AI Applications in 58.com: Voice Dialogue, Advertising Algorithms, Recommendation Systems, and Deep Learning Platform

AI technology is driving industry transformation, and 58 Group is accelerating AI adoption in real‑estate, recruitment, used‑car, local services, finance and other businesses. Deep learning has been introduced into search, recommendation, advertising, information quality, safety, and sales opportunity allocation to improve connection efficiency and user experience. New AI applications such as intelligent customer service, voice dialogue robots, voice quality inspection, and writing robots have increased labor efficiency, while a one‑stop AI algorithm platform has boosted algorithm development efficiency.

The first external technical salon, co‑hosted by the 58 Technology Committee and HR Magic Academy on October 19, 2019, featured the theme “AI Technology in 58 Life Service Scenarios.” Speakers shared practices in human‑machine voice dialogue, advertising algorithms, recommendation systems, and the AI algorithm platform.

Human‑Machine Voice Dialogue – The voice robot, built on the TEG platform’s AI Lab, supports automatic dialing, multi‑turn interaction, and intelligent intent judgment, and is used in sales, service promotion, information review, and voice notifications, delivering significant efficiency gains.

The robot consists of a basic service layer, logic layer, editing‑operation layer, access layer, and a web access platform. The basic service layer provides SIP‑based call signaling, audio codec, and speech‑recognition/synthesis interfaces. The logic layer includes intent recognition (single‑sentence and whole‑round) using BERT and TextCNN, dialogue interaction, and dialogue management with text clustering, classification, similarity, and entity recognition. The editing‑operation layer offers data labeling, effect evaluation, and analysis. The access layer handles real‑time outbound calls, message bus integration, and result feedback, while the web platform supplies visual management for accounts, scripts, tasks, and reporting.

Machine Learning in Commercial Monetization – The commercial middle‑platform serves recruitment, real‑estate, yellow‑pages, used‑car, and second‑hand goods, providing flexible monetization solutions. Smart bidding (OCPC) automatically predicts the value of each click using conversion‑rate estimates (LSTM, ESMM) and adjusts bids dynamically, aligning revenue with connection effectiveness. Advertising retrieval follows a classic “recall‑coarse‑ranking‑fine‑ranking” pipeline, employing query expansion, CLSM, collaborative filtering, matrix factorization, deep vectors, DSSM, item2vec, wide‑&‑deep, DeepFM, ESMM, DIN, etc., to improve relevance and revenue.

The intelligent creative platform automates image‑text creative generation and selection, integrating material libraries, templates, user preferences, and feedback loops to boost creative quality and efficiency.

Personalized Recommendation in 58 Recruitment – The system combines offline data mining (data, model, storage layers) and online services (algorithm layer, recommendation engine). It addresses low‑quality content detection using regex, NER, and classification; builds knowledge graphs with Bi‑LSTM+CRF; constructs user portraits via statistical rules, LR+XGBoost, and deep sequence models; and employs multi‑stage recall (content, user, global) with embedding‑based methods. Ranking evolves from LR CTR models to LR+XGBoost/wide‑&‑deep/deepFFM, adding quality, activity, and match factors. Display optimization leverages feature importance to surface hidden signals and uses NLG to generate concise job titles.

Kubernetes‑Based Deep Learning Platform – The platform unifies GPU/CPU resources via Kubernetes, supporting TensorFlow, PyTorch, and custom frameworks. It comprises hardware, cluster management, algorithm, web management, and online inference layers. Online inference uses TensorFlow‑Serving, PyTorch, Seldon, gRPC, and 58’s SCF RPC framework, with TensorRT acceleration for GPU and MKL for CPU, achieving significant performance gains.

In summary, the salon covered voice dialogue, advertising algorithms, recommendation systems, and the AI algorithm platform, sharing practical experiences, challenges, and insights. The event was successful and a second session on IM technology is planned for mid‑November.