BNU-HKBU UIC NLP Team 2 at SemEval-2019 Task 6: Detecting Offensive Language Using BERT model

Abstract

In this study we deal with the problem of identifying and categorizing offensive language in social media. Our group, BNU-HKBU UIC NLP Team2, use supervised classification along with multiple version of data generated by different ways of pre-processing the data. We then use the state-of-the-art model Bidirec-tional Encoder Representations from Transformers , or BERT (Devlin et al. (2018)), to capture linguistic, syntactic and semantic features. Long range dependencies between each part of a sentence can be captured by BERT's bidirectional encoder representations. Our results show 85.12% accuracy and 80.57% F1 scores in Subtask A (offensive language identification), 87.92% accuracy and 50% F1 scores in Subtask B (categorization of offense types), and 69.95% accuracy and 50.47% F1 score in Subtask C (offense target identification). Analysis of the results shows that distinguishing between targeted and untargeted offensive language is not a simple task. More work needs to be done on the unbalance data problem in Subtasks B and C. Some future work is also discussed.

Full text

Proceedings of the 13th International Workshop on Semantic Evaluation (SemEval-2019), pages 551–555
Minneapolis, Minnesota, USA, June 6–7, 2019. ©2019 Association for Computational Linguistics
551
BNU-HKBU UIC NLP Team 2 at SemEval-2019 Task 6: Detecting
Offensive Language Using BERT model
Zhenghao Wu Hao Zheng Jianming Wang Weifeng Su Jefferson Fong
{l630003054,l630003067,l630003049}@mail.uic.edu.hk
{wfsu,jeffersonfong}@uic.edu.hk
Computer Science and Technology, Division of Science and Technology
BNU-HKBU United International College
Zhuhai, Guangdong, China
Abstract
In this study we deal with the problem of iden-
tifying and categorizing offensive language
in social media. Our group, BNU-HKBU
UIC NLP Team2, use supervised classification
along with multiple version of data generated
by different ways of pre-processing the data.
We then use the state-of-the-art model Bidirec-
tional Encoder Representations from Trans-
formers, or BERT (Devlin et al. (2018)), to
capture linguistic, syntactic and semantic fea-
tures. Long range dependencies between each
part of a sentence can be captured by BERT’s
bidirectional encoder representations. Our
results show 85.12% accuracy and 80.57%
F1 scores in Subtask A (offensive language
identification), 87.92% accuracy and 50% F1
scores in Subtask B (categorization of offense
types), and 69.95% accuracy and 50.47% F1
score in Subtask C (offense target identifica-
tion). Analysis of the results shows that distin-
guishing between targeted and untargeted of-
fensive language is not a simple task. More
work needs to be done on the unbalance data
problem in Subtasks B and C. Some future
work is also discussed.
1 Introduction
Social media is an essential part of human com-
munication today. People can share their opinions
in this platform with anonymity. Some people
use offensive language and hate speech casually
and frequently without taking any responsibility
for their behavior. For this reason, SemEval 2019
(Zampieri et al. (2019b)) set up the task Offen-
sEval: identifying and categorizing offensive
language in social media. This task is divided into
three subtasks: offensive language identification,
automatic categorization of offensive types, and
offence target identification.
Our group uses the Natural Language Process-
ing (NLP) latest model, Bidirectional Encoder
Representations from Transformers (BERT). It
is a general-purpose “language understanding”
model trained on a large text corpus such as
Wikipedia (Devlin et al. (2018)). After fine-
tuning, the model can be used for downstream
NLP tasks. Because BERT is very complex and
is the state-of-art model, it is prudent for us not
to change its internal structure. Hence, we focus
on preprocessing the data and error analysis.
After much experimentation with the data, such
as translating emoji into words, putting more
weight on some metaphorical words, removing
the hashtag and so on, we find that using the
original data will give the best performance. The
reason for this is perhaps if we remove some
information from the sentence, some features that
affect the prediction result will be lost. So we end
up using the original data to train our model.
2 Related Work
Much research has been done in detecting of-
fensive language, aggression, and hate speech
in user-generated content. In recent years, re-
searches tend to follow several approaches: use a
simple model with logistic regression to perform
detection, use a neural network model, or use
some other methods.
For the simple model, Davidson and Warmsley
(Davidson et al. (2017)) used a sentiment lexicon
designed for social media to assign sentiment
scores to each tweet. This is an effective way to
identify potentially offensive terms. Then they
use logistic regression with L2regularization to
detect hate speech in social network.

