Ranking-based readability assessment for early primary children’s literature

2012 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 548–552, Montr´eal, Canada, June 3-8, 2012. c⃝2012 Association for Computational Linguistics Ranking-based readability assessment for early primary children’s literature Yi Ma, Eric Fosler-Lussier Dept. of Computer Science & Engineering The Ohio State University Columbus, OH 43210, USA may,fosler@cse.ohio-state.edu Robert Lofthus Xerox Corporation Rochester, NY 14604, USA Robert.Lofthus@xerox.com Abstract Determining the reading level of children’s lit- erature is an important task for providing edu- cators and parents with an appropriate reading trajectory through a curriculum. Automating this process has been a challenge addressed before in the computational linguistics litera- ture, with most studies attempting to predict the particular grade level of a text. However, guided reading levels developed by educators operate at a more fine-grained level, with mul- tiple levels corresponding to each grade. We find that ranking performs much better than classification at the fine-grained leveling task, and that features derived from the visual lay- out of a book are just as predictive as standard text features of level; including both sets of features, we find that we can predict the read- ing level up to 83% of the time on a small cor- pus of children’s books. 1 Introduction Determining the reading level of a text has received significant attention in the literature, dating back to simple arithmetic metrics to assess the reading level based on syllable counts (Flesch, 1948). In the com- putational linguistics community, several projects have attempted to determine the grade level of a text (2nd/3rd/4th/etc). However, the education commu- nity typically makes finer distinctions in reading lev- els, with each grade being covered by multiple lev- els. Moreover, there are multiple scales within the educational community; for example 1st grade is ap- proximately covered by levels 3–14 on the Reading Recovery scale,1 or levels C to H in the Fountas and Pinnell leveling system.2 For grade-level assessment, classification and regression approaches have been very promising. However, it is not clear that an increased number of classes will allow classification techniques to suc- ceed with a more fine-grained leveling system. Sim- ilarly, regression techniques may have problems if the reading levels are not linearly distributed. In this work, we investigate a ranking approach to book lev- eling, and apply this to a fine-grained leveling prob- lem for Kindergarten through 2nd grade books. The ranking approach also allows us to be more agnostic to the particular leveling system: for the vast ma- jority of pairs of books, different systems will rank the levels of the books the same way, even if the exact differences in levels are not the same. Since most previous work uses classification techniques, we compare against an SVM multi-class classifier as well as an SVM regression approach. What has not received much attention in recent research is the visual layout of the page. Yet, if one walks into a bookstore and rummages through the children’s section, it is very easy to tell the reading level of …

Content truncated to 500 words