Keywords

1 Introduction

There are approximately 70 million Internet users in Vietnam, and most of them are familiar with Facebook and Youtube. Nearly 70% of total Vietnamese people use Facebook, and spend an average of 2.5 h per day on it [24]. On social network, hate easily appears and spread. According to Kang and Hall [21], hate-speech does not reveal itself as hatred. It is just a mechanism to protect an individual’s identity and realm from the others. Hate leads to the destruction of humanity, isolating people, and debilitating society. Within the development of social network sites, the hate appears on social media as hate-speech comments, hate-speech posts, or messages, and it spreads too fast. The existence of hate speech makes the social networking spaces toxic, threatens social network users, and bewilders the community.

The automated hate speech detection task is categorized as the supervised learning task, specifically closed to the sentiment analysis task[26]. There are several state-of-the-art approaches such as deep learning and transformer models for sentiment analysis. However, to be able to make experiment on hate speech detection task, the datasets, especially large-scale datasets, play an important role. To handle this, we introduce ViHSD - a large-scale dataset used for automatically hate speech detection on Vietnamese social media texts to overcome the hate-speech problem on social networks. Then, we present the annotation process for our dataset and the method to ensure the quality of annotators. Finally, we evaluate our dataset on SOTA models and analyze the obtained empirical results to explore the advantages and disadvantages of the models on the dataset.

The content of the paper is structured as follows. Section 2 takes an overview on current researching for hate-speech detection. Section 3 shows statistical figures about our dataset as well as our annotation procedure. Section 4 presents the classification models applied for our dataset to solve the hate-speech detection problem. Section 5 describes our experiments on the dataset and the analytical results. Finally, Sect. 6 concludes and proposes future works.

2 Related Works

In English as well as other languages, there are many datasets constructed for hate speech detection. We divided them into two categories: flat labels and hierarchical labels. According to Cerri et al. [7], flat labels are treated as no relation between different labels. In contrast, hierarchical labels has a hierarchical structure which one or more labels can have sub-labels, or being grouped in super-labels. Besides, Hatebase [27] and Hurtlex [6] are two abusive words sets used for lexicon-based approaching for the hate speech detection problem.

For flat labels, we introduce two typical and large-scale datasets in English. The first dataset is provided by Waesem and Hovy [30], which contains 17,000 tweets from Twitter and has three labels: racism, sexism, and none. The second dataset is provided by Davidson et al. [12], which contains 25,000 tweets from Twitter and also has three labels including: hate, offensive, and neither. Apart from English, there are other datasets in other languages such as: Arabic [2], and Indonesian [3].

For hierarchical labels datasets, Zampieri et al. [32] provide a multi-labelled dataset for predicting offensive posts on social media in English. This dataset serves two tasks: group-directed attacking and person-directed attacking, in which each task has binary labels. Another similar multi-labelled dataset in English are provided by Basile et al. [5] in SemEval Task 5 (2019). Other multi-labelled datasets in other non-English languages are also constructed and available such as: Portuguese [17], Spanish [16], and Indonesian [19]. Besides, multilingual hate speech corpora are also constructed such as hatEval with English and Spanish [5] and CONAN with English, French, and Italian [9].

The VLSP-HSD dataset provided by Vu et al. [29] is a dataset used for the VLSP 2019 shared task about Hate speech detection on Vietnamese languageFootnote 1. However, the authors did not mention the annotation process and the method for evaluating the quality of the dataset. Besides, on the hate speech detection problem, many state-of-the-art models give optimistic results such as deep learning models [4] and transformer language models [20]. Those models require large-scale annotated datasets, which is a challenge for low-resource languages like Vietnamese. Moreover, current researches about hate speech detection do not focus on analyzing about the sentiment aspect of Vietnamese hate speech language. Those ones are our motivation to create a new dataset called ViHSD for Vietnamese with strict annotation guidelines and evaluation process to measure inter-annotator agreement between annotators.

3 Dataset Creation

3.1 Data Preparation

We collect users’ comments about entertainment, celebrities, social issues, and politics from different Vietnamese Facebook pages and YouTube videos. We select Facebook pages and YouTube channels that have a high-interactive rate, and do not restrict comments. After collected data, we remove the name entities from the comments in order to maintain the anonymity.

3.2 Annotation Guidelines

The ViHSD dataset contains three labels: HATE, OFFENSIVE, and CLEAN. Each annotator assigns one label for each comment in the dataset. In the ViHSD dataset, we have two labels denoting for hate speech comments, and one label denoting for normal comments. The detailed meanings about three labels and examples for each label are described in Table 1.

Practically, many comments in the dataset are written in informal form. Comments often contain abbreviation such as M.n (English: Everyone), mik (English: us) in Comment 1 and Dm (English: f*ck) in Comment 2, and slangs such as chịch (English: f*ck), cái lol (English: p*ssy) in Comment 2. Besides, comments has the figurative meaning instead of explicit meaning. For example, the word: lũ quan ngại (English: dummy pessimists) in Comment 3 is usually used by many Vietnamese Facebook users on social media platform to mention a group of people who always think pessimistically and posting negative contents.

