hpprc commited on
Commit
295bf19
1 Parent(s): 6b1df67

:sparkles: Update documents and improve code

Browse files
Files changed (2) hide show
  1. README.md +41 -28
  2. jsick.py +2 -2
README.md CHANGED
@@ -12,7 +12,7 @@ multilinguality:
12
  - translation
13
  pretty_name: JSICK
14
  size_categories:
15
- - 1K<n<10K
16
  source_datasets:
17
  - extended|sick
18
  tags:
@@ -207,35 +207,31 @@ A version adopting the column names of a typical NLI dataset.
207
 
208
  ### Data Splits
209
 
210
- | name | train | validation | test |
211
- | --------------- | ----: | ---------: | ---: |
212
- | base | 4500 | | 4927 |
213
- | original | 4500 | | 4927 |
214
- | stress | | | 900 |
215
- | stress-original | | | 900 |
216
-
217
 
218
 
219
  ### Annotations
220
 
221
- The annotation process for this Japanese NLI dataset involves tagging each pair (P, H) of a premise and hypothesis with a label for structural pattern and linguistic phenomenon.
222
- The structural relationship between premise and hypothesis sentences is classified into five patterns, with each pattern associated with a type of heuristic that can lead to incorrect predictions of the entailment relation.
223
- Additionally, 11 categories of Japanese linguistic phenomena and constructions are focused on for generating the five patterns of adversarial inferences.
224
-
225
- For each linguistic phenomenon, a template for the premise sentence P is fixed, and multiple templates for hypothesis sentences H are created.
226
- In total, 144 templates for (P, H) pairs are produced.
227
- Each pair of premise and hypothesis sentences is tagged with an entailment label (entailment or non-entailment), a structural pattern, and a linguistic phenomenon label.
228
 
229
- The JaNLI dataset is generated by instantiating each template 100 times, resulting in a total of 14,400 examples.
230
- The same number of entailment and non-entailment examples are generated for each phenomenon.
231
- The structural patterns are annotated with the templates for each linguistic phenomenon, and the ratio of entailment and non-entailment examples is not necessarily 1:1 for each pattern.
232
- The dataset uses a total of 158 words (nouns and verbs), which occur more than 20 times in the JSICK and JSNLI datasets.
233
 
234
 
235
  ## Additional Information
236
 
237
- - [verypluming/JaNLI](https://github.com/verypluming/JaNLI)
238
- - [Hitomi Yanaka, Koji Mineshima, Assessing the Generalization Capacity of Pre-trained Language Models through Japanese Adversarial Natural Language Inference, Proceedings of the 2021 EMNLP Workshop BlackboxNLP: Analyzing and Interpreting Neural Networks for NLP (BlackboxNLP2021), 2021.](https://aclanthology.org/2021.blackboxnlp-1.26/)
 
239
 
240
  ### Licensing Information
241
 
@@ -244,15 +240,32 @@ CC BY-SA 4.0
244
  ### Citation Information
245
 
246
  ```bibtex
247
- @InProceedings{yanaka-EtAl:2021:blackbox,
248
- author = {Yanaka, Hitomi and Mineshima, Koji},
249
- title = {Assessing the Generalization Capacity of Pre-trained Language Models through Japanese Adversarial Natural Language Inference},
250
- booktitle = {Proceedings of the 2021 EMNLP Workshop BlackboxNLP: Analyzing and Interpreting Neural Networks for NLP (BlackboxNLP2021)},
251
- url = {https://aclanthology.org/2021.blackboxnlp-1.26/},
252
- year = {2021},
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
253
  }
254
  ```
255
 
256
  ### Contributions
257
 
258
- Thanks to [Hitomi Yanaka](https://hitomiyanaka.mystrikingly.com/) and Koji Mineshima for creating this dataset.
 
