Short question-answers assessment using lexical and semantic similarity based features

Abstract

Evaluation of short answers is a challenging task. As there could be more than one way of expressing the same thing in a sentence by quite different words and phrases, evaluation through computer-based system of Short answers requires natural language understanding. Study has performed comparative analysis for short answer assessment with regression algorithms namely: Support Vector Regression, Linear Regression, Bagging Tree, Boosting Tree, Multilayer Perceptron Regressor, and Random Forest on extracted features. It proposes the combined features that take account of lexical, approximate string matching, and semantic similarity features. An empirical evaluation of feature selection is also done that further improves the results. These combined features achieved improved results as 0.71 & 0.78 for correlation and RMSE values respectively.

Subject Classification:

• Primary 68W40
• Secondary 68Q55

Keywords:

• Machine learning
• Natural language processing
• Short answers grading
• Regression
• Semantic similarity
• Word embedding