Identifying the authorship either of an anonymous or a doubtful document constitutes a cornerstone for automatic forensic applications. Moreover, it is a challenging task for both humans and computers. Clustering documents according to the linguistic style of the authors who wrote them has been a task little studied by the research community. In order to address this problem, PAN Evaluation Framework has become the first effort to promote the development of the author clustering. This article proposes a graph-based method, specifically βcompact clustering, for discovering the groups of documents written by the same author. The β-compact algorithm is based on the analysis of the similarity between documents and they belong to the same group as long as the similarity between them exceeds the threshold β and it is the maximum similarity with respect to other documents. In our proposal we evaluated different linguistic features and similarity measures presented in previous works of authorship analysis task. The training dataset was used to determine the best value of β parameter for each language. The result of the experiments was encouraging.
In the evaluation framework the task is described as follows: "Given a collection of (up to 50) short documents (paragraphs extracted from larger documents), identify authorship links and groups of documents written by the same author. All documents are single-authored, in the same language, and belong to the same genre. However, the topic or text-length of documents may vary. The number of distinct authors whose documents are included in the collection is not given. "1
One of the most used strategies for documents representation in Text Mining (TM)
applications, corresponds to the classic Bag of Words [
As a summary, in the last edition of authors clustering task, 6 papers were presented
[
With our work, we want to propose and evaluate a clustering algorithm that we have used in topic document clustering tasks in our research center, and its purpose is to group objects with the condition that for each object of the group, at least there is an object with which the similarity between them is greater than a threshold of similarity and it´s the maximum similarity with an object of the collection.
It is important to emphasize aspects of the description of the author clustering problem, such as: short texts no longer than a paragraph; the texts corresponding to the same author are of the same genre but not necessarily the same topic or homogeneous length. In addition, as part of the task, we need to obtain a ranking of similarities between objects in the same clusters. Taking these details into consideration, in the next section we propose and describe our method considering binary linguistic features and a clustering algorithm based on compact clusters. 2
We propose the use of β-compact algorithm [
In the following image (Figure 1) we expose the architecture of the implemented method and later we describe each one of the steps involved. 1 http://www.webis.de.
Both in the training and test stage, collections of documents are received and the final purpose is to obtain groups of documents, where all the documents of a group belong to the same author.
The algorithm proposed, to obtain the groups, requires the representation of each of the documents, a comparison function that allows to evaluate the similarity between a pair of documents and a threshold β to decide when two documents must belong to the same cluster.
For documents representation, we used the classic Bag of Word, and with the
training dataset we tried different types of features [
The β-compact clustering algorithm is described in the next pseudo program code.
First we need to define the concept “Graph of Maximum β similarity: It´s an oriented
graph in which the vertices are the objects and exist an arista between two vertices Oi
and Oj if Oj is β-similar with Oi and Oj is the most similar of all the rest of objects [
We performed different runs, in which the comparison function and the
representation were varied, as well as the value of β from 0 to 0.5. When the β was greater than
0.5, there were no changes in the clusters obtained. For each language, the best result
is determined by analyzing the clusters obtained by comparing them with the training
data using the FBcubed [
Finally, for each language, we have a configuration of the method, with a binary features representation to be calculated, a similarity function and a threshold β.
It is important to note that, due to the nature of the β-compact algorithm, two documents can belong to the same group, although the similarity between them not necessarily exceeds the defined β, because the only condition is that each one of them has a similarity greater than β with some document of their group. This characteristic may lead to non-necessarily spherical clusters. It can be observed in the outputs of the method in the ranking file, where similarities between documents of the same cluster will appear that are below the threshold β and therefore will appear little ranked in this list.
For the similarity ranking construction, we took into account all similarities among objects of a group and order them. The similarities that are above the threshold β were distributed in a scale of 0.5 to 1, corresponding to 0.5 the similarities that are equal to β and close to 1 the greater ones. Similarly, was realized a distribution of the similarities that did not exceed the β, which in this case were distributed from 0 to 0.5. 3
With the documents of the training dataset released and explained in [
Parameters
Features Character words lemma POS-Tagging N-gram 1-5 1-5 1-5 1-5 Language en, gr, du en, gr, du en en
β 0.05-0.5 0.05-0.5 0.05-0.5 0.05-0.5
The features used in the documents representation are those presented in the Features row, we test using n-grams character representations (n = 1, ..., 5), for English (en), Greek (gr) and Dutch (du). Representations of n-grams words (n = 1, ..., 5) for the three languages mentioned. For the representations of Lemma and Part of Speech Tagging (POS-Tagging) we only processed the English texts. For each of the configurations, evaluations were performed varying the β from 0.05 to 0.5.
As conclusions, we must emphasize that one of the essential aspects in our work is the features identification for the documents representation, and in this we could try other ideas presented in the literature. The algorithm usually obtains small groups, and we could evaluate in a future work the differences that can be reached when we use other variants such as the β-connected and β-strongly compact algorithms. The β-connected would obtain larger groups, while the strongly compact would have smaller and more compact groups than those proposed in our work.
We propose to evaluate different features weighing strategies and other comparison functions proposed in the literature. Using the results achieved in this work, take into consideration comparisons with ensemble clustering algorithms. 5