Beike Real Estate NLP Team Wins First Place in CCIR Cup 2021 Intelligent Human‑Computer Interaction Track

Recently, in the national information retrieval challenge (CCIR Cup 2021) organized by the China Conference on Information Retrieval (CCIR), Beike Real Estate’s NLP technology team achieved first place in the "Intelligent Human‑Computer Interaction Natural Language Understanding" track, demonstrating their innovation in dialogue systems.

The CCIR Cup aims to explore practical solutions for information retrieval and provides essential research data for scholars. This edition, co‑hosted by the China Computer Federation (CCF), Renmin University’s Gaoling Institute of AI, Huawei, China Mobile Research Institute, Fudan University and others, attracted top industry and academic teams. The "Intelligent Human‑Computer Interaction Natural Language Understanding" track drew over 330 teams.

In the "Intelligent Human‑Computer Interaction Natural Language Understanding" track, Beike’s NLP team tackled the small‑sample learning problem by combining semi‑supervised learning and transfer learning. For intent recognition, they proposed a semi‑supervised classifier that assigns soft labels to unlabeled data, expanding the small‑sample dataset and improving accuracy. For slot extraction, they used a classic BERT+BiLSTM+CRF architecture and applied transfer learning based on slot similarity across domains to mitigate data scarcity. Their solutions achieved first‑place scores of 84.3% and 85.3% in sub‑tracks A and B respectively, and the technologies are now widely deployed in Beike’s smart customer service, agent‑side voice interaction, and the "Xiao Hai" smart home system.

Beyond exploring cutting‑edge technologies, Beike Real Estate jointly with Tsinghua University presented the large‑scale Mandarin dialect speech benchmark system KeSpeech at NeurIPS 2021, containing recordings from nearly 30,000 speakers covering 34 cities and eight major Mandarin dialects, supporting tasks such as speech recognition, speaker verification, dialect identification, and voice conversion.