amc76/slangguard-abusive

SlangGuard abusive language detector

A small model that says whether an Indonesian text contains abusive language (kata kasar). It is the second model of SlangGuard, a portfolio project. It is separate from the sentiment model, because the two questions differ: "ujaran ini negatif?" and "ujaran ini kasar?" have different answers for friendly teasing and for polite anger.

It is a scikit-learn pipeline: TF-IDF over groups of 2 to 5 letters inside words (char_wb), then logistic regression. Letter groups let it catch spellings like k0ntol and kntl that a whole-word model has never seen. The file is 1.3 MB and runs on a CPU.

How to use

The text must be normalized like the training text: lowercase, no punctuation, no standalone amp, one space between words. The function is normalize_text in src/data/clean.py.

import joblib
from huggingface_hub import hf_hub_download

path = hf_hub_download("amc76/slangguard-abusive", "model.joblib", revision="<commit>")
model = joblib.load(path)  # a pickle: load only a revision you trust
score = model.predict_proba(["dasar goblok lu"])[:, 1]  # probability of "abusive"

A pickle can run code when it is loaded. Pin the revision and do not load files from repos you do not trust.

Training data

Abusive and clean tweets. Multi-label hate speech and abusive language detection in Indonesian Twitter (Ibrohim and Budi, 2019), column Abusive, CC BY-NC-SA 4.0. After cleaning there are 12,849 tweets (38% abusive), split 70/15/15 with seed 42: 8,994 train, 1,927 validation and 1,928 test. The text of the data is not included here.

Cleaning removed what the source added and a real user would never type: the placeholders USER, URL and RT, the escape codes \n and \x.., and non-ASCII characters. These were not noise: USER is in 58% of the tweets and 47% of them are abusive (38% overall), so a model could have learned it instead of words. Duplicates and texts with two different labels were dropped.

Extra clean text. That data is Twitter text about hate and abuse. A model trained on it alone learned that intensifiers such as "banget" mean abuse, and flagged plain praise ("enak banget" scored 0.91; 6 of 15 short praise sentences were flagged). So 6,000 more texts labelled not abusive were added to training: 1,500 from the train split of each of the four sentiment periods of SlangGuard (reviews and tweets). Texts with a word of the authors' list of 125 abusive words were left out of this sample, so that a rude tweet is not taught as clean.

Results

The setting C was chosen on a validation set only: the abusive validation tweets plus 1,200 clean texts from the sentiment validation splits (30, from a grid of 0.3 to 100). The test sets were not used to choose anything.

Set	F1 (abusive)	Precision	Recall	Macro-F1
Abusive data, validation	0.889	0.915	0.864	0.912
Abusive data, test	0.893	0.911	0.877	0.915

That test set is the same kind of text as the training data. For texts like those SlangGuard serves, here is how often the detector flags the sentiment test sets (they have no abusive label, and some of the texts really are rude, so this is not an error rate):

Period	Flagged	Flagged among positive texts	Has a word of the authors' list
P1	3.7%	0.4%	5.7%
P2	2.2%	1.8%	3.9%
P3	1.8%	2.7%	3.3%
P4	0.7%	0.9%	2.0%

A hand check of 33 sentences written for this card: 14 of 15 rude ones were flagged (anjg scored 0.40 and bgst 0.52) and none of 18 normal ones. That check is not a measurement.

Limits

• It finds abusive words, not intent. Friendly use among friends is flagged, and a polite insult is not. Context is not read.
• Very short abbreviations are weak (anjg, bgst), and some words on the list are ambiguous (babi in a menu).
• The data is Twitter from 2018 and 2019, plus the sentiment text of this project. Newer slang and other platforms will do worse.
• It is not a hate speech detector. Only the Abusive label was used.
• Do not use it alone to moderate or punish people. It is a signal for a human or a filter.

History

The first revision of this repo (ccf9a6dc) was trained on the abusive tweets alone and flagged 8.5% to 13.8% of the sentiment texts, including plain praise. It is replaced by the revision described above. Do not use ccf9a6dc.

Citation

Ibrohim, M. O. and Budi, I. 2019. Multi-label Hate Speech and Abusive Language Detection in Indonesian Twitter. In Proceedings of the Third Workshop on Abusive Language Online, pp. 46-57.

The model is licensed CC BY-NC-SA 4.0, the same as its training data: attribution, non-commercial use, and the same license for derived work.