552
Neural network models use n-gram, skip-gram
or some other methods to extract features from
the data. These features are used to train different
models. The results produced by these models
will be used as the input for training the meta-
classifier (e.g. Malmasi and Zampieri (2018))
For other methods, using bag-of-words is an
effective way to detect hate speech, but it is
difficult to distinguish hate speech from text with
offensive words that are not hate speech (Kwok
and Wang (2013)). For identifying the targets and
intensity of hate speech, syntactic features method
is a good method (Burnap and Williams (2015)).
3 Methodology and Data
Only the training data provided by the organizer
(Zampieri et al. (2019a)) are used in training our
model. The data contain 13,240 pieces of tweet
that had been desensitized (replacing the user
names and website URLs). There are three labels
that are labeled with crowdsourcing for each of
the three subtasks. Gold labels obtained through
crowdsourcing are confirmed by three annotators.
We segmented the training set by 90% for the
training set, 5% for the cross-validation set, and
5% for the test set.
Because some offensive language is subtle,
less ham-fisted, and sometimes cross sentence
boundary, the model trained for this task must
make full use of the whole sentence content in
order to extract useful linguistic, syntactic and
semantic features which may help to make a
deeper understanding of the sentences, while at
the same time less subjected by the noisiness of
speech. So, we use BERT in all three subtasks.
Unlike most of the other methods, BERT uses
bidirectional representation to make use of the left
and right context to gain a deeper understanding of
a sentence by capturing long range dependencies
between each part of the sentence.
The uncased base version of the pre-trained
model files 1is used during the entire training.
The training data are processed in many ways
to fine-tune the model. Processing methods
1BERT-Base, Uncased: https://storage.
googleapis.com/bert_models/2018_10_18/
uncased_L-12_H- 768_A-12.zip
include removing all username tags, URL tags
and symbols, converting all text to lowercase, and
translating emoji into text2. One or more of the
above methods is selected to process the training
data, and then use the processed data to train the
model.
In Subtask A, the accuracy after the various op-
erations is shown in the following table.
Preprocessing Accuracy
Original Data 0.8184
Remove tag & symbols 0.8126
Emoji translation v1 0.8081
Emoji translation v2 0.7960
Table 1: Training results for Sub-task A.
After all attempts, the best performing model
for Subtask A is the model trained by the original
data. Therefore, the original data are also used in
the training of the Subtasks B and C models.
4 Results
For Subtask A, The BERT-Base, Uncased, orig-
inal training data model get macro F1 score of
0.8057 and total accuracy of 0.8512.
For Subtask B, The BERT-Base, Uncased,
original training data model get macro F1 score of
0.50 and total accuracy of 0.8792.
For Subtask C, The BERT-Base, Uncased,
original training data model get macro F1 score of
0.5047 and total accuracy of 0.6995.
Results table and confusion matrices for Sub-
tasks A, B and C are shown below.
2In the process of translating emoji characters, v1 and v2
methods were used. v1: Translate all emoji characters into of-
ficial character name listed in the Unicode®11.0.0 Standard.
v2: In addition to ”v1” of processing of all emoji characters,
the selected 97 emotional emoji characters are translated into
manually determined emotional words.

553
System F1 (macro) Accuracy
All NOT baseline 0.4189 0.7209
All OFF baseline 0.2182 0.2790
BERT-Base,
Uncased,
original training data
0.8057 0.8512
Table 2: Results for Sub-task A.
System F1 (macro) Accuracy
All TIN baseline 0.4702 0.8875
All UNT baseline 0.1011 0.1125
BERT-Base,
Uncased,
original training data,
0.5 threshold
0.5000 0.8792
BERT-Base,
Uncased,
orginal training data,
0.65 threshold
0.4702 0.8875
Table 3: Results for Sub-task B.
System F1 (macro) Accuracy
All GRP baseline 0.1787 0.3662
All IND baseline 0.2130 0.4695
All OTH baseline 0.0941 0.1643
BERT-Base,
Uncased,
original training data
0.5047 0.6995
Table 4: Results for Sub-task C.
Figure 1: Sub-task A, BNU-HKBU UIC NLP Team 2
CodaLab 527070 BERT-Base, Uncased, original train-
ing data
Figure 2: Sub-task B, BNU-HKBU UIC NLP Team 2
CodaLab 531958 BERT-Base, Uncased, original train-
ing data
Figure 3: Sub-task C, BNU-HKBU UIC NLP Team 2
CodaLab 535873 BERT-Base, Uncased, original train-
ing data

