Jiuge: AI Poetry Writing System by Tsinghua University 160

Jiuge is an AI poetry writing system developed by the Natural Language Processing and Social Humanities Computing Laboratory at Tsinghua University. The Jiuge poetry writing system adopts the latest deep learning technology, combining multiple models specifically designed for poetry generation, and is trained on over 800,000 poems created by human poets. This enables the Jiuge system to generate high-quality poems covering various styles and genres, including five-character quatrains and seven-character regulated verses. Jiuge not only focuses on the formal beauty of poetry, but also pays attention to the artistic conception and emotional expression of the content.

Different from other poetry generation systems, Jiuge features multimodal input, multi-genre and multi-style capabilities, and a human-computer interactive creation mode. The Jiuge system and its R&D team are committed to exploring the integration of AI technology and the humanities, helping AI empower literary education, and contributing to the inheritance and development of China's outstanding poetry culture.

Features of Jiuge

• Multimodal input: Users can generate poetry by inputting keywords, sentences, paragraphs, or even uploading images. This diverse input method makes the creation process more flexible and personalized.

• Real-time generation: After users input content, the system can quickly generate candidate verses, and users can select and modify them according to their preferences.

• Poetry quality evaluation: The Jiuge system also has the function of automatically evaluating the quality of generated poetry, helping users screen out better works.

Application Scenarios of Jiuge

The Jiuge system is not only suitable for poetry creation enthusiasts, but is also applied in the field of education to help students learn and create classical Chinese poetry. By interacting with the Jiuge system, students can better understand the structure and rhythm of poetry.

Jiuge Historical Timeline:

• April 2017, the beta version of Jiuge was launched;

• August 2017, Jiuge participated in the CCTV-1 prime-time large-scale science challenge program "AI vs Human", competing with three outstanding contemporary poets in poetry creation ability; the online views of the program reached about 10 million;

• September 10, 2017, the official version of Jiuge was launched;

• September 2017, Jiuge was invited to participate in the first China Beijing International Language and Culture Expo;

• November 2017, Jiuge was invited to participate in the "AI vs Human" AI Annual Gala, and performed a sketch on stage with the famous actress Zhang Kaili;

• December 2017, Jiuge was invited to perform poetry creation at the 7th Wu Wenjun AI Science and Technology Award Ceremony;

• December 2017, Jiuge participated in the recording of the CCTV-3 program "Harvest 2017";

• May 2018, Jiuge was invited to participate in the recording of the CCTV "May Flowers" May Fourth Youth Gala; Jiuge won the Best System Demonstration Award at the 2017 China Conference on Computational Linguistics (CCL);

• Since its launch in September 2017, Jiuge has accumulated over 1 million visits.

Since its launch in 2017, the Jiuge system has created more than 4 million poems for users and has demonstrated its creative ability on many occasions, including passing the Turing test by the live audience in the CCTV program "AI vs Human" in 2017. In 2019, the Jiuge system was open-sourced, releasing the source code and dataset of its unsupervised style poetry generation model.

The R&D team of the Jiuge system is committed to exploring the integration of AI technology and the humanities, with the goal of helping to promote and develop China's outstanding traditional culture. The multimodal input and human-computer interactive creation mode of the Jiuge system enable users not only to generate poetry, but also to receive real-time feedback during the creation process.

The Natural Language Processing and Social Humanities Computing Laboratory at Tsinghua University (THUNLP) is one of the earliest and most influential scientific research institutions in China engaged in natural language processing research. Under the leadership of Professor Sun Maosong as the leader, and Associate Professors Liu Yang and Liu Zhiyuan as the backbone, the THUNLP laboratory aims at cutting-edge topics in natural language processing and conducts systematic and in-depth research, covering machine translation, knowledge graphs, question answering systems, text generation, Chinese word segmentation, syntactic analysis, etc., and has achieved a number of high-quality research results.

Related Links:

THUNLP Laboratory Homepage: http://nlp.csai.tsinghua.edu.cn

THUNLP github: https://github.com/thunlp

THUMT github: https://github.com/thunlp-mt

THUAIPoet (Jiuge) github: https://github.com/thunlp-aipoet

https://xiaoyuanyi.github.io/slides/jiuge.pdf