pietrolesci/dnc

Overview

Original dataset here.

This dataset has been proposed in Collecting Diverse Natural Language Inference Problems for Sentence Representation Evaluation.

Dataset curation

This version of the dataset does not include the type-of-inference "KG" as its label set is [1, 2, 3, 4, 5] while here we focus on NLI-related label sets, i.e. [entailed, not-entailed]. For this reason, I named the dataset DNLI for Diverse NLI, as in Liu et al 2020, instead of DNC.

This version of the dataset contains columns from the *_data.json and the *_metadata.json files available in the repo. In the original repo, each data file has the following keys and values:

• context: The context sentence for the NLI pair. The context is already tokenized.
• hypothesis: The hypothesis sentence for the NLI pair. The hypothesis is already tokenized.
• label: The label for the NLI pair
• label-set: The set of possible labels for the specific NLI pair
• binary-label: A True or False label. See the paper for details on how we convert the label into a binary label.
• split: This can be train, dev, or test.
• type-of-inference: A string indicating what type of inference is tested in this example.
• pair-id: A unique integer id for the NLI pair. The pair-id is used to find the corresponding metadata for any given NLI pair

while each metadata file has the following columns

• pair-id: A unique integer id for the NLI pair.
• corpus: The original corpus where this example came from.
• corpus-sent-id: The id of the sentence (or example) in the original dataset that we recast.
• corpus-license: The license for the data from the original dataset.
• creation-approach: Determines the method used to recast this example. Options are automatic, manual, or human-labeled.
• misc: A dictionary of other relevant information. This is an optional field.

The files are merged on the pair-id key. I do not include the misc column as it is not essential for NLI.

NOTE: the label mapping is not the custom (i.e., 3 class) for NLI tasks. They used a binary target and I encoded them with the following mapping {"not-entailed": 0, "entailed": 1}.

NOTE: some instances are present in multiple splits (matching performed by exact matching on "context", "hypothesis", and "label").

Code to create the dataset

import pandas as pd
from datasets import Dataset, ClassLabel, Value, Features, DatasetDict, Sequence
from pathlib import Path


paths = {
    "train": "<path_to_folder>/DNC-master/train",
    "dev": "<path_to_folder>/DNC-master/dev",
    "test": "<path_to_folder>/DNC-master/test",
}

# read all data files
dfs = []
for split, path in paths.items():
    for f_name in Path(path).rglob("*_data.json"):
        df = pd.read_json(str(f_name))
        df["file_split_data"] = split
        dfs.append(df)
data = pd.concat(dfs, ignore_index=False, axis=0)

# read all metadata files
meta_dfs = []
for split, path in paths.items():
    for f_name in Path(path).rglob("*_metadata.json"):
        df = pd.read_json(str(f_name))
        meta_dfs.append(df)
metadata = pd.concat(meta_dfs, ignore_index=False, axis=0)

# merge
dataset = pd.merge(data, metadata, on="pair-id", how="left")

# check that the split column reflects file splits
assert sum(dataset["split"] != dataset["file_split_data"]) == 0
dataset = dataset.drop(columns=["file_split_data"])

# fix `binary-label` column
dataset.loc[~dataset["label"].isin(["entailed", "not-entailed"]), "binary-label"] = False
dataset.loc[dataset["label"].isin(["entailed", "not-entailed"]), "binary-label"] = True

# fix datatype
dataset["corpus-sent-id"] = dataset["corpus-sent-id"].astype(str)

# order columns as shown in the README.md
columns = [
    "context",
    "hypothesis",
    "label",
    "label-set",
    "binary-label",
    "split",
    "type-of-inference",
    "pair-id",
    "corpus",
    "corpus-sent-id",
    "corpus-license",
    "creation-approach",
    "misc",
]
dataset = dataset.loc[:, columns]

# remove misc column
dataset = dataset.drop(columns=["misc"])

# remove KG for NLI
dataset.loc[(dataset["label"].isin([1, 2, 3, 4, 5])), "type-of-inference"].value_counts()
# > the only split with label-set [1, 2, 3, 4, 5], so remove as we focus on NLI
dataset = dataset.loc[~(dataset["type-of-inference"] == "KG")]

# encode labels
dataset["label"] = dataset["label"].map({"not-entailed": 0, "entailed": 1})

# fill NA in label-set
dataset["label-set"] = dataset["label-set"].ffill()

features = Features(
    {
        "context": Value(dtype="string"),
        "hypothesis": Value(dtype="string"),
        "label": ClassLabel(num_classes=2, names=["not-entailed", "entailed"]),
        "label-set": Sequence(length=2, feature=Value(dtype="string")),
        "binary-label": Value(dtype="bool"),
        "split": Value(dtype="string"),
        "type-of-inference": Value(dtype="string"),
        "pair-id": Value(dtype="int64"),
        "corpus": Value(dtype="string"),
        "corpus-sent-id": Value(dtype="string"),
        "corpus-license": Value(dtype="string"),
        "creation-approach": Value(dtype="string"),
    }
)

dataset_splits = {}
for split in ("train", "dev", "test"):
    df_split = dataset.loc[dataset["split"] == split]
    dataset_splits[split] = Dataset.from_pandas(df_split, features=features)

dataset_splits = DatasetDict(dataset_splits)
dataset_splits.push_to_hub("pietrolesci/dnli", token="<your token>")

# check overlap between splits
from itertools import combinations
for i, j in combinations(dataset_splits.keys(), 2):
    print(
        f"{i} - {j}: ",
        pd.merge(
            dataset_splits[i].to_pandas(), 
            dataset_splits[j].to_pandas(), 
            on=["context", "hypothesis", "label"], 
            how="inner",
        ).shape[0],
    )
#> train - dev:  127
#> train - test:  55
#> dev - test:  54