Datasets:
GEM
/

Tasks:
Other
Multilinguality:
unknown
Size Categories:
unknown
Language Creators:
unknown
Annotations Creators:
expert-created
Source Datasets:
original
License:
Sebastian Gehrmann commited on
Commit
6b2ce9e
1 Parent(s): 2d96299

data card.

Browse files
Files changed (1) hide show
  1. README.md +561 -232
README.md CHANGED
@@ -1,177 +1,272 @@
1
  ---
2
  annotations_creators:
3
- - expert-generated
4
- - machine-generated
5
  language_creators:
6
- - crowdsourced
7
  languages:
8
- - de
9
- - en
10
- - fi
11
- - fr
12
- - ru
13
- - sv
14
  licenses:
15
  - cc-by-nc-4.0
16
  multilinguality:
17
- - multilingual
18
- pretty_name: Opusparcus
19
  size_categories:
20
  - unknown
21
  source_datasets:
22
- - extended|open_subtitles
23
  task_categories:
24
- - conditional-text-generation
25
  task_ids:
26
- - conditional-text-generation-other-paraphrase generation
27
  ---
28
 
29
-
30
- # Dataset Card for Opusparcus
31
 
32
- **NB: This is the old format of the data set card. Generate a new one from the json in the repository!**
33
 
34
- ## Table of Contents
 
 
 
 
35
 
36
- - [Dataset Description](#dataset-description)
37
 
38
- - [Dataset Summary](#dataset-summary)
39
 
40
- - [Supported Tasks and Leaderboards](#supported-tasks-and-leaderboards)
41
 
42
- - [Languages](#languages)
43
 
44
- - [Dataset Structure](#dataset-structure)
 
 
 
 
 
45
 
46
- - [Data Instances](#data-instances)
 
47
 
48
- - [Data Fields](#data-fields)
 
49
 
50
- - [Data Splits](#data-splits)
51
 
52
- - [Dataset Creation](#dataset-creation)
53
 
54
- - [Curation Rationale](#curation-rationale)
55
 
56
- - [Source Data](#source-data)
 
 
57
 
58
- - [Annotations](#annotations)
59
 
60
- - [Personal and Sensitive Information](#personal-and-sensitive-information)
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
61
 
62
- - [Considerations for Using the Data](#considerations-for-using-the-data)
63
 
64
- - [Social Impact of Dataset](#social-impact-of-dataset)
 
 
 
65
 
66
- - [Discussion of Biases](#discussion-of-biases)
67
 
68
- - [Other Known Limitations](#other-known-limitations)
 
 
69
 
70
- - [Additional Information](#additional-information)
71
 
72
- - [Dataset Curators](#dataset-curators)
 
 
73
 
74
- - [Licensing Information](#licensing-information)
75
 
76
- - [Citation Information](#citation-information)
77
 
78
- - [Contributions](#contributions)
79
 
80
- ## Dataset Description
 
 
 
81
 
82
- - **Repository:** [Language Bank of Finland – Metashare](http://urn.fi/urn:nbn:fi:lb-2018021221)
83
 
84
- - **Paper:** [Mathias Creutz (2018): Open Subtitles Paraphrases Corpus For Six Languages](http://www.lrec-conf.org/proceedings/lrec2018/summaries/131.html)
 
 
 
85
 
86
- - **Point of Contact:** Mathias Creutz (firstname dot lastname at helsinki dot fi)
87
 
88
- ### Dataset Summary
 
 
89
 
90
- Opusparcus is a paraphrase corpus for six European languages: German,
91
- English, Finnish, French, Russian, and Swedish. The paraphrases
92
- consist of subtitles from movies and TV shows.
93
 
94
- The data in Opusparcus has been extracted from
95
- [OpenSubtitles2016](http://opus.nlpl.eu/OpenSubtitles2016.php), which
96
- is in turn based on data from http://www.opensubtitles.org/.
97
 
98
- For each target language, the Opusparcus data have been partitioned
99
- into three types of data sets: training, validation and test sets. The
100
- training sets are large, consisting of millions of sentence pairs, and
101
- have been compiled automatically, with the help of probabilistic
102
- ranking functions. The development and test sets consist of sentence
103
- pairs that have been annotated manually; each set contains
104
- approximately 1000 sentence pairs that have been verified to be
105
- acceptable paraphrases by two indepedent annotators.
106
 
107
- ### Supported Tasks and Leaderboards
108
 
109
- **Tasks:** Paraphrase detection and generation
 
 
110
 
111
- **Leaderboards:** Currently there is no Leaderboard for this dataset.
112
 
113
- ### Languages
 
 
114
 
115
- German (de), English (en), Finnish (fi), French (fr), Russian (ru), Swedish (sv)
116
 
117
- ## Dataset Structure
 
 
 
118
 
119
- When you download Opusparcus, you must always indicate the language
120
- you want to retrieve, for instance:
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
121
 
