Lossy Compressive Text Autoencoders with Semantic Preservation

A new autoencoder architecture uses residual downscaling and a discrete bottleneck to achieve 2.24 bits per byte compression on web text. The method maintains high reconstruction accuracy and semantic performance on downstream QA and similarity benchmarks, comparable to lossless compression.