In this paper, we attempt to track the dissemination of hate speech on Twitter. We argue that hate is not a blanket category but exists across multiple topics. We use topic modelling to unearth the latent topics in tweets and an ensemble classification model to capture various nuances of hate speech. We further validate our approach by manually annotating 4, 720 tweets. On analysing the mechanisms of hate speech dissemination, we find that hateful tweets garner 2.4 times more retweets than non-hateful tweets. Further, the retweet network allows us to define topical inflow and outflow vectors which we use to classify users as originators, propagators, and constrictors. We then examine how these users take part in the dissemination of topical information and observe considerable diferences in how users associate with hateful and non-hateful topics. Furthermore, the retweet network enables us to analyse the structure of the strongest connected components for diferent topics. We find some topics associate strongly with hate and users associated with these topics have high degree centrality, are densely connected, and have a large strongly connected core.
Hate Speech has taken a catalysing form in moulding the opinions of people on Online Social
Networks (OSN). It has also incited real life violence with its spread like wildfire among users
[
Prior works treat hate as a blanket topic for modelling spread. However, there are various topics in the universe of Twitter such as religion, politics, sports, among others and some of these topics are associated with the dissemination of a larger proportion of hateful content as compared to others . Hence, we model the spread of hate under the purview of the dissemination of topical information.
The characterization of users as hateful or non-hateful is prone to errors due to the subjective
and contextual nature of hate [
We are neither aiming to outperform the state of the art, nor do we wish to solely present a state-of-the-art machine learning approach for hate speech detection. Instead, we are modifying existing approaches as a base for downstream our research tasks. We have shown the eficacy of analysing hate and networks of hateful users through a topical lens on online social media platforms. Further, due to the large size of the dataset, we have utilised simple machine learning models to improve the classification speed, while also achieving decent performance on the classification task.
To summarise, we make the following contributions in this paper, a) propose an ensemble based learning approach for hate detection, b) manually annotate 4, 720 tweets as hateful or not, c) analyse hateful content present in diferent topics, d) propose a method to compute propagators, originators and constrictors for a topic and e) empirically demonstrate on a very large Twitter dataset that users associating with hateful topics are densely connected, form a dense strongly connected core, and are more central in the retweet network.
We present a detailed description of the proposed approach in this section. As the dataset we use does not have labels for tweets, we first explain how we build an ensemble classifier for hate detection which is used to label the data. We then explain how we model spread of hate along multiple topics.
Ensemble based classification models have performed better than individual classification
models [
Two classification models, A and B, are trained independently on two separate datasets. We
take the mean of the predictions as the final classification score. Both models follow Wang
et. al.[
We model the spread of hate by applying the following steps: a) topic modelling to detect latent topics, b) associating topics with hate, and c) classifying users based on their contribution to topics.
Topic Modelling: We create a corpus of documents for users by combining all the tweets of each user into one document. We then train a topic model on this corpus using Latent Dirichlet Allocation (LDA), obtaining a topic vector () (1×||) for each user and set of topics .
Associating Topics with Hate: Each topic has diferent proportions of hateful content. We quantify the association of a topic with hate using the following steps: 1. For each user, classify their tweets as hateful or non-hateful and obtain () as the fraction of tweets that are hateful. This represents the association of a user with hate. 2. Obtain the topic vector () (1×||) for each user using the topic model, 3. For each topic ∈ , obtain the set of users , having fractional interest ()( ) greater than a threshold , 4. Compute ℎ by a user . Thus, we use the outflow topic vector as the topic vector for the user. Similarly, the topic vector () (1×||) represents the distribution of topics in the content generated Defining Types of Influential Users:
role in the dissemination of a topic, namely, originators, propagators and constrictors using the approach mentioned in Algorithm 1.
Users associate diferently with various topics. Thus, we adjust the and values, which are defined as the maximum and minimum value of the diference between the outflow 1. From the retweet-induced graph, select the set of users who have been retweeted by 2. () (1×||) = 1/| |(
∑ () (1×||) ) where is a user in the set, | | is the size of the set, and || a user.
The inflow topic vector is defined as follows: =1 denotes the number of topics. and the inflow vector, respectively. The parameter represents the net topical outflow for a user.
