pietrolesci
commited on
Commit
·
448e6d5
1
Parent(s):
cda6bc6
Create README.md
Browse files
README.md
ADDED
@@ -0,0 +1,156 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
## Overview
|
2 |
+
Original dataset [here](https://github.com/decompositional-semantics-initiative/DNC).
|
3 |
+
|
4 |
+
This dataset has been proposed in [Collecting Diverse Natural Language Inference Problems for Sentence Representation Evaluation](https://www.aclweb.org/anthology/D18-1007/).
|
5 |
+
|
6 |
+
|
7 |
+
## Dataset curation
|
8 |
+
This version of the dataset does not include the `type-of-inference` "KG" as its label set is
|
9 |
+
`[1, 2, 3, 4, 5]` while here we focus on NLI-related label sets, i.e. `[entailed, not-entailed]`.
|
10 |
+
For this reason, I named the dataset DNLI for _Diverse_ NLI, as in [Liu et al 2020](https://aclanthology.org/2020.conll-1.48/), instead of DNC.
|
11 |
+
|
12 |
+
This version of the dataset contains columns from the `*_data.json` and the `*_metadata.json` files available in the repo.
|
13 |
+
In the original repo, each data file has the following keys and values:
|
14 |
+
|
15 |
+
- `context`: The context sentence for the NLI pair. The context is already tokenized.
|
16 |
+
- `hypothesis`: The hypothesis sentence for the NLI pair. The hypothesis is already tokenized.
|
17 |
+
- `label`: The label for the NLI pair
|
18 |
+
- `label-set`: The set of possible labels for the specific NLI pair
|
19 |
+
- `binary-label`: A `True` or `False` label. See the paper for details on how we convert the `label` into a binary label.
|
20 |
+
- `split`: This can be `train`, `dev`, or `test`.
|
21 |
+
- `type-of-inference`: A string indicating what type of inference is tested in this example.
|
22 |
+
- `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
|
23 |
+
|
24 |
+
while each metadata file has the following columns
|
25 |
+
|
26 |
+
- `pair-id`: A unique integer id for the NLI pair.
|
27 |
+
- `corpus`: The original corpus where this example came from.
|
28 |
+
- `corpus-sent-id`: The id of the sentence (or example) in the original dataset that we recast.
|
29 |
+
- `corpus-license`: The license for the data from the original dataset.
|
30 |
+
- `creation-approach`: Determines the method used to recast this example. Options are `automatic`, `manual`, or `human-labeled`.
|
31 |
+
- `misc`: A dictionary of other relevant information. This is an optional field.
|
32 |
+
|
33 |
+
The files are merged on the `pair-id` key. I **do not** include the `misc` column as it is not essential for NLI.
|
34 |
+
|
35 |
+
NOTE: the label mapping is **not** the custom (i.e., 3 class) for NLI tasks. They used a binary target and I encoded them
|
36 |
+
with the following mapping `{"not-entailed": 0, "entailed": 1}`.
|
37 |
+
|
38 |
+
NOTE: some instances are present in multiple splits (matching performed by exact matching on "context", "hypothesis", and "label").
|
39 |
+
|
40 |
+
## Code to create the dataset
|
41 |
+
```python
|
42 |
+
import pandas as pd
|
43 |
+
from datasets import Dataset, ClassLabel, Value, Features, DatasetDict, Sequence
|
44 |
+
from pathlib import Path
|
45 |
+
|
46 |
+
|
47 |
+
paths = {
|
48 |
+
"train": "<path_to_folder>/DNC-master/train",
|
49 |
+
"dev": "<path_to_folder>/DNC-master/dev",
|
50 |
+
"test": "<path_to_folder>/DNC-master/test",
|
51 |
+
}
|
52 |
+
|
53 |
+
# read all data files
|
54 |
+
dfs = []
|
55 |
+
for split, path in paths.items():
|
56 |
+
for f_name in Path(path).rglob("*_data.json"):
|
57 |
+
df = pd.read_json(str(f_name))
|
58 |
+
df["file_split_data"] = split
|
59 |
+
dfs.append(df)
|
60 |
+
data = pd.concat(dfs, ignore_index=False, axis=0)
