Recently, the project "Key Technologies and Applications for Natural Language Generation in Open Scenarios," completed by Tsinghua University in collaboration with TAL Education Group, won the First Prize in the Technical Invention category of the 2024 Qian Weichang Chinese Information Processing Science and Technology Award.

Open scenario natural language generation refers to tasks with high variability in output for the same input, such as dialogue generation and story generation. It is one of the most critical and challenging application scenarios today. Even the best existing large models face significant challenges in robustness, efficiency, long-text generation, and quality evaluation in open scenario generation tasks.

To address these challenges, the project systematically broke through key technologies in generation theory, generation methods, and evaluation systems, achieving a series of results applied in smart education, intelligent assistants, and real-time translation. These applications have served hundreds of millions of users and generated significant economic benefits.

In generation theory, the project analyzed the distribution bias between generated text and human text, proposed an optimization objective for generation models based on total variation distance to improve robustness to noise, and explored learning theories for non-autoregressive models, deriving important properties of training loss while introducing proxy distributions to construct a unified training framework for non-autoregressive models. In terms of generation methods, the project focused on knowledge-driven long-text natural language generation methods, conducting systematic research on knowledge representation and knowledge planning. For evaluation systems, it comprehensively built a quality evaluation system for general language generation models, covering data resources, evaluation methods, evaluation models, and application platforms.

The Qian Weichang Chinese Information Processing Science and Technology Award is the highest scientific and technological award in the Chinese information processing field. This award is granted to projects or individuals making significant innovations or breakthroughs in technology, with high technical difficulty, whose overall technical level and main technical and economic indicators have reached leading domestic levels and advanced international levels, promoting the progress of Chinese information processing technology nationwide and creating substantial economic or social benefits. The award is evaluated and conferred by the Chinese Information Processing Society of China's Qian Weichang Chinese Information Processing Science and Technology Award Office.

As an National New Generation AI Open Innovation Platform,TAL has always valued investment in technological innovation and the importance of industry-academia-research cooperation since its construction. To date, TAL has conducted in-depth cooperation with many universities, achieving a series of technical results successfully applied to various TAL products, smoothly accomplishing the integration of production, academia, research, and application.

In the future, TAL will continue to strengthen cooperation with universities and research institutions, build a symbiotic, mutually beneficial, and creative smart education ecosystem, assist in constructing a high-quality education system in China, and promote the process of educational informatization and intelligence in the country.

Related Link: Tsinghua University https://www.cs.tsinghua.edu.cn/info/1088/6488.htm