An Evaluation Model for Auto-generated Cognitive Scripts

Autonomous intelligent agents have become a very important research area in Artificial Intelligence (AI). Socio-cultural situations are one challenging area in which autonomous intelligent agents can acquire new knowledge or modify existing one. Socio-cultural situations can be best represented in the form of cognitive scripts that can allow different techniques to be used to facilitate knowledge transfer between scripts. Conceptual blending has proven successful in enhancing the social dynamics of cognitive scripts, where information is transferred from similar contextual scripts to a target script resulting in a new blended script. To the extent of our knowledge, there is no computational model available to evaluate these newly generated cognitive scripts. This work aims to develop a computational model to evaluate cognitive scripts resulting from blending two or more linear cognitive scripts. The evaluation process involves: 1) using the GloVe similarity to check if the transferred events conceptually fit the target script; 2) using the semantic view of text coherence to decide on the optimal position(s) to place the transferred event(s) in the target script. Results show that the GloVe similarity can be applied successfully to preserve the contextual meaning of cognitive scripts. Additional results show that GloVe embedding gives higher accuracy over Universal Sentence Encoder (USE) and Smooth Inverse Frequency (SIF) embedding but this comes with a high computational cost. Future work will look into reducing the computational cost and enhancing the accuracy.