TextWorld KG Dataset

Dataset proposed in the paper Building Dynamic Knowledge Graphs from Text-based Games.

This project is part of the project of Learning Dynamic Knowledge Graphs to Generalize on Text-based Games. Which we have released code in another repo.

Statistics

#Train	#Valid	#Test	Avg. Obs.	Avg. #Operations	#Vertices	#Edges	Avg. #Connections
267,031	13,442	41,865	29.3 tokens	3.1	99	10	43.1

Data Format

We have split the dataset (dataset.zip, download) into training, validation, and test sets, each of them is in a .json file.

In a .json file, each line is a json dump, which is a single data point.

Each data point consists of the following elements:

• previous_triplets: ground truth (partially observable) graph at previous game step represented by triplets;
• previous_action: text action an agent performed at previous game step;
• observation: observation at current game step, which is resulted by previous_action;
• target_commands: graph updates represented by text commands (operations).