Transforming B2B Customer Service: Table QA via Multi‑Turn Dialogue

What Is Table QA

Table QA uses high‑density tabular data as a knowledge source, allowing users to retrieve precise answers directly from cells without additional processing, which is especially suitable for e‑commerce and query scenarios.

Common Implementation Approaches

Current methods mainly include end‑to‑end NL2SQL and multi‑turn dialogue‑based SQL keyword extraction, both of which are still largely research‑focused and face challenges such as poor domain transferability, high annotation cost, uncontrollable predictions, and unsatisfactory overall performance.

End‑to‑End NL2SQL

NL2SQL converts natural language into executable SQL statements, enabling non‑technical users to query relational databases without writing SQL. Traditional pipelines rely on high‑quality syntax trees and dictionaries, while recent models combine end‑to‑end neural networks with SQL feature rules (e.g., X‑SQL, HydraNet).

SQL Keyword Extraction via Multi‑Turn Dialogue

This approach treats the user's intent as a task‑oriented dialogue, mapping common SQL query patterns to slot‑filled dialogue flows. The system extracts required SQL keywords across one or multiple turns, confirming meanings with the user before generating the final answer. However, it demands extensive pre‑configured dialogue flows, leading to high manual effort and limited domain adaptability.

Qiyu’s Table QA Solution

NetEase Qiyu transforms table QA into entity (column) and attribute (row) recognition within a multi‑turn dialogue framework. Users provide entities and attributes gradually under guided prompts, allowing the system to pinpoint the exact cell and return the answer. Entity and attribute detection leverages mature keyword and semantic similarity methods, while the dialogue design supports clarification, recommendation, and cross‑turn context reuse.

Core Application Scenarios

The solution is applied to online customer service where users query product specifications, pricing, shipping locations, and release years directly from product tables, eliminating the need for manual FAQ configuration.

Advantages

By using advanced pretrained language models for attribute recognition and a multi‑turn dialogue engine, Qiyu achieves high accuracy, supports partial information in a single turn, and can ask follow‑up questions or provide recommendations, greatly improving usability and maintenance efficiency.

Usage Flow

During a conversation, the user may mention only an entity or only an attribute; the system then asks clarifying questions to collect missing slots before executing the table lookup and delivering the answer.

Innovative Features

Qiyu’s approach combines mature entity/attribute extraction with a flexible dialogue framework, enabling context‑aware multi‑turn interactions, automatic slot filling, and real‑time table cell retrieval without requiring users to formulate full SQL queries.

Conclusion

The article reviews the state of table QA, details NL2SQL and slot‑based dialogue methods, and presents Qiyu’s practical implementation that addresses industrial pain points through entity‑attribute recognition and multi‑turn dialogue, now deployed in NetEase’s online customer service platform.

References

Qin L, Chakrabarti K, Hathi S, et al. Hybrid Ranking Network for Text-to-SQL. 2020.

He P, Mao Y, Chakrabarti K, et al. X‑SQL: reinforce schema representation with context. arXiv, 2019.

Tao Y, Li Z, Zhang Z, et al. TypeSQL: Knowledge‑Based Type‑Aware Neural Text‑to‑SQL Generation. 2018.

Zhong V, Xiong C, Socher R. Seq2SQL: Generating Structured Queries from Natural Language using Reinforcement Learning. 2017.