We reproduce the Structurally Constrained Recurrent Network (SCRN) model, and then regularize it using the existing widespread techniques, such as naïve dropout, variational dropout, and weight tying. We show that when regularized and optimized appropriately the SCRN model can achieve performance comparable with the ubiquitous LSTM model in language modeling task on English data, while outperforming it on non-English data.

Abstract

The authors replicate the Structurally Constrained Recurrent Network (SCRN) model and apply modern regularization techniques—including naïve dropout, variational dropout, and weight tying. They demonstrate that, with proper regularization and optimization, SCRN achieves performance comparable to LSTM on English language modeling tasks and even surpasses LSTM on morphologically rich languages.