Emotion Based Hate Speech Detection using Multimodal Learning

Around the world, we are seeing a disturbing groundswell of racism and intolerance — including but not limited to rising antisemitism, anti-Muslim hatred, and hate crimes against Asians. Social media and other forms of communication are being exploited as platforms for bigotry. Hate Speech is defined as any communication in speech, writing or behavior, that attacks or uses pejorative or discriminatory language with reference to a person or a group based on who they are; in other words, based on their religion, ethnicity, nationality, race, color, descent, gender or other identity factors. It can lead to incitement, which explicitly and deliberately aims at triggering discrimination, hostility, and violence¹¹1https://www.un.org/en/genocideprevention/documents/UN%20Strategy%20and%20Plan%20of%20Action%20on%20Hate%20Speech%2018%20June%20SYNOPSIS.pdf. Therefore, real-time detection and filtering of all forms of hate speech is paramount.

Hate speech can be present in various formats like text, audio, images, and video on social media. Several NLP methods have been experimented with for hate speech detection in languages, such as neural networks [1, 2, 3], n-grams [4, 5], and graph based models [6, 7]. However, there is a lack of substantial research on multimedia data. Additionally, it should be noted that hate speech in multimedia data is not simply dependent on the text but the emotional effect that the tone and delivery of speech has on the end listener. According to Patrick et al. [8], hate speech and offensive behavior are linked to the emotional and psychological state of the speaker, which is reflected in the effective emotions of their language [9]. For example, a political leader calmly speaking about immigration policies at a conference is less harmful than delivering the same speech with extreme anger and disgust towards the targeted immigrants, as the latter causes incitement against the immigrants in the country. Accounting for emotion in detecting hate speech also reduces the number of false positives detected by the systems that only consider text data as input, as emotion provides more context to the speaker’s intent.

Therefore, we propose a method to classify hate speech using a multimodal deep learning architecture that combines semantic and emotion features extracted from the speech. We manually collected a hate speech detection video dataset (HSDVD) as no such dataset exists today, to the best of our knowledge. Due to the limited size of the data, transfer learning [10] was employed to pre-train the models responsible for capturing the unimodal language and speech embedding. To summarize, there are three machine learning models. The first one detects hate speech in text and was built by pre-trainning transformer networks such as BERT and ALBERT on the existing Twitter data sets. The second one, or the emotion detection multi-task learning (MTL) model, is trained on the IEMOCAP dataset to detect the level of valence, arousal, and dominance in the audio. Due to the challenge of discrete representation of complex emotions [11] such as anger or fear, we use a dimensional representation of emotions defined by valence, arousal, and dominance attributes [12].

Lastly, for multimodal learning (MML), a multilayer perceptron model is trained on HSDVD to detect hate speech. It selects the best performing text and emotion models to generate embeddings as input. In addition, the text based models are fine-tuned and evaluated on HSDVD to create a benchmark for comparison with the MML framework. Both are tested on a holdout dataset from HSDVD. We experimented with two baselines, and the MML framework exemplified a gain in precision and recall across its respective baselines. This confirms our hypothesis on the significance of multiple modalities for the hate speech detection task.

Research on multimedia hate speech content has begun to emerge in the last couple of years, yet the prior work is limited. The hateful memes challenge by Facebook drew significant attention from researchers in 2020²²2https://ai.facebook.com/blog/hateful-memes-challenge-winners/. The top 3 winning teams used pre-trained multimodal transformer models to combine the visual features of the image with textual features of the caption with considerable success [13, 14, 15].

A much recent effort was also made in the field of video hate speech detection [16]. However, it only focuses on the text aspect of the video discarding any additional features which multimedia data provides. Another research focusing on offensive video detection collected and published a dataset in Portuguese [17]. It used transcripts along with social media features like tags and titles to detect offensiveness. Nevertheless, such models are dependent on features available after the content has reached a wider audience, rendering damage to the targeted group in the process. Therefore, a need for a more sophisticated method of hate speech detection in multimedia data arises.

