Social media sites are a rich platform for a user-generated text that can be used to identify di erent aspects of the authors. Age, gender, dialect, region, are di erent aspects of an author, which can be identi ed with the proper mining of these contents. This pro ling provides a way of identifying anonymous users. Recognizing the pro le of an anonymous user help in indirect recognition of the identity of the user. In this notebook, we describe the working of our author pro ling software submitted for FIRE 2019 which recognizes the gender, age, and dialects of Twitter users in the Arabic language. We have used Long short term memory neural network and some hand-crafted features for recognizing the age, gender, and dialects of the author of a tweet3.
Author pro ling is the technique to identify the di erent aspects of the author
from a given text. It di erentiates among the di erent classes of an author by
studying their writing style and the words used in their text. It also shows how
the behavior viewpoint is used to recognize the di erent classes of an author.
It tells about the uses of di erent writing skills and how the language is shared
by a di erent author in their text. The textual information based on di erent
features and styles helps identify the author's pro le based on di erent aspects
such as gender, age, and dialect[
Author pro ling has attracted researchers and other di erent competitions.
Researchers have studied the dependence of linguistic features and the pro le of the
author. This dependency is a subject of interest for di erent areas like linguistics,
psychology, and natural language processing [
The dataset was provided by the APDA track organizers of FIRE 20194. It
consists of Arabic tweets of 5 set of tweet data. Each data has 100 tweets of 450
users [
set of tweets. We considered each tweet as a sample, thus we have 45000 sample of tweets divided into two class. The users who were male, their tweets were labeled with the male, and similarly for females. Similarly, for age and dialect, 45000 samples were created. Finally, after getting the results, the decision is taken on the majority voting basis. For example, We have 450 tweets of 1 user in 1 document. As per our method, we will get result for all 450 tweets. Let say for 200 of the tweets, the result is Male, and female for remaining then the nal result for that document will be female. In this way, we have made the data samples and concluded the results.
These tweets are in Arabic language having Emoticons, emojis, #mentions, @mentions, and URLs. Since the provided training corpus consists of tweets are with HTML format, rstly, we extract all the Tweets from its HTML format to simple text and then applied the preprocessing step for cleaning the tweets. The preprocessing stage is useful as it reduces non-textual features to their semantic classes. We used these preprocessing steps before the extraction of features.
URLs : The URLs are deleted. @mentions : The @ mentions were deleted. Emoticons : Emoticons provide useful style-based information. It informs about the view of a speci c user. We captured their presence only. If the emoticon is present, then 1 otherwise 0. Furthermore, we apply the following normalization: Punctuation marks: The punctuation marks are split from adjacent word and captured their presence separately. The Stopwords and Punctuation were also removed from the tweets. 4
In this section, we are discussing the proposed architecture, performance measures and the results. 4.1
We applied a deep learning model for the author pro ling task. We have used long short term memory network and applied to the tweets as a classi cation problem.
Model-I Long short Term Memory(LSTMs) are designed to learn the longterm dependency of text-based problems. LSTMs remember the information for long periods and it is their default behavior5. In our LSTM model, rstly, we create a word embedding matrix, a tokenizer and a vocab for our training corpus to determine the unknown words. We randomize our training corpus to make it more e cient. We introduce three layers with di erent activation keys(relu, sigmoid) to train our training corpus on the LSTM model and test it on the test corpus. We use Binary cross-entropy as a loss function to use for binary decisions and adam as our optimizer in the compiling of our model. In sub gure 1a, we have shown the structure of the model.
(a) Model-I (b) Model-II Model-II We tried a simple variant of the Model-I discussed in 1a, by adding some extra hand-crafted features. We added the hand-crafted features in the developed model to check the performance of the system. In sub gure 4.1, we have shown the structure of the model. Below are the additional handcrafted features: { Emoji Count : It counts the total number of emoji presents in the tweet. { Word Count : It counts the total number of words present in the tweet. { Polarity of Sentiment Analysis : It gives, 1 for positive, -1 for negative and 0 for neutral sentiment. { Mean Word Length: It calculates the average length of the words present in the tweet. It is the ratio of the total length of words present in the tweet to the total no. of words present in the tweet. { Sentence length: It gives the length of the sentence in the tweet. { Special Character : It gives the total count of punctuation present in the tweet. { Unrepeated Words : It gives the total number of the words which appeared only once in the tweet. { URL Extractor : It gives 1 if URL is present in the tweet otherwise it gives 0. 4.2
We used Aravec [
For model-II, we extracted the features from the tweet data and appended the extra features to the output of the LSTM layer. After that, the nal result is extracted from the softmax layer. The Github link for the proposed framework is given below. 6 4.3
The performance of the system is evaluated based on the accuracy of the developed system. The Accuracy of a system is the ratio of the total number of instances predicted correctly to the total number of instances present in the data.
Accuracy =
P
P + Q
Where, P is the number of instances predicted correctly, and Q is the number of instances predicted incorrectly. (1)
We found the accuracy of model-I was less than model-II, thus we are reporting the performance of model-II only. In table 1, we have shown the accuracy achieved on training data, and in table 2, we have shown the cross0validation accuracy on the training data. We performed 5-fold cross-validation on the training data to check the performance of the system and reduce the generalization error. It was done because we didn't have the labels of test data before submission.
For training data, model-II is giving better results than model-I. The performance is similar on test data too. The model-II achieved an accuracy of 66.25% for gender, 22.22% for age, and 80.28% for language variety. The joint performance of the system is 0.1083. While the joint accuracy of model-I is 0.0722. In table 3, we have shown the results of the proposed systems on test data.
In this work, we have presented the task of automatic classi cation of the author's gender, age, and dialect from their writing. This work attracts several potential applications like security, forensics, and marketing, etc. We have developed an lstm-based neural network model for recognizing the age, gender, and language variety of an author by using his/her written text. We have used some style-based features for improving the performance of the lstm-based system. In the future, we will try to optimize the neural network architecture to enhance the e ciency of the system. We would also look into task-speci c handcrafted features, to improve the performance of the system. We will work on the using the properties of homonyms in our feature detection.