May 3, 2018

We propose an approach to local coherence modeling that represents how the Focus of Attention, henceforth FoA, moves through a text.  The FoA representation is defined for any two adjacent sentences by a vector that is obtained by averaging vector representations of most semantically related words of the sentences. The transition between two adjacent FoA vectors is quantified by computing the cosine similarity between them.  Finally, the local coherence of a text is encoded by a vector representing patterns of transitions that are automatically extracted by a convolutional neural network.  Our experiments demonstrate that our new approach to coherence  modeling by automatically extracting and representing FoA transition patterns offers a benefit in two evaluation tasks:
The readability ranking task in that our model achieves the state-of-the-art result; The essay scoring task in that integrating our coherence vectors into an existing essay scoring system significantly improves the performance of the essay scorer.