Reliable prediction of enzymatic kinetic parameters remains challenging due to limited structured datasets and biased evaluation strategies. We present a unified framework that bridges multimodal literature mining and enzyme kinetics prediction. zERExtractor automatically converts molecular images, tables, and text from scientific literature into structured reaction records, enabling scalable dataset construction. Building upon these data, a PLE-based multi-task model jointly predicts \({\it k}_\text{cat}\), \({\it K}_\text{m}\), and \({\it K}_\text{i}\) using protein language model representations. Experiments under sequence-clustered unbiased splits demonstrate advantages over independent single-target models. The proposed framework provides an practical workflow for integrating literature-derived knowledge with reliable enzyme kinetics prediction.