Applying natural language processing to automatically assess student conceptual understanding from textual responses

Abstract

In this study, we applied natural language processing (NLP) techniques, within an educational environment, to evaluate their usefulness for automated assessment of students’ conceptual understanding from their short answer responses. Assessing understanding provides insight into and feedback on students’ conceptual understanding, which is often overlooked in automated grading. Students and educators benefit from automated formative assessment, especially in online education and large cohorts, by providing insights into conceptual understanding as and when required. We selected the ELECTRA-small, RoBERTa-base, XLNet-base and ALBERT-base-v2 NLP machine learning models to determine the free-text validity of students’ justification and the level of confidence in their responses. These two pieces of information provide key insights into students’ conceptual understanding and the nature of their understanding. We developed a free-text validity ensemble using high performance NLP models to assess the validity of students’ justification with accuracies ranging from 91.46% to 98.66%. In addition, we proposed a general, non-question-specific confidence-in-response model that can categorise a response as high or low confidence with accuracies ranging from 93.07% to 99.46%. With the strong performance of these models being applicable to small data sets, there is a great opportunity for educators to implement these techniques within their own classes.

Implications for practice or policy:

• Students’ conceptual understanding can be accurately and automatically extracted from their short answer responses using NLP to assess the level and nature of their understanding.
• Educators and students can receive feedback on conceptual understanding as and when required through the automated assessment of conceptual understanding, without the overhead of traditional formative assessment.
• Educators can implement accurate automated assessment of conceptual understanding models with fewer than 100 student responses for their short response questions.