ronaldseoh / qanli

Source code for "Transforming Question Answering Datasets Into Natural Language Inference Datasets"

QA2D

A model for transforming questions + short answers into full answer sentences.

The dataset and the models are described in the following paper:

Demszky, D., Guu, K., & Liang, P. (2018). Transforming Question Answering Datasets Into Natural Language Inference Datasets. arXiv preprint. arXiv:1809.02922.[link]

This repo contains the code and examples for both the rule-based model and the neural model.

Data available on Codalab.

Rule-based model

We illustrate how to use the rule-based model in the designated jupyter notebook. The input sentences have to be dependency parsed. We created our example file in the following manner:

1. Save your tokenized (space-separated) questions and short answers in a file, such as examples.txt, where each sentence is a line, one example after the other (i.e. question 1 <line-break> short answer 1 <line-break> question 2 <line-break> short answer 2 <line-break> ... question N <line-break> short answer N)
2. Convert this file into CoNLL-U format, examples.conllu, with the tags and labels left empty (_).
3. POS tag the file. For ours, we used the parser by Dozat et al. (2017), which can be used as a tagger as well.
4. Dependency parse the file. We parsed ours with Dozat et al. (2017).
5. Use the resulting, tagged and parsed examples.conllu file as an input for the model, as shown in the jupyter notebook Rule-based Example.ipynb.

Neural model

Coming soon.