MUHAI Benchmark : Task 2 (Credibility of knowledge-based generated gossip stories)

Meaning and Understanding in Human-Centric AI (MUHAI) Benchmark
Task 2 (Credibility of knowledge-based generated gossip stories)

This dataset aims at investigating whether the use of Knowledge Graphs has an impact on the credibililty of automatically-generated stories.

The submission includes the following data:

1. Generated stories (.txt)
2. Story generation template 
3. A tsv file with entities and triples (to be used for generating stories)
4. Evaluation description : Questions and metrics submitted to the users

The "gossip stories" are generated with the T5 languge model fine-tuned on the WebNLG challenge. The model takes the triples (file 3) as input and generates one sentence each. A link prediction algorithm based on Jaccard's similarity learns the likelihood of two entities to be related (3). Then, the narrative continues with automatically generated celebrity background descriptions.

The credibility of the story is evaluated using a questionnaire based on Gaziano et. al. The questionnaire was filled in by the test subjects after reading each generated article. One for a KG-generated text where links were predicted using the link prediction and one for text that was generated using triples of random entities (celebrities). 

Full code available at : https://github.com/kmitd/muhai-credibility-KR