This paper presents an approach for tackling the authorship identification task. The approach is based on comparing the similarity between a given unknown document against the known documents using a number of different phrase-level and lexical-syntactic features, so that an unknown document can be classified as having been written by the same author, if the different similarity measures obtained are close to a predetermined threshold for each language in the task. The method has shown competitive results, achieving the overall 6th place in the competition ranking.
Discovering the correct features in a raw text in order to unambiguously allow to attribute the authorship of a given anonymous document is a very hard problem that recently (empowered by the continuous growing of information in Internet) has become of high interest in areas like information retrieval (IR), Natural Language Processing (NLP) and computational linguistics. Taking into account the above, the most common framework for mapping candidate authors with unknown documents is the authorship attribution problem, where, given texts of uncertain authorship and sample documents from a small, finite set of candidate authors, the task consists of mapping the uncertain texts onto their true authors among the candidates. This problem is considered as an unreasonably easy task while a more demanding problem (often presented in documents on the web) is the author verification task, where, in a given set of documents by a single author and a questioned document, the problem is to determine if the questioned document was written by that particular author or not. In this sense, the importance of finding the correct features for characterizing the signature or particular writing style of a given author is fundamental for solving the authorship problem. The results reported in this paper were obtained in the framework of the 8th International Workshop on Uncovering Plagiarism, Authorship, and Social Software Misuse (PAN 14), in particular, in the task named Author Identification. For this purpose, we have attempted an approach for representing the features that will be taken into account in the process of authorship verification using an unsupervised learning method. The proposed approaches are discussed in the following sections.
The rest of this paper is structured as follows. In Section 2 it is presented the description of the features used in the task to be tackled. Section 3 shows the Proposed approach (unsupervised algorithm) used in the experiments. The experimental setting and a discussion of the obtained results are given in Section 4. Finally, the conclusions of this research work is presented in Section 5. 2
In this work we explore a combination of different types of features in a text frecuency
vector [
In this approach are considered the following features for representing the particular
writing style of a given author:
– Phrase level features
1. Word prefixes. A group of letters added before a word or base to alter its
meaning and form a new word.
2. Word sufixes. A group of letters added after a word or base to alter its meaning
and form a new word.
3. Stopwords. A group of words that bear no content or relevant semantics in the
text.
4. Punctuation marks. Conventional signs and certain typographical devices as
aids to the understanding and correct reading, both silently and aloud, of
handwritten and printed texts.
5. Word N-grams. A contiguous sequence [
In this approach the following metrics are considered to determine the similarity of the
documents:
– Latent semantic analysis (LSA). A technique[
J S(A; B) = jA T Bj
jA S Bj vu n d(q; p) = tuX(qi i=1 pi)2 – Chebyshev Distance. A metric defined on a vector space where the distance between two vectors p and q, is the greatest of their differences along any coordinate dimension. The distance is given by:
DChebyshev(q; p) := max(j qi i pi j) – Cosine similarity. A measure of similarity between two vectors of an inner product space that measures the cosine of the angle between them. Given two vectors of attributes, p and q, the cosine similarity, is represented using a dot product and magnitude and is defined as: cos( ) =
p q k p kk q k = pPn i=1(pi)2
Pn i=1 pi
qi pPn
i=1(qi)2 (1) (2) (3) (4) end LSA1 max(vector1 [0] ; : : : ; vector1 [n]); LSA2 max(vector2 [0] ; : : : ; vector2 [n]); LSA3 max(vector3 [0] ; : : : ; vector2 [n]); M axSimilarity max(Cosine; J accard; LSA1; LSA2; LSA3) if M axSimilarity >= then
result same author else end
result end
dif f erent author Algorithm 1: Proposed approach using an unsupervised learning method
For tackling the authorship identification task we propose a methodology (see algorithm 1) in which we used an unsupervised learning method. We evaluate different feature vectors with different similarity measures in order to identify if an unknown document belongs to an author. For each language, three types of text representations are evaluated: – A frequency vector of all the vocabulary in the test documents. – A frequency vector of all the combined Lexical syntactic features of the test documents: 1. Word prefixes 2. Word sufixes 3. Stopwords 4. Punctuation marks 5. N-grams(from 1 to 5) 6. Skip-grams (from 1 to 5) 7. Vowel combination 8. Vowel permutation – A similarity vector using the LSA algorithm for each word in the test documents.
Different distance/similarity measures were tested, including the Jaccard similarity
for the vocabulary feature vector, the cosine similarity fo the Frequency vector of all
the combined Lexical syntactic features and Chebyshev Distance, Euclidean distance
and cosine similarity for the LSA vectors. The best similarities of each document of
unknown origin were obtained and a threshold (see table 1) to determine if the document
belongs to an author. The threshold was obtained using a classification tree [
The description of the six text collections used in the experiments are shown in Table 2.
As can be seen, the data set is made up of different languages/genres within the author
verification task, where, each problem consists of some (up to five) known documents
written by a single person and only one questioned document. All documents within a
single problem instance are in the same language and the documents are matched for
register, theme, and date of writing.
In Table 3 we present the results obtained by our approach using the TIRA tool [
We have presented an approach that uses an unsupervised method with lexical-syntactic features. Even if the runtime is greater than the most approaches of this competition, the performance is good. It was surprising that being a Spanish native language team, we performed better in Dutch and Greek languages, but it is a good oportunity for analyzing the text into more depth for determining the reason of this issue. As we mentioned before, we have executed the same methodology across the different languages, varying basically only the thresholds applied in the final classification. However, more experiments continue to be performed to analyze whether or not these changes introduce significant variations in the data sets. Future work is planned to observe the performance of the proposed methodology using different similarity measures.