Editing Readability (section)

=== Artificial intelligence ===

Unlike the traditional readability formulas, [[artificial intelligence]] approaches to readability assessment (also known as Automatic Readability Assessment) incorporate myriad linguistic features and construct statistical prediction models to predict text readability.<ref name="Text Readability Assessment for Sec">{{cite journal |last1=Xia |first1=Menglin |last2=Kochmar |first2=Ekaterina |last3=Briscoe |first3=Ted |date=June 2016 |title=Text Readability Assessment for Second Language Learners |url=https://www.aclweb.org/anthology/W16-0502 |journal=Proceedings of the 11th Workshop on Innovative Use of NLP for Building Educational Applications |pages=12–22 |arxiv=1906.07580 |doi=10.18653/v1/W16-0502 |doi-access=free}}</ref><ref name="aclweb.org">{{cite journal |last1=Lee |first1=Bruce W. |last2=Lee |first2=Jason |title=LXPER Index 2.0: Improving Text Readability Assessment Model for L2 English Students in Korea |journal=Proceedings of the 6th Workshop on Natural Language Processing Techniques for Educational Applications |date=Dec 2020 |pages=20–24 |doi=10.18653/v1/2020.nlptea-1.3 |arxiv=2010.13374 |url=https://www.aclweb.org/anthology/2020.nlptea-1.3}}</ref> These approaches typically consist of three steps: 1. a training corpus of individual texts, 2. a set of linguistic features to be computed from each text, and 3. a [[machine learning]] model to predict the readability, using the computed linguistic feature values.<ref>{{cite journal |last1=Feng |first1=Lijun |last2=Jansche |first2=Martin |last3=Huernerfauth |first3=Matt |last4=Elhadad |first4=Noémie |title=A Comparison of Features for Automatic Readability Assessment |journal=Coling 2010: Posters |date=August 2010 |pages=276–284 |url=https://www.aclweb.org/anthology/C10-2032}}</ref><ref name="On Improving the Accuracy of Readab">{{cite journal |last1=Vajjala |first1=Sowmya |last2=Meurers |first2=Detmar |title=On Improving the Accuracy of Readability Classification using Insights from Second Language Acquisition |journal=Proceedings of the Seventh Workshop on Building Educational Applications Using NLP |date=June 2012 |pages=163–173 |url=https://www.aclweb.org/anthology/W12-2019}}</ref><ref name="aclweb.org" />

In 2008, it was shown that syntactic complexity is correlated with longer processing times in text comprehension.<ref>{{cite journal |last1=Gibson |first1=Edward |title=Linguistic complexity: locality of syntactic dependencies |journal=Cognition |date=1998 |volume=68 |issue=1 |pages=1–76|doi=10.1016/S0010-0277(98)00034-1 |pmid=9775516 |s2cid=377292 }}</ref> It is common to use a rich set of these syntactic features to predict the readability of a text. The more advanced variants of syntactic readability features are frequently computed from [[parse tree]]. Emily Pitler ([[University of Pennsylvania]]) and Ani Nenkova (University of Pennsylvania) are considered pioneers in evaluating the parse-tree syntactic features and making it widely used in readability assessment.<ref>{{cite journal |last1=Pitler |first1=Emily |last2=Nenkova |first2=Ani |title=Revisiting Readability: A Unified Framework for Predicting Text Quality |journal=Proceedings of the 2008 Conference on Empirical Methods in Natural Language Processing |date=October 2008 |pages=186–195 |url=https://www.aclweb.org/anthology/D08-1020}}</ref><ref name="Computational assessment of text re" /> Some examples include:
*Average sentence length
*Average parse tree height
*Average number of noun phrases per sentence
*Average number of verb phrases per sentence