To perform feature extraction on text, we train a hate speech detection language model. The NLP techniques for hate speech detection have evolved through several stages in the past decade. Early approaches experimented with TFIDF [5], bag-of-words (BOW) or n-gram [4], user-specific features such as age or social media features like shares, retweets or reporting for author profiling [6, 18]. Since the maturity of deep learning approaches, much of the research in recent years has focused on neural architectures. Badjatiya et al. [19] and Gamback et al. [20] were the first to use recurrent neural networks (RNNs) and convolution neural networks (CNNs), respectively, for hate speech detection in tweets. Current state-of-the-art models fine-tune pre-trained transformers like BERT and ALBERT. Seven of the top ten teams in offensive language detection tasks (2019) used BERT with some variation in parameters and pre-processing [21, 22]. In 2020 again [23], the top ten teams used combinations of BERT, RoBERTa, or XML-RoBERTa, and the winning team used ALBERT [24].

In Speech Emotion Recognition (SER), there has been extensive research, as it has a wide variety of applications such as Human-Computer Interaction [25], Sentiment Analysis, and Enhancing film sound design [26]. Researches can be divided into two categories based on whether they classify the speech into an emotional state (happiness, sadness, anger, fear, disgust, boredom) [27] or predict the emotional attributes, that are, valence, arousal, and dominance [28]. Zisad et al. [29] has built a CNN model on the Ryerson Audio-Visual Database of Emotional Speech and Song (RAVDESS) data merged with some locally generated problem specific data collected by them to classify the speech emotion as calm, angry, fearful, disgust, happy, surprise, neutral or sad. The research uses tonal properties like MFCCs as features for the model. Zhang et al. [30] proposed that an attention-based CNN model trained on the speech spectrogram as the input instead of acoustic or statistical features can give better results. It is to be noted that the research also focuses on archetype emotion classification. On similar notes, Weiser et al. [31] compares an end-to-end learning network trained on raw audio data with a feature based network. Bojanić et al. [32], in their call center based SER system focus on both archetype emotions as well as the emotional attributes of speech. Parthasarathy and Busso [33], in their research, claim that the emotional attributes are interrelated and hence predicting them with a unified learning framework will give us better results. They make use of Multi-task learning along with Deep Neural Network models.

In this transfer learning approach, the target domain’s feature space must overlap with the source domain to ensure a positive transfer [10]. Additionally, the consistency of training data distribution in all the models is critical to the success of transfer learning [34]. Hence, we experiment with multiple hate speech detection datasets for the text models and an acoustic dataset consisting of a varying range of emotional attributes for the emotion model.

The growing interest in automated hate speech detection has led to the creation of a plethora of text datasets from sources like Yahoo, Twitter, Wikipedia comments, and Reddit [35]. Twitter datasets were selected for this task, as the microblogging platform is closest to the target multimedia domain consisting of video/audio blogs. Each dataset has its advantages but also suffers from unintended biases [36]. Therefore, multiple combinations of datasets were experimented with, and the best model was chosen for feature extraction. All the datasets were subjected to the same pre-processing steps, i.e., removing Twitter user handles, URLs, and special characters (except exclamation and question marks).

OffensEval 2019: This dataset by Zampieri et al. [37] has been used extensively in SemEval 2019 Task 6 [21] and SemEval 2020 Task 12 [23] with considerable success. The data was collected from Twitter with the help of keywords and was annotated using crowdsourcing. It consists of 13k tweets and three levels of annotation. We focus only on level A annotations classifying it as offensive (33%) or not and level B annotations classifying the targeted offense towards an individual (17.8%), a group (8.2%), or otherwise. Posts containing profane language and targeted offense, including insults, threats, or swear words, were classified as offensive.

Waseem and Hovy 2016 (W&H): A dataset of 16k TweetIDs was published by Waseem and Hovy [38]. The authors annotated the tweets into three classes, namely racism (11.7%), sexism (20.0%), or neither. An additional third-party review recorded an inter-annotator agreement of 0.84. Only 10k tweets were retrieved, as the remaining had been taken down since 2016. The class distribution of the retrieved dataset is 9.2% racism, 17.5% sexism, and 73.3% neither. It was noted by Madukwe et al. [35] that the dataset might be biased towards specific users since all the racist tweets were collected from only nine users. Mishra et al. [39] also called attention to specific tweets that lacked explicit abusive traits but were annotated as racist or sexist regardless.

