Estimation Improvement of Objective Scores of Answer Statements with Consideration of Multicollinearity and Semantic Similarity

Abstract

To eliminate mismatches between the intentions of questioners and respondents of Question and Answer (Q&A) sites, we have clarified that the impression of the statements could be captured by nine factors, and the factor scores could be estimated from the feature values of the statements. Objective scores of the statements could be estimated fairly good, and that those of subjective statements could be estimated well by taking the natural logarithm of the factor scores with the consideration of cross-validation. With the consideration of semantic similarity between Q&A, semantic similarity could be effective in estimating objective scores in the previous preliminary analysis. This paper tries to perform multiple regression analysis with the consideration of multicollinearity between feature values of semantic similarity with those other than semantic similarity. As a result of analysis, feature values of semantic similarity was obtained as a part of regression formula in 8 out of 15 trials. With the consideration of semantic similarity, average estimation error becomes smaller for three categories.