Evaluation of Featureless Strategies for Relative Simplicity Prediction
College of Engineering
Department of Computer Science
Dr. Sole Pera
Text simplification involves replacing or rephrasing (sections of) a document while minimizing meaning loss in order to generate less complex versions of said document. This task has attracted researchers from both natural language processing and information retrieval domains of study. Text simplification has a direct connection to a wide range of applications, from those that demand specific reading comprehension ease, e.g., ensuring legal or medical documents are easy to read and understand, to those focused on making reading materials further accessible for K-12 audiences or non-native speakers. Evaluating outcomes from existing text simplification techniques, in terms of pairwise correctness of label ranking predictions, is a difficult task because these techniques tend to depend upon domain specific indicators such as lexical, syntactic, morpho-syntactic, and psycholinguistic features. Our hypothesis is that we can perform pair-wise text simplification ranking in a featureless fashion by relying on state-of-the-art learning architectures. To answer our research question and manage scope, we will use sentences as a case study. We will explore diverse techniques from baselines like Naïve Bayes to well-known counterparts based on deep learning. We then will conduct an exhaustive empirical analysis on multiple datasets on feature-based and featureless strategies, in order to demonstrate the validity of our proposed approach.
Karsann, Devan, "Evaluation of Featureless Strategies for Relative Simplicity Prediction" (2019). 2019 Undergraduate Research and Scholarship Conference. 82.