Keywords

1 Introduction

Abusive and offensive language is the prime concern of technical companies nowadays due to exponential growth in number of Internet users around the world and since these people are from different walks of life and different culture. There is a fine line between hate speech and offensive language, and to detect and differentiate among them is a big challenge. In literature, researchers generally classify the text into three classes:

  • Hateful,

  • Offensive, and

  • Clean

In this paper, we showcase the study we perform on the research held in this area with some light on what can be done next in order to make it more efficient. Our objective behind carrying this work is to come up with a study of papers and research work done in this field so far.

2 Terminology

In this paper, we use the term hateful, offensive, and clean. We come to a conclusion in favor of the usage of these terms since they can have broader meaning and can be used in various contexts in user-generated content to define it first. Hateful text or speech is not a very common phrase to refer to such text in legal world but in general terms day-to-day speaking we use it quite often.

Following is the list of terms used in literature [1]:

  • abusive messages,

  • hostile messages, or

  • flames.

This will help readers to go further in literature on this topic. There is a recent trend in the NLP world that author prefers to use the word cyberbullying [2,3,4,5,6,7].

Hateful speech or hate speech is commonly referred to as conversation, communication that mocks a group of people or a single person on the grounds of social status, race, color, ethnicity, nationality, gender, sexual preferences, religions, and many others [8].

3 Literature Survey

Researchers in past have proposed various machine learning approaches and their variant to deal with the problem of offensive language. Detecting sarcasm had been the point of research for many researchers around in area of NLP or text mining, with need of hour nowadays people are more focusing on detecting the wrongs prevailing in social media. This concern of government and public leads to open new research domains as fake news detection, rumor detection, offensive language detection, etc. Many of these proposed works use feature extraction from text such as bag of words (BOW) and dictionaries. Major work in this area is focused on feature extraction from text. Dictionaries [9] and bag of words [10] were among the lexical features that were used widely by researchers to detect the offensive language or phrases. Gaydhani et al. [11] used tf-idf and N-gram as features for their classification of tweets with 95.6% accuracy.

It was found out that these features could not understand the context of sentences. Approaches that involve N-gram show better results and perform better than their counterparts [12]. Lexical features are proving to outperform other features in automatic detection of offensive language and phrases, without taking into consideration the syntactic structures as bag of word approach could not detect offensiveness if words are used in different sequences [13].

Gaydhani et al. [11] form a dataset which is the combination of three different datasets. The first dataset which they used is publicly available on Crowdflower1, which was used in [14, 15]. Dataset Crowdflower1 has tweets classified into three classes: “Hateful”, “Offensive”, and “Clean”. All the tweets in this dataset are manually annotated. The second dataset they used is crowdflower2 having tweets manually classified into same three classes. Github3 is the third dataset they integrate with other two to build their dataset for study. This third dataset consists of two columns: tweet-ID and class. “Sexism”, “Racism”, and “Neither” are the three categories or classes in which each of these tweets are classified. This dataset is used by [14, 16]. They have considered logistic regression, naive Bayes, and support vector machines for text classification. They used training of dataset on each model by performing grid search for all the combinations of feature parameters and performed 10-fold cross-validation. They analyzed performance on the basis of average score of the cross-validation.

Davidson et al. [17] reduce the dimensionality of the data using a logistic regression with L1 regularization. They show a comparative study on prior work such as logistic regression, naive Bayes, decision trees, random forests, and linear SVMs. They use fivefold cross-validation, with keeping 10% of the sample for evaluation to help prevent overfitting on all the models. Their study suggests that logistic regression and linear SVM perform slightly better than other models. They further use logistic regression with L2 regularization for the final model as it has shown better result in previous work. They use tweets from Hatebase.org which contains lexicon compiled by Internet users containing words and phrases that are considered to be hate speech. Using these words from lexicon they crawled the twitter using the Twitter API which collects tweets containing these words. They collect 33,458 user’s tweets as sample. They get these tweets annotated by CrowdFlower workers into three categories: hate speech, offensive but not hate speech, or neither offensive nor hate speech. Getting these manually annotated helps in clear tagging as they not just look for words but also context of tweets. They found majority of the tweets fall into category of offensive language. They use features from these tweets and used them to train a classifier.

Lee et al. [18] use the dataset titled “Hate and Abusive Speech on Twitter” [19] recently released. This dataset contains the tweets classified into four categories, namely, “normal”, “spam”, “hateful”, and “abusive”.

70 character dimensions using 26 lower character dimensions were used to convert the tweets into one hot encoded vector with 10 digits and special characters up to 34 including whitescape. This encoding is used for character-level representation. Before this encoding, they have removed user ID, emojis, and URLS, and replace them by special tokens.