Table 1. Annotation guidelines for annotating Vietnamese comments in the Hate Speech Detection task.
Full size table

3.3 Data Creation Process

Our annotation process contains two main phases as described in Fig. 1. The first one is the training phase, which annotators are given a detailed guidelines, and annotate for a sample of data after reading carefully. Then we compute the inter-annotator agreement by Cohen Kappa index (\(\kappa \)) [10]. If the inter-annotators agreement not good enough, we will re-train the annotator, and re-update the annotation guidelines if necessary. After all annotators are well-trained, we go to annotation phase. Our annotation phase is inspired from the IEEE peer review process of articles [15]. Two annotators annotate the entire dataset. If there are any different labels between two annotators, we let the third annotators annotate those labels. The fourth annotators annotate if all three annotators are disagreed. The final label are defined by Major voting. By this way, we guaranteed that each comment is annotated by one label and the objectivity for each comment. Therefore, the total time spent on annotating is less than four annotators doing with the same time.

Fig. 1.
figure 1

Data annotation process for the ViHSD dataset. According to Eugenio [14], \(\kappa > 0.5\) is acceptable.

Full size image

3.4 Dataset Evalutation and Discussion

We randomly take 202 comments from the dataset and give them to four different annotators, denoted as A1, A2, A3 and A4, for annotating. Table 2 shows the inter-annotator agreement between each pair of annotators. Then, we compute the average inter-annotator agreement. The final inter-annotator agreement for the dataset is \(\kappa =0.52\).

Table 2. The confusion matrix between annotators in a set of 202 comments computed by Cohen Kappa index (\(\kappa \)).
Full size table

The ViHSD dataset was crawled from the social network so they had many abbreviations, informal words, slangs, and figurative meaning. Therefore, it confuses annotators. For example, the Comment 1 contains the phrase: mik, which mean mình (English: I), and the Comment 4 in Table 1 has the profane word Dm (English: m*ther f**ker). Assume that two annotators assign label for Comment 4, and it contains the word Dm written in abbreviation form. The first annotator knew about this word before, thus he/she annotates this comment as hate. The second annotator instead, annotates this comment as clean because he/she do not understand that word. The next example is the phrase lũ quan ngại (English: dummy pessimists) in the Comment 3 in Table 1. Two annotators assign label for Comment 3. The first annotator does not understand the real meaning of that phase, thus he/she marks this comment as clean. In contrast, the second annotator knows what is the real meaning of this word (See Sect. 3.2 for the meaning of that phase) so he/she knows the abusive meaning of Comment 3 and annotates it as hate. Although the guidelines have clearly definition about the CLEAN, OFFENSIVE, and HATE labels, the annotation process is mostly impacted by the knowledge and subjective of annotators. Thus, it is necessary to re-train annotators and improve the guidelines continuously to increase the quality of annotators and the inter-annotator agreement.

3.5 Dataset Overview

Table 3. Several examples extracted from the ViHSD dataset.
Full size table
Fig. 2.
figure 2

The distributions of three labels on the train, dev, and test sets.

Full size image

The ViHSD contains 33,400 comments. Each one was labelled as CLEAN (0), OFFENSIVE (1), and HATE (2). Table 3 displays some examples from the ViHSD dataset. Then we divided our dataset into training (train), development (dev), and test sets, respectively, with proportion: 7-1-2. Figure 2 describes the distribution of data on three labels on those sets. According to Fig. 2, the distribution of data labels on the training, development, and test sets are the same, and the data are skewed to the CLEAN label.

4 Baseline Models

The problem of text classification is defined according to Aggarwal and Zhai [1] as given a set of training texts as training data \(D=\{X_1, X_2, ..., X_n\}\), in which each \(X_i \in D\) has one of the label in the label set \(\{1..k\}\). The training data is used to build the classification model. Then, for unlabeled data coming, the classification model predicts the label for it. In this section, we introduce two approaches for constructing prediction models on the ViHSD dataset.

4.1 Deep Neural Network Models (DNN Models)

Convolutional Neural Network (CNN) uses particular layers called the CONV layer for extracting local features in the image [23]. However, although invented for computer vision, CNN can also be applied to Natural Language Processing (NLP), in which a filter W relevant to a window of h words [22]. Besides, the pre-trained word vectors also influence the performance of the CNN model [22]. Besides, Gated Recurrent Unit (GRU) is a variant of the RNN network. It contains two recurrent networks: encoding sequences of texts into a fixed-length word vector representation, and another for decoding word vector representation back to raw input [8].

We implement the Text-CNN and the GRU models, and evaluate them on the ViHSD dataset with the fasttextFootnote 2 pre-trained word embedding of 157 different languages provided by Grave et al. [18]. This embedding transforms a word into a 300 dimension vector.