Davidson 2017: Davidson et al. [40] published a dataset of 24k tweets, which were labeled as hate speech (5.77%), offensive (77.43%), or neither. Tweets were collected using a lexicon compiled by hatebase.org³³3https://hatebase.org/ containing hateful words and phrases. It was further annotated by CrowdFlower workers (now known as Appen⁴⁴4https://appen.com/), and an inter-annotator agreement of 0.92 was reported. Madukwe et al. [35] noted inconsistencies in the labels and the lack of a diverse group of annotations in this dataset.

A combination of the aforementioned datasets was used to create three subsets of data for the text model experiments, as shown in Table I. OffensEval being the most reliable data source was used in dataset A. For dataset B, W&H was combined with a subset of the Davidson dataset. To avoid class imbalance, Davidson was sub-sampled to retain only 3.5k offensive tweets. The racist tweets from the W&H dataset were dropped in dataset C to further reduce noisy data.

IEMOCAP Dataset 2007: The interactive emotional dyadic motion capture database, collected by the Speech Analysis and Interpretation Laboratory (SAIL) at the University of Southern California, is an audio-visual database [41]. A total of ten actors were recorded while performing a given script, but to maintain some originality in the emotions they were also asked to improvise certain hypothetical scenarios. It has approximately 12 hours of recordings. In total, the corpus contained 10039 samples (scripted session: 5255 samples; spontaneous sessions: 4784 samples) of an average duration of 4.5 seconds. Annotators were asked to evaluate the corpus in terms of the attributes valence (1-negative, 5-positive), arousal (1-calm, 5-excited), and dominance (1-weak, 5-strong). Two different annotators were asked to evaluate a single sample, and then speaker dependent z-normalization was used to compensate for inter-evaluator variation.

Hate Speech Detection Video Dataset (HSDVD): Multimedia data consists of various types of media like text, images, audio, video, and animation. No existing hate speech dataset currently incorporates video and audio data, which accounts for a significant amount of content produced on social media platforms today. For this purpose, we compiled the Hate Speech Detection Video Dataset (HSDVD). It was collected from Twitter and YouTube using Twitter API and PyTube, respectively. Hate speech comprises abuses targeting a broad category of ethnicity, gender, or other identities. Since HSDVD will be used only to compare relative performances and due to the availability of limited resources, we collected data concerning specific groups so that each group has sufficient training data for the model. We began by creating a lexicon of sexist and ethnic slurs⁵⁵5https://en.wikipedia.org/wiki/List_of_ethnic_slurs_by_ethnicity focusing on these groups of people. Tweets with video posts were searched using the lexicon and Twitter API. On the other hand, YouTube videos primarily consist of viral social media posts when original copies were taken down. Therefore, we searched by combining the lexicon with phrases like went viral or racist rant. Finally, 1k records of Tweets IDs and YouTube links were collected and classified as hate speech (25%) or not. Both authors labeled each record upon pre-deciding on the definition of hate speech and the target groups. We computed the Cohen’s kappa coefficient of 0.92, indicating a high agreement.

The collected dataset focuses on hate speech targeted towards a gender, sexual orientation, autistic minority, Muslim, Jews, Sikh, Latino, Native Americans, and Asians. Although various definitions of hate speech exist, the common elements we considered are [42] :

• •

  use of sexist, ethnic, or racial slurs intended to incite violence or hate against a target group

• •

  intended to threaten or harm a target group

• •

  abusive or offensive words being used to attack a target group

• •

  humor being used to attack a target group

• •

  intended to be derogatory or humiliating to a target group

• •

  negative stereotyping of a target group

It is important to note that some terms are offensive when used to hurt a target group but can also be used casually by the same group, such as n**ga. Other terms can be used abusively but also in raps or among friends such as f**k and bi**h. Almost all terms that can be used to cause offense, can also be used for informational purposes to spread awareness. Moreover, some videos can be implicitly offensive without containing any explicit slurs. Such nuances are often difficult to differentiate using machine learning models. However, it could be more easily distinguishable in audio/video where emotions and intent are obvious, unlike text content, hence the model can benefit from both audio and text modalities.

