Investigating cross-lingual training for offensive language detection

Pelicon, Andraž; Shekhar, Ravi; Škrlj, Blaž; Purver, Matthew; Pollak, Senja

doi:10.7717/peerj-cs.559

Cited by 13 publications

(17 citation statements)

References 46 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…Overall, multilingual models trained or finetuned directly on the target languages, i.e. in manyshot settings, are still consistently found to perform best (Aluru et al, 2020;Pelicon et al, 2021). MHC's functional tests are model-agnostic and can be used to evaluate multilingual hate speech detection models trained on any amount of data.…”

Section: Multilingual Hate Speech Detectionmentioning

confidence: 99%

HateCheck: Functional Tests for Hate Speech Detection Models

Röttger¹,

Vidgen²,

Nguyen³

et al. 2021

Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Confer

102

View full text Add to dashboard Cite

Detecting online hate is a difficult task that even state-of-the-art models struggle with. Typically, hate speech detection models are evaluated by measuring their performance on held-out test data using metrics such as accuracy and F1 score. However, this approach makes it difficult to identify specific model weak points. It also risks overestimating generalisable model performance due to increasingly well-evidenced systematic gaps and biases in hate speech datasets. To enable more targeted diagnostic insights, we introduce HATECHECK, a suite of functional tests for hate speech detection models. We specify 29 model functionalities motivated by a review of previous research and a series of interviews with civil society stakeholders. We craft test cases for each functionality and validate their quality through a structured annotation process. To illustrate HATECHECK's utility, we test near-state-of-the-art transformer models as well as two popular commercial models, revealing critical model weaknesses.

show abstract

Section: Multilingual Hate Speech Detectionmentioning

confidence: 99%

HateCheck: Functional Tests for Hate Speech Detection Models

Röttger¹,

Vidgen²,

Nguyen³

et al. 2021

Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Confer

102

View full text Add to dashboard Cite

show abstract

“…Additionally, several works use BERT in zero-shot cross-lingual setups. Pelicon et al (2021) and Nozza (2021) use mBERT, the multilingual version of BERT, without any intermediate steps between source-language training and target-language testing. In this work we use mBERT.…”

Section: Hate Speech Data Scarcity and Cross-lingual Transfermentioning

confidence: 99%

Label modification and bootstrapping for zero-shot cross-lingual hate speech detection

Bigoulaeva

Hangya²,

Gurevych

et al. 2023

Lang Resources & Evaluation

View full text Add to dashboard Cite

The goal of hate speech detection is to filter negative online content aiming at certain groups of people. Due to the easy accessibility and multilinguality of social media platforms, it is crucial to protect everyone which requires building hate speech detection systems for a wide range of languages. However, the available labeled hate speech datasets are limited, making it difficult to build systems for many languages. In this paper we focus on cross-lingual transfer learning to support hate speech detection in low-resource languages, while highlighting label issues across application scenarios, such as inconsistent label sets of corpora or differing hate speech definitions, which hinder the application of such methods. We leverage cross-lingual word embeddings to train our neural network systems on the source language and apply them to the target language, which lacks labeled examples, and show that good performance can be achieved. We then incorporate unlabeled target language data for further model improvements by bootstrapping labels using an ensemble of different model architectures. Furthermore, we investigate the issue of label imbalance in hate speech datasets, since the high ratio of non-hate examples compared to hate examples often leads to low model performance. We test simple data undersampling and oversampling techniques and show their effectiveness.

show abstract

“…A typical form of toxic disinhibition is flaming behavior, which involves using hostile expressions to refer to other users in online communication. Textual features of flaming behavior include harsh language, negative connotations, sexual harassment, and disrespectful expressions ( Pelicon et al, 2021 ). The definition of toxic disinhibition, or toxic behavior, varies based on the users, the communities, and the types of interactions ( Shores et al, 2014 ).…”

Section: Literature Reviewmentioning

confidence: 99%

Investigating toxicity changes of cross-community redditors from 2 billion posts and comments

Almerekhi

Kwak

Jansen

2022

PeerJ Computer Science

View full text Add to dashboard Cite

This research investigates changes in online behavior of users who publish in multiple communities on Reddit by measuring their toxicity at two levels. With the aid of crowdsourcing, we built a labeled dataset of 10,083 Reddit comments, then used the dataset to train and fine-tune a Bidirectional Encoder Representations from Transformers (BERT) neural network model. The model predicted the toxicity levels of 87,376,912 posts from 577,835 users and 2,205,581,786 comments from 890,913 users on Reddit over 16 years, from 2005 to 2020. This study utilized the toxicity levels of user content to identify toxicity changes by the user within the same community, across multiple communities, and over time. As for the toxicity detection performance, the BERT model achieved a 91.27% classification accuracy and an area under the receiver operating characteristic curve (AUC) score of 0.963 and outperformed several baseline machine learning and neural network models. The user behavior toxicity analysis showed that 16.11% of users publish toxic posts, and 13.28% of users publish toxic comments. However, results showed that 30.68% of users publishing posts and 81.67% of users publishing comments exhibit changes in their toxicity across different communities, indicating that users adapt their behavior to the communities’ norms. Furthermore, time series analysis with the Granger causality test of the volume of links and toxicity in user content showed that toxic comments are Granger caused by links in comments.

show abstract

Investigating cross-lingual training for offensive language detection

Cited by 13 publications

References 46 publications

HateCheck: Functional Tests for Hate Speech Detection Models

HateCheck: Functional Tests for Hate Speech Detection Models

Label modification and bootstrapping for zero-shot cross-lingual hate speech detection

Investigating toxicity changes of cross-community redditors from 2 billion posts and comments

Contact Info

Product

Resources

About