Ant Financial’s ZhiXiaoBao Team Achieves Human-Level Scores on SQuAD 2.0 and Advances Machine Reading Comprehension

Understanding and responding like humans remains a major challenge for AI, especially in natural‑language processing where semantics are highly abstract and rich.

SQuAD 2.0, the Stanford Question Answering Dataset, is a benchmark for machine reading comprehension built from Wikipedia; a recent Ant Financial (ZhiXiaoBao) team topped the leaderboard with scores exceeding human performance.

According to senior expert Lu Xin, this breakthrough enables significant efficiency gains in Ant’s business scenarios by mining and producing knowledge points, partially replacing human effort.

However, surpassing human scores does not mean machines can fully replace humans in open‑domain or highly specialized tasks; in finance, additional capabilities are still required for comprehensive advisory services.

Team lead Dong Yang explained that SQuAD 2.0 was chosen because its questions closely resemble ZhiXiaoBao’s business needs, the competition’s participant quality reflects industry advances, and the host’s authority adds credibility.

The team’s “extra heat from the stove” analogy describes how they use spare capacity to fine‑tune models for competitions while serving business needs.

Technical Deconstruction: “The Residual Heat of Roasting Sweet Potatoes”

Improved start‑end span ranking by adding extensive ranking logic and features.

Designed coarse‑to‑fine passage retrieval algorithms to handle overly long or short documents.

Applied robustness‑enhancing training such as adversarial text augmentation.

Employed data‑augmentation methods (back‑translation, EDA, CMRC, DRCD) to compensate for limited dataset size.

Integrated pretrained embeddings and AutoML for model architecture and hyper‑parameter search.

In the highly regulated financial domain, the team ensures compliance by using professional financial texts for training and subjecting outputs to audit before release.

Initially, the knowledge base was manually authored by experts, limiting scalability; the team’s machine‑reading approach now feeds large volumes of articles and real‑time user queries into the model, with expert review, dramatically increasing efficiency and user perception of intelligence.

Beyond reading comprehension, the team innovated human‑machine collaboration by automatically generating dialogue scripts from “every‑person‑talk” scenarios, extending the robot’s conversational capabilities.

The team comprises about 30 members, half of whom are algorithm engineers with strong academic backgrounds (30% PhDs, >95% from top Chinese universities), contributing to research papers across NLP and dialogue understanding.

Current deployment has produced over 10,000 knowledge points covering more than 300 products, achieving a “question‑answer‑correct” service model.

Metrics show a ~30% increase in click‑through rate for dialogue‑structured learning, ~50% boost in personalized recommendation clicks, and continuous improvement in answer accuracy and user satisfaction.

The technology also assists financial advisors by automating repetitive Q&A, improving their efficiency, with roughly 20% of user queries now supported by the system.

Future plans include enhancing semantic, numerical, and commonsense reasoning, integrating multi‑turn context, scaling low‑resource cross‑document reading, and opening the capabilities to institutions for broader fintech services.