Update README.md

README.md

@@ -78,22 +78,64 @@ eurlex = Dataset.from_hub("AutoIntent/eurlex")
 This dataset is taken from `coastalcph/multi_eurlex` and formatted with our [AutoIntent Library](https://deeppavlov.github.io/AutoIntent/index.html):
 
 ```python
-from datasets import load_dataset
+import datasets
 from autointent import Dataset
 
-eurlex = load_dataset("coastalcph/multi_eurlex", "en", trust_remote_code=True)
-labels = []
-def transform(example: dict):
-    for intent in example["labels"]:
-        labels.append(intent)
-    return {"utterance": example["text"], "label": example["labels"]}
-
-labels = [{"id": label, "name": None} for label in set(labels)]
-multilabel_eurlex_train = eurlex["train"].map(transform, remove_columns=eurlex["train"].features.keys())
-multilabel_eurlex_test = eurlex["test"].map(transform, remove_columns=eurlex["test"].features.keys())
-eurlex_converted = Dataset.from_dict({
-    "intents": labels,
-    "test": multilabel_eurlex_test.to_list(),
-    "train": multilabel_eurlex_train.to_list()
-})
+
+def get_number_of_classes(ds: datasets.Dataset) -> int:
+    return len(set(i for example in ds for labels in example for i in labels))
+
+
+def parse(ds: datasets.Dataset, n_classes: int) -> datasets.Dataset:
+    def transform(example: dict):
+        return {"utterance": example["text"], "label": [int(i in example["labels"]) for i in range(n_classes)]}
+    return ds.map(transform, remove_columns=ds.features.keys())
+
+
+def get_low_resource_classes_mask(ds: datasets.Dataset, n_classes: int, fraction_thresh: float = 0.01) -> list[bool]:
+    res = [0] * n_classes
+    for sample in ds:
+        for i, indicator in enumerate(sample["label"]):
+            res[i] += indicator
+    for i in range(n_classes):
+        res[i] /= len(ds)
+    return [(frac < fraction_thresh) for frac in res]
+
+
+def remove_low_resource_classes(ds: datasets.Dataset, mask: list[bool]) -> list[dict]:
+    res = []
+    for sample in ds:
+        if sum(sample["label"]) == 1 and mask[sample["label"].index(1)]:
+            continue
+        sample["label"] = [
+            indicator for indicator, low_resource in
+            zip(sample["label"], mask, strict=True) if not low_resource
+        ]
+        res.append(sample)
+    return res
+
+
+def remove_oos(ds: list[dict]):
+    return [sample for sample in ds if sum(sample["label"]) != 0]
+
+
+if __name__ == "__main__":
+    eurlex = datasets.load_dataset("coastalcph/multi_eurlex", "en", trust_remote_code=True)
+
+    n_classes = get_number_of_classes(eurlex["train"])
+
+    train = parse(eurlex["train"], n_classes)
+    test = parse(eurlex["test"], n_classes)
+    validation = parse(eurlex["validation"], n_classes)
+
+    mask = get_low_resource_classes_mask(train, n_classes)
+    train = remove_oos(remove_low_resource_classes(train, mask))
+    test = remove_oos(remove_low_resource_classes(test, mask))
+    validation = remove_oos(remove_low_resource_classes(validation, mask))
+
+    eurlex_converted = Dataset.from_dict({
+        "train": train,
+        "test": test,
+        "validation": validation,
+    })
 ```