At testing time, we do not need the distant supervision signals anymore. At training time, we use distant supervision labels to train the neural classifier to predict the four edit types. The main differences between testing and training is the distant supervision labels generated through preprocessor. If you are interested in training your ABCD model, we provide scripts of doing so.

You could test this pre-trained model on your data. We provide ABCD model trained on MinWiki (Wikipedia Text). Depending on your applications and data, ABCD could be flexibly trained to keep (or drop) connectives, or to perform simplification. It first constructs a sentence graph using dependency parsing information, and edits the graph into subgraphs by a neural classifier with four graph operations: A(accept), B(break), C(copy) and D(drop). IntroductionДЄBCD is a linguistically motivated sentence editor that decomposes a complex sentence into N simple sentences, where N corresponds to the number of predicates in the complex sentence. You could also find the slides from my oral presentation. Please cite our paper if you are using ABCD (BibTex at the end). This is the github repository for the ACL 2021 paper: ABCD: A Graph Framework to Convert Complex Sentences to a Covering Set of Simple Sentences.

ABCD: A Graph Framework to Convert Complex Sentences to a Covering Set of Simple Sentences

Abcd 2 wiki