We define originators as users whose value is higher than at least half (50%) of the maximum net topical outflow across all the users ( ). For a particular topic and user , as the sum || of the outflow vector ∑ ()( ) = 1, a high value of implies a high value of the fractional =1 interest ()( ). This suggests that the user’s content mostly contains information discussing the particular topic.
Similarly, propagators are users whose value is less than those of originators but greater than at least 10% of . The lower limit of 10% is empirically set to account for neutral users whose value is quite low.
Further, we define constrictors as users who restrict the flow of topical information. Thus, the value for these users is always negative and lower than a threshold of 10% of the minimum net topical outflow across all the users ( ). This threshold serves the same purpose as explained above.
As per our observations, the values of were always between 0 and +1 and the values of were always between 0 and −1.
Dataset: We use the dataset provided by [
Out of the 100,386 users, labels (hateful or normal) are available for 4, 972 users, out of which 544 users are labelled as hateful and the rest as normal. This dataset does not have labels for the tweet content.
Though the original dataset does not have labels for the tweet content, we use it for two reasons, a) it is the largest collection of users marked as hateful and b) it gives us an opportunity to examine the retweet-induced graph structure.
Preprocessing: We first convert tweet text to lower case and then remove the stop words. We also remove numbers, punctuation, special characters and emoticons, HTTP links, user mentions such as @{user} and RT{user}, trailing spaces, and condense instances of trailing letters to a single occurrence, such as lmaooo to lmao and lolll to lol.
Annotation: As the dataset we use does not have labels for the tweets, we manually annotate the tweets as hateful or not to create a validation set for our ensemble classification model. Annotating 19 tweets is very expensive and time consuming process. Hence, we annotate a subset of tweets. We pick the 10 original tweets produced by each of the 544 users labelled as hateful in the dataset for two main reasons. Firstly, annotating the tweets of hateful users ensures that we get an adequate number of hateful tweets. Secondly, this annotated set serves as a validation ground truth for downstream tasks in our research. After preprocessing the tweets, we were left with 4, 720 original tweets.
The tweets were annotated by a group of 4 independent annotators whose primary language is English. The Inter-Annotator Agreement (IAA) score (Cohen ) was 0.87. A high score was obtained due to a singular class - hateful or not.
We use an ensemble based classification model to label the dataset we use for our analysis. We use two separate models, A and B for our ensemble model. Model A learns from a dataset of toxic comments, while model B learns a dataset of hate speech against immigrants and women. Both the models consist of a logistic regression classifier with Naive Bayes log count ratio as features. We implement the models using the scikit-learn1 library. Further, we split both the datasets into a 80:20 train-test split. We empirically tune the parameters for the model to maximize the F1 score.
Model A is borrowed from a kernel from Kaggle posted for the competition ”Toxic Comment Classification Challenge” 2. The dataset contains 160 labelled (hateful or non-hateful) comments. We add tf-idf with unigrams and bigrams, set minimum document frequency to 33 and maximum document frequency to 0.94. The model achieves an accuracy of 95% with a F1 score of 0.8 on the test data. This model achieves an accuracy of 72% with a F1 score of 0.69 on our annotated dataset (validation set).
Model B follows the same principal as model A. However, we modify the feature space by not specifying the minimum and maximum document frequency. Model B is trained on a dataset from a Semeval challenge called ”Multilingual detection of hate speech against immigrants and women in Twitter (hatEval)”5. This dataset consists of 9, 000 labelled tweets. The model achieves accuracy of 80% with a F1 score of 0.7 on the test data. This model achieves an accuracy of 78% with a F1 score of 0.73 on our annotated dataset (validation set).
The final predictions are performed separately using both the models and the mean confidence score is used to predict a class (hateful or non-hateful). The ensemble model achieves an accuracy of 82% with a F1 score of 0.75 on the annotated dataset (validation set).
We then proceed to classify our whole corpus of 19 tweets using our ensemble model. From the classification results, we infer that out of 1.6 quotes, 12.4% are hateful, out of 7.23 retweets, 8.36% are hateful and out of 10.7
original tweets, 7% are hateful. This suggests that quoting and retweeting are popular mechanisms of propagating hateful content. On comparing the influence of tweets we find that, on an average, hateful tweets garner
than non-hateful tweets.
Topic Modelling: We analyse the temporal spread of tweets to ascertain if they represent the user’s interests adequately. We find that 80% users have a temporal spread greater than a week, leading us to the conclusion that for most users, the interest in topics is not transitory in nature.
