Deep Text Eval (DTE) - Project in the Cognitive Systems Master Program

by Shlomi Hod, Maximilian Seidler and Vageesh Saxena.

Problem

A differentiable function that measures text complexity/readability will be beneficial as part of the loss function of neural network text simplification system. However, to our best knowledge, there are not such a differentiable function.

Basic Idea

Designing a differentiable readability measure, either by classic machine learning models or neural networks.

Learning Task

Leveraging a leveled corpus of articles into a real number readability score.

Paper

Notebooks (partial list, please explore the repo)

1. Corpora Analysis
2. End2End Design
3. Stacked Design
   1. Features Analysis
   2. Features2Prediction - Model Selection
      1. Classification Models
      2. Regression Models
      3. Ordinal Regression
   3. Feature Selection
   4. Final Model in Tensorflow version

Data

In the paper we used the WeeBit corpus, however, we have also prepared other corpora, in English and German. All the corpora are described here. For access to the texts, please send an email to the authors.

Structure

Each corpus has its own HD5 file with the key text_df that contains the articles, with two these columns:

1. text - The text itself.
2. y - The level label (integer, stating from 0), as described below.

Train-Test Split

80%-20% with shuffle

Usage in Python (Example with the Weebit Corpus)

import pandas as pd

with pd.HDFStore('data/weebit/weebit.h5') as weebit_store:
	text_df = weebit_store['text_df']
	train_df = weebit_store['train_df']
	test_df = weebit_store['test_df']

Corpora

Crude word count:

1. Weebit - 458,497 words
2. GEO - 1,188,247 words in total
3. DW - 934,734 words in total

Weebit

Sowmya Vajjala and Detmar Meurers: "On Improving the Accuracy of
Readability Classification using Insights from Second Language
Acquisition". Proceedings of the 7th Workshop on Innovative Use of NLP
for Building Educational Applications (BEA7), Association for
Computational Linguistics. 2012.

Levels

0 - WRLevel2 (Age 7-8)
1 - WRLevel3 (Age 8-9)
2 - WRLevel4 (Age 9-10)
3 - BitKS3 (Age 11-14)
4 - BitGCSE (Age 14-16)

Train-Test Datasets

The BitGCSE class has ~x5 texts than the other classes, therefore it was downsampled to 800 texts after all the cleaning and preprocessing.

level  #text  %text  #train  %train  #test  %test
0      607    16.69  486     16.71   121    16.62
1      788    21.67  630     21.66   158    21.70
2      798    21.95  638     21.94   160    21.98
3      643    17.68  514     17.68   129    17.72
4      800    22.00  640     22.01   160    21.98

Newsla

Levels

Based on Lexile

Continuous y

300 - 1670

Discrete y_cat

2.0, 3.0, 4.0, 5.0, 6.0, 7.0, 8.0, 9.0, 12.0

Removed Data by small y_cat

10.0    23
11.0     2
13.0     1

Train-Test Datasets

Categorial y

      #text  %text  #train  %train  #test  %test
2.0     921   4.25     737    4.25    184   4.25
3.0    2681  12.37    2145   12.37    536  12.37
4.0    2582  11.92    2065   11.91    517  11.93
5.0    3449  15.92    2759   15.92    690  15.92
6.0    2291  10.57    1833   10.57    458  10.57
7.0    2570  11.86    2056   11.86    514  11.86
8.0    1650   7.61    1320    7.62    330   7.61
9.0    1804   8.33    1443    8.32    361   8.33
12.0   3720  17.17    2976   17.17    744  17.17

Continuous y_lexile

    name      #  min   max        mean         std
0   text  21668  300  1670  939.476186  268.819219
1  train  17334  300  1670  939.390793  268.471205
2   test   4334  310  1660  939.817720  270.237459
train-test Kolmogorov-Smirnov p-value 0.9999167905571691

GEO

Levels

0 - GEOlino
1 - GEO

Deutsche Welle

Deutsche Welle - Deutsche Lernen news articles leveled by CEFR.

Levels

0 - B1
1 - B2, C1