Disclaimer: This project contains real comments that could be considered profane, offensive, or abusive.

Dataset Card for "ViHOS - Vietnamese Hate and Offensive Spans Dataset"

Dataset Motivation

The rise in hateful and offensive language directed at other users is one of the adverse side effects of the increased use of social networking platforms. This could make it difficult for human moderators to review tagged comments filtered by classification systems.

To help address this issue, we present the ViHOS (Vietnamese Hate and Offensive Spans) dataset, the first human-annotated corpus containing 26k spans on 11k online comments.

Our goal is to create a dataset that contains comprehensive hate and offensive thoughts, meanings, or opinions within the comments rather than just a lexicon of hate and offensive terms.

We also provide definitions of hateful and offensive spans in Vietnamese comments as well as detailed annotation guidelines. Futhermore, our solutions to deal with nine different online foul linguistic phenomena are also provided in the paper (e.g. Teencodes; Metaphors, metonymies; Hyponyms; Puns...).

We hope that this dataset will be useful for researchers and practitioners in the field of hate speech detection in general and hate spans detection in particular.

Dataset Summary

ViHOS contains 26,476 human-annotated spans on 11,056 comments (5,360 comments have hate and offensive spans, and 5,696 comments do not)

It is splitted into train, dev, and test set with following information:

Train set: 8,844 comments
Dev set: 1,106 comments
Test set: 1,106 comments

Data Instance

An span extraction-based (see Data Structure for more details) example of 'test' looks as follows:

{
  "content": "Thối CC chỉ không ngửi đuợc thôi",
  'index_spans': "[0, 1, 2, 3, 5, 6]"
}

An sequence labeling-based (see Data Structure for more details) example of 'test' looks as follows:

{
  "content": "Thối CC chỉ không ngửi đuợc thôi",
  'index_spans': ["B-T", "I-T", "O", "O", "O", "O", "O"]
}

Data Structure

Here is our data folder structure!

.
└── data/
    ├── train_sequence_labeling/
    │   ├── syllable/
    │   │   ├── dev_BIO_syllable.csv
    │   │   ├── test_BIO_syllable.csv
    │   │   └── train_BIO_syllable.csv
    │   └── word/
    │       ├── dev_BIO_Word.csv
    │       ├── test_BIO_Word.csv
    │       └── train_BIO_Word.csv
    ├── train_span_extraction/
    │   ├── dev.csv
    │   └── train.csv
    └── test/
        └── test.csv

Sequence labeling-based version

Syllable

Description:

This folder contains the data for the sequence labeling-based version of the task. The data is divided into two files: train, and dev. Each file contains the following columns:
- index: The id of the word.
- word: Words in the sentence after the processing of tokenization using VnCoreNLP tokenizer followed by underscore tokenization. The reason for this is that some words are in bad format: e.g. "điện.thoại của tôi" is split into ["điện.thoại", "của", "tôi"] instead of ["điện", "thoại", "của", "tôi"] if we use space tokenization, which is not in the right format of Syllable. As that, we used VnCoreNLP to tokenize first and then split words into tokens. e.g. "điện.thoại của tôi" ---(VnCoreNLP)---> ["điện_thoại", "của", "tôi"] ---(split by "_")---> ["điện", "thoại", "của", "tôi"].
- tag: The tag of the word. The tag is either B-T (beginning of a word), I-T (inside of a word), or O (outside of a word).
The train_BIO_syllable and dev_BIO_syllable file are used for training and validation for XLMR model, respectively.
The test_BIO_syllable file is used for reference only. It is not used for testing the model. Please use the test.csv file in the Testdata folder for testing the model.

Word

Description:

This folder contains the data for the sequence labeling-based version of the task. The data is divided into two files: train, and dev. Each file contains the following columns:
- index: The id of the word.
- word: Words in the sentence after the processing of tokenization using VnCoreNLP tokenizer
- tag: The tag of the word. The tag is either B-T (beginning of a word), I-T (inside of a word), or O (outside of a word).
The train_BIO_Word and dev_BIO_Word file are used for training and validation for PhoBERT model, respectively.
The test_BIO_Word file is used for reference only. It is not used for testing the model. Please use the test.csv file in the data/test folder for testing the model.

Span Extraction-based version

Description:

This folder contains the data for the span extraction-based version of the task. The data is divided into two files: train and dev. Each file contains the following columns:
- content: The content of the sentence.
- span_ids: The index of the hate and offensive spans in the sentence. The index is in the format of [start, end] where start is the index of the first character of the hate and offensive span and end is the index of the last character of the hate and offensive span.
The train and dev file are used for training and validation for BiLSTM-CRF model, respectively.

Citation Information

@inproceedings{hoang-etal-2023-vihos,
    title = "{V}i{HOS}: Hate Speech Spans Detection for {V}ietnamese",
    author = "Hoang, Phu Gia  and
      Luu, Canh Duc  and
      Tran, Khanh Quoc  and
      Nguyen, Kiet Van  and
      Nguyen, Ngan Luu-Thuy",
    booktitle = "Proceedings of the 17th Conference of the European Chapter of the Association for Computational Linguistics",
    month = may,
    year = "2023",
    address = "Dubrovnik, Croatia",
    publisher = "Association for Computational Linguistics",
    url = "https://aclanthology.org/2023.eacl-main.47",
    doi = "10.18653/v1/2023.eacl-main.47",
    pages = "652--669",
    abstract = "The rise in hateful and offensive language directed at other users is one of the adverse side effects of the increased use of social networking platforms. This could make it difficult for human moderators to review tagged comments filtered by classification systems. To help address this issue, we present the ViHOS (Vietnamese Hate and Offensive Spans) dataset, the first human-annotated corpus containing 26k spans on 11k comments. We also provide definitions of hateful and offensive spans in Vietnamese comments as well as detailed annotation guidelines. Besides, we conduct experiments with various state-of-the-art models. Specifically, XLM-R{\_}Large achieved the best F1-scores in Single span detection and All spans detection, while PhoBERT{\_}Large obtained the highest in Multiple spans detection. Finally, our error analysis demonstrates the difficulties in detecting specific types of spans in our data for future research. Our dataset is released on GitHub.",
}