應用潛在語意分析於試題相似度比較之可行性

Abstract

本研究旨在應用潛在語意分析(Latent semantic analysis，LSA)模型於試題相似度之判斷，並探討不同的評分函式對於結果的影響，同時根據試題關鍵字的特性，與LSA模型處理詞彙共現(Lexically Co-occur)的特性，提出訓練文件可採用相關文件來提高判斷的精確率。研究結果使用dice或內積為評分函式較接近專家評鑑結果，對於專家相似度評鑑比較一致的試題，有高達0.9的相關程度，而平均相關值也有0.7以上的相關程度，因此潛在語意分析應用於試題相似度是可行的技術。The purpose of this study is to apply latent semantic analysis (LSA) to analyze item similarity , and discuss the result of using different score function. The feature of LSA model is “Lexically Co-occur” detection , in other words, LSA model can analyze many documents, and find synonyms , but synonyms rarely exist in the same item , so LSA model needs to be trained by documents which are related to this item . This study revealed that the result using dice measure or inner product measure correlates more closely with expert’s scores. For the items which is more agreeable of expert’s scores than others , the maximum correlation is up to 0.9, and the mean of correlation is up to 0.7, so applying latent semantic analysis to analyze item similarity is a feasible technology.