Training Shallow Phrase Structure Parsers From Semantic Frames and Simple NLG

Anonymous

Keywords: parsing, human in the loop, sequence tagging

TL;DR: standoff sematic representations from human in loop plus simple NLG yields fast and easy way to create training data for NLU models

Abstract:

Determining the meaning of customer utterances is an important part
of fulfilling customer requests in task-oriented dialogue.
Natural Language Understanding (NLU) models can determine this meaning,
but typically require many customer utterances that are hand-annotated with
meaning representations, which are difficult to obtain and must be
repeated for each new target domain. One way to reduce the labor
involved in hand annotation is to have the human annotate a meaning
representation (a ``semantic frame'' representation) separate from the
corresponding utterance. In this work, we investigate the use of this
approach in conjunction with several simple natural language
generation (NLG) approaches in order to train shallow parsers
to extract phrase structure representations from customer
utterances. Our results show the effectivness of this approach for
training NLU models.