This paper presents the current state of a work in progress, whose objective is to better understand the effects of factors that significantly influence the performance of the Latent Semantic Analysis (LSA). A difficult task, which is answering to biology MCQ, was used to test the semantic properties of truncated singular space and to study the relative influence of several parameters. An original and dedicated software eLSA1 has been used to fine tune the LSA semantic space for MCQ purpose. With the parameters of best configuration, the performances of our model were equal or superior to 7th and 8th grades students. Besides, global entropy weighting of answers was an important factor in the model's success.
View on arXiv