Ant Financial Research Highlights at WWW2020: Enhanced‑RCNN, IntentDial, Captcha Solver, Billion‑Scale Knapsack, and EET Loss

From April 20‑24, 2020, the prestigious WWW2020 conference was held online due to the COVID‑19 pandemic, receiving 1,129 submissions and accepting 217 papers (19.2% acceptance rate). Ant Financial had multiple papers selected, showcasing its advances in intelligent services and cognitive computing.

Enhanced‑RCNN: An Efficient Method for Comparing Sentence Similarity – This paper proposes a novel model that improves upon the ESIM architecture to better capture information within and between two texts. Tested on the Quora Question Pair and Ant Financial datasets, Enhanced‑RCNN achieves competitive results comparable to BERT while using far fewer parameters, making it suitable for online deployment. Knowledge distillation can further boost its accuracy by treating BERT‑Base as a teacher model.

IntentDial: An Intent Graph‑Based Multi‑Turn Dialogue System with Reasoning Path Visualization – The work introduces a dialogue system that incorporates a graph of user intents and reinforcement learning. The graph injects domain knowledge to accelerate learning, and the generated reasoning paths provide high interpretability, aiding both model analysis and downstream entity recognition.

A Generic Solver Combining Unsupervised Learning and Representation Learning for Breaking Text‑Based Captchas – Addressing the security challenge of text captchas, the authors design a method that leverages unsupervised and representation learning to reduce reliance on large labeled datasets. With only 500 labeled samples, the approach can break most mainstream website captchas, revealing vulnerabilities in current captcha designs.

Solving Billion‑Scale Knapsack Problems – The paper presents a formulation and algorithm for ultra‑large knapsack problems common in internet scenarios such as red‑packet marketing and traffic allocation. It supports multi‑dimensional capacities and constraints, enabling efficient decision‑making over billions of variables.

EET Loss for Image Retrieval: Equal‑Distance and Equal‑Distribution Triplet Loss – To overcome limitations of traditional triplet loss, the authors propose the EET method, which adds equal‑distance constraints for matching pairs and equal‑distribution constraints for non‑matching pairs, improving global feature space alignment and achieving superior performance on multiple retrieval benchmarks.

The article concludes by noting Ant Financial’s ongoing contributions to top‑tier conferences, bridging cutting‑edge AI research with real‑world applications.