Penilaian esei berbantukan komputer menggunakan teknik Bayesian dan pengunduran linear berganda

Abstract

Disagreement of grade given by two human judges, time consuming and high evaluation cost became a reason of research on Computer-based Assessment System (CbAS) been studied. The main key is CbAS assessment must be closest to human assessment. Based on UPSR Essay Assessment Schema, there are three main assessment components consists of language, discourse element and style. Recently, Fuzzy Logic is used to determine and classify the discourse element while Stepwise Linear Regression Algorithm (SLR) is used to make closest prediction for style of writing. Both of them have its weakness. Fuzzy Logic did not measure the form of linguistic features and required a huge size of training data. SLR Algorithm derive prediction of writing style using un-standardize feature set and size of features set not clearly defined and no warranty of significance in contribute to get closest grade prediction. This study emphasized on optimization of prediction on discourse elements and writing style that leading to the development of CbAS through four phases of research methodology. (1) Pre-processing and data extraction phase where essay will be parsed into word (token) and implemented Word Correction Algorithm to re-correct the misspell word. (2) Training process of determination and classification of discourse elements using Multivariate Bernoulli Model (MMB) Technique. It considers both presence and absence features thus it measured the form of linguistic features that reflected essay quality. MMB Technique only required a small size of training data. (3) Prediction process of writing style using Multiple Linear Regression (MLR) Algorithm. MLR Algorithm applied six fixed features (based on previous research) to ensure the prediction is more standardize and feature set is more significant. (4) Test the performance agreement derived from the combination of MMB, MLR and data of language component (taken from human assessment) and compared it to human assessment for five cycles of cross-validation. The outcome shows performance is consistent with 95.2% agreement. Thus, the experiment has shown by utilizing both techniques (MMB and MLR), better prediction or essay assessment has been achieved compared to the one’s implemented using Fuzzy Logic and SLR Algorithm.