122
  ```
123
  data = load_dataset("GEM/opusparcus", lang="de")
124
  ```
125
 
126
- The above command will download the validation and test sets for
127
- German. If additionally, you want to retrieve training data, you need
128
- to specify the level of quality you desire, such as "French, with 90%
129
- quality of the training data":
130
 
131
  ```
132
  data = load_dataset("GEM/opusparcus", lang="fr", quality=90)
133
  ```
134
 
135
- The entries in the training sets have been ranked automatically by how
136
- likely they are paraphrases, best first, worst last. The quality
137
- parameter indicates the estimated proportion (in percent) of true
138
- paraphrases in the training set. Allowed quality values range between
139
- 60 and 100, in increments of 5 (60, 65, 70, ..., 100). A value of 60
140
- means that 60% of the sentence pairs in the training set are estimated
141
- to be true paraphrases (and the remaining 40% are not). A higher value
142
- produces a smaller but cleaner set. The smaller sets are subsets of
143
- the larger sets, such that the `quality=95` set is a subset of
144
- `quality=90`, which is a subset of `quality=85`, and so on.
145
-
146
- The default `quality` value, if omitted, is 100. This matches no
147
- training data at all, which can be convenient, if you are only
148
- interested in the validation and test sets, which are considerably
149
  smaller, but manually annotated.
150
 
151
- Note that an alternative to typing the parameter values explicitly,
152
- you can use configuration names instead. The following commands are
153
- equivalent to the ones above:
154
 
155
  ```
156
  data = load_dataset("GEM/opusparcus", "de.100")
157
  data = load_dataset("GEM/opusparcus", "fr.90")
158
  ```
159
 
160
- Remark regarding the optimal choice of training set qualities:
161
- Previous work suggests that a larger and noisier set is better than a
162
- smaller and clean set. See Sjöblom et al. (2018). [Paraphrase
163
- Detection on Noisy Subtitles in Six
164
- Languages](http://noisy-text.github.io/2018/pdf/W-NUT20189.pdf). In
165
- *Proceedings of the 2018 EMNLP Workshop W-NUT: The 4th Workshop on
166
- Noisy User-generated Text*, and Vahtola et al. (2021). [Coping with
167
- Noisy Training Data Labels in Paraphrase
168
- Detection](https://aclanthology.org/2021.wnut-1.32/). In *Proceedings
169
- of the 7th Workshop on Noisy User-generated Text*.
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
170
 
171
- ### Data Instances
 
 
 
 
 
 
 
 
 
172
 
173
- As a concrete example, loading the English data requesting 95% quality of
174
- the train split produces the following:
175
 
176
  ```
177
  >>> data = load_dataset("GEM/opusparcus", lang="en", quality=95)
@@ -208,188 +303,422 @@ DatasetDict({
208
 
209
  >>> data["train"][1000]
210
  {'annot_score': 0.0, 'gem_id': 'gem-opusparcus-train-12501001', 'lang': 'en', 'sent1': 'Am I beautiful ?', 'sent2': 'Am I pretty ?'}
211
- ```
212
 
213
- ### Data Fields
214
 
215
- `sent1`: a tokenized sentence
 
 
216
 
217
- `sent2`: another tokenized sentence, which is potentially a paraphrase of `sent1`.
 
 
218
 
219
- `annot_score`: a value between 1.0 and 4.0 indicating how good an example of paraphrases `sent1` and `sent2` are. (For the training sets, the value is 0.0, which indicates that no manual annotation has taken place.)
220
 
221
- `lang`: language of this dataset
222
 
223
- `gem_id`: unique identifier of this entry
224
 
225
- **Additional information about the annotation scheme:**
226
 
227
- The annotation scores given by an individual annotator are:
228
 
229
- 4: Good example of paraphrases (Dark green button in the annotation
230
- tool): The two sentences can be used in the same situation and
231
- essentially "mean the same thing".
232
 
233
- 3: Mostly good example of paraphrases (Light green button in the
234
- annotation tool): It is acceptable to think that the two sentences
235
- refer to the same thing, although one sentence might be more specific
236
- than the other one, or there are differences in style, such as polite
237
- form versus familiar form.
238
 
239
- 2: Mostly bad example of paraphrases (Yellow button in the annotation
240
- tool): There is some connection between the sentences that explains
241
- why they occur together, but one would not really consider them to
242
- mean the same thing.
243
 
244
- 1: Bad example of paraphrases (Red button in the annotation tool):
245
- There is no obvious connection. The sentences mean different things.
 
246
 
247
- If the two annotators fully agreed on the category, the value in the
248
- `annot_score` field is 4.0, 3.0, 2.0 or 1.0. If the two annotators
249
- chose adjacent categories, the value in this field will be 3.5, 2.5 or
250
- 1.5. For instance, a value of 2.5 means that one annotator gave a
251
- score of 3 ("mostly good"), indicating a possible paraphrase pair,
252
- whereas the other annotator scored this as a 2 ("mostly bad"), that
253
- is, unlikely to be a paraphrase pair. If the annotators disagreed by
254
- more than one category, the sentence pair was discarded and won't show
255
- up in the datasets.
256
 
257
- The training sets were not annotated manually. This is indicated by
258
- the value 0.0 in the `annot_score` field.
 
