Neural Lexical Cohesion in English
Abstract
Morris and Hirst (1991) describe that lexical cohesion is the result of relationships between related words that contribute to the continuity of lexical meaning. In this talk, I present modeling lexical cohesion in neural models. In particular, I mainly introduce our two projects regarding neural essay scoring, which is one of the typical applications in modeling text coherence.
Recent neural essay scoring systems outperform non-neural systems on a standard benchmark dataset. However, the latest works reached a plateau in performance. While they emphasize novel aspects of the neural architecture, they have neither shown significant improvement nor helped interpreting the scores assigned by humans. I first review the recent neural models for essay scoring, and analysis the main factor of their performance. Based on our observation, I introduce a simple lexical cohesion modeling to alleviate the limitation of the neural models. Lastly, I introduce two future works to deal with other limitations of neural models in other applications.
