58.com AI Algorithm Competition: Winning Teams and Their Technical Solutions

58.com has built an intelligent customer‑service system based on AI, focusing on text matching and classification, which are core problems in natural language understanding. In 2017 the company launched its first AI algorithm competition, providing real‑world customer‑service data for participants.

The competition lasted 50 days, attracted 158 teams from 60 universities and 30 companies, and selected ten finalist teams. Awards included a first‑place prize of ¥25,000, second‑place ¥10,000, third‑place ¥5,000, and certificates for all finalists.

First place – OUCERS (Ocean University of China) : The team presented a BERT‑based text classification approach, treating the task either as sentence‑pair similarity or single‑sentence classification. They combined expanded and standard question data, re‑weighted standard questions, and fine‑tuned a pre‑trained BERT model after removing the NSP task and using only MLM loss.

Second place – Online Be Da Lao (Xi'an Jiaotong University) : Their solution addressed class imbalance by pre‑training BERT and ELECTRA on a large corpus, then fine‑tuning with a fused loss of focal loss and F1‑score to improve recognition of difficult samples.

Third place – Rookie Team (Harbin Institute of Technology) : They explored the ELECTRA model for fast pre‑training, applied it to both text classification (861 classes) and text matching, and experimented with three pipelines: pure classification, pure matching, and a two‑stage classification‑then‑matching approach.

Fourth place – Multi‑Model Fusion : The team combined multiple models (e.g., SPTM, ELMo, DSSM, Bi‑LSTM, Seq2Seq) for both classification and similarity matching, achieving better performance than any single model.

The award ceremony was livestreamed on September 16, 2020, with speeches from senior executives and technical leaders. Winners shared their solution videos and PPT files, which can be downloaded by contacting the 58 Technology public account and mentioning “AI Algorithm Competition PPT”.

Additional resources include video recaps of each winning solution (live 1‑4) and links to ongoing competition registration, updates, and an open‑source QA‑match tool for future participants.