259
 
260
- For an assessment of of inter-annotator agreement, see Aulamo et
261
- al. (2019). [Annotation of subtitle paraphrases using a new web
262
- tool.](http://ceur-ws.org/Vol-2364/3_paper.pdf) In *Proceedings of the
263
- Digital Humanities in the Nordic Countries 4th Conference*,
264
- Copenhagen, Denmark.
265
 
266
- ### Data Splits
 
 
267
 
268
- The data is split into training, validation and test sets. The
269
- validation and test sets come in two versions, the regular validation
270
- and test sets and the full sets, called validation.full and
271
- test.full. The full sets contain all sentence pairs successfully
272
- annotated by the annotators, including the sentence pairs that were
273
- rejected as paraphrases. The annotation scores of the full sets thus
274
- range between 1.0 and 4.0. The regular validation and test sets only
275
- contain sentence pairs that qualify as paraphrases, scored between 3.0
276
- and 4.0 by the annotators.
277
 
278
- The number of sentence pairs in the data splits are as follows for
279
- each of the languages. The range between the smallest (`quality=95`)
280
- and largest (`quality=60`) train configuration have been shown.
281
 
282
- | | train | valid | test | valid.full | test.full |
283
- | ----- | ------ | ----- | ---- | ---------- | --------- |
284
- | de | 0.59M .. 13M | 1013 | 1047 | 1582 | 1586 |
285
- | en | 1.0M .. 35M | 1015 | 982 | 1455 | 1445 |
286
- | fi | 0.48M .. 8.9M | 963 | 958 | 1760 | 1749 |
287
- | fr | 0.94M .. 22M | 997 | 1007 | 1630 | 1674 |
288
- | ru | 0.15M .. 15M | 1020 | 1068 | 1854 | 1855 |
289
- | sv | 0.24M .. 4.5M | 984 | 947 | 1887 | 1901 |
290
 
 
291
 
292
- ## Dataset Creation
293
 
294
- ### Curation Rationale
 
 
295
 
296
- Opusparcus was created in order to produce a *sentential* paraphrase corpus
297
- for multiple languages containing *colloquial* language (as opposed to
298
- news or religious text, for instance).
299
 
300
- ### Source Data
 
 
301
 
302
- #### Initial Data Collection and Normalization
303
 
304
- The data in Opusparcus has been extracted from
305
- [OpenSubtitles2016](http://opus.nlpl.eu/OpenSubtitles2016.php), which
306
- is in turn based on data from http://www.opensubtitles.org/.
307
 
308
- The sentences have been tokenized.
309
 
310
- #### Who are the source language producers?
311
 
312
- The texts consist of subtitles that have been produced using
313
- crowdsourcing.
 
314
 
315
- ### Annotations
316
 
317
- #### Annotation process
 
 
318
 
319
- The development and test sets consist of sentence
320
- pairs that have been annotated manually; each set contains
321
- approximately 1000 sentence pairs that have been verified to be
322
- acceptable paraphrases by two indepedent annotators.
323
 
324
- The `annot_score` field reflects the judgments made by the annotators.
325
- If the annnotators fully agreed on the category (4.0: dark green, 3.0:
326
- light green, 2.0: yellow, 1.0: red), the value of `annot_score` is
327
- 4.0, 3.0, 2.0 or 1.0. If the annotators chose adjacent categories,
328
- the value in this field will be 3.5, 2.5 or 1.5. For instance, a
329
- value of 2.5 means that one annotator gave a score of 3 ("mostly
330
- good"), indicating a possible paraphrase pair, whereas the other
331
- annotator scored this as a 2 ("mostly bad"), that is, unlikely to be a
332
- paraphrase pair. If the annotators disagreed by more than one
333
- category, the sentence pair was discarded and won't show up in the
334
- datasets.
335
 
336
- #### Who are the annotators?
337
 
338
- Students and staff at the University of Helsinki (native or very
339
- proficient speakers of the target languages)
340
 
341
- ### Personal and Sensitive Information
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
342
 
343
- The datasets do not contain any personal or sensitive information.
344
 
345
- ## Considerations for Using the Data
346
 
347
- ### Social Impact of Dataset
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
348
 
349
- The goal of Opusparcus is to promote the support for colloquial language.
350
 
351
  ### Discussion of Biases
352
 
353
- The data reflect the biases present in the movies and TV shows that
354
- have been subtitled.
355
 
356
- ### Other Known Limitations
 
 
357
 
358
- The sentence pairs in the validation and test sets have been selected
359
- in such a manner that their Levenshtein distance (minimum edit
360
- distance) exceeds a certain theshold. This guarantees that the manual
361
- annotation effort focuses on "interesting" sentence pairs rather than
362
- trivial variations (such as "It is good." vs. "It's good."). The
363
- training sets, however, have not been prefiltered in this manner and
364
- thus also contain highly similar sentences.
365
 
