Authorship Attribution Helper

Tools to make authorship attribution faster.

Dependencies

Libraries required:

numpy
sklearn
joblib
keras
matplotlib

How to Use

Currently available classifiers are SVM and CNN. To use SVM:

from classifiers import SVM
svm = SVM(C=1, analyzer="word", ngram_range=(1,2))

SVM takes 3 arguments.

• C: The original C value to use.
• analyzer: The type of the analyzer. Either "word" or "char".
• ngram_range: The range of ngrams to use.

To use CNN:

from classifeirs import CNN
cnn = CNN(max_length=500, vector_size=150, minibatch_size=50, ngram=4, epochs=10) 

CNN takes 5 arguments.

• max_length: How many input nodes.
• vector_size: the embedding dimensions.
• minibatch_size: The size of the minibatch to use.
• ngram: Which ngram size to use. Currently can only take one number, not a range.
• epochs: How many epochs to train.

After the classifer has been imported and created, the next step is to create a Handler to handle the rest.

from wrapper import Handler
attributer = Handler(train_data=train_data, test_data=test_data, positive_class=["Author 1"], split_size=500, split_type="char", data_type="book_split", classifier=svm, threads=10, iteration=-1)

Here the Handler takes in the data and multiple different arguments.

• train_data: The train data. Needs to be a JSON dictionary, where the key is the name of the author, and the value is a dictionary of books / manuscripts from the author. e.g. train_data = {"auth1": {"book1": "book text"}}
• test_data: The test data. Same format as train_data.
• positive_class: List of authors to consider the positive class.
• split_size: The size of splits to do, if any. -1 if no splits.
• split_type: Whether to split by charcters or words.
• data_type: How to do the splitting. "book": No splitting, one book is one training sample. "single": Books are split into parts and they are classified independent of each other. "book_split": Books are split into parts, but they are classified together.
• classifier: The classifier to use. SVM or CNN.
• threads: How many threads to use with SVM. CNN uses all by default.
• iteration: Current iteration. If -1, do everything in one go, but otherwise only perform one classification. Useful if classifying in parallel.

Handler has multiple functions that can be called to perform different tasks.

attributer.optimize_C([2,4,6,8])

Optimizes the SVM C value using Leave-one-out Cross Validation. C values to test are give as argument.

attributer.cross_validate()

Performs Cross Validation to get values for the train data samples as well.

attributer.attribute.test_data()

Attributes the test data samples.

attributer.print_results(normalize=True)

Print results. Normalizing scales the values to be between 1 and -1, where 0 is the threshold.

attributer.plot_values(scale=True, title="Title")

Plots values. Scale scales the values. Title for the plot.

attributer.get_best_features()

Extract best features (Requires SVM).