This is a t5-small pretrained model, fine-tuned on the task of generating QANom QAs.

"QANom" stands for "QASRL for Nominalizations", which is an adaptation of QASRL (Question-Answer driven Semantic Role Labeling) for the nominal predicates domain. See the QANom paper for details about the task. The QANom Dataset official site is a Google drive, but we also wrapped it into a Huggingface Dataset, which is easier to plug-and-play with (check out our HF profile for other related datasets, such as QASRL, QAMR, QADiscourse, and QA-Align).

Visit our demo for interactively exploring our model!

The model and tokenizer can be downloaded as simply as running:

However, the model fine-tuning procedure involves input preprocessing (marking the predicate in the sentence, T5's "task prefix", incorporating the predicate type and/or the verbal for of the nominalization) and output postprocessing (parsing the sequence into a list of QASRL-formatted QAs).
In order to use the model for QANom parsing easily, we suggest downloading the pipeline.py file from this repository, and then use the QASRL_Pipeline class:

Which will output:

You can learn more about using transformers.pipelines in the official docs.

Notice that you need to specify which word in the sentence is the predicate, about which the question will interrogate. By default, you should precede the predicate with the <predicate> symbol, but you can also specify your own predicate marker:

In addition, you can specify additional kwargs for controling the model's decoding algorithm: