Exploiting Peer Trust and Semantic Similarities in the Assignment Assessment Process

In many scenarios, the assessment by a single expert of all the content produced by an individual may be impractical due to the overall vast amount of content to be assessed by the expert. Examples are, for instance, online education services with thousands of students or scientific papers submitted to a conference that have to be assessed by program chairs in a short time period. Leveraging peer evaluations is a crucial strategy to mitigate assessment burdens and reduce the time required to deliver the expected results.
This paper revisits the foundational concept of Personalised Automated Assessment (PAAS), which seeks to approximate the assessments of a particular community member, known as the leader, by integrating the peer assessments among the other community members of their answers to an assignment. Our extension of PAAS enhances its machine learning capabilities by integrating in the algorithm the semantic similarity among peer assessments to improve its prediction power. Experimental validation using synthetic and real-world datasets shows the efficacy of our extension, reducing prediction errors and increasing accuracy, especially in scenarios where the several assignments are significantly similar with one another.