Multi-Task NLP on Noisy Multilingual E-Commerce Data: BiLSTM vs AutoML Without Large-Scale Pretraining

arXiv·low signal

Manurung, Al-Kahfi, and Rizqi benchmark multi-task BiLSTM against AutoML approaches for simultaneous sentiment and five-class emotion classification on 5,400 Indonesian marketplace reviews mixing standard vocabulary with slang, regional loanwords, numeric shorthands, and emoji. The work demonstrates that lightweight multi-task architectures handle noisy multilingual text without large-scale pretraining — relevant for teams building NLP systems for non-English markets with messy real-world user data.