Update README.md

README.md

@@ -17,59 +17,29 @@ pipeline_tag: text-classification
 
 ---
 
-# HateBERTimbau
+# HateBERTimbau-YouTube
 
-**HateBERTimbau** is a foundation, large language model for European **Portuguese** from **Portugal** for Hate Speech content.
-
-It is an **encoder** of the BERT family, based on the neural architecture Transformer and 
-developed over the [BERTimbau](https://huggingface.co/neuralmind/bert-large-portuguese-cased) model, retrained on a dataset of 229,103 tweets specifically focused on potential hate speech.
+**HateBERTimbau-YouTube** is a transformer-based encoder model for identifying hate speech in Portuguese social media text. It is a fine-tuned version of [HateBERTimbau](https://huggingface.co/knowhate/HateBERTimbau) model, retrained on a dataset of 23,912 YouTube comments specifically focused on hate speech.
 
 ## Model Description
 
 - **Developed by:** [kNOwHATE: kNOwing online HATE speech: knowledge + awareness = TacklingHate](https://knowhate.eu)
 - **Funded by:** [European Union](https://ec.europa.eu/info/funding-tenders/opportunities/portal/screen/opportunities/topic-details/cerv-2021-equal)
-- **Model type:** Transformer-based model retrained for Hate Speech in Portuguese social media text
+- **Model type:** Transformer-based text classification model fine-tuned for Hate Speech detection in Portuguese social media text
 - **Language:** Portuguese
-- **Retrained from model:** [neuralmind/bert-large-portuguese-cased](https://huggingface.co/neuralmind/bert-large-portuguese-cased)
-
-Several models were developed by fine-tuning Base HateBERTimbau for Hate Speech detection present in the table bellow:
-
-| HateBERTimbau's Family of Models                                                                             | 
-|---------------------------------------------------------------------------------------------------------|
-| [**HateBERTimbau YouTube**](https://huggingface.co/knowhate/HateBERTimbau-youtube)        |
-| [**HateBERTimbau Twitter**](https://huggingface.co/knowhate/HateBERTimbau-twitter)        |
-| [**HateBERTimbau YouTube+Twitter**](https://huggingface.co/knowhate/HateBERTimbau-yt-tt)|
+- **Fine-tuned from model:** [knowhate/HateBERTimbau](https://huggingface.co/knowhate/HateBERTimbau)
 
 # Uses
 
-You can use this model directly with a pipeline for masked language modeling:
+You can use this model directly with a pipeline for text classification:
 
 ```python
 from transformers import pipeline
-unmasker = pipeline('fill-mask', model='knowhate/HateBERTimbau')
-
-unmasker("Os [MASK] são todos uns animais, deviam voltar para a sua terra.")
-
-[{'score': 0.6771652698516846,
-  'token': 12714,
-  'token_str': 'africanos',
-  'sequence': 'Os africanos são todos uns animais, deviam voltar para a sua terra.'},
- {'score': 0.08679857850074768,
-  'token': 15389,
-  'token_str': 'homossexuais',
-  'sequence': 'Os homossexuais são todos uns animais, deviam voltar para a sua terra.'},
- {'score': 0.03806231543421745,
-  'token': 4966,
-  'token_str': 'portugueses',
-  'sequence': 'Os portugueses são todos uns animais, deviam voltar para a sua terra.'},
- {'score': 0.035253893584012985,
-  'token': 16773,
-  'token_str': 'Portugueses',
-  'sequence': 'Os Portugueses são todos uns animais, deviam voltar para a sua terra.'},
- {'score': 0.023521048948168755,
-  'token': 8618,
-  'token_str': 'brancos',
-  'sequence': 'Os brancos são todos uns animais, deviam voltar para a sua terra.'}]
+classifier = pipeline('text-classification', model='knowhate/HateBERTimbau-youtube')
+
+classifier("as pessoas tem que perceber que ser 'panasca' não é deixar de ser homem, é deixar de ser humano 😂😂")
+
+[{'label': 'Hate Speech', 'score': 0.9228119850158691}]
 