366
- ## Additional Information
 
 
367
 
368
- ### Dataset Curators
369
 
370
- Mathias Creutz, University of Helsinki, Finland
371
 
372
- ### Licensing Information
373
 
374
- CC-BY-NC 4.0
375
 
376
- ### Citation Information
377
 
378
- ```
379
- @InProceedings{creutz:lrec2018,
380
- title = {Open Subtitles Paraphrase Corpus for Six Languages},
381
- author={Mathias Creutz},
382
- booktitle={Proceedings of the 11th edition of the Language Resources and Evaluation Conference (LREC 2018)},
383
- year={2018},
384
- month = {May 7-12},
385
- address = {Miyazaki, Japan},
386
- editor = {Nicoletta Calzolari (Conference chair) and Khalid Choukri and Christopher Cieri and Thierry Declerck and Sara Goggi and Koiti Hasida and Hitoshi Isahara and Bente Maegaard and Joseph Mariani and Hélène Mazo and Asuncion Moreno and Jan Odijk and Stelios Piperidis and Takenobu Tokunaga},
387
- publisher = {European Language Resources Association (ELRA)},
388
- isbn = {979-10-95546-00-9},
389
- language = {english},
390
- url={http://www.lrec-conf.org/proceedings/lrec2018/pdf/131.pdf}
391
- ```
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
392
 
393
- ### Contributions
394
 
395
- Thanks to [@mathiascreutz](https://github.com/mathiascreutz) for adding this dataset.
1
  ---
2
  annotations_creators:
3
+ - expert-created
 
4
  language_creators:
5
+ - unknown
6
  languages:
7
+ - unknown
 
 
 
 
 
8
  licenses:
9
  - cc-by-nc-4.0
10
  multilinguality:
11
+ - unknown
12
+ pretty_name: opusparcus
13
  size_categories:
14
  - unknown
15
  source_datasets:
16
+ - original
17
  task_categories:
18
+ - paraphrasing
19
  task_ids:
20
+ - unknown
21
  ---
22
 
23
+ # Dataset Card for GEM/opusparcus
 
24
 
25
+ ## Dataset Description
26
 
27
+ - **Homepage:** http://urn.fi/urn:nbn:fi:lb-2018021221
28
+ - **Repository:** http://urn.fi/urn:nbn:fi:lb-2018021221
29
+ - **Paper:** http://www.lrec-conf.org/proceedings/lrec2018/pdf/131.pdf
30
+ - **Leaderboard:** N/A
31
+ - **Point of Contact:** Mathias Creutz
32
 
33
+ ### Link to Main Data Card
34
 
35
+ You can find the main data card on the [GEM Website](https://gem-benchmark.com/data_cards/opusparcus).
36
 
37
+ ### Dataset Summary
38
 
39
+ Opusparcus is a paraphrase corpus for six European languages: German, English, Finnish, French, Russian, and Swedish. The paraphrases consist of subtitles from movies and TV shows.
40
 
41
+ You can load the dataset via:
42
+ ```
43
+ import datasets
44
+ data = datasets.load_dataset('GEM/opusparcus')
45
+ ```
46
+ The data loader can be found [here](https://huggingface.co/datasets/GEM/opusparcus).
47
 
48
+ #### website
49
+ [Website](http://urn.fi/urn:nbn:fi:lb-2018021221)
50
 
51
+ #### paper
52
+ [LREC](http://www.lrec-conf.org/proceedings/lrec2018/pdf/131.pdf)
53
 
54
+ ## Dataset Overview
55
 
56
+ ### Where to find the Data and its Documentation
57
 
58
+ #### Webpage
59
 
60
+ <!-- info: What is the webpage for the dataset (if it exists)? -->
61
+ <!-- scope: telescope -->
62
+ [Website](http://urn.fi/urn:nbn:fi:lb-2018021221)
63
 
64
+ #### Download
65
 
66
+ <!-- info: What is the link to where the original dataset is hosted? -->
67
+ <!-- scope: telescope -->
68
+ [Website](http://urn.fi/urn:nbn:fi:lb-2018021221)
69
+
70
+ #### Paper
71
+
72
+ <!-- info: What is the link to the paper describing the dataset (open access preferred)? -->
73
+ <!-- scope: telescope -->
74
+ [LREC](http://www.lrec-conf.org/proceedings/lrec2018/pdf/131.pdf)
75
+
76
+ #### BibTex
77
+
78
+ <!-- info: Provide the BibTex-formatted reference for the dataset. Please use the correct published version (ACL anthology, etc.) instead of google scholar created Bibtex. -->
79
+ <!-- scope: microscope -->
80
+ ```
81
+ @InProceedings{creutz:lrec2018,
82
+ title = {Open Subtitles Paraphrase Corpus for Six Languages},
83
+ author={Mathias Creutz},
84
+ booktitle={Proceedings of the 11th edition of the Language Resources and Evaluation Conference (LREC 2018)},
85
+ year={2018},
86
+ month = {May 7-12},
87
+ address = {Miyazaki, Japan},
88
+ editor = {Nicoletta Calzolari (Conference chair) and Khalid Choukri and Christopher Cieri and Thierry Declerck and Sara Goggi and Koiti Hasida and Hitoshi Isahara and Bente Maegaard and Joseph Mariani and Hélène Mazo and Asuncion Moreno and Jan Odijk and Stelios Piperidis and Takenobu Tokunaga},
89
+ publisher = {European Language Resources Association (ELRA)},
90
+ isbn = {979-10-95546-00-9},
91
+ language = {english},
92
+ url={http://www.lrec-conf.org/proceedings/lrec2018/pdf/131.pdf}
93
+ ```
94
 
