Sequence-to-Sequence Generation for Spoken Dialogue via Deep Syntax Trees and Strings
This work addresses natural language generation for spoken dialogue systems, offering an incremental improvement by optimizing the generation pipeline.
The authors tackled the problem of natural language generation for spoken dialogue by comparing two-step (separate sentence planning and surface realization) versus joint one-step approaches using sequence-to-sequence models trained on deep syntax trees and strings. The joint setup achieved better performance, surpassing state-of-the-art n-gram-based scores with more relevant outputs while requiring very little training data.
We present a natural language generator based on the sequence-to-sequence approach that can be trained to produce natural language strings as well as deep syntax dependency trees from input dialogue acts, and we use it to directly compare two-step generation with separate sentence planning and surface realization stages to a joint, one-step approach. We were able to train both setups successfully using very little training data. The joint setup offers better performance, surpassing state-of-the-art with regards to n-gram-based scores while providing more relevant outputs.