Harassment detection: a benchmark on the #HackHarassment dataset

CERC​2016  Harassment detection: a benchmark on the #HackHarassment dataset  Alexei Bastidas, Edward Dixon, Chris Loo, John Ryan   ​Intel   e­mail:  edward.dixon@intel.com  Keywords:​e.g.Machine Learning, Natural Language Processing, Cyberbullying  Introduction  Online harassment has been a problem to a greater or lesser extent since the early days of the                                     internet. Previous work has applied anti­spam techniques like machine­learning based text                       classification (Reynolds, 2011) to detecting harassing messages. However, existing public datasets                       are limited in size, with labels of varying quality. The #HackHarassment initiative (an alliance of                               1 tech companies and NGOs devoted to fighting bullying on the internet) has begun to address this                                 issue by creating a new dataset superior to its predecssors in terms of both size and quality. As we                                       (#HackHarassment) complete further rounds of labelling, later iterations of this dataset will                         increase the available samples by at least an order of magnitude, enabling corresponding                           improvements in the quality of machine learning models for harassment detection. In this paper, we                               introduce the first models built on the #HackHarassment dataset v1.0 (a new open dataset, which                               we are delighted to share with any interested researcherss) as a benchmark for future research.  Related Work  Previous work in the area by Bayzik 2011 showed that machine learing and natural language                               processing could be successfully applied to detect bullying messages on an online forum. However,                             the same work also made clear that the limiting factor on such models was the availability of a                                     suitable quantity of labeled examples. For example, the Bayzick work relied of a dataset of 2,696                                 samples, only 196 of which were found to be examples of bullying behaviour. Additionally, this                               work relied on model types like J48 and JRIP (types of decision tree), and k­nearest neighbours                                 classifiers like IBk, as opposed to popular modern ensemble methods or deep neural­network­based                           approaches.  Methodology  Our work was carried out using the #HackHarassment Verison 1 dataset, the first iteration of which  consists exclusively of Reddit posts.  An initially random selection of posts, in which harassing  content occured at a rate of between 5% and 7% was culled of benign content using models training  on a combination of existing cyberbullying datasets (Reynolds 2001, also “Improved cyberbullying  detection through personal profiles). Each post is labelled independently by at least five Intel  Security Web Analysts.   (a post is considered “bullying” if it labelled as such by 20% or more of  the human labelers ­ as shown in the following histogram, a perfect consensus is relatively rare, and  so we rate a post as “harassing” if 20% ­ 2 of our 5 raters ­ consider it to be harassing).  This is a  relatively balanced dataset, with 1,280 non­bullying/harassing posts,, and 1,118 bullying/harassing  examples.  1 "Hack Harassment." 2016. 26 Jul. 2016 <​http://www.hackharassment.com/​> 

CERC​2016    All pre­processing, training and evaluation was carried out in Python, using the popular                           SciKit­Learn library (for feature engineering and linear models) in combination with Numpy (for                           2 3 matrix operations), Keras  and TensorFlow  (for models based on deep neural networks ­ DNNs).    4 5 For the linear models, features were generated by tokenizing the text (breaking it aparting into                               words), hashing the resulting unigrams, bigrams and trigrams (collectiojns of one, two, or three                             adjacent words) and computing at TF/IDF for each hashed value. The resulting feature vectors                             were used to train and test Logistic Regressioin, Support Vector Machine and Gradient Boosted                             Tree models, with 80% of data used for training and 20% held out for testing (results given are                                     based on the held­out 20%).  For the DNN­based approach, a similar approach was taken to tokenization, both bigram and                             trigram hashes were computed; these were one­hot encoded, and dense representations of these                           features were learned during training, as per Joulin 2016.  2 “scikit­learn: machine learning in Python — scikit­learn 0.17.1 …” 2011. 29 Jul. 2016 <​http://scikit­learn.org/​>  3 “NumPy — Numpy.” 2002. 29 Jul. 2016 <​http:/

This content is AI-processed based on open access ArXiv data.