95
+ #### Contact Name
96
 
97
+ <!-- quick -->
98
+ <!-- info: If known, provide the name of at least one person the reader can contact for questions about the dataset. -->
99
+ <!-- scope: periscope -->
100
+ Mathias Creutz
101
 
102
+ #### Contact Email
103
 
104
+ <!-- info: If known, provide the email of at least one person the reader can contact for questions about the dataset. -->
105
+ <!-- scope: periscope -->
106
+ firstname dot lastname at helsinki dot fi
107
 
108
+ #### Has a Leaderboard?
109
 
110
+ <!-- info: Does the dataset have an active leaderboard? -->
111
+ <!-- scope: telescope -->
112
+ no
113
 
 
114
 
115
+ ### Languages and Intended Use
116
 
117
+ #### Multilingual?
118
 
119
+ <!-- quick -->
120
+ <!-- info: Is the dataset multilingual? -->
121
+ <!-- scope: telescope -->
122
+ yes
123
 
124
+ #### Covered Languages
125
 
126
+ <!-- quick -->
127
+ <!-- info: What languages/dialects are covered in the dataset? -->
128
+ <!-- scope: telescope -->
129
+ `German`, `English`, `Finnish`, `French`, `Russian`, `Swedish`
130
 
131
+ #### Whose Language?
132
 
133
+ <!-- info: Whose language is in the dataset? -->
134
+ <!-- scope: periscope -->
135
+ Opusparcus is a paraphrase corpus for six European languages: German, English, Finnish, French, Russian, and Swedish. The paraphrases consist of subtitles from movies and TV shows.
136
 
137
+ The data in Opusparcus has been extracted from [OpenSubtitles2016](http://opus.nlpl.eu/OpenSubtitles2016.php), which is in turn based on data from http://www.opensubtitles.org/.
 
 
138
 
139
+ #### License
 
 
140
 
141
+ <!-- quick -->
142
+ <!-- info: What is the license of the dataset? -->
143
+ <!-- scope: telescope -->
144
+ cc-by-nc-4.0: Creative Commons Attribution Non Commercial 4.0 International
 
 
 
 
145
 
146
+ #### Intended Use
147
 
148
+ <!-- info: What is the intended use of the dataset? -->
149
+ <!-- scope: microscope -->
150
+ Opusparcus is a sentential paraphrase corpus for multiple languages containing colloquial language.
151
 
152
+ #### Primary Task
153
 
154
+ <!-- info: What primary task does the dataset support? -->
155
+ <!-- scope: telescope -->
156
+ Paraphrasing
157
 
158
+ #### Communicative Goal
159
 
160
+ <!-- quick -->
161
+ <!-- info: Provide a short description of the communicative goal of a model trained for this task on this dataset. -->
162
+ <!-- scope: periscope -->
163
+ Models can be trained, e.g., for paraphrase detection and generation, that is, determining whether two given sentences mean the same thing or generating new paraphrases for a given sentence.
164
 
165
+
166
+ ### Credit
167
+
168
+ #### Who added the Dataset to GEM?
169
+
170
+ <!-- info: Who contributed to the data card and adding the dataset to GEM? List the people+affiliations involved in creating this data card and who helped integrate this dataset into GEM. -->
171
+ <!-- scope: microscope -->
172
+ Mathias Creutz (University of Helsinki)
173
+
174
+
175
+ ### Dataset Structure
176
+
177
+ #### Data Fields
178
+
179
+ <!-- info: List and describe the fields present in the dataset. -->
180
+ <!-- scope: telescope -->
181
+ - `sent1`: a tokenized sentence
182
+ - `sent2`: another tokenized sentence, which is potentially a paraphrase of `sent1`.
183
+ - `annot_score`: a value between 1.0 and 4.0 indicating how good an example of paraphrases `sent1` and `sent2` are. (For the training sets, the value is 0.0, which indicates that no manual annotation has taken place.)
184
+ - `lang`: language of this dataset
185
+ - `gem_id`: unique identifier of this entry
186
+
187
+ All fields are strings except `annot_score`, which is a float.
188
+
189
+ #### Reason for Structure
190
+
191
+ <!-- info: How was the dataset structure determined? -->
192
+ <!-- scope: microscope -->
193
+ For each target language, the Opusparcus data have been partitioned into three types of data sets: training, validation and test sets. The training sets are large, consisting of millions of sentence pairs, and have been compiled automatically, with the help of probabilistic ranking functions. The development and test sets consist of sentence pairs that have been annotated manually; each set contains approximately 1000 sentence pairs that have been verified to be acceptable paraphrases by two independent annotators.
194
+
195
+ When you download Opusparcus, you must always indicate the language you want to retrieve, for instance:
196
 
