Update README.md
Browse files
README.md
CHANGED
@@ -19,3 +19,109 @@ size_categories:
|
|
19 |
---
|
20 |
|
21 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
19 |
---
|
20 |
|
21 |
|
22 |
+
# MLSNT: Multi-Lingual Social Network Toxicity Dataset
|
23 |
+
|
24 |
+
**MLSNT** is a multi-lingual dataset for toxicity detection created through a large language model-assisted label transfer pipeline. It enables efficient and scalable moderation across languages and platforms, and is built to support span-level and category-specific classification for toxic content.
|
25 |
+
|
26 |
+
This dataset is introduced in the following paper:
|
27 |
+
|
28 |
+
> **Unified Game Moderation: Soft-Prompting and LLM-Assisted Label Transfer for Resource-Efficient Toxicity Detection**
|
29 |
+
> π Accepted at **KDD 2025**, Applied Data Science Track
|
30 |
+
|
31 |
+
---
|
32 |
+
|
33 |
+
## π§© Overview
|
34 |
+
|
35 |
+
MLSNT harmonizes 15 publicly available toxicity datasets across **7 languages** using GPT-4o-mini to create consistent binary and fine-grained labels. It is suitable for both training and evaluating toxicity classifiers in multi-lingual, real-world moderation systems.
|
36 |
+
|
37 |
+
---
|
38 |
+
|
39 |
+
## π Supported Languages
|
40 |
+
|
41 |
+
- π«π· French (`fr`)
|
42 |
+
- π©πͺ German (`de`)
|
43 |
+
- π΅πΉ Portuguese (`pt`)
|
44 |
+
- π·πΊ Russian (`ru`)
|
45 |
+
- π¨π³ Simplified Chinese (`zh-cn`)
|
46 |
+
- πΉπΌ Traditional Chinese (`zh-tw`)
|
47 |
+
- π―π΅ Japanese (`ja`)
|
48 |
+
|
49 |
+
---
|
50 |
+
|
51 |
+
## ποΈ Construction Method
|
52 |
+
|
53 |
+
1. **Source Datasets**
|
54 |
+
15 human-annotated datasets were gathered from `hatespeechdata.com` and peer-reviewed publications.
|
55 |
+
|
56 |
+
2. **LLM-Assisted Label Transfer**
|
57 |
+
GPT-4o-mini was prompted to re-annotate each instance into a unified label schema. Only examples where human and LLM annotations agreed were retained.
|
58 |
+
|
59 |
+
3. **Toxicity Categories**
|
60 |
+
Labels are fine-grained categories (e.g., `threat`, `hate`, `harassment`).
|
61 |
+
|
62 |
+
---
|
63 |
+
|
64 |
+
## π Dataset Statistics
|
65 |
+
|
66 |
+
| Language | Total Samples | % Discarded | Toxic % (Processed) |
|
67 |
+
|----------------------|----------------|-------------|----------------------|
|
68 |
+
| German (HASOC, etc.) | ~13,800 | 28β69% | 32β56% |
|
69 |
+
| French (MLMA) | ~3,200 | 20% | 94% |
|
70 |
+
| Russian | ~14,300 | ~40% | 33β54% |
|
71 |
+
| Portuguese | ~21,000 | 20β44% | 26β50% |
|
72 |
+
| Japanese | ~2,000 | 10β25% | 17β45% |
|
73 |
+
| Chinese (Simplified) | ~34,000 | 29β46% | 48β61% |
|
74 |
+
| Chinese (Traditional)| ~65,000 | 37% | ~9% |
|
75 |
+
|
76 |
+
---
|
77 |
+
|
78 |
+
## πΎ Format
|
79 |
+
|
80 |
+
Each row in the dataset includes:
|
81 |
+
|
82 |
+
- `full_text`: The original utterance or message.
|
83 |
+
- `start_string_index`: A list of start string indices (start positions of toxic spans).
|
84 |
+
- `end_string_index`: A list of end string indices (end positions of toxic spans).
|
85 |
+
- `category_id`: A list of toxic category IDs (integer values).
|
86 |
+
- `final_label`: A list of toxic category names (string values).
|
87 |
+
- `min_category_id`: The minimum toxic category ID in the row (used as the primary label).
|
88 |
+
- `match_id`: A unique identifier composed of the original dataset name and a row-level ID.
|
89 |
+
|
90 |
+
---
|
91 |
+
|
92 |
+
## ποΈ Category ID Mapping
|
93 |
+
|
94 |
+
| ID | Friendly Name |
|
95 |
+
|-----|---------------------------------------------|
|
96 |
+
| 0 | Non Toxic |
|
97 |
+
| 1 | Threats (Life Threatening) |
|
98 |
+
| 2 | Minor Endangerment |
|
99 |
+
| 3 | Threats (Non-Life Threatening) |
|
100 |
+
| 4 | Hate |
|
101 |
+
| 5 | Sexual Content / Harassment |
|
102 |
+
| 6 | Extremism |
|
103 |
+
| 7 | Insults |
|
104 |
+
| 8 | Controversial / Potentially Toxic Topic |
|
105 |
+
|
106 |
+
---
|
107 |
+
|
108 |
+
## π¬ Applications
|
109 |
+
|
110 |
+
- Fine-tuning multi-lingual moderation systems
|
111 |
+
- Cross-lingual toxicity benchmarking
|
112 |
+
- Training span-level and category-specific toxicity detectors
|
113 |
+
- Studying LLM label transfer reliability and agreement filtering
|
114 |
+
|
115 |
+
---
|
116 |
+
|
117 |
+
## π Citation
|
118 |
+
|
119 |
+
If you use MLSNT in academic work, please cite:
|
120 |
+
|
121 |
+
```bibtex
|
122 |
+
@inproceedings{yang2025mlsnt,
|
123 |
+
title={Unified Game Moderation: Soft-Prompting and LLM-Assisted Label Transfer for Resource-Efficient Toxicity Detection},
|
124 |
+
author={Zachary Yang and Domenico Tullo and Reihaneh Rabbany},
|
125 |
+
booktitle={Proceedings of the 30th ACM SIGKDD Conference on Knowledge Discovery and Data Mining (KDD)},
|
126 |
+
year={2025}
|
127 |
+
}
|