This paper describes our entry for the Author Identification task at PAN 2013. The Author Identification task was performed using a combination of Vector Space Model [1] (VSM) and Similarity Overlap Metric [3] (SOM) on the character n-grams extracted from the documents related to an author and the document of question. A combination of the VSM and SOM provided an overall F-measure, precision and recall values of 0.576 each.
The author identification task is to identify whether the given document is written by the author of the known document(s). The initial challenge was to identify what features would help in identifying the author. Based on the following intuition, we identified that character 3-grams would help in identifying the author. For example in English, some authors like to write-up either in past tense, present tense, future tense or continuous tense. All these tenses in English can be identified using clear indicators such as continuous tense is always indicated by Jerund format ie., words with characters ’ing’ at the suffix. The tense will clearly indicate the style of the author. Further to this, an author might use descriptive words (i.e., adjectives) to represent something in their storyline. These details can be easily identified from a character 3-gram or 4-gram. The intuition was considered with expertise in English language only. Based on this intuition, we decided to use character n-gram model for the system. 2.1
For the author identification task, we were provided with the following training data by PAN.
– 10 English, 20 Greek and 5 Spanish folders – Each folder had upto 10 documents written by a known author – Each folder also had a document written by an unknown author – A gold standard was also provided for all the folders The idea was to classify whether the document written by the unknown author is written by the known author or not. 2.2
The author identification task being closely related to intrinsic plagiarism detection,
we considered author style as the only feature for classification. Therefore, we used a
combination of VSM [
Vector Space Model The known author’s documents were split into 3-gram
characters and considered as a vector and unknown author’s document was split into 3-gram
characters and was considered as a different vector. Based on the VSM, a similarity
measure was computed which represented the cosine of the angle between two vectors.
As mentioned in [
consider known documents vector be!x = (x1; x2; ::; xn) and unknown document vector be !y = (y1; y2; ::; yn)
cos(!x ; !y) = j!x!xj;:!jy!yj = pPin=P1inx=i21pxiPyiin=1 yi2
Similarity Overlap On initial evaluation, it was identified that just the cosine values
produced by VSM were not enough to classify whether the unknown document was
written by the author of the known documents or not. Therefore, we required another
approach for the classification. The similarity overlap metric [
consider known documents set X = set(x1; x2; ::; xn) consider unknown document set Y = set(y1; y2; ::; yn)
X T Y
Overlap(X; Y ) = min(X;Y )
The cosine and overlap values always range between 0 and 1, therefore we used thresold
values to classify the unknown documents. The following table represents the threshold
values used for classification.
Based on the said threshold values, the unknown document was classified as known
document or unknown document. Although the initial intuition was to use this
methodology for English dataset, we attempted the same for all other datasets as well. The
following results and discussion section provides the insight based on the results
obtained for the said approach.
The baseline precision, recall and f-measure values were set at 0.500 for the author
identification task. The results available at [
Data Set Overall English Greek Baseline Spanish As we get into the scores for different language sets, the approach did not work well for spanish dataset. This suggests that we need to look for better language specific features to classify author. Although the approach was carried out based on intuition for English language, the results suggests that the same language features work better for Greek than for English. Since we have used two very simple approaches to identify the author, the time complexity of the system proves to be the best of all. 3
The author identification task was completed with better than baseline results for all the datasets, while the precision and recall values for spanish dataset were lower than baseline. Therefore the features for spanish need to be selected carefully understanding the language specification and flavour. Further to this, the current system produced promising results with a combination of vector space model and similarity overlap metric, which can further be experimented with diferent character grams. Moreover, we will be scrutinizing other language specific features which may be useful to identify the document’s author.