A NAACL-HLT 2018 ACCEPTED PUBLICATION

Zero-Shot Question Generation from Knowledge Graphs for Unseen Predicates and Entity Types

Authors Hady Elsahar, Christophe Gravier and Frederique Laforest.

Preprint

The Github Repo consists of the following folders:

• baselines : containing all the baselines RTransE,IR,SELECT
• models: containing code for the encoder-decoder baseline triples2seq.py and our-model: tripletext2seq.py
• kbembeddings: code forked and adapted from https://github.com/thunlp/KB2E
• evaluator: code for calculating BLEU, METEOR, ROUGE forked and adapted from https://github.com/tylin/coco-caption
• data: containing data for training and testing as well as data.py python modules that is responsible for creating training/test sets in zeroshot or normal setups.

Running Model

** Prep Repo **

From the repository base folder run the script that downloads extra files and setup the repository . prepare_repo.sh

The file run.py takes some arguments described in the begining of the file and responsible for running training the our-models or the encoder-decoder baseline.

Experiment params:

• epochs: number of epochs to train
• setup: either zeroshot or normal train/test/valid splits using the regular splits from SimpleQuestions
• criteria: the zeroshot criteria pred for unseen pred, subtype or objtype
• min: minimum number of samples for each predicate to keep
• datapath: path of the folder with the preprocessed files
• kfolds : number of kfolds
• fold: 0-9 which cross validation split

picking model params:

• pred: add textual context for predicates (flag with no values)
• subtype : add textual context for subject types (flag with no values)
• objtype : add textual context for object types (flag with no values)

Examples:

How to run zeroshot experiments for unseen-predicates :

Encoder-Decoder: No textual contexts or copy actions:

python run.py -epochs 10 -min 50 -setup zeroshot -criteria pred -datapath ./data/nocopy -fold 1

Our-Model the model with textual contexts but no copy actions:

python run.py -epochs 10 -min 50 -setup zeroshot -criteria pred -fold 1 -datapath ./data/nocopy/ -subtype -objtype -pred

Our-Model+Copy full model with textual contexts and copy actions:

python run.py -epochs 10 -min 50 -setup zeroshot -criteria pred -fold 1 datapath ./data/copy/ -subtype -objtype -pred

Browsing predicate textual evidence:

The file ./data/pred_textual_contexts.csv contains textual contexts collected for the freebase relations.