According to Poletto et al. [36], most of the existing hate speech datasets might be prone to unintended annotator bias and topic bias, including the above mentioned text datasets and HSDVD. However, our purpose is to demonstrate the effect of multiple modalities when detecting hate speech in multimedia and compare only the models’ relative performances. Hence, our results will most likely to be independent of biases that may exist in the annotations.

We propose three deep learning models, namely text based hate speech classification, speech based emotion attribute prediction, and finally, a multimodal deep learning model to classify hate speech based on text and emotion.

Modality refers to a way in which something is experienced or expressed. Humans perceive the world around us through multiple modalities such as sound, image, language, smell, and taste. In the context of hate speech, an individual determines something as racist or sexist after combining visual, audio, and textual information from a particular instance. Therefore, a progressive artificial intelligence system should also consider the information in different formats while determining hate speech. In this paper, we focus on language, which captures the semantic information and vocal signal that encodes paraverbal information through the tone, pitch, and pacing of the voice. Although visual features have the potential to provide additional perspectives to hate speech detection, its efficacy was finite when learning over the limited scope of the current dataset. For example, only 15% of the dataset contains distinguishable facial expressions, while in 10% of the videos, religious objects could be linked to hate speech comments. Therefore, an extensive analysis with additional resources and extending HSDVD can be highly beneficial when included in future work. Nevertheless, with the current findings, we hope to build the foundation and observe the advantages of these two modalities compared to one in the multimedia hate speech detection task.

The most critical aspect for the performance of multimodal learning specific to this task is the representation and fusion of information for prediction, which can be achieved together using deep neural networks. We use a joint representation technique that combines unimodal signals into a common representation space. There are other kinds of techniques called coordinated representations that process unimodal signals separately but under certain similarity constraints to project them onto a coordinated space [43]. The joint technique is used for this task, as all the modalities are present during both training and inference steps. Mathematically, it is expressed as:

where the multimodal representation $x_{m}$ is calculated using a function $f$ (a neural network activation function) that combines unimodal representations $x_{1},\;x_{2},\;...,\;x_{n}$.

For unimodal representations of text and audio data, deep neural networks have become increasingly effective in the last decade [44, 45, 46]. We train two separate neural networks to learn text and audio embeddings. Since each successive layer in deep learning captures more abstract features, we use the last layer for final representations [47]. To construct the multimodal representation, the individual embeddings of both modalities are then fed to a common neural layer that projects these modalities into a joint space [48, 49]. Multiple hidden layers can be used for further training from the multimodal representations before a non-linear classification layer is added to make the predictions (as seen in Figure 2).

As neural networks require a large amount of labeled training data, it is common to pre-train embeddings. Since HSDVD contains only 1000 records, we utilize the technique of transfer learning to train the unimodal embeddings. Transfer learning helps in improving the performance of a model on the target domain by transferring the knowledge from a different but related source domain [10]. In this approach of homogeneous transfer learning X_S = X_T, where X is the space of all possible feature vectors in a domain D [50]. As previously mentioned, this type of transfer learning is sensitive to differences in the domain and distribution of data. Therefore, the main objective is to reduce the distribution difference between source and target domains [34]. In the pre-trained text model, both the input feature space and labels extensively overlap with that of the target domain. Whereas for the speech model, the feature space is the same, and the classification objectives are different.

In this paper, we proposed a multimodal learning framework that considers both the tone and the words of the speaker in a video or audio to determine hate speech. The model benefits from the mutually beneficial relationship that exists between these modalities in the real world. Our experiments demonstrated that adding the speech features is more beneficial than just relying on the text features when detecting hate speech in the multimedia domain. This initial evidence suggests that including multiple other modalities and adding knowledge from different perspectives can further enhance the model’s performance. For instance, we expect hate speech detection to also benefit from visual features such as detecting religious attires/objects, violence, or facial expressions. Finally, this also calls attention to the need for a different kind of system for hate speech detection in multimedia data which accounts for a large portion of the internet today and encourages new research avenues for this task.