 ```
 
@@ -79,8 +49,8 @@ Or this model can be used by fine-tuning it for a specific task/dataset:
 from transformers import AutoTokenizer, AutoModelForSequenceClassification, TrainingArguments, Trainer
 from datasets import load_dataset
 
-tokenizer = AutoTokenizer.from_pretrained("knowhate/HateBERTimbau")
-model = AutoModelForSequenceClassification.from_pretrained("knowhate/HateBERTimbau")
+tokenizer = AutoTokenizer.from_pretrained("knowhate/HateBERTimbau-youtube")
+model = AutoModelForSequenceClassification.from_pretrained("knowhate/HateBERTimbau-youtube")
 dataset = load_dataset("knowhate/youtube-train")
 
 def tokenize_function(examples):
@@ -104,42 +74,33 @@ trainer.train()
 
 ## Data
 
-229,103 tweets associated with offensive content were used to retrain the base model.
+23,912 YouTube comments associated with offensive content were used to fine-tune the base model.
 
 ## Training Hyperparameters
 
-- Batch Size: 4 samples
-- Epochs: 100
-- Learning Rate: 5e-5 with Adam optimizer
-- Maximum Sequence Length: 512 sentence pieces
+- Batch Size: 32
+- Epochs: 3
+- Learning Rate: 2e-5 with Adam optimizer
+- Maximum Sequence Length: 350 tokens
 
 # Testing
 
 ## Data
 
-We used two different datasets for testing, one for YouTube comments [here](https://huggingface.co/datasets/knowhate/youtube-test) and another for Tweets [here](https://huggingface.co/datasets/knowhate/twitter-test).
+The dataset used to test this model was: [knowhate/youtube-test](https://huggingface.co/datasets/knowhate/youtube-test)
 
-## Hate Speech Classification Results (with no fine-tuning)
+## Results
 
 | Dataset         | Precision  | Recall    | F1-score     |
 |:----------------|:-----------|:----------|:-------------|
-| **YouTube**     | 0.928      | 0.108     | **0.193**    | 
-| **Twitter**     | 0.686      | 0.211     | **0.323**    |
+| **YouTube**     | 0.856      | 0.892     | **0.874**    | 
 
 # BibTeX Citation
 
+Currently in Peer Review
+
 ``` latex
-@mastersthesis{Matos-Automatic-Hate-Speech-Detection-in-Portuguese-Social-Media-Text,
-title = {{Automatic Hate Speech Detection in Portuguese Social Media Text}},
-author = {Matos, Bernardo Cunha},
-month = nov,
-year = {2022},
-abstract = {{Online Hate Speech (HS) has been growing dramatically on social media and its uncontrolled spread has motivated researchers to develop a diversity of methods for its automated detection. However, the detection of online HS in Portuguese still merits further research. To fill this gap, we explored different models that proved to be successful in the literature to address this task. In particular, we have explored models that use the BERT architecture. Beyond testing single-task models we also explored multitask models that use the information on other related categories to learn HS. To better capture the semantics of this type of texts, we developed HateBERTimbau, a retrained version of BERTimbau more directed to social media language including potential HS targeting African descent, Roma, and LGBTQI+ communities. The performed experiments were based on CO-HATE and FIGHT, corpora of social media messages posted by the Portuguese online community that were labelled regarding the presence of HS among other categories.
-The results achieved show the importance of considering the annotator's agreement on the data used to develop HS detection models. Comparing different subsets of data used for the training of the models it was shown that, in general, a higher agreement on the data leads to better results.
-HATEBERTimbau consistently outperformed BERTimbau on both datasets confirming that further pre-training of BERTimbau was a successful strategy to obtain a language model more suitable for online HS detection in Portuguese. 
-The implementation of target-specific models, and multitask learning have shown potential in obtaining better results.}},
-language = {eng},
-copyright = {embargoed-access},
+
 }
 ```