PaperlayerView exact revision on arXiv

Source-linked AI summary

DeepHate: Hate Speech Detection via Multi-Faceted Text Representations

Rui Cao, Roy Ka-Wei Lee, Tuan-Anh Hoang

arXiv:2103.11799v1cs.CLcs.SI

TL;DR

Hate speech on social media is linked to community discord and violent hate crimes, while many detection methods use only one type of text feature. DeepHate combines semantic, sentiment, and topical representations; experiments on real-world datasets report that it outperforms state-of-the-art baselines.

  • Problem

    Many existing hate-speech detection methods consider only one type of textual feature, despite the importance of detecting harmful speech on social media.

    Source: §1 · p. 1

  • Method

    DeepHate combines semantic, sentiment, and topical representations of social-media posts for hate-speech classification.

    Source: §1 · pp. 1–2

  • Results

    DeepHate outperformed state-of-the-art baselines in experiments on three publicly available real-world datasets.

    Source: §1 · p. 2, §6 · p. 8

  • Takeaways & Limitations

    Salient-feature analyses provided insights into hate-speech detection and some explanation of DeepHate’s classification decisions.

    Source: §1 · p. 2, §6 · p. 8

    The approach focuses on textual content, excluding user metadata that may be scarce or incomplete and could introduce bias through personal information.

    Source: §2 · p. 2

Abstract

from arXiv · show

Online hate speech is an important issue that breaks the cohesiveness of online social communities and even raises public safety concerns in our societies. Motivated by this rising issue, researchers have developed many traditional machine learning and deep learning methods to detect hate speech in online social platforms automatically. However, most of these methods have only considered single type textual feature, e.g., term frequency, or using word embeddings. Such approaches neglect the other rich textual information that could be utilized to improve hate speech detection. In this paper, we propose DeepHate, a novel deep learning model that combines multi-faceted text representations such as word embeddings, sentiments, and topical information, to detect hate speech in online social platforms. We conduct extensive experiments and evaluate DeepHate on three large publicly available real-world datasets. Our experiment results show that DeepHate outperforms the state-of-the-art baselines on the hate speech detection task. We also perform case studies to provide insights into the salient features that best aid in detecting hate speech in online social platforms.

1 INTRODUCTION

Hate speech spreads through social media, while manual moderation is labor-intensive and difficult to scale. DeepHate combines semantic, sentiment, and topical text representations, and the paper reports stronger benchmark performance and salient-feature analyses.

1 INTRODUCTION

  • Manual platform responses to hate speech—including policies, user reporting, and content moderation—are labor-intensive, time-consuming, and difficult to sustain at scale.The introduction links hate speech on social media to online discord and violent hate crimes.Source: §1 · p. 1
  • Existing methods often use only one type of textual feature, and deep learning methods offer limited explanations for individual hate-speech classifications.The paper identifies both limitations as motivation for its approach.Source: §1 · p. 1
  • DeepHate combines semantic, sentiment, and topical representations to classify social-media posts and analyze salient features.It uses pretrained word embeddings for semantics, a sentiment-analysis tool and neural network for sentiment embeddings, and LDA for topic representations.Source: §1 · pp. 1–2
  • DeepHate consistently outperformed state-of-the-art methods across three publicly available real-world datasets.The paper also reports empirical analyses of salient features that aided detection.Source: §1 · p. 2

2 RELATED WORK

Prior hate-speech detection work includes both traditional machine-learning methods using extracted features and deep-learning methods using learned representations. DeepHate differs by combining multiple textual representations, while this paper excludes user metadata from its detection approach.

2 RELATED WORK

  • Traditional methods use extracted textual features with classifiers, while some studies also use profile or network metadata.Common textual features include TF-IDF, bag-of-words, and linguistic attributes; classifiers include Logistic Regression, SVM, and Random Forest.Source: §2 · p. 2
  • This paper uses textual content alone because user metadata can be scarce or incomplete and may bias detection through personal information.Source: §2 · p. 2
  • Deep-learning studies have used word or character embeddings and architectures including RNNs, CNNs, LSTMs, and hybrids.DeepHate is presented as an alternative that combines multi-faceted text representations.Source: §2 · p. 2

3 PROPOSED MODEL

DeepHate encodes semantic, sentiment, and topical information from posts, combines their latent representations, and predicts hate-speech classes. Its components use pretrained word embeddings, sentiment-specific embeddings, sparse topic distributions, and a C-LSTM-Att encoder.