We use MALLET6 to train a LDA based topic model on the corpus ( documents for the users). We vary the number of topics || from 10 to 80 as 10, 20, 40, 80. We empirically tune the topic density parameter ( ) and the number of topics to maximize the coherence score and minimize the overlap between topics. We chose the number of topics to be 40 and to 0.01. A smaller value of leads to more concentrated topical distributions. We obtain topic vectors () for each user , represented as a fraction of interest in each of the ∈ ∀ ∈ || topics. The topic vector matrix is a × ||
matrix, where is the number of users and || is the number of topics.
Associating hate with each topic: We calculate the ℎ ) as described in Section 2 for each ∈ . We set the threshold to 0.5, which implies that the fractional interest ()( ) of a user for a particular topic is more than 50%. To validate our approach, we first find the top 10 topics discussed by the 544 labelled ground truth hateful users. For this, we find the topic that has the max ()( ) ∀ ∈ || , for each user in the set. The top 10 topics are ascertained by 2https://www.kaggle.com/jhoward/nb-svm-strong-linear-baseline 3This implies removing the terms which appear in less than 3 documents 4This implies removing the terms which appear in more than 90% of the documents 5https://competitions.codalab.org/competitions/19935 sorting the number of occurrences of these in a descending order. Finally, we analyse the overlap of these top 10 topics with the top 10 hateful topics calculated based on descending ℎ _ ( ) values.
We find that 7 out of the 10 topics are common and the top 5 topics are same for both the approaches. Hence, we use the top 5 topics as hateful topics for further analysis. Further, we ifnd that there are considerable diferences in the top words for the top 5 hateful topics and the top 5 least hateful topics as shown in Figure 1.
Inflow and Outflow Results: Figure 2 shows the average inflow and outflow distributions for the top 5 hateful and non-hateful topics. We observe that, at lower values of contribution, there are more users that associate with non-hateful topics. However, as the values of contribution increase, very few users associate with non-hateful topics, whereas, a considerably higher number of people associate with hateful topics.
Node Influence Metrics: Figure 3 shows the distribution of originators, propagators and constrictors for each topic. We observe that there are a considerably higher number of propagators for hateful topics as compared to non-hateful topics.
Static Network Structure Related Properties: For each topic ∈ ∀ ∈ || , we create a topic graph in following manner: • Obtain the set of users , having fractional interest ()( ) greater than a threshold = 0.5 (we set this value empirically to include only those users who associate strongly with a particular topic) • For each of these users , obtain their ego network, containing edges ( , ) and ( , ) from retweet-induced graph = ( , ) • Take the union of the ego networks to create a topic graph
We analyse centrality metrics, average path lengths, and the structure of strongly connected
components (SCCs) for the topic graphs. From Figure 4, we observe that users associated with
hateful topics have 2.67 times the average degree centrality that non-hateful topics. Moreover,
we find that non-hateful topics have 1.33 times the average path lengths of hateful topics. This
indicates a more central position in the network and supplements the results of the study
conducted by Ribeiro et. al. [
From Figure 5, we observe that the SCC of a hateful topic graph has a strong core while it is relatively sparse for the non-hateful topic. Hateful topics also have more users in the largest SCC; the mean number of users for the hateful topics is 11, 875 compared to 4, 050 for non-hateful topics. Moreover, we observe that hateful topics correlate with denser networks; the mean average degree of the SCC of the hateful topics is 35.28 as compared 8.54 for non-hateful topics.
We present a distinctive approach to study the dissemination of hate speech on Twitter. We combine a topic model with an ensemble based learning approach to detect hate speech in tweets. This allows us to capture the diverse nuances of hateful content. Our results indicate that propagation of hateful content can be efectively studied through a topic-based analysis of tweets. On analysing the roles of diferent types of users, we observe that there are a significantly higher number of propagators in topics that associate with hate. Further, examining the strongly connected components of the topical subgraphs provides insights into the community structure of hateful and non-hateful topics; observing that hateful topics have a denser and larger core, and hence we assume that information has a more lucid flow as compared to non hateful topics. This study, however, lacks in a temporal analysis of hate spread along multiple topics present. In the future, we plan to study the temporal spread of multiple forms of hate. Using the methodology and findings of this study, further analysis and experiments can reproach hateful content on OSNs rigorously.