|
61 |
+
|
62 |
+
# read all metadata files
|
63 |
+
meta_dfs = []
|
64 |
+
for split, path in paths.items():
|
65 |
+
for f_name in Path(path).rglob("*_metadata.json"):
|
66 |
+
df = pd.read_json(str(f_name))
|
67 |
+
meta_dfs.append(df)
|
68 |
+
metadata = pd.concat(meta_dfs, ignore_index=False, axis=0)
|
69 |
+
|
70 |
+
# merge
|
71 |
+
dataset = pd.merge(data, metadata, on="pair-id", how="left")
|
72 |
+
|
73 |
+
# check that the split column reflects file splits
|
74 |
+
assert sum(dataset["split"] != dataset["file_split_data"]) == 0
|
75 |
+
dataset = dataset.drop(columns=["file_split_data"])
|
76 |
+
|
77 |
+
# fix `binary-label` column
|
78 |
+
dataset.loc[~dataset["label"].isin(["entailed", "not-entailed"]), "binary-label"] = False
|
79 |
+
dataset.loc[dataset["label"].isin(["entailed", "not-entailed"]), "binary-label"] = True
|
80 |
+
|
81 |
+
# fix datatype
|
82 |
+
dataset["corpus-sent-id"] = dataset["corpus-sent-id"].astype(str)
|
83 |
+
|
84 |
+
# order columns as shown in the README.md
|
85 |
+
columns = [
|
86 |
+
"context",
|
87 |
+
"hypothesis",
|
88 |
+
"label",
|
89 |
+
"label-set",
|
90 |
+
"binary-label",
|
91 |
+
"split",
|
92 |
+
"type-of-inference",
|
93 |
+
"pair-id",
|
94 |
+
"corpus",
|
95 |
+
"corpus-sent-id",
|
96 |
+
"corpus-license",
|
97 |
+
"creation-approach",
|
98 |
+
"misc",
|
99 |
+
]
|
100 |
+
dataset = dataset.loc[:, columns]
|
101 |
+
|
102 |
+
# remove misc column
|
103 |
+
dataset = dataset.drop(columns=["misc"])
|
104 |
+
|
105 |
+
# remove KG for NLI
|
106 |
+
dataset.loc[(dataset["label"].isin([1, 2, 3, 4, 5])), "type-of-inference"].value_counts()
|
107 |
+
# > the only split with label-set [1, 2, 3, 4, 5], so remove as we focus on NLI
|
108 |
+
dataset = dataset.loc[~(dataset["type-of-inference"] == "KG")]
|
109 |
+
|
110 |
+
# encode labels
|
111 |
+
dataset["label"] = dataset["label"].map({"not-entailed": 0, "entailed": 1})
|
112 |
+
|
113 |
+
# fill NA in label-set
|
114 |
+
dataset["label-set"] = dataset["label-set"].ffill()
|
115 |
+
|
116 |
+
features = Features(
|
117 |
+
{
|
118 |
+
"context": Value(dtype="string"),
|
119 |
+
"hypothesis": Value(dtype="string"),
|
120 |
+
"label": ClassLabel(num_classes=2, names=["not-entailed", "entailed"]),
|
121 |
+
"label-set": Sequence(length=2, feature=Value(dtype="string")),
|
122 |
+
"binary-label": Value(dtype="bool"),
|
123 |
+
"split": Value(dtype="string"),
|
124 |
+
"type-of-inference": Value(dtype="string"),
|
125 |
+
"pair-id": Value(dtype="int64"),
|
126 |
+
"corpus": Value(dtype="string"),
|
127 |
+
"corpus-sent-id": Value(dtype="string"),
|
128 |
+
"corpus-license": Value(dtype="string"),
|
129 |
+
"creation-approach": Value(dtype="string"),
|
130 |
+
}
|
131 |
+
)
|
132 |
+
|
133 |
+
dataset_splits = {}
|
134 |
+
for split in ("train", "dev", "test"):
|
135 |
+
df_split = dataset.loc[dataset["split"] == split]
|
136 |
+
dataset_splits[split] = Dataset.from_pandas(df_split, features=features)
|
137 |
+
|
138 |
+
dataset_splits = DatasetDict(dataset_splits)
|
139 |
+
dataset_splits.push_to_hub("pietrolesci/dnli", token="<your token>")
|
140 |
+
|
141 |
+
# check overlap between splits
|
142 |
+
from itertools import combinations
|
143 |
+
for i, j in combinations(dataset_splits.keys(), 2):
|
144 |
+
print(
|
145 |
+
f"{i} - {j}: ",
|
146 |
+
pd.merge(
|
147 |
+
dataset_splits[i].to_pandas(),
|
148 |
+
dataset_splits[j].to_pandas(),
|
149 |
+
on=["context", "hypothesis", "label"],
|
150 |
+
how="inner",
|
151 |
+
).shape[0],
|
152 |
+
)
|
153 |
+
#> train - dev: 127
|
154 |
+
#> train - test: 55
|
155 |
+
#> dev - test: 54
|
156 |
+
```
|