Generalizable Implicit Hate Speech Detection Using Contrastive Learning
4645Authors
nan
Abstract
AbstractHate speech detection has gained increasing attention with the growing prevalence of hateful contents. When a text contains an obvious hate word or expression, it is fairly easy to detect it. However, it is challenging to identify implicit hate speech in nuance or context when there are insufficient lexical cues. Recently, there are several attempts to detect implicit hate speech leveraging pre-trained language models such as BERT and HateBERT. Fine-tuning on an implicit hate speech dataset shows satisfactory performance when evaluated on the test set of the dataset used for training. However, we empirically confirm that the performance drops at least 12.5%p in F1 score when tested on the dataset that is different from the one used for training. We tackle this cross-dataset underperforming problem using contrastive learning. Based on our observation of common underlying implications in various forms of hate posts, we propose a novel contrastive learning method, ImpCon, that pulls an implication and its corresponding posts close in representation space. We evaluate the effectiveness of ImpCon by running cross-dataset evaluation on three implicit hate speech benchmarks. The experimental results on cross-dataset show that ImpCon improves at most 9.10% on BERT, and 8.71% on HateBERT.
Keywords
Full Content
Proceedings of the 29th International Conference on Computational Linguistics, pages 684–693 October 12–17, 2022. 684 DialogueEIN: Emotion Interaction Network for Dialogue Affective Analysis Yuchen Liu1, Jinming Zhao2, Jingwen Hu1, Ruichen Li1, Qin Jin1∗ 1 School of Information, Renmin University of China 2 Qiyuan Lab, Beijing, China {liuyuchen_alfred, hujingwen_benja, ruichen, qjin}@ruc.edu.cn zhaojinming@qiyuanlab.com Abstract Emotion Recognition in Conversation (ERC) has attracted increasing attention in the affec- tive computing research field. Previous works have mainly focused on modeling the seman- tic interactions in the dialogue and implicitly inferring the evolution of the speakers’ emo- tional states. Few works have considered the emotional interactions, which directly reflect the emotional evolution of speakers in the dia- logue. According to psychological and behav- ioral studies, the emotional inertia and emo- tional stimulus are important factors that affect the speaker’s emotional state in conversations. In this work, we propose a novel Dialogue Emo- tion Interaction Network, DialogueEIN, to ex- plicitly model the intra-speaker, inter-speaker, global and local emotional interactions to re- spectively simulate the emotional inertia, emo- tional stimulus, global and local emotional evolution in dialogues. Extensive experiments on four ERC benchmark datasets, IEMOCAP, MELD, EmoryNLP and DailyDialog, show that our proposed DialogueEIN considering emotional interaction factors can achieve su- perior or competitive performance compared to state-of-the-art methods. Our codes and models are released1. 1 Introduction Emotion Recognition in Conversation (ERC), aim- ing to recognize the emotional status of each ut- terance in a conversation, has attracted increasing research attention in recent years. It has rich ap- plication potentials in emotional support, mental health, and legal trials etc (Poria et al., 2019). Unlike traditional emotion recognition based on isolated utterances, conversation context modeling is very important for the ERC task (Poria et al., 2019). Different approaches have been proposed *Corresponding Author 1https://github.com/AIM3-RUC/DialogueEIN Male Female Neutral You are not sorry you came? No, I’m not sorry. But I’m not going to stay. Look, you saw the way that she acted and then… Well, I mean, you, you were acting kind of… What? I don’t know, embarrassed of her since I came. I mean, that’s the reason why I asked you here. Well. I love you. I guess that’s why I came. Neutral Happy Happy Happy Happy Sad Sad Sad Sad Emotional Stimulus The emotional change of the male stimulates the female to become happy. Emotional Inertia No obvious stimulus and the emotion remains unchanged. Figure 1: Illustration of emotional interaction in an ex- ample dialogue from the IEMOCAP dataset. The grey arrow line indicates that the speaker’s emotional state remains unchanged, and the orange arrow line indicates that the speaker is stimulated and the emotion changes, corresponding to emotional inertia and emotional stimu- lus respectively. to model the context of a conversation. For ex- ample, CMN (Hazarika et al., 2018b) and ICON (Hazarika et al., 2018a) use speaker-specific and global recurrent networks to model the semantic context in a dialogue, and multi-hop memory net- works are used to generate the summaries for pre- diction. DialogueGCN (Ghosal et al.) …
Content truncated to 500 words
Labeled Text
AbstractHate speech detection has gained increasing attention with the growing prevalence of hateful contents. When a text contains an obvious hate word or expression, it is fairly easy to detect it. However, it is challenging to identify implicit hate speech in nuance or context when there are insufficient lexical cues. Recently, there are several attempts to detect implicit hate speech leveraging pre-trained language models such as BERT and HateBERT. Fine-tuning on an implicit hate speech dataset shows satisfactory performance when evaluated on the test set of the dataset used for training. However, we empirically confirm that the performance drops at least 12.5%p in F1 score when tested on the dataset that is different from the one used for training. We tackle this cross-dataset underperforming problem using contrastive learning. Based on our observation of common underlying implications in various forms of hate posts, we propose a novel contrastive learning method, ImpCon, that pulls an implication and its corresponding posts close in representation space. We evaluate the effectiveness of ImpCon by running cross-dataset evaluation on three implicit hate speech benchmarks. The experimental results on cross-dataset show that ImpCon improves at most 9.10% on BERT, and 8.71% on HateBERT.
Publication Details
- Paper ID:
- 4645
- Year:
- 2022
- Publication:
- TIPSTER
- Publication Type:
- conf
- DOI:
- nan