dstc3

This is a text classification dataset. It is intended for machine learning research and experimentation.

This dataset is obtained via formatting another publicly available data to be compatible with our AutoIntent Library.

Usage

It is intended to be used with our AutoIntent Library:

from autointent import Dataset

dstc3 = Dataset.from_hub("AutoIntent/dstc3")

Source

This dataset is taken from marcel-gohsen/dstc3 and formatted with our AutoIntent Library:

import datasets
from autointent import Dataset
from autointent.context.data_handler import split_dataset


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
    )