Alibaba’s PAI Platform Powers Three Groundbreaking ACL 2023 AI Papers

Paper Summaries

FashionKLIP: Enhancing E‑Commerce Image‑Text Retrieval with a Fashion Multi‑Modal Conceptual Knowledge Graph

Image‑text retrieval is crucial for e‑commerce. The paper introduces FashionKLIP, which builds a large‑scale e‑commerce multimodal knowledge graph (FashionMMKG) and aligns image‑text representations with visual prototypes of fashion concepts, achieving superior zero‑shot performance on Alibaba’s product search platform.

ConaCLIP: Distilling a Dual‑Encoder Model for Lightweight Text‑Image Retrieval

The authors propose a distillation framework that uses a fully‑connected knowledge interaction graph to transfer knowledge from large dual‑encoder models to compact student networks, enabling efficient retrieval on edge devices while maintaining accuracy.

Rapid Diffusion: Domain‑Specific Text‑to‑Image Synthesis with Fast Inference

This work injects domain knowledge into CLIP’s text encoder via a knowledge graph and combines a diffusion model with an ESRGAN up‑sampler, accelerated by FlashAttention and the BladeDISC compiler, delivering high‑quality, fast‑inference image generation for Chinese e‑commerce scenarios.

All three algorithms will be open‑sourced in Alibaba’s EasyNLP framework on GitHub, providing the community with ready‑to‑use implementations.