How Xianyu Boosted Sales with an AI-Powered Auto-Reply Chatbot

Project Background

Xianyu is a second‑hand sharing marketplace that encourages interaction between buyers and sellers to increase transaction fun. However, several problems hinder smooth communication: sellers often reply late, answer quality varies, repetitive answers are common, and price negotiations are inefficient.

Key Issues Identified

Seller delayed replies: average first response time exceeds 6 hours, while replies within 1 hour raise conversion by over 75%.

Inconsistent answer quality: sellers may give thorough or perfunctory replies, causing potential buyers to be missed.

Redundant answers: over 40% of chat content concerns product information that could be extracted from the listing itself.

Negotiation friction: sellers dislike repeated haggling and may become upset, harming the experience.

Technical Framework

To address these issues, Xianyu introduced a chatbot that automatically answers buyer queries when sellers are offline. The system builds a knowledge base composed of product‑related question‑answer key‑value pairs through two pipelines:

Attribute Initialization Module – When a seller enables auto‑reply, the system extracts generic Q&A pairs (e.g., shipping, payment) from the seller’s profile and from product attributes, storing them in a configuration service.

User‑Dialogue Extraction Module – During real conversations, the system extracts new Q&A pairs from the seller’s actual replies, enriching the knowledge base with missing product details.

Core Algorithms

The chatbot relies on two main AI components provided by AliNLP: an intent‑recognition module and a response‑generation module.

Intent Recognition

Buyer intents are ambiguous (e.g., “apple” could mean fruit or phone). Approximately 20% of dialogues fall into four major categories: price, bargaining, product attributes, and transaction details. Xianyu built 24 specialized models covering common buyer intents, achieving >90% accuracy for most attributes.

High‑precision attributes use rule‑based pre‑filters.

Diverse semantic attributes use deep‑learning models.

Two deep‑learning approaches are employed:

Dependency Classification Model – Combines word embeddings, part‑of‑speech tags, and dependency parsing with an attention mechanism to boost recall from 44% to 74% for attribute extraction.

BERT‑based Model – Fine‑tuned on dialogue data, incorporates speaker role information, and improves overall classification accuracy by 4 percentage points.

Response Generation

The generation pipeline consists of three steps:

Identify buyer intent and map it to the relevant attribute.

Apply sequence labeling to extract attribute values from the seller’s historical replies.

Fill a predefined template with the attribute and its value to produce a complete answer.

Application and Effects

Two conversation scenarios illustrate the impact. In the first, a buyer waited 12.5 hours for a manual reply and the deal failed. In the second, the chatbot responded instantly with price and shipping information, guided the buyer, and the seller closed the sale shortly after coming online.

Business metrics after several months of rollout show substantial improvements:

Timely response rate (within 2 hours) increased by 30%.

Number of chat rounds for successful transactions grew by ~20%.

Seller reply cost reduced by thousands of hours (assuming 20 seconds per reply).

Funnel efficiency improved at every stage, boosting overall product turnover.

7‑day sales velocity rose by ~30% for items with the chatbot enabled.

Conversion rate for interactions where the seller replied after >2 hours increased by ~30% when the chatbot was active.