Bullying Traces v1.0

*****************************************************************************
As per Terms of Services of Twitter, the original tweets are not included in 
this dataset.  Instead, we release proceeded data files only.  Specifically,
we release the vocabulary, feature vectors, and labels.  It is impossible to 
fully reconstruct the original tweets from this data.  However, we believe
the dataset is still useful for the scientific study of bullying.
*****************************************************************************

This dataset contains data collected from Twitter stream API and labeled by 
experienced annotators for the study of bullying traces in social media.  We 
collected tweets using the public Twitter stream API, such that each tweet 
contains at least one of the following keywords: "bully, bullied, bullying". 
We further removed re-tweets by excluding tweets containing the acronym "RT." 
Our annotators labeled 1762 tweets uniformly sampled from the ones collected 
by the above procedure on August 6, 2011.  The tweets are cased-folded and 
tokenized, but without any stemming or stopword removal. Any user mentions 
preceded by a ``@'' were replaced by the anonymized user name ``@USERNAME''. 
Any URLs starting with ``http'' were replaced by the token ``HTTPLINK''. 
Hashtags (compound words following ``#'') were not split and were treated as 
a single token. Emoticons, such as ``:)'' or ``:D'', were also included as 
tokens. Our features include both unigrams and bigrams that appear at least 
twice in the 1762 tweets. 

To cite this dataset:

Learning from bullying traces in social media
Jun-Ming Xu, Kwang-Sung Jun, Xiaojin Zhu, and Amy Bellmore
In North American Chapter of the Association for Computational Linguistics - 
Human Language Technologies (NAACL HLT)
Montreal, Canada, 2012

Contact: Jun-Ming Xu (xujm@cs.wisc.edu), Xiaojin Zhu (jerryzhu@cs.wisc.edu)
April 2012


----------------------------------Format-----------------------------------

The feature vectors are written in a sparse vector format as commonly used 
in SVM-light.  Each line corresponds to one tweet and has the format:

Label featureIndex:value featureIndex:value featureIndex:value ...

Only features with a nonzero value are listed.  Each vector is normalized
to have norm 1.



----------------------------------Content----------------------------------

vocab
	The vocabulary file for all features. Each line is a token and the 
	index is the line number, starting from 1.

tweetType
	This corresponds to "NLP Task A: Text Categorization" in the paper.
	It contains 1762 feature vectors for all the labeled tweets. 
	Labels: 1 bullying trace, -1 not bullying trace
	
authorRole
	This corresponds to "NLP Task B: Role Labeling / Author's Roles". 
	It contains the 684 feature vectors for the bullying traces only.
	Labels:	
		1 Accuser
		2 Bully
		3 Reporter
		4 Victim
		5 Other

teasing
	This corresponds to "NLP Task C: Sentiment Analysis" in the paper.
	It contains the 684 feature vectors for the bullying traces only.
	Label: 1 Teasing, -1 Not teasing