To Plan or not to Plan? Discourse planning in slot-value informed sequence to sequence models for language generation

Neha Nayak, Dilek Hakkani-Tur, Marilyn Walker, Larry Heck

Abstract: Natural language generation for task-oriented dialogue systemsaims to effectively realize system dialogue actions. All natu-ral language generators (NLGs) must realize grammatical, nat-ural and appropriate output, but in addition, generators for task-oriented dialogue must faithfully perform a specific dialogue actthat conveys specific semantic information, as dictated by thedialogue policy of the system dialogue manager. Most previouswork on deep learning methods for task-oriented NLG assumesthat generation output can be an utterance skeleton. Utterancesare delexicalized, with variable names for slots, which are thenreplaced with actual values as part of post-processing. How-ever, the value of slots do, in fact, influence the lexical selec-tion in the surrounding context as well as the overall sentenceplan. To model this effect, we investigate sequence-to-sequence(seq2seq) models in which slot values are included as part ofthe input sequence and the output surface form. Furthermore,we study whether a separate sentence planning module that de-cides on grouping of slot value mentions as input to the seq2seqmodel results in more natural sentences than a seq2seq modelthat aims to jointly learn the plan and the surface realization.