Author pro ling is used to identify di erent personality traits of an author from the given text. Di erent applications of author pro ling are marketing, security, forensics, spam detection etc. In this paper, we describe the methodology used in author pro ling for the FIRE'18MAPonSMS shared task. Our main objective is to identify age and gender of an author from the multilingual (English and Roman Urdu) text. We incorporated di erent style, vocabulary and emoticon based features to build our proposed system using the training data provided by the task organizers. Furthermore, we used a combination of di erent extracted features and investigated di erent ML algorithms (i.e. Random Forest, Nave Bayes and Support Vector Machines) for the classi cation problem. Results show that our proposed system achieved 73% accuracy in gender identi cation and 53% accuracy in age prediction tasks.
Author pro ling task is to identify di erent personality traits of an author i.e. age, gender, personality type, native language etc. from their social dialect. In other words, author pro ling is to detect di erent attributes of an author's personality from the written text. The task can be further categorized into monolingual and multilingual author pro ling. In monolingual, the pro le text is in one language while in multilingual, the pro le text is in more than one language. Author pro ling has vast applications in marketing, security and for the detection of fake pro les.
In this research work, we used style, vocabulary and emoticon based features to detect age and gender of a person from a multilingual author pro ling corpus provided for the MAPonSMS'18 task. Our major contribution are as follows: { We used 83 distinct features from the multilingual author pro ling corpus, grouped into three main categories (1) style based features (2) vocabulary based features and (3) emoticon based features. { We used an ensemble approach to combine di erent features for the classi cation task. { For classi cation, we investigated three Machine Learning (ML) algorithms i.e. Random Forest (RF), Support Vector Machine (SVM), and Nave Bayes (NB). { We performed better than the baseline and achieved 73% accuracy in gender identi cation, 53% accuracy in age prediction and an overall 38% joint accuracy.
The rest of paper is organized as follows. Section 2 summarizes existing work on author pro ling. Section 3 explains the multilingual corpus and its statistics. We then present the proposed system and details of the three types of features in Section 4. Section 5 discusses results and their analysis and Section 6 concludes the paper. 2
With the evolution of mobile technology, the Short Message Service (SMS) is considered as one of the most widely used source of communication. Extracting useful information from SMS data is very crucial task and it attracted many researches.
In recent past, a lot of work has been done on author pro ling using data
from social media i.e. Twitter and Facebook data, but less work has been done
on multilingual author pro ling using SMS. In [
Argamon et al. in [
In [
Hernandez et al. in [
Cruz et al. in [
As can be seen from the above discussion, researchers have used Facebook
and Twitter data for the author pro ling task but less work has been done
for multilingual author pro ling using SMS data. Fatima et al. in [
For the task of FIRE'18-MAPonSMS organizers provided a total of 500 author
pro les from the SMS-AP-18 corpus [
We propose an author pro ling system to detect age and gender from the multilingual text for the MAPonSMS'18 task. Our system has three main steps 1) feature extraction, 2) combination of features and 3) classi cation. 4.1
In the feature extraction step, we extracted three types of features i.e. style based features, vocabulary based features and emoticon based features. Using these three types, a total of 83 features were extracted 3. In the following sections, we describe each of the feature types in detail.
Style based Features We used a total of 24 style based features which were extracted from the training corpus. This set of 24 features has 17 special characters features and 7 count of occurrences features. The details are given below. { number of tokens { number of distinct words { average words per line { number of lines { number of capital letters { number of small letters { number of digits used { number of commas [,] { number of full stops [.] { number of commercial at [@] { number of start braces [(] { number of end braces [)] { number of exclamation marks [!] { number of dashes [-] 3 Before extracting these features, the text was preprocessed by converting it into lower case. { number of question marks [?] { number of percentages [%] { number of ampersands [&] { number of underscores [ ] { number of hashes [#] { number of equal-tos [=] { number of colons [;] { number of semicolons [;] { number of spaces [] ] { number of forward slashes [/] Vocabulary Based Features We also used a number of vocabulary based features i.e. abbreviations, academic terms, contractions and slang words. Each of these are explained in the next sections.
Word aoa w.s w8 ty tc msg y ia os lol wow pls btw Abbreviations Abbreviations are de ned as shortened form of a word. In SMS conversations, people prefer to use short form of words. Considering this fact, we used a list of Roman Urdu and English abbreviations to extract 14 features from the training corpus. As the corpus data comes from the SMS messages, people often use di erent forms of a single word (abbreviations) in SMS communication. Therefore, we further normalized these word variations to their single form. For example, \aoa" is abbreviated in many ways i.e. \a.o.a", \a.a.w.w" and \a.w". We replaced all of these variants with a single abbreviation \aoa".