197
  ```
198
  data = load_dataset("GEM/opusparcus", lang="de")
199
  ```
200
 
201
+ The above command will download the validation and test sets for German. If additionally, you want to retrieve training data, you need to specify the level of quality you desire, such as "French, with 90% quality of the training data":
 
 
 
202
 
203
  ```
204
  data = load_dataset("GEM/opusparcus", lang="fr", quality=90)
205
  ```
206
 
207
+ The entries in the training sets have been ranked automatically by how likely they are paraphrases, best first, worst last. The quality parameter indicates the estimated proportion (in percent) of true
208
+ paraphrases in the training set. Allowed quality values range between 60 and 100, in increments of 5 (60, 65, 70, ..., 100). A value of 60 means that 60% of the sentence pairs in the training set are estimated to be true paraphrases (and the remaining 40% are not). A higher value produces a smaller but cleaner set. The smaller sets are subsets of the larger sets, such that the `quality=95` set is a subset of `quality=90`, which is a subset of `quality=85`, and so on.
209
+
210
+ The default `quality` value, if omitted, is 100. This matches no training data at all, which can be convenient, if you are only interested in the validation and test sets, which are considerably
 
 
 
 
 
 
 
 
 
 
211
  smaller, but manually annotated.
212
 
213
+ Note that an alternative to typing the parameter values explicitly, you can use configuration names instead. The following commands are equivalent to the ones above:
 
 
214
 
215
  ```
216
  data = load_dataset("GEM/opusparcus", "de.100")
217
  data = load_dataset("GEM/opusparcus", "fr.90")
218
  ```
219
 
220
+ #### How were labels chosen?
221
+
222
+ <!-- info: How were the labels chosen? -->
223
+ <!-- scope: microscope -->
224
+ Annotators have used the following scores to label sentence pairs in the test and validation sets:
225
+
226
+ 4: Good example of paraphrases (Dark green button in the annotation tool): The two sentences can be used in the same situation and essentially "mean the same thing".
227
+
228
+ 3: Mostly good example of paraphrases (Light green button in the annotation tool): It is acceptable to think that the two sentences refer to the same thing, although one sentence might be more specific
229
+ than the other one, or there are differences in style, such as polite form versus familiar form.
230
+
231
+ 2: Mostly bad example of paraphrases (Yellow button in the annotation tool): There is some connection between the sentences that explains why they occur together, but one would not really consider them to mean the same thing.
232
+
233
+ 1: Bad example of paraphrases (Red button in the annotation tool): There is no obvious connection. The sentences mean different things.
234
+
235
+ If the two annotators fully agreed on the category, the value in the `annot_score` field is 4.0, 3.0, 2.0 or 1.0. If the two annotators chose adjacent categories, the value in this field will be 3.5, 2.5 or
236
+ 1.5. For instance, a value of 2.5 means that one annotator gave a score of 3 ("mostly good"), indicating a possible paraphrase pair, whereas the other annotator scored this as a 2 ("mostly bad"), that is, unlikely to be a paraphrase pair. If the annotators disagreed by more than one category, the sentence pair was discarded and won't show up in the datasets.
237
+
238
+ The training sets were not annotated manually. This is indicated by
239
+ the value 0.0 in the `annot_score` field.
240
+
241
+ For an assessment of of inter-annotator agreement, see Aulamo et al. (2019). [Annotation of subtitle paraphrases using a new web tool.](http://ceur-ws.org/Vol-2364/3_paper.pdf) In *Proceedings of the
242
+ Digital Humanities in the Nordic Countries 4th Conference*, Copenhagen, Denmark.
243
+
244
+ #### Example Instance
245
+
246
+ <!-- info: Provide a JSON formatted example of a typical instance in the dataset. -->
247
+ <!-- scope: periscope -->
248
+ ```
249
+ {'annot_score': 4.0, 'gem_id': 'gem-opusparcus-test-1587', 'lang': 'en', 'sent1': "I haven 't been contacted by anybody .", 'sent2': "Nobody 's contacted me ."}
250
+ ```
251
+
252
+ #### Data Splits
253
+
254
+ <!-- info: Describe and name the splits in the dataset if there are more than one. -->
255
+ <!-- scope: periscope -->
256
+ The data is split into training, validation and test sets. The validation and test sets come in two versions, the regular validation and test sets and the full sets, called validation.full and test.full. The full sets contain all sentence pairs successfully annotated by the annotators, including the sentence pairs that were rejected as paraphrases. The annotation scores of the full sets thus range between 1.0 and 4.0. The regular validation and test sets only contain sentence pairs that qualify as paraphrases, scored between 3.0 and 4.0 by the annotators.
257
 
