This paper presents our experience implementing three approaches for the 'PAN 2014 Author Identification' [3,1] task using the same representation for the author's style. Two of our approaches extend previous successful approaches: naive Bayes [4] and impostor [8] methods. The third approach is based on original research on sparse representation for text documents. We present results with the official development and test corpora.
Author verification has multiple applications on several areas including information
retrieval and computational linguistics, and has an impact in fields such as law and
journalism [
Given a small set (no more than 5, possibly as few as one) of “known” documents by a single person and a “questioned” document, the task is to determine whether the questioned document was written by the same person who wrote the known document set.
This year the documents were in four languages and four genres with the following combinations: essays and reviews in Dutch, essays and novels in English and articles in Greek and Spanish.
In this work we present the implementation of three approaches to perform the
authorship verification task based on the same document representation. In particular, we
repeatedly use a vector space representation of documents as described in our approach
last edition [
The representation for an author’s style is generated in two stages. First, we represent
the documents from an author using the vector space model [
d = (w1; w2; : : : ; wm) where wi is a frequency or weight of a word of the vocabulary of size m for the word i. A common representation of a vector space model is the bag of words model, in which the words represent actual words of the document and the frequency count occurrences of such words in the represented document.
Additionally to the bag of words we use the following feature frequencies to represent the documents:
Bigram Frequencies of two consecutive words.
Trigram Frequencies of three consecutive words.
Prefix Frequencies of prefixes of words.
Suffix Frequencies of suffixes words.
Prefix bigram Frequencies of two consecutive prefixes of words.
Suffix bigram Frequencies of two consecutive suffixes words.
Stop words Frequencies of stop words.
Stop words bigram Frequencies of two consecutive stop words.
Punctuation Frequencies of punctuations.
Words per sentence Frequencies of words per sentence.
In the second stage an author is represented by the sum of the vectors representing the documents written by her or him. This cumulative vector is normalized by the number of documents by the author and the resulting normalized vector represents the style of the author. This representation is not novel, however many approaches on author verification optimize the representation on the domain, in our experiments we keep the same representation independently of the domain. Some of the implemented approaches require an instance of a document by a certain author, to accomplish this it we sample a document from the author’s style representation. 3
This method consists on iteratively compare the vector distance between the author’s
document to the questioned document versus the distances between several impostor
documents to the questioned document. With these distances a score is built up based
on how many times the author and questioned documents are closer than the impostor
and questioned documents. For this approach we follow the description of the method
by Seidman (2013) [
This method consists of sampling from the author and the impostor style representations two document instances for each. A probability score is then calculated using the common term between the questioned document and the author’s documents. On the other hand, an alternative score is calculated between the questioned document and the impostor documents. These scores are derived using Bayes2. The purpose of the score is to capture the probability that the document was created by the same author, if the score for the author is higher than the impostor, we consider it as evidence of authorship. We iterate n times over this method to calculate the probability of authorship. 3.3
This methodology has been successfully applied to the face recognition task, in which
the identity of a face image has to be determined from a set of known faces [
Ax = y (1) Where y is the questioned document, A is the matrix of n samples from different m candidate authors (impostors), and x is the variable to minimize which represent the contribution from each candidate. So that multiplying the samples by the contribution we could generate the questioned document. From the resulting variable x0 we can quantify the residuals given by Ax versus y and decide which author contributes with more components. We adapt this method to produce a probability as result by iterating k times over the full method. 4
In the preparation of the authorship verification systems we implemented three approaches: impostor, n-gram and sparse methodologies. During development we tested all of them on the same representation for the author’s style and the same parameter for the four languages and three genres of the task. During development and testing the best results were achieved using the sparse methodology which is interesting to us since it is the first time such method is applied to the task of authorship verification.