essay_scorer

An automated essay scorer for english language learner essays.

Description

Extracts a set of linguistic features and compares them to a model trained on over 3000 essays to predict the score on a 40 point scale.

Installation

pip install essay-scorer

Usage/Tutorial

Command line usage

Accepts either a directory of .txt files or a single .txt file.

(You'll have to locate the bin where the script is saved by pip to use the command line like this.)

For a directory of text files

python3 essay_scorer.py path/to/essays/

Bonus

python3 essay_scorer.py path/to/essays/ >> output.csv

For a single text file

python3 essay_scorer.py path/to/essays/test.txt

Importing in a python script

import essay_scorer

text = open('test.txt', 'r').read()
feat_set = essay_scorer.get_feats(text)
pred_score = essay_scorer.gbr_model.predict(feats)[0]
print('predicted score', pred_score)

About

This automated essay scoring system is based on Travis Moore's master's thesis work.

Moore's master's thesis, which lays the theoretical groundwork for this project can be found here.

This model is best used on english learner essays that are between 140-300 words. Due to the distribution of scores in the model, this model tends to make better predictions around the median of actual scores from the dataset which seems to be around 20. There is more variance in predictions of outlier scores.

The model itself is a GradientBoostingRegressor model. Here are the parameters and results of it testing itself on its own data:

`model.fit` results:
GradientBoostingRegressor(alpha=0.9, criterion='friedman_mse', init=None,
             learning_rate=0.02, loss='ls', max_depth=4, max_features=0.3,
             max_leaf_nodes=None, min_impurity_decrease=0.0,
             min_impurity_split=None, min_samples_leaf=9,
             min_samples_split=2, min_weight_fraction_leaf=0.0,
             n_estimators=500, n_iter_no_change=None, presort='auto',
             random_state=0, subsample=1.0, tol=0.0001,
             validation_fraction=0.1, verbose=0, warm_start=False)
Mean Absolute Error:
Train error:	2.360482423992901
Test error:	2.32169958721344

r2 scores of both train/test:
r2_train:	0.8341340712062068
r2_test:	0.8575492864872207

License

GNU GPLv3 - see LICENSE file for details.2

Contact

@mkylemartin on Twitter, GitHub

Note

The pickled data file in this version does not include the age_bracket or language_id. The 49 features extracted are the following (in alphabetical order).

['ari', # readability index
 'avg_len_word',  	# average word length
 'cli', # readability measure
 'conjunctions', 
 'cttr', # corrected type to token ratio
 'dcrs', # dale chall readability score
 'determiners', 
 'dw', # difficult words
 'english_usage', # number of english words used
 'fkg', # flesch_kincaid_grade
 'fre',  # flesch_reading_ease
 'function_ttr', 
 'gf', # gunning_fog
 'grammar_chk', # checks for 2000+ grammar errors
 'lwf', # linsear_write_formula
 'n_bigram_lemma_types', 
 'n_bigram_lemmas', 
 'n_trigram_lemma_types',
 'n_trigram_lemmas', 
 'ncontent_words', 
 'nfunction_words', 
 'nlemma_types',
 'nlemmas', 
 'noun_ttr', 
 'num_tokens', 
 'num_types', 
 'pct_rel_trigrams',
 'pct_transitions', 
 'rank_avg', 
 'rank_total', 
 's1',  # negation stages (the next several features)
 's1a', 
 's1b', 
 's1c',
 's2', 
 's2a', 
 's2b', 
 's2c', 
 's3', 
 's3a', 
 's3b', 
 's3c', 
 's4', 
 's4a',
 's4b', 
 's4c', 
 'sent_density', # average words per sentence
 'spelling_perc',  # what percentage of words spelled correctly
 'ttr' # type token ratio ]