SBIR Phase I: Handheld Devices for Practical Simultaneous Translation

这项小型企业创新研究（SBIR）第一阶段项目对手持式同声翻译的广泛影响推动了自然语言处理和机器学习加速方面的最新技术。除了技术上的贡献，这种新颖的翻译器可能会对广泛的领域产生重大影响，例如部署在外国的军事人员、参加多语种会议的商人、医疗服务提供者、执法部门、客户支持服务、外交官和政治代表、用首选语言向公民提供服务的地方政府实体，以及国际游客。因此，该项目将在提高经济竞争力、促进健康和福利以及提高美国的军事能力方面产生更广泛的影响。此外，该项目有助于提高教育公平和STEM素养，使不同背景的人，特别是移民、妇女和一些国家/地区代表性不足的少数民族，更好地获得教育资源，这些人以前由于先前教育机会有限或获得外语课程的机会有限而处于不利地位。这个小企业创新研究（SBIR）一期项目的重点是研究和开发创新的算法优化和专用翻译设备，以实现快速、准确和低功耗的同声翻译推理。总体方法是利用不同层次的语言元素之间不言而喻的隐含联系来指导学习模式。具体技术在音素级、工作级和句子级进行了研究。总的来说，这些创新旨在将同声翻译的计算复杂度降低几个数量级，同时提高翻译精度。一个系统和全面的方法也正在建立，允许通过高级综合快速实现推理硬件，并报告翻译准确性，延迟，功率和面积的详细统计数据，以促进对研究的彻底评估。该奖项反映了美国国家科学基金会的法定使命，并通过使用基金会的知识价值和更广泛的影响审查标准进行评估，被认为值得支持。

The broader impact of this Small Business Innovation Research (SBIR) Phase I project on a handheld simultaneous translator advances the state-of-the-art technologies in natural language processing and machine learning acceleration. Beyond technical contributions, the novel translator will likely have a significant impact on a wide range of domains, such as military personnel deployed in foreign countries, businesspeople participating in multilingual meetings, medical service providers, law enforcement, customer support services, diplomats and political representatives, local governmental entities providing services to citizens in the preferred language, and international tourists. As such, the project will have broader impacts in increasing economic competitiveness, advancing health and welfare, and improving military capabilities of the United State. Furthermore, the project helps to enhance equity in education and STEM literacy by enabling better access to educational resources to people of diverse backgrounds, especially immigrants, women, and underrepresented minorities in some countries/regions, who were previously disadvantaged due to limited prior educational access or limited access to foreign language courses.This Small Business Innovation Research (SBIR) Phase I project focuses on the research and development of innovative algorithmic optimizations and a purpose-built translation device to enable fast, accurate and low-power inference for simultaneous translation. The overall approach is to exploit the unspoken but implied connections among language elements at various levels to guide the learning model. Specific techniques are investigated at the phoneme-level, work-level, and sentence-level. Collectively, these innovations aim to reduce the computation complexity of simultaneous translation by orders of magnitude while increasing translation accuracy. A systematic and comprehensive methodology is also being established that allows fast implementation of the inference hardware via high-level synthesis and reports detailed statistics on translation accuracy, latency, power, and area to facilitate a thorough evaluation of the research.This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.