init

README.md

@@ -0,0 +1,72 @@
+---
+language:
+- en
+license:
+- other
+multilinguality:
+- monolingual
+size_categories:
+- 1K<n<10K
+pretty_name: ConceptNet
+---
+# Dataset Card for "relbert/conceptnet"
+## Dataset Description
+- **Repository:** [RelBERT](https://github.com/asahi417/relbert)
+- **Paper:** [https://ojs.aaai.org/index.php/AAAI/article/view/11164](https://ojs.aaai.org/index.php/AAAI/article/view/11164)
+- **Dataset:** ConceptNet5
+
+### Dataset Summary
+ConceptNet5, which compiled to fine-tune [RelBERT](https://github.com/asahi417/relbert) model.
+
+## Dataset Structure
+### Data Instances
+An example of `train` looks as follows.
+```
+{
+    "relation_type": "AtLocation",
+    "positives": [["fish", "water"], ["cloud", "sky"], ["child", "school"], ... ],
+    "negatives": [["pen", "write"], ["sex", "fun"], ["soccer", "sport"], ["fish", "school"], ... ]
+}
+```
+
+### Data Splits
+|  name   |train|validation|
+|---------|----:|---------:|
+|conceptnet| 25 |      24|
+
+### Number of Positive/Negative Word-pairs in each Split
+
+| relation_type    |   positive (train) |   negative (train) |   positive (validation) |   negative (validation) |
+|:-----------------|-------------------:|-------------------:|------------------------:|------------------------:|
+| AtLocation       |                383 |                356 |                      97 |                      80 |
+| CapableOf        |                195 |                190 |                      73 |                      78 |
+| Causes           |                 71 |                 73 |                      26 |                      26 |
+| CausesDesire     |                  9 |                  7 |                      11 |                      11 |
+| CreatedBy        |                  2 |                  3 |                       0 |                       0 |
+| DefinedAs        |                  0 |                  0 |                       2 |                       2 |
+| Desires          |                 16 |                 15 |                      12 |                      12 |
+| HasA             |                 67 |                 86 |                      17 |                      17 |
+| HasFirstSubevent |                  2 |                  3 |                       0 |                       0 |
+| HasLastSubevent  |                  2 |                  3 |                       3 |                       1 |
+| HasPrerequisite  |                168 |                176 |                      57 |                      54 |
+| HasProperty      |                 94 |                100 |                      39 |                      49 |
+| HasSubevent      |                125 |                128 |                      40 |                      54 |
+| IsA              |                310 |                279 |                      98 |                     106 |
+| MadeOf           |                 17 |                 15 |                       7 |                       5 |
+| MotivatedByGoal  |                 14 |                 15 |                      11 |                      11 |
+| NotCapableOf     |                 15 |                 13 |                       0 |                       0 |
+| NotDesires       |                  4 |                  4 |                       4 |                       1 |
+| PartOf           |                 34 |                 40 |                       7 |                       5 |
+| ReceivesAction   |                 18 |                 16 |                       8 |                       6 |
+| SymbolOf         |                  0 |                  0 |                       2 |                       3 |
+| UsedFor          |                249 |                269 |                      81 |                      74 |
+
+### Citation Information
+```
+@inproceedings{speer2017conceptnet,
+  title={Conceptnet 5.5: An open multilingual graph of general knowledge},
+  author={Speer, Robyn and Chin, Joshua and Havasi, Catherine},
+  booktitle={Thirty-first AAAI conference on artificial intelligence},
+  year={2017}
+}
+```

process.py

@@ -0,0 +1,52 @@
+import os
+import json
+from tqdm import tqdm
+
+from datasets import load_dataset
+
+export_dir = 'dataset'
+os.makedirs(export_dir, exist_ok=True)
+dataset = load_dataset("conceptnet5", "conceptnet5", split="train")
+
+
+def check(example):
+    if example['sentence'] == '':
+        return False
+    if example['lang'] != 'en':
+        return False
+    if example['rel'] == 'None':
+        return False
+    atom_1 = os.path.basename(example['arg1'])
+    atom_2 = os.path.basename(example['arg2'])
+    for atom in [atom_1, atom_2]:
+        if len(atom) <= 2:  # condition on the number of characters
+            return False
+        if len(atom.split(' ')) != 1:  # condition on the number of words
+            return False
+        if len(atom.split('_')) != 1:  # condition on the number of words
+            return False
+    return True
+
+
+dataset = dataset.filter(lambda example: check(example))
+relations = list(set(dataset["rel"]))
+
+for r in tqdm(relations):
+
+    _dataset = dataset.filter(lambda example: example['rel'] == r)
+    _dataset = _dataset.shuffle(0)
+    train_size = int(len(_dataset) * 0.7)
+
+    with open(f"{export_dir}/train.{os.path.basename(r)}.jsonl", 'w') as f:
+        f.write(json.dumps({
+            'relation_type': os.path.basename(r),
+            'positives': [[i['arg1'], i['arg2']] for i in _dataset[:train_size]],
+            'negatives': []
+        }))
+
+    with open(f"{export_dir}/valid.{os.path.basename(r)}.jsonl", 'w') as f:
+        f.write(json.dumps({
+            'relation_type': os.path.basename(r),
+            'positives': [[i['arg1'], i['arg2']] for i in _dataset[train_size:]],
+            'negatives': []
+        }))