Alipay’s SIGIR 2019 Papers: Reinforcement Learning for User Intent Prediction and Unsupervised QUEST for Complex Question Answering

On July 21, 2019, the SIGIR 2019 conference opened in Paris, showcasing top‑level research in information retrieval. Alipay had several papers accepted, covering topics such as intelligent customer service and dynamic text retrieval, to share its AI‑driven information‑retrieval achievements with the industry.

With the rapid growth of digital technologies, everyday life is flooded with massive amounts of information, demanding continuously updated methods to process data efficiently and uncover value.

This article highlights the key contributions of Alipay’s accepted papers.

Foreseeing: Using Reinforcement Learning to Predict User Intent – Since 2017, Alipay’s intelligent customer service has featured a “pre‑answer” capability that guesses a user’s question before it is asked. User‑intent prediction, based on historical behavior and state, is treated as a top‑N recommendation problem. Traditional approaches ignored relationships among candidate questions. In the paper “Reinforcement Learning for User Intent Prediction in Customer Service Bots,” Alipay engineers model intent prediction as an N‑step sequential decision process and employ reinforcement learning to discover optimal recommendation strategies, dynamically updating the policy as question popularity and user behavior evolve.

Instant Answers to Complex Questions: Introducing the Unsupervised QUEST Method – For text‑based QA, systems must answer directly even when questions involve multiple entities and relations across documents. Knowledge‑graph‑based QA can be limited by incompleteness and latency. In the paper “Answering Complex Questions by Joining Multi‑Document Evidence with Quasi Knowledge Graphs,” Alipay researchers propose QUEST, which constructs a noisy quasi‑knowledge graph from real‑time text sources, enriches it with entity types and semantic alignment, and applies a Group Steiner Tree algorithm to find the best answer. QUEST is unsupervised, avoiding training‑data bottlenecks, and demonstrates superior performance on complex questions.

Readers can click the “Read Original” link at the bottom left to visit the Ant Financial Technology website for more details.