first stab at multilingual hatespeech task

mteb/tasks/Classification/__init__.py

@@ -23,7 +23,6 @@
 from .hin.HindiDiscourseClassification import *
 from .hrv.CroatianSentimentClassification import *
 from .ind.IndonesianIdClickbaitClassification import *
-from .ita.ItaHateClassification import *
 from .jpn.WRIMEClassification import *
 from .mkd.MacedonianTweetSentimentClassification import *
 from .mlt.MalteseSentimentClassification import *
@@ -34,6 +33,7 @@
 from .multilingual.MassiveScenarioClassification import *
 from .multilingual.MTOPDomainClassification import *
 from .multilingual.MTOPIntentClassification import *
+from .multilingual.MultiHateClassification import *
 from .multilingual.NordicLangClassification import *
 from .multilingual.ScalaClassification import *
 from .nld.DutchBookReviewSentimentClassification import *

mteb/tasks/Classification/ita/ItaHateClassification.py → mteb/tasks/Classification/multilingual/MultiHateClassification.py

@@ -1,27 +1,82 @@
 from __future__ import annotations
 
-from mteb.abstasks.AbsTaskClassification import AbsTaskClassification
+import datasets
+
+from mteb.abstasks import AbsTaskClassification, MultilingualTask
 from mteb.abstasks.TaskMetadata import TaskMetadata
 
+_LANGUAGES = {
+    "ara": ["ara-Arab"],
+    "cmn": ["cmn-Hans"],
+    "eng": ["eng-Latn"],
+    "deu": ["deu-Latn"],
+    "fra": ["fra-Latn"],
+    "hin": ["hin-Deva"],
+    "ita": ["ita-Latn"],
+    "nld": ["nld-Latn"],
+    "por": ["por-Latn"],
+    "spa": ["spa-Latn"],
+}
+
+_HF_AFFIX = {
+    "ara": "arabic",
+    "cmn": "mandarin",
+    "eng": "",
+    "deu": "german",
+    "fra": "french",
+    "hin": "hindi",
+    "ita": "italian",
+    "nld": "dutch",
+    "por": "portuguese",
+    "spa": "spanish",
+}
+
+_REVISION_DICT = {
+    "ara": "65eb7455a05cb77b3ae0c69d444569a8eee54628",
+    "cmn": "617d3e9fccd186277297cc305f6588af7384b008",
+    "eng": "9d2ac89df04254e5c427bcc8d61b6d6c83a1f59b",
+    "deu": "5229a5cc475f36c08d03ca52f0ccb005705e60d2",
+    "fra": "5d3085f2129139abc10d2b58becd4d4f2978e5d5",
+    "hin": "e9e68e1a4db04726b9278192377049d0f9693012",
+    "ita": "21e3d5c827cb60619a89988b24979850a7af85a5",
+    "nld": "d622427417d37a8d74e110e6289bc29af4ba4056",
+    "por": "323bdf67e0fbd3d7f8086fad0971b5bd5a62524b",
+    "spa": "a7ea759535bb9fad6361cca151cf94a46e88edf3",
+}
+
+
+def _transform(dataset):
+    keep_cols = ["test_case", "label_gold"]
+    rename_dict = dict(zip(keep_cols, ["text", "label"]))
+    remove_cols = [col for col in dataset["test"].column_names if col not in keep_cols]
+    dataset = dataset.rename_columns(rename_dict)
+    dataset = dataset.class_encode_column("label")
+    dataset = dataset.class_encode_column("functionality")
+    dataset = dataset["test"].train_test_split(
+        test_size=0.5, seed=42, stratify_by_column="functionality"
+    )  # balanced sampling across types of hate speech
+    dataset = dataset.remove_columns(remove_cols)
+    return dataset
+
 
-class ItaHateClassification(AbsTaskClassification):
+class MultiHateClassification(MultilingualTask, AbsTaskClassification):
     metadata = TaskMetadata(
-        name="ItaHateClassification",
+        name="MultiHateClassification",
         dataset={
-            "path": "Paul/hatecheck-italian",
-            "revision": "21e3d5c827cb60619a89988b24979850a7af85a5",
+            "path": "Paul/hatecheck",
+            "revision": _REVISION_DICT,  # dynamic
         },
         description="""Hate speech detection dataset with binary
                        (hateful vs non-hateful) labels. Includes 25+ distinct types of hate
-                       and challenging non-hate. Multilingual datase released as 10 unilingual models
+                       and challenging non-hate, and 10 languages.
                      """,
         reference="https://aclanthology.org/2022.woah-1.15/",
         type="Classification",
         category="s2s",
         eval_splits=["test"],
-        eval_langs=["ita-Latn"],
+        eval_langs=_LANGUAGES,
         main_score="accuracy",
-        date=("2021-11-01", "2022-02-28"),
+        date=("2020-11-23", "2022-02-28"),
         form=["written"],
         domains=["Constructed"],
         task_subtypes=["Sentiment/Hate speech"],
@@ -54,20 +109,23 @@ class ItaHateClassification(AbsTaskClassification):
             abstract = "Hate speech detection models are typically evaluated on held-out test sets. However, this risks painting an incomplete and potentially misleading picture of model performance because of increasingly well-documented systematic gaps and biases in hate speech datasets. To enable more targeted diagnostic insights, recent research has thus introduced functional tests for hate speech detection models. However, these tests currently only exist for English-language content, which means that they cannot support the development of more effective models in other languages spoken by billions across the world. To help address this issue, we introduce Multilingual HateCheck (MHC), a suite of functional tests for multilingual hate speech detection models. MHC covers 34 functionalities across ten languages, which is more languages than any other hate speech dataset. To illustrate MHC{'}s utility, we train and test a high-performing multilingual hate speech detection model, and reveal critical model weaknesses for monolingual and cross-lingual applications.",
         }
         """,
-        n_samples={"test": 1845},
-        avg_character_length={"test": 50.4},
+        n_samples={"test": 18250},
+        avg_character_length={"test": 45.9},
     )
 
-    def dataset_transform(self):
-        keep_cols = ["test_case", "label_gold"]
-        rename_dict = dict(zip(keep_cols, ["text", "label"]))
-        remove_cols = [
-            col for col in self.dataset["test"].column_names if col not in keep_cols
-        ]
-        self.dataset = self.dataset.rename_columns(rename_dict)
-        self.dataset = self.dataset.class_encode_column("label")
-        self.dataset = self.dataset.class_encode_column("functionality")
-        self.dataset = self.dataset["test"].train_test_split(
-            test_size=0.5, seed=42, stratify_by_column="functionality"
-        )  # balanced sampling across types of hate speech
-        self.dataset = self.dataset.remove_columns(remove_cols)
+    def load_data(self, **kwargs):
+        """
+        Load dataset from HuggingFace hub
+        """
+        if self.data_loaded:
+            return
+        self.dataset = {}
+        for lang in self.langs:
+            metadata = self.metadata_dict.get("dataset", None)
+            path = f"{metadata['path']}-{_HF_AFFIX[lang]}".rstrip("-")
+            dataset = datasets.load_dataset(
+                path=path, revision=metadata["revision"][lang]
+            )
+            self.dataset[lang] = _transform(dataset)
+        self.dataset_transform()
+        self.data_loaded = True