Automatic Quality Assessment of Documents with Application to Essay Grading

Abstract

In this paper, we focus on automatic quality assessment for intelligent essay grading. Our devised system grades essays without depending upon completely overlapping essays in training data. This increases the scope of devised system due to list dependency on highly topic focused labeled data for automatic essay grading. Instead of depending upon direct topic specific matching w.r.t., training data, the devised system judge the quality of essay by exploiting knowledgebase documents and SentiWordNet, etc. To achieve this goal, we concentrate on five different features: (1) relevance of information, (2) presence of sparsely connected words, (3) statistical and semantic role of words, (4) presence of talkative terms and (5) length of essay. We extract all these features by using word graph of text, populated with statistical, semantic and topical relation between words. Next, we use graph theoretical techniques, like: weighted all pair shortest paths, Ego-Networks, entropy based measures for effectiveness of nodes in weighted graph and statistical and probabilistic techniques like: total correlation score and Point wise Mutual Information (PMI) etc. Our experimental result on standard dataset shows that our devised system performs better than state-of-the-Art systems of this area.