With the ubiquity of social media, although one's thoughts and opinions can be expressed through virtual platforms effortlessly, there have been numerous cases of posts that threaten the security of a certain community, caste, or religion or spread false propaganda against a certain group of people. Developments in the fields of Natural Language Processing and Machine Learning have paved the way to the concept of author profiling, which helps identify an author's age, demographics, and gender details. The Author Profiling in Arabic Tweets task of FIRE 2019 aims to monitor Arabic Twitter posts and profile their authors concerning their age, gender, and language variety using learning concepts. The task of Deception Detection in Arabic texts focuses on monitoring Twitter and News headlines and detect deceptive texts: Posts that are drafted to seem authentic but suggest other ulterior motives. We have adopted the concept of SGD Optimized Support Vector Machine classification with AraVec word embedding for both the tasks and have achieved a joint F-1 score of 0.3403 for Author Profiling and an average score of 0.7598 for the Deception detection task.
The APDA task of FIRE 2019 funded by ARAP Qatar aims to enhance
cybersecurity using Machine Learning. Social media such as Twitter, Facebook,
Instagram, etc. have gained popularity over the past decade where users can post
images, and text online, no questions asked. Many news magazines and papers
have shifted their base to virtual media as well. Although these practices have
uncovered many societal goods, they have engendered many drawbacks such as
the rapid propagation of offensive posts and unreliable news that threaten the
security of certain communities or individuals. Author profiling is the process of
predicting the chatacterestics an author (gender, age, and location), based on
their stylistic writing features. This not only forms a security layer, but also
provides interesting linguistic information for targeted advertising, and marketing.
The second task, Deception detection aimed at classifying news headlines and
snippets as deceptive or non-deceptive, based on their language and word usage.
This was done in [
The APDA task of FIRE 2019 was divided into two individual tasks, Author Profiling of Twitter posts and Deception Detection of news headlines and tweets. The corpora of the Author Profiling task consisted of a total of 2,250 users released over five days as groups of 450 users per day, which had even distribution across all classes. Each user’s posts were collectively classified into three of the following classes: Gender, Age and language variety.
– Gender - Male or Female – Age - Under 25, Between 25 and 34, Above 34 – Language variety - Algeria, Egypt, Iraq, Kuwait, Lebanon-Syria, Lybia, Morocco, Oman, Palestine-Jordan, Qatar, Saudi Arabia, Sudan, Tunisia, UAE, Yemen.
The training dataset of the Deception detection in Arabic texts task was divided into Twitter headlines and News headlines, each of which had two classifications based on whether or not the sentence was deceptive – Truth or Lie. 2.2
Data cleaning was a major task for the Arabic data set, as the language was written from right to left, and had different rules of sentence termination, all – Links : Twitter posts contain lots of hyperlinks containing “Http://”, “www.”, “.com”, etc. All of these were adding noise to the textual data and hence were removed using regular expressions. – Hashtags: English hashtags were removed in their entirety, whereas Arabic hashtags were maintained with only the removal of the hashtag to prevent loss of useful linguistic information. – Non - Arabic Words: As it was impractical to run language models for both Arabic and English due to the sparse density of English posts, all English words were removed from the text. – Twitter handles: Social Media handles starting with “@” were eliminated from the text – Special Characters: To further level the contents of the text, all special characters, erroneous blank spaces, numbers, and empty strings were removed – Emojis: A major challenge for data cleanup for both tasks, was the removal of emojis, which had characters outside of the basic multilingual plane. These characters were extracted using Unicode conversion and used regular expressions to remove them. – Stop Words: To remove Arabic stopwords, the NLTK platform’s Arabic stopword list was taken, against which every sentence was filtered.
Word embedding [
We have implemented an SGD optimized SVM classifier with Aravec word
embedding for our model. In Stochastic Gradient Descent optimization[
The performance of our model for deception detection was remarkably better
than that for author profiling. The F1 scores were 0.34 for author profiling
and 0.76 for deception detection [
Author profiling and deception detection have numerous applications given the amount of data exchanges over the internet each day. Although numerous NLP models are available for the English language, these need to be extended to languages such as Arabic, which is spoken by a vast majority of the world, which is what APDA@FIRE aims to achieve. This paper aims to predict the personality of authors based on their style of word usage and to ensure the credibility of news by classifying the snippets as true or false. We have used word embedding and a traditional learning model to implement the same. Future scope includes comparing the accuracy of different types of traditional models such as Decision trees, Random Forrest, and Bayesian classifiers, and studying the choice of other word embedding tools that can capture the minute differences in written Arabic amongst age groups.