Source-linked AI summary

KUISAIL at SemEval-2020 Task 12: BERT-CNN for Offensive Speech Identification in Social Media

Ali Safaya, Moutasem Abdullatif, Deniz Yuret

arXiv:2007.13184v1cs.CL

TL;DR

Offensive-language detection on multilingual social-media text requires models that can use textual context across languages. The paper combines pretrained BERT models with CNNs, trains ArabicBERT models, and reports top-four rankings for Arabic, Greek, and Turkish. The authors conclude that BERT-CNN outperformed BERT alone and that language-specific pretrained models were advantageous in their experiments.

  • Problem

    The paper addresses offensive-language identification in Twitter data for OffensEval2020 Subtask-A across Arabic, Greek, and Turkish.

  • Method

    The authors combine contextualized representations from pretrained BERT models with CNN layers and train ArabicBERT models because no pretrained Arabic BERT was available.

  • Results

    The BERT-CNN system outperformed BERT alone, while language-specific pretrained models showed an advantage over multilingual models in the experiments.

  • Takeaways & Limitations

    The proposed model achieved top-four rankings across all three studied languages in OffensEval2020.

  • Takeaways & Limitations

    The model relies on the assumption that concatenating BERT’s last four hidden layers encodes more information than using the top layer alone.

Abstract

from arXiv · show

In this paper, we describe our approach to utilize pre-trained BERT models with Convolutional Neural Networks for sub-task A of the Multilingual Offensive Language Identification shared task (OffensEval 2020), which is a part of the SemEval 2020. We show that combining CNN with BERT is better than using BERT on its own, and we emphasize the importance of utilizing pre-trained language models for downstream tasks. Our system, ranked 4th with macro averaged F1-Score of 0.897 in Arabic, 4th with score of 0.843 in Greek, and 3rd with score of 0.814 in Turkish. Additionally, we present ArabicBERT, a set of pre-trained transformer language models for Arabic that we share with the community.

1 Introduction

The paper addresses offensive-language detection on Twitter across Arabic, Greek, and Turkish, using a BERT-CNN approach with minimal text preprocessing. The system ranked among the top four participants for each language.

  • Existing solutions range from manual control to rule-based filtering, but they can be time-consuming or error-prone when full textual context is needed.
  • OffensEval2020 Subtask-A focused on detecting offensive language in Twitter data across five languages, of which the authors studied Arabic, Greek, and Turkish.
  • The proposed approach combines pre-trained bidirectional BERT representations with CNNs for text classification and uses minimal text preprocessing.
  • The system ranked 4th in Arabic, 4th in Greek, and 3rd in Turkish among more than 40 participants.

2 Background

Prior offensive-language identification research has used lexical, linguistic, metadata, machine-learning, and deep-neural approaches. The paper builds on datasets and models including OLID, SVM, BiLSTM, and CNN, while also releasing source code.

  • Offensive-language identification has employed lexical resources, linguistic features, metadata, and machine-learning models.
  • The authors provide source code for the main model and additional experiments.
  • Deep-neural approaches include CNNs, LSTMs, and their derivatives for offensive-language classification.
  • OLID is a tweet dataset annotated for offensive content and was evaluated with models such as SVM, BiLSTM, and CNN.

3 Data

The task data consists of tweets labeled as offensive or non-offensive, with the training data split into training and development subsets. Preprocessing converts hashtags, applies Greek-specific normalization, and tokenizes text with language-specific BERT WordPiece tokenizers.

  • The task dataset contains tweets annotated as Offensive or Non-offensive, with positive and negative tweet counts reported for each set.
  • Because the provided training data lacked a development split, the authors allocated 90% for training and 10% for development.
  • Hashtags were converted into raw words by splitting hashtag strings into separate tokens.
  • For Greek text, preprocessing lowercased letters and removed Greek diacritics before tokenization with the corresponding language-specific BERT WordPiece tokenizer.

4 Model Description

The model feeds contextualized representations from BERT into a CNN classifier, using language-specific or multilingual pretrained models and ArabicBERT trained for this task. Its convolutional outputs are pooled and classified into binary labels.

  • 4 Model Description: The proposed model feeds contextualized embeddings from BERT’s last four hidden layers into convolutional filters, then uses a dense layer for prediction.
  • 4 Model Description: The experiments use GreekBERT, BERTurk, ArabicBERT, and multilingual BERT as pretrained language models.
  • 4.3 ArabicBERT: Because no pretrained Arabic BERT model was available, the authors trained four Arabic BERT models from scratch and released them publicly.
  • 4.3 ArabicBERT: ArabicBERT models of Large, Base, Medium, and Mini sizes were trained with masked language modeling and whole-word masking for 4M steps.
  • 4 Model Description: BERT provides contextualized vector representations, while CNN serves as the classifier in the two-part architecture.
  • 4.4 BERT-CNN Model Structure: For 64-token tweets, the last four BERT layers were concatenated into 768x4x64 representations and processed by 160 filters across five kernel sizes.
  • 4.4 BERT-CNN Model Structure: Global max-pooling outputs were concatenated, flattened, passed through a dense layer and Sigmoid function, and converted into binary labels.
  • 4.4 BERT-CNN Model Structure: The model was trained for 10 epochs with learning rate 2e-5, retaining the version with the best development-set macro-averaged F1-Score.

5 Experiments and Results

Experiments evaluated the proposed system and comparison models using macro-averaged F1 on shared data splits. Results indicate that combining BERT with CNN and using language-specific pre-trained models improved average performance.

  • Evaluation setup: Macro-averaged F1-Score was used to evaluate submissions and comparison experiments on test data.All models used the same train/dev/test splits.
  • Comparison models: The baseline used TF-IDF with an SVM and a 3000-feature Count Vectorizer.
  • Comparison models: CNN-Text used randomly initialized 300-dimensional embeddings instead of pre-trained BERT, and the difference was significant.
  • Comparison models: Two-layer BiLSTM models with hidden size 128 and randomly initialized 300-dimensional embeddings were also evaluated, but CNN-Text outperformed them on average.
  • Results: Combining BERT with CNN improved average results over BERT alone, while language-specific pre-trained models outperformed multilingual ones.

6 Conclusion

The paper describes BERT-CNN for offensive speech identification and reports that combining BERT with CNN outperformed BERT alone. It also explains ArabicBERT pre-training and reports high rankings across OffensEval2020 languages.

  • Conclusion: The paper describes and compares BERT-CNN with other models for identifying offensive speech in social media.
  • Conclusion: Combining BERT with CNN yielded better results than using BERT alone.
  • Conclusion: The paper explains ArabicBERT pre-training and reports strong average results with minimal text preprocessing.
  • Conclusion: The team ranked among the four highest participating teams for every language in OffensEval2020.
Loading 2007.13184v1…