Data Driven and Psycholinguistics Motivated Approaches to Hate Speech Detection

Silva, Samuel Caetano da; Ferreira, Thiago Castro; Ramos, Ricelli Moreira Silva; Paraboni, Ivandré

doi:10.13053/cys-24-3-3478

Cited by 12 publications

(3 citation statements)

References 4 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…Regarding the text portion of the model, the current task may benefit from multiple, well-known NLP methods and applications. Among these, we may consider the use of hate speech detection methods (Basile et al, 2019;Mishra et al, 2019;da Silva et al, 2020), authorship attribution (Custódio and Paraboni, 2021; Barlas and Stamatatos, 2021), or author profiling (López-Santillán et al, 2020;Rangel et al, 2020). The latter, comprising the computational task of determining individuals' demographics from text, may help determine their stance towards a particular topic by taking into account, for instance, information regarding their political orientation (Flores et al, 2022), personality traits (Verhoeven et al, 2016;dos Santos et al, 2019), moral values (dos Santos andPavan et al, 2023), and others.…”

Section: Final Remarksmentioning

confidence: 99%

Stance Prediction from Multimodal Social Media Data

C. L. Cavalheiro,

C. Pavan,

Paraboni

2023

Proceedings of the Conference Recent Advances in Natural Language Processing - Large Language Models for Natural Language Proce

View full text Add to dashboard Cite

Stance prediction -the computational task of inferring attitudes towards a given target topic of interest -relies heavily on text data provided by social media or similar sources, but it may also benefit from non-text information such as demographics (e.g., users' gender, age, etc.), network structure (e.g., friends, followers, etc.), interactions (e.g., mentions, replies, etc.) and other non-text properties (e.g., time information, etc.). However, so-called hybrid (or in some cases multimodal) approaches to stance prediction have only been developed for a small set of target languages, and often making use of count-based text models (e.g., bag-of-words) and time-honoured classification methods (e.g., support vector machines). As a means to further research in the field, in this work we introduce a number of text-and non-text models for stance prediction in the Portuguese language, which make use of more recent methods based on BERT and an ensemble architecture, and ask whether a BERT stance classifier may be enhanced with different kinds of network-related information.

show abstract

Section: Final Remarksmentioning

confidence: 99%

Stance Prediction from Multimodal Social Media Data

C. L. Cavalheiro,

C. Pavan,

Paraboni

2023

Proceedings of the Conference Recent Advances in Natural Language Processing - Large Language Models for Natural Language Proce

View full text Add to dashboard Cite

show abstract

“…A detecc ¸ão transtorno depressivo com base em dados textuais (e.g., provenientes de redes sociais ou outras fontes) é tipicamente modelada na forma de um problema de aprendizado de máquina supervisionado, ou seja, fazendo uso de córpus de textos rotuladas com informac ¸ões relativas ao estado de saúde mental de seus autores (e.g., usuários de redes sociais) para treino e teste de classificadores. Sob esta perspectiva, a tarefa pode ser vista como uma instância do problema de caracterizac ¸ão autoral [dos Santos et al 2020b, Flores et al 2022] combinado à detecc ¸ão de linguagem afetiva [da Silva et al 2020]. Um levantamento de estudos recentes deste tipo é apresentado na Tabela 1, com indicac ¸ão do gênero de texto considerado (Reddit, Twitter), a forma de representac ¸ão dos dados textuais (b=bag of words, BERT [Devlin et al 2019], d=características de domínio, e=embeddings, h=horário da publicac ¸ão, i=imagens, l=atributos LIWC [Pennebaker et al 2001], m=metadados, n=informac ¸ões de rede, p=part-of-speech, s=atributos afetivos, t=tópicos, u=informac ¸ões demográficas), e métodos computacionais (e.g., CNN=redes neurais convolucionais, LSTM=long short-term neural networks, LR=regressão logística, RF=Random Forest, etc.…”

Section: Trabalhos Relacionadosunclassified

Predição de transtorno depressivo em redes sociais: BERT supervisionado ou ChatGPT zero-shot?

dos Santos,

Paraboni

2023

Anais Do XIV Simpósio Brasileiro De Tecnologia Da Informação E Da Linguagem Humana (STIL 2023)

View full text Add to dashboard Cite

Este artigo apresenta um primeiro estudo sobre o uso do sistema de dialogo ChatGPT em uma aplicação complexa e sensível: a predição computacional de transtornos de saúde mental a partir de textos provenientes de redes sociais. Para esse fim, foi conduzido um experimento comparando uma abordagem supervisionada tradicional baseada em BERT com uma estratégia zero-shot baseada em prompts em língua natural submetidos diretamente ao sistema de diálogo. Resultados desta avaliação, levando em conta a acurácia da tarefa de classificação face à necessidade de anotação prévia de córpus da abordagem supervisionada, destacam diferentes vantagens de cada alternativa.

show abstract

“…Supervised training based methods rely on large labeled datasets to train models from classical machine learning (Badjatiya et al, 2017), deep learning such as RNNs (Pavlopoulos et al, 2017), CNNs (Zhang et al, 2018) or BERT and GPT-2 transformers (D'sa et al, 2019;Caselli et al, 2020;Silva et al, 2020). Specifically transformers have been fine-tuned for hate-speech detection (Caselli et al, 2020) as well as for detection of targeted offensive language (Rosenthal et al, 2021) and bias .…”

Section: Related Workmentioning

confidence: 99%

Few-shot Instruction Prompts for Pretrained Language Models to Detect Social Biases

Prabhumoye¹,

Kocielnik²,

Shoeybi³

et al. 2021

Preprint

View full text Add to dashboard Cite

Warning: this paper contains content that may be offensive or upsetting.Detecting social bias in text is challenging due to nuance, subjectivity, and difficulty in obtaining good quality labeled datasets at scale, especially given the evolving nature of social biases and society. To address these challenges, we propose a few-shot instructionbased method for prompting pre-trained language models (LMs). We select a few labelbalanced exemplars from a small support repository that are closest to the query to be labeled in the embedding space. We then provide the LM with instruction that consists of this subset of labeled exemplars, the query text to be classified, a definition of bias, and prompt it to make a decision. We demonstrate that large LMs used in a few-shot context can detect different types of fine-grained biases with similar and sometimes superior accuracy to fine-tuned models. We observe that the largest 530B parameter model is significantly more effective in detecting social bias compared to smaller models (achieving at least 20% improvement in AUC metric compared to other models). It also maintains a high AUC (dropping less than 5%) in a few-shot setting with a labeled repository reduced to as few as 100 samples. Large pretrained language models thus make it easier and quicker to build new bias detectors.

show abstract

Data Driven and Psycholinguistics Motivated Approaches to Hate Speech Detection

Cited by 12 publications

References 4 publications

Stance Prediction from Multimodal Social Media Data

Stance Prediction from Multimodal Social Media Data

Predição de transtorno depressivo em redes sociais: BERT supervisionado ou ChatGPT zero-shot?

Few-shot Instruction Prompts for Pretrained Language Models to Detect Social Biases

Contact Info

Product

Resources

About