الفهرس | Only 14 pages are availabe for public view |
Abstract The manual correction of short answer questions is a tedious and time consuming task. Thus, the automatic correction for short answer questions is one of the important text similarity applications. It grades a student answer via calculating its similarity degree with the model answer. Several Short Answer Correction (SAC) models exist in the literature. However, the current SAC models suffer from being inaccurate where the best accuracy degree measured by the correlation coefficient reaches 0.59 when tested on Texas data set. Thus, in this dissertation, we propose a deep learning approach for correcting short answer questions using three different models.The first one is concatenation trainable model which is implemented using the multichannel CNN with 50-dimension embedding size, multichannel CNN with 300-dimension, Bidirectional LSTM with 300-dimension embedding size, Bidirectional GRU with 300-dimension embedding size, and CNNLSTM with 300-dimension embedding size.The second model is the concatenation nontrainable model, which is implemented using Global Vectors, Glove (50-d), on multichannel CNN and CNN-LSTM. The last model is Siamese trainable model which is implemented using multichannel CNN with 300-dimension embedding size, and CNNLSTM with 300-dimension embedding size |