Implicitly Abusive Comparisons – A New Dataset and Linguistic Analysis

Wiegand, Michael; Geulig, Maja; Ruppenhofer, Josef

doi:10.18653/v1/2021.eacl-main.27

Cited by 12 publications

(23 citation statements)

References 31 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…This observation is in line with findings by Fortuna et al (2021), and suggests that generalization should be evaluated on implicit and explicit abuse separately. However, due to complexities of annotation of abusive content, curating separate implicit and explicit test sets is too costly (Wiegand et al, 2021). Instead, we propose to adapt the Testing Concept Activation Vector (TCAV) algorithm, originally developed for image classification (Kim et al, 2018), to calculate the classifiers' sensitivity to explicit and implicit COVID-related racism, using only a small set of examples.…”

Section: Sensitivity To Implicit and Explicit Abuse To Explain Genera...mentioning

confidence: 99%

See 1 more Smart Citation

Improving Generalizability in Implicitly Abusive Language Detection with Concept Activation Vectors

Nejadgholi¹,

Fraser²,

Kiritchenko³

2022

Preprint

View full text Add to dashboard Cite

Robustness of machine learning models on ever-changing real-world data is critical, especially for applications affecting human wellbeing such as content moderation. New kinds of abusive language continually emerge in online discussions in response to current events (e.g., COVID-19), and the deployed abuse detection systems should be updated regularly to remain accurate. In this paper, we show that general abusive language classifiers tend to be fairly reliable in detecting out-of-domain explicitly abusive utterances but fail to detect new types of more subtle, implicit abuse. Next, we propose an interpretability technique, based on the Testing Concept Activation Vector (TCAV) method from computer vision, to quantify the sensitivity of a trained model to the human-defined concepts of explicit and implicit abusive language, and use that to explain the generalizability of the model on new data, in this case, COVID-related anti-Asian hate speech. Extending this technique, we introduce a novel metric, Degree of Explicitness, for a single instance and show that the new metric is beneficial in suggesting out-ofdomain unlabeled examples to effectively enrich the training data with informative, implicitly abusive texts.

show abstract

Section: Sensitivity To Implicit and Explicit Abuse To Explain Genera...mentioning

confidence: 99%

“…An important factor in this study is whether the text expresses explicit or implicit abuse (Waseem et al, 2017;Caselli et al, 2020;Wiegand et al, 2021). Explicit abuse refers to utterances that include direct insults or strong rudeness, often involving profanities, whereas implicit abuse involves more indirect and nuanced language.…”

Section: Introductionmentioning

confidence: 99%

Improving Generalizability in Implicitly Abusive Language Detection with Concept Activation Vectors

Nejadgholi¹,

Fraser²,

Kiritchenko³

2022

Preprint

View full text Add to dashboard Cite

show abstract

“…Past work has partitioned offensive comments into explicitly offensive (those that include profanity-swear words, taboo words, or hate terms) and implicitly offensive (those that do not include profanity) (Waseem et al, 2017;Caselli et al, 2020a;Wiegand et al, 2021). Some other past work has defined explicitly and implicitly offensive instances a little differently: Sap et al (2020) considered factors such as obviousness, intent to offend and biased implications, Breitfeller et al (2019) considered factors such as the context and the person annotating the instance, and Razo and Kübler (2020) considered the kind of lexicon used.…”

Section: Offensive Language Datasetsmentioning

confidence: 99%

“…It is interesting to note that bin 4 contains some instances of implicit offensive language such as 'You look like a lesbian mechanic who has a shell collection'. In their paper, Wiegand et al (2021) explore the category of such "implicity abusive comparisons", in depth. More examples of implicitly offensive comments present in our dataset can be found in table 2 and table 6 (in Appendix A.3).…”

Section: Distribution Of Scoresmentioning

confidence: 99%

“…The most common strategy used is key-word based sampling. This results in datasets that are rich in explicit offensive language (language that is unambiguous in its potential to be offensive, such as those using slurs or swear words (Waseem et al, 2017)) but lack cases of implicit offensive language (language with its true offensive nature obscured due to lack of unambiguous swear words, usage of sarcasm or offensive analogies, and others (Waseem et al, 2017;Wiegand et al, 2021)) (Waseem, 2016;Wiegand et al, 2019). key-word based sampling often results in spurious correlations (e.g., sports-related expressions such as announcer and sport occur very frequently in offensive tweets).…”

Section: Introductionmentioning

confidence: 99%

See 1 more Smart Citation

Ruddit: Norms of Offensiveness for English Reddit Comments

Hada¹,

Sudhir²,

Mishra³

et al. 2021

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

View full text Add to dashboard Cite

Warning: This paper contains comments that may be offensive or upsetting.On social media platforms, hateful and offensive language negatively impact the mental well-being of users and the participation of people from diverse backgrounds. Automatic methods to detect offensive language have largely relied on datasets with categorical labels. However, comments can vary in their degree of offensiveness. We create the first dataset of English language Reddit comments that has fine-grained, real-valued scores between -1 (maximally supportive) and 1 (maximally offensive). The dataset was annotated using Best-Worst Scaling, a form of comparative annotation that has been shown to alleviate known biases of using rating scales. We show that the method produces highly reliable offensiveness scores. Finally, we evaluate the ability of widely-used neural models to predict offensiveness scores on this new dataset.

show abstract

Reading Between the Lines: Machine Learning Ensemble and Deep Learning for Implied Threat Detection in Textual Data

Raza,

Meghji,

Mahoto

et al. 2024

Int J Comput Intell Syst

View full text Add to dashboard Cite

With the increase in the generation and spread of textual content on social media, natural language processing (NLP) has become an important area of research for detecting underlying threats, racial abuse, violence, and implied warnings in the content. The subtlety and ambiguity of language make the development of effective models for detecting threats in text a challenging task. This task is further complicated when the threat is not explicitly conveyed. This study focuses on the task of implied threat detection using an explicitly designed machine-generated dataset with both linguistic and lexical features. We evaluated the performance of different machine learning algorithms on these features including Support Vector Machines, Logistic Regression, Naive Bayes, Decision Tree, and K-nearest neighbors. The ensembling approaches of Adaboost, Random Forest, and Gradient Boosting were also explored. Deep learning modeling was performed using Long Short-Term Memory, Deep Neural Networks (DNN), and Bidirectional Long Short-Term Memory (BiLSTM). Based on the evaluation, it was observed that classical and ensemble models overfit while working with linguistic features. The performance of these models improved when working with lexical features. The model based on logistic regression exhibited superior performance with an F1 score of 77.13%. While experimenting with deep learning models, DNN achieved an F1 score of 91.49% while the BiLSTM achieved an F1 score of 91.61% while working with lexical features. The current study provides a baseline for future research in the domain of implied threat detection.

show abstract

Implicitly Abusive Comparisons – A New Dataset and Linguistic Analysis

Cited by 12 publications

References 31 publications

Improving Generalizability in Implicitly Abusive Language Detection with Concept Activation Vectors

Improving Generalizability in Implicitly Abusive Language Detection with Concept Activation Vectors

Ruddit: Norms of Offensiveness for English Reddit Comments

Reading Between the Lines: Machine Learning Ensemble and Deep Learning for Implied Threat Detection in Textual Data

Contact Info

Product

Resources

About