Table 1 lists the abbreviations along with their English translations that we extracted from the training corpus.
Academic terms We also used the academia related terms as features for the classi er. Academic terms such as mam (madam), sir, assignment, class, quiz, o ce, teacher, test, uni (university) and clg (college) were extracted from the training corpus. Similar to the abbreviations (See Section 4.1) di erent spelling variants of a word were normalized to their single word form e.g. mam was used for ma'am, maam and ma'm.
Contractions Contractions are shorten form of words excluding some internal letters. We used the most commonly used 15 contractions as features from the training corpus. The list of contractions is shown below.
{ I'm { I am { you're { You are { we're { We are { they're { They are { Who're { Who are { I'll { I shall { She'll { She will { It'll { It will { Can't {Can not { Don't { Do not { Isn't { Is not { Aren't { Are not { Doesn't { Does not { Wasn't { Was not { Didn't { Did not Slang Words Slangs are those informal words that people use in their conversation especially on the Web and di erent social media forums. We developed a small dictionary of these commonly used slang words in Roman Urdu language and their English translations. Some of the examples from the dictionary are shown in below.
{ chuss { Unpleasant { scene { Situation { chill { Enjoyment { pagal { Mad { miss { Remembering { mood { Temper { burger { Impersonate { bhai { Brother { drama { Acting { t { Healthy { chawal { Stupid { yar { Friend { dfa { Get lost
It should be noted that our dictionary also converts di erent variations of these Roman Urdu words into their normalized forms.
Emoticon based Features: Emoticons are small images or icons that are used to express moods, expressions and feelings in day-to-day communication. We collected seven main types of emoticons. The following list shows all types of emoticons extracted from the training corpus.
{ Happy { :) { Sad { :( { Cry { :'( { Unsure { :/ { Squint { - { Kiss { :* { Wink { ;) 4.2
Combination of Features { All style features: In this combination set, we combined all style based features together (See Section 4.1). The set consists of 24 features which were used together to train the classi er for prediction of age and gender. { All vocabulary features: For the second set, we combined all vocabulary based features (See Section 4.1). The set consists of 14 abbreviations, 10 academic terms, 13 slang words and 15 contractions. { All emoticon features: The third set is the combination of all emoticon features which are 7 in number (See Section 4.1). { All features combined: In this combination set, we combined all style, vocabulary and emoticon features together for the classi cation task. The combination set contains all 83 features. 4.3
For the classi cation task of predicting age and gender from the multilingual text, we explored di erent ML classi cation algorithms including Random Forest (RF), Support Vector Machine (SVM) and Nave Bayes (NB). The feature sets from Section 4.2 were provided as input to train the classi er. Based on the feature sets (See Section 4.2) and ML algorithms (See Section 4.3), we trained separate models for both gender identi cation and age prediction tasks. We applied 10-fold cross validation and evaluated the performance on the basis of accuracy.
From Table 2 it can be seen that RF classi er has outperformed in gender identi cation with `All features' set with the highest accuracy of 78.57%. This shows that combining all distinct features has resulted in better predicting gender of an author from the multilingual text. The gender prediction also highlights that the max contribution to the best accuracy is by 'Style' based features (accuracy = 75%). On the other hand, for age prediction, SVM performs overall best with the highest accuracy of 55.42% again using `All features'. Furthermore, all the three individual sets of features used in the age prediction task has reported almost the same results. Conclusively, we can say that the ensemble approach that we have used to combine the di erent individual features together has resulted in the better performance of di erent classi ers.
We submitted our system to the MAPonSMS'18 task and it was evaluated on the test dataset. Our system performed better than the baseline with the achieved results are 73% accuracy in gender identi cation task while 53% accuracy in age prediction task. 6
In this paper we used style, vocabulary and emoticon based features for gender identi cation and age prediction tasks on the multilingual author pro ling corpus provided by the MAPonSMS'18 task. We trained a number of ML classi ers i.e. RF, SVM and NB using di erent combinations of the above mentioned sets of features. Though our system (gender = 73.57%, age = 53.42% and joint 38%) beat the baseline accuracies (gender = 0.60, age = 0.51 and joint 0.32), in comparison to the top rank system, there is still much room for improvement. If the task is o ered in the future, we will explore more sophisticated methods to improve our results.