Create README.md

README.md

@@ -0,0 +1,156 @@
+## Overview
+Original dataset available [here](https://wellecks.github.io/dialogue_nli/).
+
+
+## Dataset curation
+Original `label` column is renamed `original_label`. The original classes are renamed as follows
+
+```
+{"positive": "entailment", "negative": "contradiction", "neutral": "neutral"})
+```
+
+and encoded with the following mapping
+
+```
+{"entailment": 0, "neutral": 1, "contradiction": 2}
+```
+
+and stored in the newly created column `label`.
+
+
+The following splits and the corresponding columns are present in the original files
+
+```
+train {'dtype', 'id', 'sentence1', 'sentence2', 'original_label', 'label', 'triple2', 'triple1'}
+dev {'dtype', 'id', 'sentence1', 'sentence2', 'original_label', 'label', 'triple2', 'triple1'}
+test {'dtype', 'id', 'sentence1', 'sentence2', 'original_label', 'label', 'triple2', 'triple1'}
+verified_test {'dtype', 'annotation3', 'sentence1', 'sentence2', 'annotation1', 'annotation2', 'original_label', 'label', 'triple2', 'triple1'}
+extra_test {'dtype', 'id', 'sentence1', 'sentence2', 'original_label', 'label', 'triple2', 'triple1'}
+extra_dev {'dtype', 'id', 'sentence1', 'sentence2', 'original_label', 'label', 'triple2', 'triple1'}
+extra_train {'dtype', 'id', 'sentence1', 'sentence2', 'original_label', 'label', 'triple2', 'triple1'}
+valid_havenot {'dtype', 'id', 'sentence1', 'sentence2', 'original_label', 'label', 'triple2', 'triple1'}
+valid_attributes {'dtype', 'id', 'sentence1', 'sentence2', 'original_label', 'label', 'triple2', 'triple1'}
+valid_likedislike {'dtype', 'id', 'sentence1', 'sentence2', 'original_label', 'label', 'triple2', 'triple1'}
+```
+
+Note that I only keep the common columns, which means that I drop "annotation{1, 2, 3" from `verified_test`.
+Note that there are some splits with the same instances, as found by matching on "original_label", "sentence1", "sentence2".
+
+
+## Code to create dataset
+```python
+import pandas as pd
+from pathlib import Path
+import json
+from datasets import Features, Value, ClassLabel, Dataset, DatasetDict, Sequence
+
+
+# load data
+ds = {}
+for path in Path(".").rglob("<path to folder>/*.jsonl"):
+    print(path, flush=True)
+    with path.open("r") as fl:
+        data = fl.read()
+    try:
+        d = json.loads(data, encoding="utf-8")
+    except json.JSONDecodeError as error:
+        print(error)
+    
+    df = pd.DataFrame(d)
+
+    # encode labels
+    df["original_label"] = df["label"]
+    df["label"] = df["label"].map({"positive": "entailment", "negative": "contradiction", "neutral": "neutral"})
+    df["label"] = df["label"].map({"entailment": 0, "neutral": 1, "contradiction": 2})
+
+    ds[path.name.split(".")[0]] = df
+    
+    
+# prettify names of data splits
+datasets = {
+    k.replace("dialogue_nli_", "").replace("uu_", "").lower(): v
+    for k, v in ds.items()
+}
+datasets.keys()
+#> dict_keys(['train', 'dev', 'test', 'verified_test', 'extra_test', 'extra_dev', 'extra_train', 'valid_havenot', 'valid_attributes', 'valid_likedislike'])
+
+
+# cast to datasets using only common columns
+features = Features({
+    "label": ClassLabel(num_classes=3, names=["entailment", "neutral", "contradiction"]),
+    "sentence1": Value(dtype="string", id=None),
+    "sentence2": Value(dtype="string", id=None),
+    "triple1": Sequence(feature=Value(dtype="string", id=None), length=3),
+    "triple2": Sequence(feature=Value(dtype="string", id=None), length=3),
+    "dtype": Value(dtype="string", id=None),
+    "id": Value(dtype="string", id=None),
+    "original_label": Value(dtype="string", id=None),
+})
+
+ds = {}
+for name, df in datasets.items():
+    if "id" not in df.columns:
+        df["id"] = ""
+    ds[name] = Dataset.from_pandas(df.loc[:, list(features.keys())], features=features)
+ds = DatasetDict(ds)
+ds.push_to_hub("dialogue_nli", token="<token>")
+
+
+# check overlap between splits
+from itertools import combinations
+for i, j in combinations(ds.keys(), 2):
+    print(
+        f"{i} - {j}: ",
+        pd.merge(
+            ds[i].to_pandas(), 
+            ds[j].to_pandas(), 
+            on=["original_label", "sentence1", "sentence2"], 
+            how="inner",
+        ).shape[0],
+    )
+#> train - dev:  58
+#> train - test:  98
+#> train - verified_test:  90
+#> train - extra_test:  0
+#> train - extra_dev:  0
+#> train - extra_train:  0
+#> train - valid_havenot:  0
+#> train - valid_attributes:  0
+#> train - valid_likedislike:  0
+#> dev - test:  19
+#> dev - verified_test:  19
+#> dev - extra_test:  0
+#> dev - extra_dev:  75
+#> dev - extra_train:  75
+#> dev - valid_havenot:  75
+#> dev - valid_attributes:  75
+#> dev - valid_likedislike:  75
+#> test - verified_test:  12524
+#> test - extra_test:  34
+#> test - extra_dev:  0
+#> test - extra_train:  0
+#> test - valid_havenot:  0
+#> test - valid_attributes:  0
+#> test - valid_likedislike:  0
+#> verified_test - extra_test:  29
+#> verified_test - extra_dev:  0
+#> verified_test - extra_train:  0
+#> verified_test - valid_havenot:  0
+#> verified_test - valid_attributes:  0
+#> verified_test - valid_likedislike:  0
+#> extra_test - extra_dev:  0
+#> extra_test - extra_train:  0
+#> extra_test - valid_havenot:  0
+#> extra_test - valid_attributes:  0
+#> extra_test - valid_likedislike:  0
+#> extra_dev - extra_train:  250946
+#> extra_dev - valid_havenot:  250946
+#> extra_dev - valid_attributes:  250946
+#> extra_dev - valid_likedislike:  250946
+#> extra_train - valid_havenot:  250946
+#> extra_train - valid_attributes:  250946
+#> extra_train - valid_likedislike:  250946
+#> valid_havenot - valid_attributes:  250946
+#> valid_havenot - valid_likedislike:  250946
+#> valid_attributes - valid_likedislike:  250946
+```