New Breakthrough in Multi-Turn Conversational Reading Comprehension, Zhuiyi Technology Tops CoQA Leaderboard
AI has once again broken new ground in challenging human capabilities — this time in multi-turn conversational reading comprehension.
Recently, the AI Lab team at Zhuiyi Technology surpassed Microsoft to take first place on the leaderboard of CoQA (Conversational Question Answering Challenge), a dialogue-based question-answering competition launched by Stanford University, breaking the previous CoQA records set by Microsoft and other teams. Notably, in the technical solution submitted by the team, the single-model performance across all metrics exceeded human performance for the first time.
Leaderboard: https://stanfordnlp.github.io/coqa/
Reading comprehension is currently the most closely watched and fastest-advancing direction in language AI. Two top-tier global competitions draw the most attention: SQuAD, which focuses on single-question single-answer formats, and CoQA, which emphasizes multi-turn conversational question answering. Compared to SQuAD, CoQA specifically challenges models' ability to answer interrelated questions within a conversation. Moreover, the conversational QA dataset consists of dialogue questions drawn from articles across different domains, with very free-form answers — all of which significantly raise the difficulty of the challenge.
Prior to this, well-known enterprises and institutions including Microsoft, iFlytek, and Sogou had all reached the top of the leaderboard. Among them, Microsoft had become the best record holder with its ensemble model, drawing considerable industry attention.
Building on the large-scale pre-trained model RoBERTa, the Zhuiyi Technology team incorporated adversarial training and knowledge distillation, while also accounting for the specific characteristics of the CoQA dataset. Experimental results showed that these methods significantly improved system performance. The team's RoBERTa + AT + KD system, submitted on September 5, 2019, ranked first on the CoQA leaderboard, with its single-model performance exceeding human evaluation metrics across the board for the first time.
This CoQA achievement is also one of the results of Zhuiyi Technology's sustained efforts in AI technology. As an AI company driven by technological innovation, Zhuiyi Technology has consistently focused on advancing frontier AI semantic interaction and algorithm technologies, making continuous breakthroughs in intent recognition, sentiment analysis, knowledge graphs, reading comprehension, and multi-turn dialogue. The company won the 2018 CMRC Machine Reading Comprehension Competition and the 2019 CCF Language and Intelligence Technology Competition's "Knowledge-Driven Dialogue" championship.
Beyond technical breakthroughs, Zhuiyi Technology has consistently adhered to a mutually reinforcing approach between R&D and application, scaling AI deployment. Zhuiyi Technology's intelligent products and solutions have been deployed across finance, telecom operators, government and enterprise, and broad internet sectors, serving more than 300 industry benchmark clients. At the same time, Zhuiyi Technology is committed to the application, mass production, and standardization of innovative technologies. For example, reading comprehension and multimodal interactive digital human technologies and products have been applied in scenarios with select securities firms and banking partners, achieving industry-first breakthroughs.
As breakthroughs in NLP and other AI technologies arrive and the era of industrial intelligence takes hold, multi-industry scenario-based demands will pose greater challenges to technological innovation and productization capabilities. Zhuiyi Technology will continue to focus on innovation, advancing the technological frontier to empower more industry partners through technological innovation!
Zhuiyi Technology Closes $41 Million Series C to Accelerate Industrial Intelligence Upgrade
YITU Claims Another World Championship, This Time in "Voiceprint Recognition"
