This repo is for Measuring and Reducing Model Update Regression in Structured Prediction for NLP. NeurIPS2022
Our experiments are on two typical structured prediction tasks, dependency parsing and conversational semantic parsing.
The folder dp
contains the code and instructions to train different dependency parsers and get the parses from them.
The folder top
contains the code and instructions to train different conversational semantic parsers and get the parses from them.
The folder main
contains the code to analyze parses. The jupyter notebooks there directly reproduce the numbers shown in our papers.
The readme file in each folder gives more details.