Tan, Bowei Zou, Chang Liu, Yujia Hu, Xing Xie, Xiaoyuan Yi, Jing Yao, Chaojun Wang, Long Li, Rui Liu, Huiyao Liu, Koji Inoue, Ryuichi Sumida, Tatsuya Kawahara, Fan Xu, Lingyu Ye, Wei Tian, Dongjun Kim, Jimin Jung, Jaehyung Seo, Nadya Yuki Wangsajaya, Pham Minh Duc, Ojasva Saxena, Palash Nandi, Xiyan Tao, Wiwik Karlina, Tuan Luong, Keertana Arun Vasan, Roy Ka-Wei Lee, Nancy F. Chen. Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2026.

The standard training algorithm in neural machine translation (NMT) suffers from exposure bias, and alternative algorithms have been proposed to mitigate this.However, the practical impact of exposure bias is under debate.In this paper, we link exposure bias to another well-known problem in NMT, namely the tendency to generate hallucinations under domain shift.In experiments on three datasets with multiple test domains, we show that exposure bias is partially to blame for hallucinations, and that training with Minimum Risk Training, which avoids exposure bias, can mitigate this.Our analysis explains why exposure bias is more problematic under domain shift, and also links exposure bias to the beam search problem, i.e. performance deterioration with increasing beam size.Our results provide a new justification for methods that reduce exposure bias: even if they do not increase performance on in-domain test sets, they can increase model robustness to domain shift.

Telemedicine provides health care services at a distance using information and communication technologies, which intends to be a solution to the challenges faced by current health care systems with growing numbers of population, increased demands from patients, and shortages in human resources. Recent advances in telemedicine, especially in wearable electrocardiogram (ECG) monitors, call for more intelligent and efficient automatic ECG analysis and diagnostic systems. We present a streaming architecture implemented on Field-Programmable Gate Arrays (FPGAs) to accelerate real-time ECG signal analysis and diagnosis in a pipelining and parallel way. Association-rule mining is employed to generate early diagnostic results by matching features of ECG with generated association rules. To improve performance of the processing, we propose a hardware-oriented data-mining algorithm named Bit_Q_Apriori . The corresponding hardware implementation indicates a good scalability and outperforms other hardware designs in terms of performance, throughput, and hardware cost.

Current aging society has seen huge increases in portable devices sending massive volume of signals to medical servers. To address inefficient and unscalable signal processing on generic servers and clients, in this paper, we present an FPGA(field programmable gate array)-assisted cloud system providing an efficient framework for electrocardiogram (ECG) telemedicine including real-time data acquisition, transmission and analyzing over the Internet. We explore the requirements of massive cloud signal processing supporting a large number of connections or channels from cloud clients. A prototype system was composed of a client-server platform and an FPGA hardware system with PCI-E endpoint and ECG R-peak detection modules. A streaming micro-architecture for hardware system is proposed to detect ECG pattern from scalable number of channels. Evaluation results show that our streaming system design has good performance in terms of real-time, scalability and latency.

The conventional time-domain back-projection (BP) imaging method typically suffers from performance degradation when the echo signals are interfered by noise. In this letter, a virtual-echo projection imaging method is proposed for noise suppression, including the suppression of the additive white Gaussian noise and the irrelevant colored noise. First, the proposed method calculates a number of coarse imaging maps by projecting the realistic echoes on the virtual echoes. Second, the fine imaging maps are calculated by accumulating all the coarse imaging maps coherently. The performance of the proposed method is numerically investigated in comparison to that of the traditional BP method. The results indicate that the proposed method outperforms the conventional BP method in suppressing both the additive white Gaussian noise and the colored noise.