38 papers from Alibaba Cloud were accepted by the top conference ACL, and the Tongyi team disclosed a number of cutting-edge technologies for large models
2024-08-14
한어Русский языкEnglishFrançaisIndonesianSanskrit日本語DeutschPortuguêsΕλληνικάespañolItalianoSuomalainenLatina
Zhang Yunshan, reporter of Chao News Client
The ACL 2024 annual meeting, the top academic conference in the field of computational linguistics and natural language processing, was held in Thailand this week. A total of 38 high-level papers from Alibaba Cloud were accepted by the conference, including several large-model related papers from the Tongyi Qianwen team, covering topics such as large-model SFT technology, LLM role-playing capabilities, and multimodal model evaluation benchmarks. The Tongyi booth at the Bangkok conference was surrounded by NLP researchers and developers from all over the world, and Tongyi became the most popular Chinese large model at the conference.
ACL 2024 Annual Conference held in Thailand this week
The ACL Annual Meeting (Annual Meeting of the Association for Computational Linguistics) is organized by the Association for Computational Linguistics and is the top academic conference in the field of computational linguistics and natural language processing. ACL 2024 is the 62nd meeting of the association. This year's conference focuses on the topic of "Promoting reproducible natural language processing research with open science, open data, and open models."
Alibaba Cloud has always been one of the technology companies with the highest number of papers selected for ACL, with a total of 38 papers included this year, including 16 main conference papers. As a representative of the "open source" power of China's big model, the Tongyi Big Model team disclosed a number of cutting-edge big model technologies at this conference and had face-to-face exchanges with NLP R&D personnel and developers from all over the world.
"Large Language Models are Superpositions of All Characters: Attaining Arbitrary Role-play via Self-Alignment" first proposed the self-alignment strategy DITTO for large model role-playing, which significantly improved the role-playing ability of LLM. The Tongyi team has open-sourced DITTO; "AIR-Bench: Benchmarking Large Audio-Language Models via Generative Comprehension" introduced the audio language understanding model evaluation benchmark AIR-Bench launched by the Tongyi team, which is used to evaluate the generative instruction following ability of the model, filling the gap in the industry; "How Abilities in Large Language Models are Affected by Supervised Fine-tuning Data Composition" demonstrated through a series of experiments how SFT data affects the mathematics, code, and human alignment capabilities of LLM, which can provide reference for researchers and developers' SFT work.
Overseas developers trained Thai and Southeast Asian language models based on Qwen
Since August 2023, Tongyi Qianwen has open-sourced dozens of LLM, multimodal and special capability models, and the Qwen series of open-source models have been downloaded more than 20 million times.
In Southeast Asia, Tongyi Qianwen open source models also have many loyal users, and large models such as Thai, Vietnamese, and Southeast Asian languages trained based on Qwen are often seen in the open source community. For example, Singaporean engineer Liu Qian trained the popular Southeast Asian language large model Sailor based on Qwen1.5, covering a full set of sizes such as 0.5B, 1.8B, 4B, 7B, and 14B; Vietnamese engineer Nguyen Quan developed a Vietnamese large model. He said: "According to our internal benchmark evaluation, the Qwen2 basic model surpasses all closed-source large language models currently on the market."
Appendix: List of Alibaba Cloud papers included in ACL 2024
List of Alibaba Cloud papers accepted by ACL 2024
"Please indicate the source when reprinting"