Lijun Feng proposed some cognitively-motivated features (mostly lexical) in 2009. This was during her [[doctorate]] study at the [[City University of New York]].<ref>{{cite book |last1=Feng |first1=Lijun |last2=Elhadad |first2=Noémie |last3=Huenerfauth |first3=Matt |chapter=Cognitively motivated features for readability assessment |title=Proceedings of the 12th Conference of the European Chapter of the Association for Computational Linguistics on – EACL '09  |date=March 2009 |pages=229–237 |doi=10.3115/1609067.1609092 |s2cid=13888774 |chapter-url=https://dl.acm.org/doi/10.5555/1609067.1609092|doi-access=free }}</ref> The cognitively-motivated Although cognitively-motivated features were originally designed to ensure comprehension by adults with [[intellectual disability]], Feng showed such features also improved reading comprehension among the general population. In combination with a [[logistic regression]] model, cognitively-motivated features can correct the average error of [[Flesch–Kincaid readability tests|Flesch–Kincaid grade-level]] by more than 70%. The newly discovered features by Feng include:
*Number of [[lexical chain]]s in document
*Average number of unique entities per sentence
*Average number of entity mentions per sentence
*Total number of unique entities in document
*Total number of entity mentions in document
*Average lexical chain length
*Average lexical chain span

In 2012, Sowmya Vajjala at the [[University of Tübingen]] created the WeeBit corpus by combining educational articles from the [[Weekly Reader]] website and [[BBC Bitesize]] website, which provide texts for different age groups.<ref name="On Improving the Accuracy of Readab" /> In total, there are 3125 articles that are divided into five readability levels (from age 7 to 16). Weebit corpus has been used in several AI-based readability assessment research.<ref name="Computational assessment of text re">{{cite journal |last1=Collins-Thompson |first1=Kevyn |title=Computational assessment of text readability: A survey of current and future research |journal=International Journal of Applied Linguistics |date=2015 |volume=165 |issue=2 |pages=97–135|doi=10.1075/itl.165.2.01col |s2cid=17571866 }}</ref>

Wei Xu ([[University of Pennsylvania]]), Chris Callison-Burch ([[University of Pennsylvania]]), and Courtney Napoles ([[Johns Hopkins University]]) introduced the [[Newsela]] corpus to the academic field in 2015.<ref>{{cite journal |last1=Xu |first1=Wei |last2=Callison-Burch |first2=Chris |last3=Napoles |first3=Courtney |title=Problems in Current Text Simplification Research: New Data Can Help |journal=Transactions of the Association for Computational Linguistics |date=2015 |volume=3 |pages=283–297|doi=10.1162/tacl_a_00139 |s2cid=17817489 |doi-access=free }}</ref> The corpus is a collection of thousands of news articles professionally leveled to different reading complexities by professional editors at [[Newsela]].  The corpus was originally introduced for [[text simplification]] research, but was also used for text readability assessment.<ref>{{cite journal |last1=Deutsch |first1=Tovly |last2=Jasbi |first2=Masoud |last3=Shieber |first3=Stuart |title=Linguistic Features for Readability Assessment |journal=Proceedings of the Fifteenth Workshop on Innovative Use of NLP for Building Educational Applications |date=July 2020 |pages=1–17 |doi=10.18653/v1/2020.bea-1.1 |arxiv=2006.00377 |url=https://www.aclweb.org/anthology/2020.bea-1.1|doi-access=free }}</ref>

Advanced semantic or semantic features' influence on text readability was pioneered by Bruce W. Lee during his study at the ([[University of Pennsylvania]]), in 2021. Whilst introducing his features hybridization method, he also explored handcrafted advanced semantic features which aim to measure the amount of knowledge contained in a given text.<ref>{{cite book |last1=Lee |first1=Bruce W. |last2=Jang |first2=Yoo Sung |last3=Lee |first3=Jason Hyung-Jong |chapter=Pushing on Text Readability Assessment: A Transformer Meets Handcrafted Linguistic Features |title=Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing |series=EMNLP '21 |date=November 2021 |pages=10669–10686 |doi=10.18653/v1/2021.emnlp-main.834 |s2cid=237940206 |chapter-url=https://aclanthology.org/2021.emnlp-main.834/|arxiv=2109.12258 }}</ref>
*Semantic Richness : <math>\sum_{i=1}^{n}  p_i \cdot i</math>
*Semantic Clarity : <math>\frac{1}{n} \cdot \sum_{i=1}^{n}  max(p) - p_{i}</math>
*Semantic Noise : <math>n \cdot \frac{\sum_{i=1}^{n} (p_i - \bar{p})^4}{(\sum_{i=1}^{n} (p_i - \bar{p})^2)^2}</math>
where the count of discovered topics (n) and topic probability (p)