Designing and Deploying a Real‑Estate Dialogue System: Architecture, Challenges, and Practices

This presentation, originally delivered at the 2018 WOT Global AI Technology Summit by senior algorithm expert Chen Kaijiang, describes Beike's practical applications of semantic understanding, dialogue systems, voice assistants, and VR‑assisted house‑viewing.

The speaker first frames the current AI landscape, noting that while recommendation, search, and advertising dominate commercial AI, natural‑language processing (NLP) remains under‑exploited despite its long history and potential to enable human‑machine interaction.

Dialogue systems are categorized into task‑oriented, chitchat, and knowledge‑Q&A types; the talk focuses on task‑oriented and knowledge‑Q&A for the real‑estate domain, emphasizing that users primarily seek fast, effective problem‑solving and often still prefer human assistance.

Five key difficulties of dialogue systems are identified: (1) no single model can solve all problems, (2) scarcity of high‑quality, low‑cost annotated data, (3) lack of common‑sense knowledge, (4) evaluation challenges, and (5) limited generalizability across industries. The real‑estate sector naturally mitigates several of these issues.

The system aims to empower agents rather than replace them, establishing trust, standardizing efficiency, and supporting rational decision‑making. It operates through three layers: an access layer (handling error correction, sentiment analysis, and basic preprocessing), a central‑control layer (intent recognition and routing), and a task layer (micro‑services that fulfill specific intents). Supporting layers store analysis data and manage model configurations.

Data processing begins with billions of chat messages, from which roughly one‑third are text. Q&A pairs are extracted via rule‑based methods, then clustered after sentence‑to‑vector conversion. Human annotators refine clusters into a small set of intents (<10) and define slot schemas for each intent.

Modeling choices include a customized FastText classifier for intent detection, enhanced with character‑level bigrams and weighted loss; a Bi‑LSTM‑CRF architecture for sequence labeling of multi‑intent dialogues; and a GBDT model to rank candidate answers from multiple intent‑specific services. Slot extraction feeds parameters into template‑based answer generation, which agents can further polish.

Multi‑intent handling uses top‑3 intent probabilities, merges responses via a recommendation‑style ranking, and returns the highest‑scoring answer to the agent. The system also incorporates extensive manual labeling, template management, and fast matching libraries to support hundreds of models.

Future work includes 4D house‑viewing (adding voice assistance to 3D/VR), exposing the AI dialogue platform internally for other scenarios, and building a comprehensive industry knowledge graph to power richer interactions.

In conclusion, the speaker stresses that artificial intelligence must work hand‑in‑hand with human expertise—templates, domain knowledge, and data all originate from people, and the technology ultimately serves to augment human capabilities.