554
5 Conclusions
Our model for Subtask A ranked 6 out of the
103 groups. This shows that BERT can identify
offensive language. However, the results for
Subtasks B and C are not as good. We will try to
explain the possible reasons using error analysis.
First, some speech may be miss-labeled by
annotators. In our test set, some predictions were
judged as wrong, but our manual examination
shows the predictions seem correct. For example,
“@USER B***hhhhh I’m jell” and “@USER
Crazy Russian dude owns all your data” are
both labeled NOT (not offensive). The model, as
well as our manual examination, deem these as
offensive.
Second, we also notice a problem is that it is
hard for our model to understanding some specific
noun such as people name when our training data
is not enough. For example, our model predict
sentence “Hitler will be so proud of David Hogg”
as not offensive. The word “Hitler” has a very
special meaning that can makes an otherwise
innocent sentence to be offensive. Our model
presently can’t detect this.
Another problem is emoji characters in offen-
sive languages, which usually contains strong
emotions. And may be used to express irony. So
emoji characters are translated by two methods2
to help BERT model understand the meaning
of tweet posts. But the results show that both
translation methods lead to a drop in accuracy.
The main reason should be that some emoji
characters contain different meanings in different
contexts. For example, (Slightly Smiling Face)
can contain emotion of happy but also banter
as well. Thus, it is difficult to understand the
meaning of emoji characters in context.
Moreover, unbalanced data is a big problem.
In Subtask B, few sentences are predicted as
untargeted, and in Subtask C, no sentence is
predicted as in the Others category. This leads to a
low F1 score in these subtasks. Over-sampling in
less numerous categories would not work not well
in our task, and threshold moving only slightly
raises the F1 score. To deal with this problem as
future work, we may have to remove the labels
and use unsupervised learning.
Figure 4
For future work, we notice that offensive
languages often contain strong emotions such as
angry, banter or taunt. This emotion and other use-
ful contents may be improved by using DeepMoji
(Felbo et al. (2017)), which translates a sentence
into an emoji list to express a sentence’s hidden
information, such as sentiment and sarcasm. A
list of emoji related to the meaning of a sentence
produced by DeepMoji can be used to help BERT
to better classify the sentence categories, as show
in the Figure 4. The last step is to put the original
sentence and the encoded new sentence as input
for BERT’s sentence-pair classification task.
References
Pete Burnap and Matthew L. Williams. 2015. Cyber
hate speech on twitter: An application of machine
classification and statistical modeling for policy and
decision making.
Thomas Davidson, Dana Warmsley, Michael W. Macy,
and Ingmar Weber. 2017. Automated hate speech
detection and the problem of offensive language. In
ICWSM.
Jacob Devlin, Ming-Wei Chang, Kenton Lee, and
Kristina Toutanova. 2018. Bert: Pre-training of deep
bidirectional transformers for language understand-
ing. CoRR, abs/1810.04805.
Bjarke Felbo, Alan Mislove, Anders Søgaard, Iyad
Rahwan, and Sune Lehmann. 2017. Using millions

555
of emoji occurrences to learn any-domain represen-
tations for detecting sentiment, emotion and sar-
casm. arXiv preprint arXiv:1708.00524.
Irene Kwok and Yuzhou Wang. 2013. Locate the hate:
Detecting tweets against blacks. In AAAI.
Shervin Malmasi and Marcos Zampieri. 2018. Chal-
lenges in discriminating profanity from hate speech.
J. Exp. Theor. Artif. Intell., 30:187–202.
Marcos Zampieri, Shervin Malmasi, Preslav Nakov,
Sara Rosenthal, Noura Farra, and Ritesh Kumar.
2019a. Predicting the Type and Target of Offensive
Posts in Social Media. In Proceedings of NAACL.
Marcos Zampieri, Shervin Malmasi, Preslav Nakov,
Sara Rosenthal, Noura Farra, and Ritesh Kumar.
2019b. SemEval-2019 Task 6: Identifying and Cat-
egorizing Offensive Language in Social Media (Of-
fensEval). In Proceedings of The 13th International
Workshop on Semantic Evaluation (SemEval).