readme

README.md

@@ -33,3 +33,17 @@ Since MT models are not perfect, some messages are not entirely translated or no
 To check for obvious errors in pipeline, a general language detection model is used to prune non french texts.
 
 Language detection model : papluca/xlm-roberta-base-language-detection
+
+### Annotation
+
+Since "hate speech" dimension is highly subjective, and datasets comes with different annotations types, a conventional labeling stategy is required.
+
+Each sample is annotated with "0" if negative sample and "1" if positive sample.
+
+### Filtering rules :
+
+- FTR dataset : 
+- MLMA dataset :
+- CAA dataset :
+- "Annotated Corpus" dataset : 
+- UC-Berkeley Measuring Hate Speech dataset : average hate_speech_score > 0 => 1