Racism and COVID-19

• ANR
  XTCOVIF
• Semantic
  Shift
• Predicting the Spread
  of COVID-19
• Trust in
  Government
• Racism and
  COVID-19
• Population
  Mobility
• Publications and
  Datasets

Large scale analysis of mobility data and language during the COVID19 pandemic

Racism and COVID-19

Introduction

In late 2019, a highly infectious new virus, SARS-CoV-2, started spreading in Wuhan, China. In early 2020, the virus had spread to most countries around the world causing the pandemic of the COVID-19 disease. The fact that COVID-19 originated in China triggered racist and xenophobic violence towards Asians and people of Asian descent. In the context of this project, we are interested in detecting racism behaviors and hate speech on social media and further in analyzing their spread pattern.

The problem of hate speech has grown a lot in recent years, mainly due to the large adoption and use of social media. There is no widely-accepted definition of hate speech, however, hate speech mainly refers to some sort of communication that denigrates people on the basis of their membership to a particular group. Hate speech detection (i.e., the task of detecting hate speech) has attracted a lot of attention recently [Schmidt and Wiegand, 2017]. However, hate speech detection is a very challenging task since, as mentioned above, there is no clear definition of hate speech. Social media platforms have been making an effort to ban hateful content, sparking some communities to create a new vocabulary to bypass the regulations [MacAvaney et al., 2019]. Lately, the spread of racism in social media towards the Chinese community has started being studied [Ziems et al., 2020; Cheng and Conca-Cheng, 2020; Devakumar et al., 2020]. However, all these studies have focused on tweets written in English. In an effort to promote French natural language processing, we have created a simple hand labeled dataset for general offensiveness detection in Twitter data during the COVID-19 pandemic.

Dataset

We first created a dataset for the task of offensiveness detection. More specifically, we manually annotated 5,786 tweets of which 1,301 were labeled as offensive, while the rest of them were labeled as not offensive. As offensiveness can be subjective in certain cases, during the annotation process, we considered a tweet as being offensive in case it targeted another person specifically. For instance, the tweet “The chinese virus is tiring” would not be considered as offensive, while the following tweet “I hate the chinese for bringing us the chinese virus” would instead be regarded as offensive.

Methodology

To deal with this task, we capitalized on recent advances in representation learning in the field of natural language processing (NLP) and specifically on BERT [Devlin et al., 2019], a method for generating contextual word embeddings. Since its inception, BERT has brought a revolution to representation learning in NLP, and has led to impressive results in several NLP tasks such as machine translation, text categorization, and question answering. Due to its recent success, we decided to use BERT as our starting point. It is important to note that, as we worked with a French corpora, we used CamemBERT [Martin et al., 2020], a French variation of BERT, in order to produce appropriate word representations.

The CamemBERT model is also capable of generating document representations, and we used this model to obtain a vector for each tweet of our dataset. We then fed these tweet representations to a logistic regression classifier which was trained to identify offensiveness. Along with the tweet representations, we also used some linguistic features, such as the length of a tweet. Furthermore, we provided the classifier with a list of terms that can be considered as being systematically offensive. This corresponds to a linguistic feature that can help the classifier in the task of identifying offensive tweets. To that end we selected words that had a history of always being used in offensive scenarios to support the classifier, with different level of virulence. All these semantic features were found to improve the performance of the classifier.

Results & Discussion

The performance of the classifier in detecting offensive tweets is illustrated in the following Table.

As we can see, the classifier produces fairly accurate results that are on par with similar classifiers that have been developed for the English language. We can thus use the classifier to extract tweets that contain hate speech towards the Chinese community.