12
  - translation
13
  pretty_name: JSICK
14
  size_categories:
15
+ - 10K<n<100K
16
  source_datasets:
17
  - extended|sick
18
  tags:
 
207
 
208
  ### Data Splits
209
 
210
+ | name | train | validation | test |
211
+ | --------------- | ----: | ---------: | ----: |
212
+ | base | 4,500 | | 4,927 |
213
+ | original | 4,500 | | 4,927 |
214
+ | stress | | | 900 |
215
+ | stress-original | | | 900 |
 
216
 
217
 
218
  ### Annotations
219
 
220
+ To annotate the JSICK dataset, they used the crowdsourcing platform "Lancers" to re-annotate entailment labels and similarity scores for JSICK.
221
+ They had six native Japanese speakers as annotators, who were randomly selected from the platform.
222
+ The annotators were asked to fully understand the guidelines and provide the same labels as gold labels for ten test questions.
 
 
 
 
223
 
224
+ For entailment labels, they adopted annotations that were agreed upon by a majority vote as gold labels and checked whether the majority judgment vote was semantically valid for each example.
225
+ For similarity scores, they used the average of the annotation results as gold scores.
226
+ The raw annotations with the JSICK dataset are [publicly available](https://github.com/verypluming/JSICK/blob/main/jsick/jsick-all-annotations.tsv).
227
+ The average annotation time was 1 minute per pair, and Krippendorff's alpha for the entailment labels was 0.65.
228
 
229
 
230
  ## Additional Information
231
 
232
+ - [verypluming/JSICK](https://github.com/verypluming/JSICK)
233
+ - [Compositional Evaluation on Japanese Textual Entailment and Similarity](https://direct.mit.edu/tacl/article/doi/10.1162/tacl_a_00518/113850/Compositional-Evaluation-on-Japanese-Textual)
234
+ - [JSICK: 日本語構成的推論・類似度データセットの構築](https://www.jstage.jst.go.jp/article/pjsai/JSAI2021/0/JSAI2021_4J3GS6f02/_article/-char/ja)
235
 
236
  ### Licensing Information
237
 
 
240
  ### Citation Information
241
 
242
  ```bibtex
243
+ @article{yanaka-mineshima-2022-compositional,
244
+ title = "Compositional Evaluation on {J}apanese Textual Entailment and Similarity",
245
+ author = "Yanaka, Hitomi and
246
+ Mineshima, Koji",
247
+ journal = "Transactions of the Association for Computational Linguistics",
248
+ volume = "10",
249
+ year = "2022",
250
+ address = "Cambridge, MA",
251
+ publisher = "MIT Press",
252
+ url = "https://aclanthology.org/2022.tacl-1.73",
253
+ doi = "10.1162/tacl_a_00518",
254
+ pages = "1266--1284",
255
+ }
256
+
257
+ @article{谷中 瞳2021,
258
+ title={JSICK: 日本語構成的推論・類似度データセットの構築},
259
+ author={谷中 瞳 and 峯島 宏次},
260
+ journal={人工知能学会全国大会論文集},
261
+ volume={JSAI2021},
262
+ number={ },
263
+ pages={4J3GS6f02-4J3GS6f02},
264
+ year={2021},
265
+ doi={10.11517/pjsai.JSAI2021.0_4J3GS6f02}
266
  }
267
  ```
268
 
269
  ### Contributions
270
 
271
+ Thanks to [Hitomi Yanaka](https://hitomiyanaka.mystrikingly.com/) and [Koji Mineshima](https://abelard.flet.keio.ac.jp/person/minesima/index-j.html) for creating this dataset.
jsick.py CHANGED
@@ -51,12 +51,12 @@ class JSICKDataset(ds.GeneratorBasedBuilder):
51
  ds.BuilderConfig(
52
  name="stress",
53
  version=VERSION,
54
- description="fuga",
55
  ),
56
  ds.BuilderConfig(
57
  name="stress-original",
58
  version=VERSION,
59
- description="fuga",
60
  ),
61
  ]
62
 
 
51
  ds.BuilderConfig(
52
  name="stress",
53
  version=VERSION,
54
+ description="The dataset to investigate whether models capture word order and case particles in Japanese.",
55
  ),
56
  ds.BuilderConfig(
57
  name="stress-original",
58
  version=VERSION,
59
+ description="The original version of JSICK-stress Test set retaining the unaltered column names.",
60
  ),
61
  ]
62