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

In sequence-to-sequence tasks, sentences with heterogeneous semantics or grammatical structures may lead to some difficulties in the model's convergence. To resolve this problem, we introduce a feature-concentrated model that focuses on each of the heterogeneous features in the input-output sequences. Building upon the encoder-decoder architecture, we design a latent-enhanced multi-filter sequence-to-sequence model (LMS2S) that analyzes the features preserved by latent space representations and constructs the outputs accordingly. We divide the latent space into subspaces using a clustering algorithm and train a set of decoders in which each decoder only concentrates on the features from its corresponding subspace. We then design a self-enhancing mechanism that uses reinforcement learning to optimize the clustering algorithm. We perform two sets of experiments on semantic parsing and machine translation. We empirically demonstrate the advantages of the multi-filter architecture and show the performance improvement made by the self-enhancing mechanism.