258
+ The number of sentence pairs in the data splits are as follows for each of the languages. The range between the smallest (`quality=95`) and largest (`quality=60`) train configuration have been shown.
259
+
260
+ | | train | valid | test | valid.full | test.full |
261
+ | ----- | ------ | ----- | ---- | ---------- | --------- |
262
+ | de | 0.59M .. 13M | 1013 | 1047 | 1582 | 1586 |
263
+ | en | 1.0M .. 35M | 1015 | 982 | 1455 | 1445 |
264
+ | fi | 0.48M .. 8.9M | 963 | 958 | 1760 | 1749 |
265
+ | fr | 0.94M .. 22M | 997 | 1007 | 1630 | 1674 |
266
+ | ru | 0.15M .. 15M | 1020 | 1068 | 1854 | 1855 |
267
+ | sv | 0.24M .. 4.5M | 984 | 947 | 1887 | 1901 |
268
 
269
+ As a concrete example, loading the English data requesting 95% quality of the train split produces the following:
 
270
 
271
  ```
272
  >>> data = load_dataset("GEM/opusparcus", lang="en", quality=95)
303
 
304
  >>> data["train"][1000]
305
  {'annot_score': 0.0, 'gem_id': 'gem-opusparcus-train-12501001', 'lang': 'en', 'sent1': 'Am I beautiful ?', 'sent2': 'Am I pretty ?'}
 
306
 
307
+ #### Splitting Criteria
308
 
309
+ <!-- info: Describe any criteria for splitting the data, if used. If there are differences between the splits (e.g., if the training annotations are machine-generated and the dev and test ones are created by humans, or if different numbers of annotators contributed to each example), describe them here. -->
310
+ <!-- scope: microscope -->
311
+ The validation and test sets have been annotated manually, but the training sets have been produced using automatic scoring and come in different size configurations depending on the desired quality level. (See above descriptions and examples for more details.)
312
 
313
+ Please note that previous work suggests that a larger and noisier training set is better than a
314
+ smaller and clean set. See Sjöblom et al. (2018). [Paraphrase Detection on Noisy Subtitles in Six
315
+ Languages](http://noisy-text.github.io/2018/pdf/W-NUT20189.pdf). In *Proceedings of the 2018 EMNLP Workshop W-NUT: The 4th Workshop on Noisy User-generated Text*, and Vahtola et al. (2021). [Coping with Noisy Training Data Labels in Paraphrase Detection](https://aclanthology.org/2021.wnut-1.32/). In *Proceedings of the 7th Workshop on Noisy User-generated Text*.
316
 
 
317
 
 
318
 
 
319
 
320
+ ## Dataset in GEM
321
 
322
+ ### Rationale for Inclusion in GEM
323
 
324
+ #### Why is the Dataset in GEM?
 
 
325
 
326
+ <!-- info: What does this dataset contribute toward better generation evaluation and why is it part of GEM? -->
327
+ <!-- scope: microscope -->
328
+ Opusparcus provides examples of sentences that mean the same thing or have very similar meaning. Sentences are available in six languages and the style is colloquial language.
 
 
329
 
330
+ #### Similar Datasets
 
 
 
331
 
332
+ <!-- info: Do other datasets for the high level task exist? -->
333
+ <!-- scope: telescope -->
334
+ yes
335
 
336
+ #### Unique Language Coverage
 
 
 
 
 
 
 
 
337
 
338
+ <!-- info: Does this dataset cover other languages than other datasets for the same task? -->
339
+ <!-- scope: periscope -->
340
+ yes
341
 
342
+ #### Difference from other GEM datasets
 
 
 
 
343
 
344
+ <!-- info: What else sets this dataset apart from other similar datasets in GEM? -->
345
+ <!-- scope: microscope -->
346
+ There is another data set containing manually labeled Finnish paraphrases.
347
 
348
+ #### Ability that the Dataset measures
 
 
 
 
 
 
 
 
349
 
350
+ <!-- info: What aspect of model ability can be measured with this dataset? -->
351
+ <!-- scope: periscope -->
352
+ Sentence meaning
353
 
 
 
 
 
 
 
 
 
354
 
355
+ ### GEM-Specific Curation
356
 
357
+ #### Modificatied for GEM?
358
 
359
+ <!-- info: Has the GEM version of the dataset been modified in any way (data, processing, splits) from the original curated data? -->
360
+ <!-- scope: telescope -->
361
+ yes
362
 
363
+ #### GEM Modifications
 
 
364
 
365
+ <!-- info: What changes have been made to he original dataset? -->
366
+ <!-- scope: periscope -->
367
+ `other`
368
 
369
+ #### Modification Details
370
 
371
+ <!-- info: For each of these changes, described them in more details and provided the intended purpose of the modification -->
372
+ <!-- scope: microscope -->
373
+ Training sets have been prepared for each the "quality levels" 60% – 95%.
374
 
375
+ In the original release, this task was left to the user of the data.
376
 
377
+ #### Additional Splits?
378
 
379
