readability_classification_w122019_y121059

Introduction

Readability is the ease with which a reader can understand a written text, a score to know how it will be accessible to readers. From a linguistic perspective, the readability of a text is determined by much more than a few superficial textual features. For example, does the reader know most of the words? Does the text contain complex grammatical structures? Are there enough connectives to explain the flow of the text? Is the text about a lot of different concepts? Measuring the readability of a text has a long history, mainly in the educational domain. Second Language Acquisition (SLA) is applied as a base method of combining Lexical and Syntactic features.

Our article: Readability Classification Article

Dataset

We use 2 dataset in English.

OneStopCorpusEnglish This dataset is collected from ACL anthology and created by Sowmya Vajjala and Ivana Lucic from the Iowa State University, USA.

It consists of about 509 documents classified into 3 reading levels for English as Second Language learners:

• Advanced
• Intermediate
• Elementary

The original articles of this dataset is collected from The Guardian newspaper and had been rewritten to suit the three levels of English learners on onestopenglish.com, a website run by MacMillan Education.

More Information: W18-0535

Download: OneStopEnglishCorpus - SeparatedByReadingLevel

Cambridge Readability Dataset This dataset is created by Menglin Xia, Ekaterina Kochmar, and Ted Briscoe from University of Cambridge. It composes of reading passages from the 5 reading levels which target at ESL learners:

• CPE
• CAE
• FCE
• PET
• KET

and contains about 282 documents.

More Information and Download: Cambridge Readability Dataset

Evaluation

Accuracy: this is the percentage let us know how well the model is. It is calculated by dividing the number of matching label by the total number of reference label. For each of the dataset, our method is calculating the accuracy of each level and then get the average number to get the final accuracy.

Our method:

We apply some feature and build a coefficient set, based on analizing our dataset.

Some examples:

Lexical Features

Lexical Density: Lexical density is a concept that measures the complexity of a sentence or a document in linguistic by using functional words like nouns, adjectives, verbs, and adverbs. It also provides us the meaning and the information regarding what the sentence or the document is about.

Proportion of words in AWL: Academic Word List (AWL) is a list contains 570 words that appear with high frequency in academic text. It was developed by Averil Coxhead at the School of Linguistics and Applied Language Studies at Victoria University of Wellington, New Zealand. The words are divided into 10 groups. Group 1 is the most frequent words and group 10 is the least frequent words.

Type-Token Ratio: The total number of words in a docu- ment are called token, each token may repeated many times.

Syntactic Features

Mean Word Length: the average number of characters per word in a sentence.

Result:

OneStopCorpusEnglish best performance: 81.71%

Cambridge Readability Dataset best performance : 70.56%

How to run

Required packages:

• nltk
• numpy
• statistics

Guide for MacOS.

cd Readability-Classification-W122019-Y121059
alias runn='bash run/run.sh'
runn