Practical Implementation of Conversational Chatbots at Guazi: Architecture, Techniques, and Challenges

This article is edited from Wang Wenbin’s talk at DataFun “AI+” Talk titled “Application of AI in Second Hand Market”, where he shared the practice of dialogue robots at Guazi.

The presentation covers four parts: an introduction to dialogue robots, the technology selection process, the designed algorithmic and system architecture, and finally the online performance and challenges faced at Guazi.

Dialogue robots are popular for three reasons: (1) Turing regarded them as a hallmark of AI, (2) deep‑learning advances have made them industrially viable, and (3) many companies have built on existing intelligent‑customer‑service solutions. User interactions generate data (feedback, conversation, human‑service logs) that can be fed back to CRM and analytics.

Guazi’s goal for the chatbot is to improve efficiency by reducing manpower and training costs, providing 24/7 service, and ensuring controllable quality. This leads to a digital‑, data‑, and intelligence‑driven online service model that enables traceable communication, optimization, differentiation, and fine‑grained operations.

Online robots act as both entry and exit points of the service loop; users express needs via chat, while search and recommendation act as passive components, and IM serves as an active request portal.

Chatbot categories include open‑domain (e.g., Microsoft Xiaoice) and task‑oriented (e.g., flight booking). From a product perspective, three views are needed: customer view (dialogue content, UI, recommendations), service view (context, user profile, order info), and manager view (console, knowledge base).

The classic dialogue flow consists of voice wake‑up, speech‑to‑text, semantic understanding (often involving a knowledge base), dialogue management to select a response, text‑to‑speech conversion, and output.

An example of intent and slot filling: the sentence “Book a flight from Beijing to Shanghai tomorrow at 10 am” has intent “book flight” and slots {origin: Beijing, destination: Shanghai, time: tomorrow 10 am}.

Technology selection evolved from keyword‑based and template methods (high precision but low generalization) to search‑based bots (better coverage but lower accuracy) and finally to deep‑learning models for intent recognition (higher accuracy but data‑intensive). Templates can be auto‑generated but still require human review; search‑based routing forwards queries to specialized sub‑bots; deep‑learning models perform multi‑intent classification.

A template example: the query “Is the transfer from Quanzhou to Xiamen troublesome?” would not match any existing template and thus must be added to the knowledge base.

The system later emphasizes operations: a management platform (knowledge base) allows operators to maintain template‑answer mappings, while a search pipeline breaks queries into terms, ranks results, and returns the most similar answer, requiring large‑scale data for correct ranking.

Deep‑learning is applied to core NLP tasks such as tokenization, POS tagging, entity recognition, and similarity scoring in search. The overall architecture is a deep‑learning pipeline, a current research hotspot.

For single‑turn dialogs, intent classification is a multi‑class problem; the challenge is to automate data collection and model evolution while keeping quality controllable. Generation‑based models are also explored for chit‑chat scenarios.

Semantic understanding follows a funnel: fast keyword/template matching for high‑precision cases, followed by model‑based recognition, then a search‑based recall to improve coverage, achieving over 90 % accuracy in practice.

Multi‑turn dialogue focuses on slot filling, scene management (maintaining large conversation histories), and configurability (operators can define business rules). The DM (dialogue manager) is business‑agnostic and executes based on configured results.

Common risks include: model selection bias, stale historical data, model‑business mismatch, intent explosion, subjective evaluation standards, and lagging model updates. Proactive monitoring and A/B testing are suggested mitigation strategies.

The system architecture consists of a front‑end chat window, a message server (similar to IM), a central dialogue manager, a knowledge‑base console, and backend services (e.g., IP‑to‑city lookup, semantic understanding, CRM integration).

Online performance shows a single‑turn dialog that accurately handles loan queries and a dynamic API resembling a knowledge‑graph for more complex interactions.

Key challenges at Guazi include data scarcity, the need for self‑learning mechanisms to keep up with business changes, and productization details such as data pipelines, embedding‑based intent recognition, and a rigorous content‑approval workflow.

Future directions aim to integrate CRM and other internal systems, leverage data for business intelligence across pre‑sale, sale, and post‑sale scenarios, enable precise marketing through multi‑turn dialogs, and ultimately achieve enterprise‑wide digital and intelligent transformation.

Author: Wang Wenbin, Head of NLP at Guazi, Peking University alumnus with experience at Meituan, Baidu, and expertise in compilers, browsers, IM, big data, search, QA, data mining, ML, and NLP.

Team: Guazi NLP team focuses on chatbots and related products to improve efficiency and service quality, forming a core capability for Guazi’s online transformation.

——END——

Recruitment: Positions for NLP Algorithm Engineer and Data Mining Engineer are open; interested candidates can apply via internal referral to [email protected].

Other related articles: Knowledge Graph Practice at Beike , Robot & Human‑Computer Interaction , and others.