This paper describes a statistical approach to the task of gender and age range classi cation in SMS. Our approach started developing our own implementation of Low Dimensional Representation (LDR) method, with the idea to add some other statistics which had not been used in the original implementation, such as skewness, kurtosis and central moments. The proposed method calculates term frequencies and uses 3 statistics per class: mean, standard deviation and skewness.
The Author pro ling task is widely studied and some new ideas arise from time
to time [
The MAPonSMS Author Pro ling Train corpus is composed by one single collection of SMS messages written in Roman Urdu and English languages mixed. The goal of the task is to classify the user by gender (Male/Female) and by age range (15-19, 20-24, 25-xx). We were provided with SMS messages from 350 di erent users, being the number of SMS messages variable. So, we had not the same number of SMS messages for each user. In regard of gender, the classes were not balanced, being 60% of the SMS labeled as Male, thus 40% Female. Age range classes were again not balanced. 30.86% of the users belong to 15-19 class, 50.29% to 20-24 class and 18.85% to 25-xx class.
Our goal was to develop a method that was not language dependent, and that required no prior knowledge of the language. We started implementing TF representation for each user in the dataset, counting how many times each word appears in each user and globally for all users. We decided to use TF since this way we could represent a priori class dependent probability for each term for each class simply by counting the number of times a term occurs for each class, and then dividing this amount by the total number of times this term shows for all classes. For the sake of understanding we wanted to be sure we were dealing with probabilities. In addition, calculating TF is less time and resource consuming than calculating Tf-Idf (as in the original LDR). We build a vocabulary set including each word we have seen in the training corpus. We decided to discard the words that appear less than 5 times in the corpus, which highly reduces the size of the resulting dictionaries. Then, we went over the training corpus, one user each time, checking for words for this user and writing down into an array the related a priori probability for each class. Finally, we got one vector per class per user (2 for gender, 3 for age range) with the a priori probability of each word to pertain to each class. Then, we can calculate the di erent statistics from these a priori probabilities arrays that represent the text used by each user. Once we had the average, standard deviation and skewness for every user in the training dataset we used these characteristics to train a LinearSVC machine learning algorithm from Python's Sklearn library. This is a Support Vector Machine with Linear kernel, where multiclass support is handled according to a one-vs-the-rest scheme. It is important to notice that when we want the model to include more users or SMS messages to build the a priori probability vectors, we only have to run the procedure with new labeled SMS messages. This new vector is what we will use to predict new incoming SMS messages. The whole procedure is simple and fast, and can be done in parallel. Once the new vectors are built, we only need to point the algorithm to these new vectors. This is an easy way of keeping the vectors up to date, and more importantly, this is a clean, fast and reliable updating procedure. The code used in this task can be found at https://github.com/OscarGariboiOrts/MAPonSMS18 4
In Table 3 we present our results vs Baseline. Baseline was built by controlling class. Our results improve Baseline for gender by almost 30% and age range by almost 12%. We have stated that we were removing from the vocabulary all words appearing in less than 5 users. This is a parameter that can be adjusted depending on the dataset characteristics and on the task involved. We would have liked to submit more than 1 single submission, as was stated in the task, since we have seen that removing more or less words might have a big in uence in the algorithm accuracy. Even more, this is important in such a small dataset as the one used in this task. Our classi cation method ends up with the accuracies shown in Table 3.
The method is using statistical parameters calculated from a priori
probabilities of each word to belong to one of the classes. The average, standard deviation
and skewness for these distributions should re ect the fact that di erent classes
use language in di erent ways. At the end, we are relying on the mathematical
concept the more similar 2 distributions are, the more similar their statistics will
be. Hence, feeding the Support Vector Machine (SVM) with these statistics SVM
should be able to nd an hyperplane to linearly separate the related classes. In
[
Works on Multilingual Author Pro ling are rare. "Multilingual author pro ling
on Facebook" [
In this paper we presented a method to classify Gender and Age rank in the MAPonSMS International Author Pro ling Shared Task at Forum for Information Retrieval Evaluation (FIRE'18). We have shown that codifying the text with the statistical features extracted from a priori class dependent arrays is an easy, cheap and language independent method to classify texts with di erent features. We are always relying on the fact that di erent classes will use the language in di erent ways that our algorithm can extract information from. A bigger dataset should give us a clearer idea of how good the performance can be. We need to check how critical in the performance eliminating words can be. For age range we could consider not removing any word. In fact, during task works we saw that eliminating words that appeared less than 3 times made us get a test set prediction with the same % of Male and Female labels as we got in the training set. But as 1 single submission was nally allowed, we decided to rest in the number we had already made some tests in-house. We could try to introduce some other statistics as central moments, minimums, maximums, etc, to see whether these new features a ect to the algorithm performance.