Quantifying the contribution of different knowledge sources in narrative-based text understanding

Authors: Steels. L, Verheyen. L.

BNAIC/BeNeLearn 2022. To appear in Springer’s CCIS series

  • Date:

    November 2022

Narrative-based text understanding is the process of making a rich model of the situations being evoked by a text. This process goes significantly beyond coarse-grained natural language processing and even beyond the finegrained linguistic analysis of a text based on lexicons and grammars. It requires integrating information gleaned from ontologies and common sense knowledge bases, logical inference, qualitative mental simulation, image interpretation (if the text is accompanied with images), discourse modeling, quantitative physical simulation, sensori-motor action and inclusion of extra-textual context. This paper introduces a novel way to integrate and measure the contributions from different knowledge sources in narrative-based understanding based on framing understanding as a process of raising questions, finding answers to questions, and interlinking questions. The proposed measures can be used to evaluate narrativebased understanding, evaluate the quality of texts, and provide a feedback signal for machine learning to improve the efficiency and efficacy of the understanding
process. We illustrate the paper with an operational system that goes some way
to understanding cooking recipes.