On the Normalization of the Experimental Design of Multilingual Hate Speech Detection

PhD Thesis Proposal Defence


Title: "On the Normalization of the Experimental Design of Multilingual 
Hate Speech Detection"

by

Miss Nedjma Djouhra OUSIDHOUM


Abstract:

With the expanding use of social media platforms and the amount of text 
data generated online, hate speech has been proven to negatively affect 
individuals in general, and marginalized communities in particular. In 
order to improve the online moderation process, there has been an 
increasing need for filters of superior quality, and therefore better data 
to train classifiers. However, there is no universal definition of hate 
speech which makes the data collection hard and the training corpora 
sparse and challenging for nowadays machine learning techniques. In this 
thesis, we address the problem of automatic hate speech detection through 
two main aspects: (1) the construction of resources lacking from robust 
hate speech detection and classification systems, and (2) bias in hate 
speech classifiers.

To tackle the lack of data, we start by building a new multilingual 
multi-aspect hate speech dataset in English, French, and Arabic. We 
provide a detailed annotation scheme which indicates (a) whether a tweet 
is direct or indirect; (b) if it is offensive, disrespectful, hateful, 
fearful out of ignorance, abusive, or normal; (c) the attribute based on 
which it discriminates against an individual or a group of people; (d) the 
name of this group; and (e) how annotators feel about this tweet given a 
range of negative to neutral sentiments. We define classification tasks 
based on each aspect and use multi-task learning to investigate how such a 
paradigm can boost the classification performance.

Second, we examine misclassification instances due to bias and equivocal 
interpretations of hate speech. In contrast to work on bias which 
typically focuses on the classification performance, we look into the 
frequently neglected selection bias caused by the data collection process. 
We present two language and label-agnostic evaluation metrics based on 
topic models and semantic similarity measures to assess the extent of 
selection bias on various hate speech datasets. Furthermore, since we 
generally focus on English and overlook other languages, we notice a gap 
in content moderation across languages and cultures, especially in low 
resource settings. Hence, we leverage the observed differences and nuances 
across languages, datasets, and annotation schemes to carry a study on 
multilingual hate speech data and how people react to it.

Finally, we propose to assess harmful biases using simple classifiers in 
large pre-trained language models which are at the core of NLP systems, in 
order to avoid replicating them. Then, we discuss future work and ways to 
mitigate bias in different contexts.


Date:			Wednesday, 31 March 2021

Time:                  	2:00pm - 4:00pm

Zoom Meeting: 
https://hkust.zoom.us/j/98682650205?pwd=YXZHR0s4blJvRjR6czRiUGRpRmNkUT09

Committee Members:	Dr. Yangqiu Song (Supervisor)
 			Prof. Dit-Yan Yeung (Supervisor)
  			Prof. Nevin Zhang (Chairperson)
 			Prof. Pascale Fung (ECE)


**** ALL are Welcome ****