查看文章

We have set out to create a system that gives essay feedback by identifying stronger and weaker parts of a given essay. In order to do this, we have created a two step model: the first step uses an LSTM to turn a chunk of an essay into an encoding vector. The second step uses a feed forward network to assign a score to an encoding. We believe that if the feed forward network is able to predict the score of an essay based on the average of the encodings of the chunks, then it should be able to distinguish between the stronger and weaker chunks as well. We have achieved an RMSE of 0.17 on the task of predicting the score of an essay. While this is not comparable to human grading, it is good enough to be able to determine when there is significant variation in quality between different parts of the same essay.