Table 1 shows the distribution of tweets among four categories, which are discussed below:

Table 1 Distribution of categories among tweets
Full size table

Aken et al. [20] consider two datasets to evaluate their proposed algorithm: one of the datasets they pick from Kaggle’s second challenge on toxic comment classification which contains comments on Wikipedia talk pages presented by Google Jigsaw and other datasets they consider are based on Twitter by Davidson et al. [21]. Class distribution of both datasets is shown in Tables 2 and 3. 24,783 tweets were extracted from Twitter which constitute to dataset of Davidson et al. [21], and all these tweets were annotated by CrowdFlower workers with the labels “hate speech”, “offensive but not hate speech”, and “neither offensive nor hate speech”.

Table 2 Wikipedia comment dataset
Full size table
Table 3 Twitter dataset
Full size table

They propose an ensemble to figure out that a single classifier is most effective on certain kind of comment. The ensemble classifier analyzes the features from comments, weights, and for a given feature combination it identifies the suitable single classifier. To attain the goal of identifying the classifier using gradient boosting decision tree, they perform validation across the average final predictions on five trained models.

The most valuable contribution by Aken et al. [20] is Error Classes of False Negatives they have defined. These classes are as such Doubtful labels, and these are the labels that cannot be clearly identified as toxic because for a particular user it is toxic but there are users or annotators that consider it as nontoxic. Second class of false-negative error is Tweets that contain toxicity without any kind of hate words or swear words that this class of error needs to overcome which will require investigating some semantic embeddings for obtaining better classification on different paradigmatic contexts. Third class of error identified by the author is Rhetorical Questions and these are the kind of text sentences that does not contain any toxic words but have sarcastic questions in it, usually such text contains question marks and question words. Other classes they introduced are Metaphors and comparisons, and idiom that can be twisted in meaning by looking at context which are difficult to see in short text and such text usually requires knowledge about the implications of language or some additional contextual knowledge. Aken et al. [20] find that different approaches fail in identifying different texts and make errors, but this can be combined into an ensemble with F1-measure. They find some combination of shallow learners with deep neural networks showing remarkable results and proved it to be very effective.

Mathur et al. [22] explore the usage of mixed language in their work and identify the offensive text or hate speech. They choose Hinglish as their subject because of its ease in communication and being popular on Twitter due to its reachability to larger audience in native language. They faced difficulty as this mix of two languages has inherent variations of spellings and absence of grammar induces considerable amount of ambiguity to text and makes the problem even harder to disambiguate and understand the true meaning of text. They proposed the multi-input multichannel transfer learning (MIMCT)-based model is used to identify and detect the hate speeches and offensive language in Hinglish tweets. They use the dataset proposed by them and named it as Hinglish Offensive Tweet (HOT) dataset. Their proposed learning model uses multiple feature inputs using transfer learning. They employed word embedding with secondary extracted features as input to train their multichannel CNN LSTM which is pretrained on English tweets.

Table 4 shows the distribution of tweets among different classes in HOT dataset.

Table 4 Hinglish offensive tweet (HOT) dataset
Full size table

Pitsilis et al. [23] address the effectiveness of identifying the class (being offensive or not offensive) of new tweet or post, using the identity and history of user who has posted the tweet and other tweets posted by him or by other user related to him. They use LSTM for classification and classify the tweets into three classes, namely, neutral, racism, and sexism. The dataset they used is proposed by Waseem et al. [24] and contains about 16,000 short messages collected across Twitter (Table 5).

Table 5 Waseem Twitter dataset
Full size table

The biggest issue with this dataset is of dual labeled tweets in the dataset. The number of these tweets is not that small that they can ignore them. Being more precisely, there are 42 tweets that are annotated as both “Neutral” and “Sexism”, while 06 tweets were classified as “Racism” and “Neutral” both. According to the dataset providers, the labeling was performed manually.

Wiedemann et al. [25] explore different techniques for automatic detection of offensive text or hate speech on Tweets written in German language. They also employ deep learning for this task and use a series BiLSTM and CNN neural network in sequence. They improve the accuracy of three learning transfer task for improving the classification performance using context and historical data. They compare supervised categories such as near offensive to weakly supervised categories that contain emojis, and they also show comparison to unsupervised category using tweets of same topic by clustering them using latent Dirichlet allocation (LDA).

4 Conclusion

In this paper, we try to present the work done recently in this field of automatic detection of offensive language. We show that how research goes from using tf-idf to popular classifiers such as naïve Bayes, support vector machine (SVM), logistic regression, and then research work goes to variant of these classifiers such as linear SVM, logistic regression with L2, and from here researchers further explore ensemble classifiers using the combination of these classifiers by decomposing the task into subtasks, and then lastly the usage of deep learning and we found many researchers using approaches such as LSTM, CNN, and RNN. Each of these techniques has their own advantages and for classification accuracy LSTM models have outperformed others.