This paper talks about (and reports) experimental data

Automatic discrimination of useless papers via machine learning of abstracts.

The main method follows the next pipeline:

Training mode

• Parse abstracts from two input files (classA and classB; see files format at the data/ directory)
• Transform abstracts into their TFIDF sparse representations
• Train Support Vector Machines with different parameters by using GridSearch
• Select the best estimator and save it at model_binClass/svm_model.pkl (default)
• Save TFIDF transformation for keeping the training vocabulary (stored at model_binClass/tfidf_model.pkl)

Prediction mode

• Parse abstracts from a unique input file
• Transform abstracts into their TFIDF sparse representations
• Transform TFIDF representations into their 200-dimensional SVD approximation
• Predict useless/useful papers by means of their abstracts using pretrained Support Vector Machines

Usage

For filtering unknown abstracts run

$ python filter_abstracts_binClass.py --input data/test_abstracts.txt

The predictions will be stored by default at filter_output/, unless a different directory is specified by means of the --out option. The default names containing the predicitons are

• filter_output/useful.out
• filter_output/useless.out

The format of each file is:

<PMID> \t <text of the abstract>
...
<PMID> \t <text of the abstract>

For training a new model set the list of parameters at model_params_binClass.conf and then run

$ python filter_abstracts_binClass.py --classA data/ecoli_abstracts/not_useful_abstracts.txt --classB data/ecoli_abstracts/useful_abstracts.txt

where --classA and --classB (the useful papers) are used to specify input training files. In this example data/ecoli_abstracts/useful_abstracts.txt is the training files containing abstracts of papers reporting experimental data (the desired or useful class for us).