Attention Regularized Sequence-to-Sequence Learning for E2E NLG Challenge

Abstract

This paper describes our system used for the end-to-end (E2E) natural language generation (NLG) challenge. The challenge collects a novel dataset for spoken dialogue system in the restaurant domain, which shows more lexical richness and syntactic variation and requires content selection (Novikova et al., 2017). To solve this challenge, we employ the CAEncoder-enhanced sequence-tosequence learning model (Zhang et al., 2017) and propose an attention regularizer to spread attention weights across input words as well as control the overfitting problem. Without any specific designation, our system yields very promising performance. Particularly, our system achieves a ROUGE-L score of 0.7083, the best result among all submitted primary systems.

Tasks

Reproductions