A two-stage contagious Naive Bayes classi er for detecting sociolinguistic features in text

Abstract

Online platforms allow users to masquerade themselves; making virtual interactions anonymous or misleading recipients of the interactions. It also facilitates an environment for cybercrimes, allowing users to take advantage of others and commit heinous acts. An important concern on social media usage, in particular, has to do with the security of under-age users that have access to the Internet. Children are more vulnerable to threatening situations, such as harassment [3], cyberbullying [7], and inappropriate conversations [8]. Natural language processing (NLP) techniques can be used to process and understand social media data [1]. In the area of sociolinguistics, there is evidence that links natural word use to personality and social uctuations [5]. In NLP, the term burstiness is used to describe the tendency of word recurrence. The burstiness phenomenon is frequently exhibited in real text, in which an informative word is more likely to occur if it has already appeared in the text [2]. State-of-the-art NLP models, such as the multinomial Naive Bayes model, are often used to model text documents [4].

-

Introduction One application area of NLP is sociolinguistics which can be de ned as the relationship between social factors and linguistics. Sociolinguistics aims to isolate features to determine linguistic variation in social conditions [ 6 ]. This paper investigates classi cation models which can model the burstiness, or contagious e ects of text as the text that we are interested in are manifestations of di erent social groups of people. For example, a teenager vs. an adult impersonating as a teenager will have di erent sentence structures. More formally de ned, the purpose of this work is to: 1. Learn the linguistic patterns among di erent social groups of people, and classifying unknown authors according to these patterns; and 2. Represent these patterns as Bayesian networks to gain an understanding of the dependency structure of words used among di erent social groups. To identify these linguistic patterns, a comparison is made between two classi cation techniques, namely the Naive Bayes (NB) classi er and a contagious counterpart thereof. The NB classi er assumes that words occurring in a document are independent of each other. On the other hand, the contagious classi er captures the burstiness phenomenon. To go one step further, we investigate the dependencies between words using a Bayesian network. This allows us to understand why certain word patterns results in a classi cation. 2.1

Data Application This paper presents a comparison between the baseline NB classi er and the proposed contagious counterpart thereof. Two data sets will be used to evaluate the performance of each method, namely the IMDB data set and the PAN 2012 data set. The IMDB data set consist of movie reviews, with binary sentiment classi cation. The PAN 2012 data set is originally used to identify potential predators in online conversations, with 66 927 conversations. The problem is addressed with a two-stage solution, where stage 1 is based on text classi cation techniques and stage 2 makes use of Bayesian networks to understand the structural dependencies among words in a document. The evaluation of stage 1 is typical classi cation performance, whereas the visual structural learning of the Bayesian network provides for exploratory data analysis in order to understand the conditional dependencies between words.

References

1. Chowdhury , G.G. : Natural language processing . Annual review of information science and technology 37(1) , 51 { 89 ( 2003 )

2. Doyle , G. , Elkan , C. : Accounting for burstiness in topic models . In: Proceedings of the 26th Annual International Conference on Machine Learning . pp. 281 { 288 . ACM ( 2009 )

3. Kennedy , G. , McCollough , A. , Dixon , E. , Bastidas , A. , Ryan , J. , Loo , C. , Sahay , S. : Technology solutions to combat online harassment . In: Proceedings of the rst workshop on abusive language online . pp. 73 { 77 ( 2017 )

4. Madsen , R.E. , Kauchak , D. , Elkan , C. : Modeling word burstiness using the dirichlet distribution . In: Proceedings of the 22nd international conference on Machine learning . pp. 545 { 552 . ACM ( 2005 )

5. Pennebaker , J.W. , Mehl , M.R. , Niederho er, K.G.: Psychological aspects of natural language use: Our words, our selves . Annual review of psychology 54(1) , 547 { 577 ( 2003 )

6. Spolsky , B. , Widdowson , H. , et al.: Sociolinguistics , vol. 1 . Oxford University Press ( 1998 )

7. Van Hee , C. , Lefever , E. , Verhoeven , B. , Mennes , J. , Desmet , B. , De Pauw , G. , Daelemans , W. , Hoste , V. : Automatic detection and prevention of cyberbullying . In: International Conference on Human and Social Analytics (HUSO 2015 ). pp. 13 { 18 . IARIA ( 2015 )

8. Yenala , H. , Jhanwar , A. , Chinnakotla , M.K. , Goyal , J.: Deep learning for detecting inappropriate content in text . International Journal of Data Science and Analytics 6 ( 4 ), 273 { 286 ( 2018 )