Update README.md

README.md

@@ -73,43 +73,66 @@ dstc3 = Dataset.from_hub("AutoIntent/dstc3")
 This dataset is taken from `marcel-gohsen/dstc3` 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
 from autointent.context.data_handler import split_dataset
 
-# load original data
-dstc3 = load_dataset("marcel-gohsen/dstc3")
-
-# extract intent names
-dstc3["test"] = dstc3["test"].filter(lambda example: example["transcript"] != "")
-intent_names = sorted(
-    set(name for intents in dstc3["test"]["intent"] for name in intents)
-)
-intent_names.remove("reqmore")
-dstc3["test"].filter(lambda example: "reqmore" in example["intent"])
-name_to_id = {name: i for i, name in enumerate(intent_names)}
-
-# parse complicated dstc format
-def transform(example: dict):
-    return {
-        "utterance": example["transcript"],
-        "label": [int(name in example["intent"]) for name in intent_names],
-    }
-dstc_converted = dstc3["test"].map(
-    transform, remove_columns=dstc3["test"].features.keys()
-)
-
-# format to autointent.Dataset
-intents = [{"id": i, "name": name} for i, name in enumerate(intent_names)]
-utterances = []
-for rec in dstc_converted.to_list():
-    if sum(rec["label"]) == 0:
-        rec.pop("label")
-    else:
-        utterances.append(rec)
-
-dstc_converted = Dataset.from_dict({"intents": intents, "train": utterances})
-dstc_converted["train"], dstc_converted["test"] = split_dataset(
-    dstc_converted, split="train", test_size=0.2, random_seed=42
-)
+
+def extract_intent_info(ds: datasets.Dataset) -> list[str]:
+    ds = ds.filter(lambda example: example["transcript"] != "")
+    intent_names = sorted(
+        set(name for intents in ds["intent"] for name in intents)
+    )
+    intent_names.remove("reqmore")
+    ds.filter(lambda example: "reqmore" in example["intent"])
+    return intent_names
+
+def parse(ds: datasets.Dataset, intent_names: list[str]):
+    def transform(example: dict):
+        return {
+            "utterance": example["transcript"],
+            "label": [int(name in example["intent"]) for name in intent_names],
+        }
+    return ds.map(
+        transform, remove_columns=ds.features.keys()
+    )
+
+def calc_fractions(ds: datasets.Dataset, intent_names: list[str]) -> list[float]:
+    res = [0] * len(intent_names)
+    for sample in ds:
+        for i, indicator in enumerate(sample["label"]):
+            res[i] += indicator
+    for i in range(len(intent_names)):
+        res[i] /= len(ds)
+    return res
+
+def remove_low_resource_classes(ds: datasets.Dataset, intent_names: list[str], fraction_thresh: float = 0.01) -> tuple[list[dict], list[str]]:
+    remove_or_not = [(frac < fraction_thresh) for frac in calc_fractions(ds, intent_names)]
+    intent_names = [name for i, name in enumerate(intent_names) if not remove_or_not[i]]
+    res = []
+    for sample in ds:
+        if sum(sample["label"]) == 1 and remove_or_not[sample["label"].index(1)]:
+            continue
+        sample["label"] = [
+            indicator for indicator, low_resource in
+            zip(sample["label"], remove_or_not, strict=True) if not low_resource
+        ]
+        res.append(sample)
+    return res, intent_names
+
+def remove_oos(ds: datasets.Dataset):
+    return ds.filter(lambda sample: sum(sample["label"]) != 0)
+
+
+if __name__ == "__main__":
+    dstc3 = datasets.load_dataset("marcel-gohsen/dstc3")
+    
+    intent_names = extract_intent_info(dstc3["test"])
+    parsed = parse(dstc3["test"], intent_names)
+    filtered, intent_names = remove_low_resource_classes(remove_oos(parsed), intent_names)
+    intents = [{"id": i, "name": name} for i, name in enumerate(intent_names)]
+    dstc_final = Dataset.from_dict({"intents": intents, "train": filtered})
+    dstc_final["train"], dstc_final["test"] = split_dataset(
+        dstc_final, split="train", test_size=0.2, random_seed=42
+    )
 ```