Prepare dataset
Dataset
In this tutorial, we train a model to identify argument structure constructions (ASCs) in English by assigning construction labels to verbs.1
Data format
The data have already been prepared for this tutorial. Each line contains one sentence, and each word is followed by | and its label:
And|O we|O packed|INTRAN_RES up|O .|O
After|O all|O this|O problems|O I|O became|INTRAN_RES dissappointed|O .|O
Verbs annotated with an ASC category receive a label such as INTRAN_RES or TRAN_S. All other tokens receive O, indicating that no ASC label is assigned. The original spelling and grammatical errors are preserved. This task is called token classification: the model learns to predict a label for each token. In the first example, the target label for packed is INTRAN_RES, while the target label for every other token is O.
Download the data
Download asc-simple.zip and extract it. The archive contains three files:
train.txt: used to train the model.dev.txt: used to evaluate the model during development and guide model selection.test.txt: used to evaluate the final model on held-out data.
-
The data are from the ASC Treebank and were used in Sung and Kyle’s (2024), Leveraging pre-trained language models for linguistic analysis: A case of argument structure constructions. ↩