3 PROPOSED MODEL

  • DeepHate encodes semantic, sentiment, and topic representations, fuses them in a feed-forward network, and uses softmax to predict post classes.The architecture first learns each representation from feature embeddings, then combines them for classification.Source: §3 · p. 2

3.1 Semantic Representation

  • Three pretrained word-embedding representations are assigned attention weights and summed element-wise to form each post’s latent semantic representation.The embeddings are GloVe, word2vec trained on wiki news, and Paragram; their vectors are 300-dimensional.Source: §3.1 · p. 3

3.2 Sentiment Representation

  • Sentiment-specific word embeddings are trained to predict VADER-generated tweet polarity labels, then encoded into a post-level sentiment representation.VADER labels each tweet as negative, positive, or neutral according to its highest polarity score; the sentiment embeddings are 300-dimensional and frozen during hate-speech training.Source: §3.2 · p. 3

3.3 Topic Representation

  • LDA treats each post as a document and uses sparsity regularization to produce a sparse distribution over topics.The choice reflects sparse shared hashtags, posting times, and user histories in the datasets.Source: §3.3 · p. 3

3.4 C-LSTM-Att Encoder

  • C-LSTM-Att adds word attention to a CNN–LSTM encoder, selecting informative word features from the learned hidden states.Convolutional outputs are passed to the LSTM without pooling so their sequence organization is retained; attention uses the final hidden state to weight word states.Source: §3.4 · p. 4

3.5 Representation Fusion

  • Sentiment and topic representations are gated through interactions with the semantic representation, then summed with it to form a joint post representation.The joint representation is used as features for hate-speech multi-class classification in the softmax layer.Source: §3.5 · p. 5

3.6 Implementation Details

  • Implementation uses 300-dimensional embeddings, three CNN filter sizes with 50 filters each, 200 LSTM hidden states, and ADAM at learning rate 0.001.The reported implementation also applies dropout after the embedding and fully connected layers.Source: §3.6 · p. 5

4 DATASET

DeepHate is evaluated on three public datasets and a merged dataset, which vary in size and composition. The section describes each dataset’s labels and the adjustments made for these experiments.

Dataset overview

  • The four evaluation datasets are WZ-LS, DT, FOUNTA, and COMBINED; the authors describe them as diverse in size and nature.The merged dataset was formed by concatenating the three original datasets and removing spam tweets.Source: §4 · p. 5, Table 1 · p. 5

WZ-LS dataset

  • WZ-LS distinguishes racism, sexism, both, and neither, but the recovered tweet set is slightly smaller than the original because some tweets were deleted.The authors retrieved tweet text through Twitter’s APIs because the released data contained only tweet IDs.Source: §4 · p. 5

DT dataset

  • DT labels tweets as offensive, hate, or neither, distinguishing offensive content from hate speech.The dataset’s authors argued that tweets containing hateful words do not necessarily meet the threshold for hate speech.Source: §4 · p. 5

FOUNTA dataset

  • FOUNTA’s final labels are normal, spam, hate, and abusive; retweets were removed for these experiments.The dataset used two annotation rounds, with the second refining labels for tweets initially marked inappropriate.Source: §4 · p. 5

COMBINED dataset

  • The COMBINED dataset aggregates normal tweets with inappropriate tweets, including offensive and hate tweets, from the three datasets.The authors suggest it is closer to real-world use because platforms may want to detect and reduce both hate and offensive tweets.Source: §4 · p. 5

5 EXPERIMENT

DeepHate combines semantic, sentiment, and topical representations to detect hate speech, and the experiments compare it with established baselines and test its components. The section reports that it outperformed state-of-the-art baselines and examines salient features and sentiment-related errors.

5.2 Experiment Results

5.1 Experiment Setup

  • The evaluation used 80/20 train-test splits, five-fold cross-validation, micro-averaged precision, recall, and F1, VADER sentiment labels, and dataset-specific topic counts.Source: §5.1 · pp. 6–7

5.3 Ablation Study

5.4 Salient Feature Analysis

  • Saliency highlights label-descriptive words, including terms distinguishing sexism, racism, and offensive from hateful content.Darker red indicates greater saliency; highlighted keywords also provide a form of explanation for classification decisions.Source: §5.4 · p. 8

5.5 Case Studies

  • Case studies suggest sentiment may help reduce false negatives on non-hate posts, but inaccurate VADER labels can also accompany errors.The authors identify better sentiment classification techniques as a direction for future work.Source: §5.5 · p. 8

6 CONCLUSION AND FUTURE WORK

DeepHate combines multi-faceted text representations for automatic hate-speech detection and outperformed state-of-the-art baselines across three publicly available real-world datasets. The authors also report salient-feature analysis and identify non-textual features and improved sentiment and topic representations as future work.

6 CONCLUSION AND FUTURE WORK

  • DeepHate outperformed state-of-the-art baselines in experiments on three publicly available real-world datasets.The authors also report analyzing the model's salient features.Source: §6 · p. 8
  • Salient-feature analysis provided some explanation of DeepHate's hate-classification decisions.Source: §6 · p. 8
  • Future work will incorporate non-textual features and improve posts' sentiment and topic representations with more advanced techniques.Source: §6 · p. 8

Related papers

Loading 2103.11799v1…