4.2 Transformer Models

The transformer model [28] is a deep neural network architecture based entirely on the attention mechanism, replaced the recurrent layers in auto encoder-decoder architectures with special called multi-head self-attention layers. Yang et al. [31] found that the transformer blocks improved the performance of the classification model. In this paper, we implement BERT[13] - the SOTA transformer model with multilingual pre-trainingFootnote 3 such as bert-base-multilingual-uncased (m-BERT uncased) and bert-base-multilingual-cased (m-BERT cased), DistilBERT [25] - a lighter but faster variant of BERT model with multilingual cased pre-trained modelFootnote 4, and XLM-R[11] - a cross-lingual language model with xlm-roberta-base pre-trainedFootnote 5. Those multilingual pre-trained models are trained on various of languages including Vietnamese.

5 Experiment

5.1 Experiment Settings

First of all, we pre-process our dataset as belows: (1) Word-segmentation texts into words by the pyvi toolFootnote 6, (2) Removing stopwordsFootnote 7, (3) Changing all texts into lower casesFootnote 8, and (4) Removing special characters such as hashtags, urls, and mention tags.

Next, we run the Text-CNN model with 50 epochs, batch size equal to 256, sequence length equal to 100, and the dropout ratio is 0.5. Our model uses 2D Convolution Layer with 32 filters and size 2, 3, 5, respectively. Then, we run the GRU model with 50 epochs, sequence length equal to 100, the dropout ratio is 0.5, and the bidirectional GRU layer. We use the Adam optimizer for both Text-CNN and GRU. Finally, we implement transformer models includes BERT, XLM-R, and DistilBERT with the batch size equal to 16 for both training and evaluation, 4 epochs, sequence length equal to 100, and manual seed equal to 4.

5.2 Experiment Results

Table 4 illustrates the results of deep neural models and transformer models on the ViHSD dataset. The results are measured by Accuracy and macro-averaged F1-score. According to Table 4, Text-CNN achieves 86.69% in accuracy and 61.11% in F1-score, which is better than GRU. Transformer models such as BERT, XLM-R, and DistilBERT give better results than deep neural models in F1-score. The BERT with bert-base-multilingual-cased model (m-bert cased) obtained best result in both Accuracy and F1-score with 86.88% and 62.69% respectively on the ViHSD dataset.

Table 4. Empirical results of baseline models on the test set
Full size table

Overall, the performance of transformer models are better than deep neural models, indicating the power of BERT and its variants on text classification task, especially on hate speech detection even if they were trained on various languages. Additionally, there is a large gap between the accuracy score and the F1-score, which caused by the imbalance in the dataset, as described in Sect. 3.

5.3 Error Analysis

Figure 3 shows the confusion matrix of the m-BERT cased model. Most of the offensive comments in the dataset are predicted as clean comments. Besides, Table 5 shows incorrect predictions by the m-BERT cased model. The comments number 1, 2, and 3 had many special words, which are only used on the social network such as: “dell”, “coin card”, and “éo”. These special words make those comments had wrong predicting labels, misclassified from offensive labels to clean labels. Moreover, the comments number 4 and 5 have profane words and are written in abbreviation form and teen codes such as “cc”, “lol”. Specifically, for the fifth comments, in which “3” represents for “father” in English, combined which other bad words such as: “Cc” - profane word and “m” - represents for “you” in English. Generally, the fifth comment has bad meaning (see Table 5 for the English meaning) thus it is the hate speech comment. However, the classification model predicts that comment as an offensive label because it cannot identify the target objects and the profane words written in irregular form.

Fig. 3.
figure 3

Confusion matrix of the m-BERT cased model on the ViHSD dataset

Full size image
Table 5. Wrong prediction samples in the ViHSD dataset
Full size table

In general, the wrong prediction samples and most of the ViHSD dataset comments are written with irregular words, abbreviations, slangs, and teencodes. Therefore, in the pre-processing process, we need to handle those characteristic to enhance the performance of classification models. For abbreviations and slangs, we can build a Vietnamese slangs dictionary for slangs replacement and a Vietnamese abbreviation dictionary to replace the abbreviations in the comments. Besides, for words that are not found in mainstream media, we can try to normalize them to the regular words. For example, words like should be normalized to (what) and (ok). In addition, the emoji icons are also a polarity feature to define whether a comment is negative or positive, which can support for detecting the hate and offensive content in the comments.

6 Conclusion

We constructed a large-scale dataset called the ViHSD dataset for hate speech detection on Vietnamese social media texts. The dataset contains 33,400 comments annotated by humans and achieves 62.69% by the Macro F1-score with the BERT model. We also proposed an annotation process to save time for data annotating.

The current inter-annotator agreement of the ViHSD dataset is just in moderate level. Therefore, our next studies focus on improving the quality of the dataset based on the data annotation process. Besides, the best baseline model on the dataset is 62.69%, and this is a challenge for future researches to improve the performance of classification models for Vietnamese hate-speech detection task. From the error analysis, we found that it is difficult to detect the hate speech on Vietnamese social media texts due to their characteristics. Hence, we will improve the pre-processing technique for social media texts such as lexicon-based approach for teen codes and normalizing acronyms on next studies to increase the performance of this task.