ADIMA: Abuse Detection In Multilingual Audio

Gupta, Vikram; Sharon, Rini A.; Sawhney, Ravinder Singh; Mukherjee, Debdoot

doi:10.48550/arxiv.2202.07991

Cited by 2 publications

(6 citation statements)

References 0 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…In Figure 2, we show the t-SNE plot of the emotion embeddings for the utterances and observe that the emotion embeddings of abusive and non-abusive words are fairly separable. This also explains the strong performance reported by ADIMA [34] while using features extracted using a CNN based model trained to perform sound classification instead of human speech understanding. ADIMA dataset contains audios collected from over 6k users across 10 languages which shows that this correlation holds for multiple users and languages.…”

Section: Abusive Behaviour and Emotionsmentioning

confidence: 79%

“…We train our unimodal and multimodal networks using binary cross entropy loss. We use the train/test splits (70:30 for each language) provided by [34] across all the experiments for fair comparison and report accuracy and F1 score on the test set.…”

Section: Resultsmentioning

confidence: 99%

“…We compare the performance of our proposed method (MADA) with the baselines reported in [34] and the binary classification results obtained using the two stage process (TSP) on the ADIMA dataset. ADIMA experimented with different backbones and feature processing techniques.…”

Section: Classification Resultsmentioning

confidence: 99%

“…After extracting the features from audio utterance, we train a binary classifier which uses these representations as input and learns to classify them into abusive and non-abusive categories. We experiment with the classifier architecture used by [34] and also propose a stack classifier(SC) architecture [35]. ADIMA classifier (AC): AC consists of two fully connected layers(512 → 256 → 128 → 2) with ReLU activation.…”

Section: Classification Modulementioning

confidence: 99%

“…We experiment with Abuse Detection in Multilingual Audio (ADIMA) [34] dataset. ADIMA is a well balanced multilingual dataset comprising of 10 Indic languages, Hindi (Hi), Bengali (Be), Punjabi (Pu), Haryanvi (Ha) , Kannada (Ka), Odia (Od), Bhojpuri (Bh), Gujarati (Gu), Tamil (Ta) and Malyalam (Ma), with 1200 utterances for each language.…”

Section: Dataset and Evaluationmentioning

confidence: 99%

See 4 more Smart Citations

Multilingual and Multimodal Abuse Detection

Sharon¹,

Shah²,

Mukherjee³

et al. 2022

Preprint

Self Cite

View full text Add to dashboard Cite

The presence of abusive content on social media platforms is undesirable as it severely impedes healthy and safe social media interactions. While automatic abuse detection has been widely explored in textual domain, audio abuse detection still remains unexplored. In this paper, we attempt abuse detection in conversational audio from a multimodal perspective in a multilingual social media setting. Our key hypothesis is that along with the modelling of audio, incorporating discriminative information from other modalities can be highly beneficial for this task. Our proposed method, MADA, explicitly focuses on two modalities other than the audio itself, namely, the underlying emotions expressed in the abusive audio and the semantic information encapsulated in the corresponding textual form. Observations prove that MADA demonstrates gains over audio-only approaches on the ADIMA dataset. We test the proposed approach on 10 different languages and observe consistent gains in the range 0.6%-5.2% by leveraging multiple modalities. We also perform extensive ablation experiments for studying the contributions of every modality and observe the best results while leveraging all the modalities together. Additionally, we perform experiments to empirically confirm that there is a strong correlation between underlying emotions and abusive behaviour. 1

show abstract

Section: Abusive Behaviour and Emotionsmentioning

confidence: 79%

Section: Resultsmentioning

confidence: 99%

Section: Classification Resultsmentioning

confidence: 99%

Section: Classification Modulementioning

confidence: 99%

Section: Dataset and Evaluationmentioning

confidence: 99%

See 3 more Smart Citations

Multilingual and Multimodal Abuse Detection

Sharon¹,

Shah²,

Mukherjee³

et al. 2022

Preprint

Self Cite

View full text Add to dashboard Cite

show abstract

Data-Efficient Methods For Improving Hate Speech Detection

Roychowdhury,

Gupta

2023

Findings of the Association for Computational Linguistics: EACL 2023

View full text Add to dashboard Cite

Scarcity of large-scale datasets, especially for resource-impoverished languages encouraged exploration of data-efficient methods for hate speech detection. In this work, we progress implicit and explicit hate speech detection using an input-level data augmentation technique, task reformulation using entailment and crosslearning across five languages. Our proposed data augmentation technique EasyMixup, improves the F1 performance across languages by 0.5-9%. We also observe substantial F1 gains of 1-8% by reformulating hate speech detection as Entailment-style problem. We further probe the contextual models and observe that higher layers encode implicit hate while lower layers focus on explicit hate, highlighting the importance of token-level understanding for explicit and context-level for implicit hate speech detection. 1

show abstract

ADIMA: Abuse Detection In Multilingual Audio

Cited by 2 publications

References 0 publications

Multilingual and Multimodal Abuse Detection

Multilingual and Multimodal Abuse Detection

Data-Efficient Methods For Improving Hate Speech Detection

Contact Info

Product

Resources

About