Self-Enhancing Multi-filter Sequence-to-Sequence Model

Representation learning is important for solving sequence-to-sequence problems in natural language processing. transforms raw data into vector-form representations while preserving their features. However, with significantly different features leads to heterogeneity representations, which may increase the difficulty of convergence. We design a multi-filter encoder-decoder model resolve problem tasks. The divides latent space subspaces using clustering algorithm and trains set decoders (filters) each decoder only concentrates on from its corresponding subspace. As main contribution, we self-enhancing mechanism that uses reinforcement optimize without additional training data. run semantic parsing machine translation experiments indicate proposed can outperform most benchmarks by at least 5%. also empirically show improve performance over 10% provide evidence demonstrate positive correlation between model's clustering.