This repository contains the synthetic data for the paper: Overcoming Conflicting Data when Updating a Neural Semantic Parser, which was presented at the Workshop on NLP for Conversational AI (NLP4ConvAI) at EMNLP 2021.
The data is derived from, and presented in the same format as the data in: Semantic Parsing for Task Oriented Dialog using Hierarchical Representations
The generated data is present in the top_data/remodeling directory. Each of the 5 synthesized updates are in a separate subfolder of this directory, and each update folder is broken down into the V1, V2, and evaluation data for that update. All experiments in the paper are done on a targeted evaluation split specific to each update, which is stored in the devset folders in this repository. The original TOP testset is included in the testset folders, but was not used for numbers reported in the paper. Note that within the V1_trainset partitions, the data in the changed_relabeled folders represent "oracle" labels after the update and should not be used for training models.