Affiliations: [a] Center for Computing Research (CIC), Instituto Poitécnico Nacional (IPN), Av. Juan de Dios Bátiz, Zacatenco, Mexico City, Mexico
| [b] Department of Computer Science, COMSATS University Islamabad Lahore Campus, Pakistan
Correspondence:
[*]
Corresponding author. Grigori Sidorov, Center for Computing Research (CIC), Instituto Poitécnico Nacional (IPN), Av. Juan de Dios Bátiz, Zacatenco, Mexico City, 07738, Mexico. E-mail: [email protected].
Abstract: The process of automatic identification of an author’s demographic traits like gender, age, native language, geographical location, personality type and others from his/her written text is termed as author profiling (AP). Currently, it has engaged the research community due to its promising uses in security, marketing, forensic, bogus account identification on public networks. A variety of benchmark corpora (English text) released by PAN shared task is used to perform our experiments. This study presents a Content-based approach for detection of author’s traits (age group and gender) for same-genre author profiles. In our proposed method, we used a different set of features including syntactic n-grams of part-of-speech tags, traditional n-grams of part-of-speech tags, the combination of word n-grams and combination of character n-grams. We tried a range of classifier for several profile sizes. We used the word uni-grams and character tri-grams as our baseline approaches. We achieved best accuracy of 0.496 and 0.734 for both traits, i.e., age group and gender respectively, by applying the combination of word n-grams of various sizes. Experimental results signify that the combination of word n-grams can produce good results on benchmark corpora.
Keywords: Author profiling, machine learning, syntactic n-grams, traditional n-grams, part-of-epeech
