This paper describes our approach to tackle the Author Profiling task at PAN 2019. The objective is to distinguish between bot and human users and for human users it is also necessary to detect their gender. We are given only Twitter messages in two languages (Spanish and English). Our preprocessing stage includes data cleaning as well as the extraction of features using character bi-grams. We experimented with several feature representations and machine learning algorithms ( Support Vector Machines (SVM) from libSVM). For both languages we use the same methods of feature extraction and classification.
Thanks to artificial intelligence, learning using computer is possible, because with each interaction in technology, it can learn more from us to give us more comfort in some tasks or to provide us with solutions, which are more according to our tastes or interests. Actually with the help of artificial intelligence, what we want to do is to model the human intelligence [11].
Currently the use of artificial intelligence to make predictions is very involved in
most streaming services or social networks, to mention some internet services. They are
constantly learning about users to give them the best service according to their interests,
for the streaming services we can consider artificial intelligence to predict what a user
may like and in this way invite him to continue using the services. On the other hand,
social networks are also used to show news, pages, forums, friends or simply to meet
new people. In this new generation of web 2.0, social networks are a great double-edged
sword, since both companies and users with a more direct interaction [
One of the main reasons to study bots is the impact they generate on social networks
through opinions [
Then PAN workshop is organized every year since 2011 with aim of promoting
research on authorship analysis which includes authorship attribution, author profiling,
and plagiarism detection, among others [
The task proposed by PAN is to predict if a user is a bot or not, if it is a human user then it is also necessary to predict the gender of the user. The released dataset contains two languages: Spanish and English. It is important to mention that each user is represented by 100 tweets, which will be analyzed and separated depending on the language. The dataset contains only tweets in which each file corresponds to a user. This dataset was mainly used for training a system, which was tested with other datasets for PAN evaluations. 3 3.1
Having only the tweets of the users it is necessary to do a preprocessing considering some features that the tweets can have: Digits For the part of the digits we decided to remove them since we considered that they were not necessary for text feature representation.
URLs Since the links are resource identifiers in this case are Internet pages are not necessary for the bi-gram structures either. @Mentions Mentions refer to other Twitter users with whom they interact in the message; they are important to quote on Twitter but in our case they will not be necessary so we will also eliminate them.
Emoticons There are messages that contain emoticons but for the structure that we use they are not necessary, however, we consider them not to be helpful for the classification.
Considering the data, a procedure for standardization is necessary: Punctuation marks For the case of our selection of characteristics, we will not need to use punctuation marks. We extracted them to have our data as clean as possible. 3.2
First for the preprocessing of the data we removed punctuation marks, since in the experiment we will not use them as a feature, we also removed the references to other Twitter users as well as links, numbers and emojis contained in the messages as well as characters that are not inside of the Standard ASCII (American Standard Code for Information Interchange).
Since we have the data somehow clean, it is necessary to eliminate the spaces between the words by the following procedures.
The main idea of the extraction of features is to obtain particular features of the
object so we can then compare those features with others and consider some patterns that
have in common. So one way in which we can obtain these characteristics or features
is with the use of character n-grams [
Now we have the characteristics obtaining the frequencies of characterbi-gram per user, we need a method in which we can organize the data of all the documents with respect to their characteristics; because of this we created a vector space model.
The main idea of using vector space model is to represent the characteristics of each object with its corresponding object but in an organized manner where the objects can be compared later [10].
We proceed to organize our data in a table Term-Document Matrix [11] where for the part of the columns we have the document and in the part of the rows is a description of the character bi-gram, in this way the content of the table will be the difference of the character bi-gram in the analyzed file. If a character bi-gram is not found in the document, the value of the box must be 0.
Having this matrix you have all the documents with character bi-grams in an orderly way and can be analyzed in a much more efficient way. 3.4
Thanks to the structure of the organized matrix, they were tested with several classifiers and evaluated the accuracy to know which could be the best classifier for this task. All the n-grams tested were of character since with them we had much better accuracy than with other structures to obtain characteristics.
After having the results of the classification between humans and bot, we only use humans for the classification of gender using same methods of extraction of characteristics and the same classification models. In this paper, we present an approach to get the solution for the Task "Bots and Gender Profiling" of PAN at CLEF 2019. Our final system for the classification between bots and humans followed by classifying users who are human by extracting characteristics from the tweets and placing them in a structure formed by character bi-grams. In this way, a term-document matrix is formed in which the entire data set is ordered to pass through a classification process. With respect to the tests carried out, we decided to use Support Vector Machine as a classifier with cross validation with 10 boxes for training the model and later use it with the PAN tests. We realized that for the Spanish and English languages it did not differ much in the value of the accuracy for the classifications, so we used the same method for the extraction of characteristics: as well as to determine between human users and bots and the gender classification in the human users. In the same way we use the same classifier for both languages [8].
For future work due to the good performance between the classification of humans and bots we would like to try with the different characteristics that social networks allow to introduce in messages (for example, using the 250 characters that Twitter allows in each Tweet); perhaps we can find more efficient ways to classify between humans and bots using natural language processing techniques. 8. Rangel, F., R.P.F.M.: Proceedings of the 17th International Conference on Intelligent Text Processing and Computational Linguistics. In: A Low Dimensionality Representation for Language Variety Identification, pp. 156–169. Springer-Verlag (2018) 9. Sapkota, U., Bethard, S., Montes-y-Gómez, M., Solorio, T.: Not all character n-grams are created equal: A study in authorship attribution. In: Proceedings of the 2015 Annual Conference of the North American Chapter of the ACL: Human Language Technologies.
NAACL-HLT ’15, Association for Computational Linguistics (2015) 10. Sidorov, G.: N-gramas sintácticos y su uso en la lingüística computacional. In: Vectores de investigación, 6(6). pp. 1–15. SpringerBriefs in Computer Science, Springer (2013) 11. Sidorov, G.: Formalization in computational linguistics. In: Syntactic n-grams in Computational Linguistics. SpringerBriefs in Computer Science, Springer (2016)