+ <!-- info: Does GEM provide additional splits to the dataset? -->
380
+ <!-- scope: telescope -->
381
+ yes
382
 
383
+ #### Split Information
384
 
385
+ <!-- info: Describe how the new splits were created -->
386
+ <!-- scope: periscope -->
387
+ There are two versions of the validations and test sets: the regular sets which only contain positive examples of paraphrases and the full sets containing all examples.
388
 
389
+ #### Split Motivation
 
 
 
390
 
391
+ <!-- info: What aspects of the model's generation capacities were the splits created to test? -->
392
+ <!-- scope: periscope -->
393
+ In the original release, only the full validation and test sets were supplied. The "regular sets" have been added in order to make it easier to test on true parapahrases only.
 
 
 
 
 
 
 
 
394
 
 
395
 
396
+ ### Getting Started with the Task
 
397
 
398
+ #### Pointers to Resources
399
+
400
+ <!-- info: Getting started with in-depth research on the task. Add relevant pointers to resources that researchers can consult when they want to get started digging deeper into the task. -->
401
+ <!-- scope: microscope -->
402
+ Creutz (2018). [Open Subtitles Paraphrase Corpus for Six Languages](http://www.lrec-conf.org/proceedings/lrec2018/pdf/131.pdf), Proceedings of the 11th edition of the Language Resources and Evaluation Conference (LREC 2018).
403
+
404
+ Sjöblom et al. (2018). [Paraphrase Detection on Noisy Subtitles in Six Languages](http://noisy-text.github.io/2018/pdf/W-NUT20189.pdf). In Proceedings of the 2018 EMNLP Workshop W-NUT: The 4th Workshop on Noisy User-generated Text.
405
+
406
+ Aulamo et al. (2019). [Annotation of subtitle paraphrases using a new web tool.](http://ceur-ws.org/Vol-2364/3_paper.pdf) In Proceedings of the Digital Humanities in the Nordic Countries 4th Conference.
407
+
408
+ Sjöblom et al. (2020). [Paraphrase Generation and Evaluation on Colloquial-Style Sentences](https://aclanthology.org/2020.lrec-1.224/), Proceedings of the 12th Language Resources and Evaluation Conference (LREC).
409
+
410
+ Vahtola et al. (2021). [Coping with Noisy Training Data Labels in Paraphrase Detection](https://aclanthology.org/2021.wnut-1.32/). In Proceedings of the 7th Workshop on Noisy User-generated Text.
411
+
412
+
413
+
414
+
415
+ ## Previous Results
416
+
417
+ ### Previous Results
418
+
419
+ #### Measured Model Abilities
420
+
421
+ <!-- info: What aspect of model ability can be measured with this dataset? -->
422
+ <!-- scope: telescope -->
423
+ Sentence meaning
424
+
425
+ In a scenario of paraphrase detection, the model determines whether two given sentences carry approximately the same meaning.
426
+
427
+ In a scenario of paraphrase generation, the model generates a potential paraphrase of a given sentence.
428
+
429
+ #### Metrics
430
+
431
+ <!-- info: What metrics are typically used for this task? -->
432
+ <!-- scope: periscope -->
433
+ `BLEU`, `BERT-Score`, `Other: Other Metrics`
434
+
435
+ #### Other Metrics
436
+
437
+ <!-- info: Definitions of other metrics -->
438
+ <!-- scope: periscope -->
439
+ PINC
440
+
441
+ #### Proposed Evaluation
442
+
443
+ <!-- info: List and describe the purpose of the metrics and evaluation methodology (including human evaluation) that the dataset creators used when introducing this task. -->
444
+ <!-- scope: microscope -->
445
+ The metrics mentioned above can be used to assess how well a generated paraphrase corresponds to a given reference sentence. The PINC score additionally assesses how different the surface forms are.
446
+
447
+ #### Previous results available?
448
+
449
+ <!-- info: Are previous results available? -->
450
+ <!-- scope: telescope -->
451
+ yes
452
+
453
+ #### Other Evaluation Approaches
454
+
455
+ <!-- info: What evaluation approaches have others used? -->
456
+ <!-- scope: periscope -->
457
+ See publications on using Opusparcus
458
+
459
+ #### Relevant Previous Results
460
+
461
+ <!-- info: What are the most relevant previous results for this task/dataset? -->
462
+ <!-- scope: microscope -->
463
+ Sjöblom et al. (2020). [Paraphrase Generation and Evaluation on Colloquial-Style Sentences](https://aclanthology.org/2020.lrec-1.224/), Proceedings of the 12th Language Resources and Evaluation Conference (LREC).
464
 
 
465
 
 
466
 
467
+ ## Dataset Curation
468
+
469
+ ### Original Curation
470
+
471
+ #### Original Curation Rationale
472
+
473
+ <!-- info: Original curation rationale -->
474
+ <!-- scope: telescope -->
475
+ Opusparcus was created in order to produce a *sentential* paraphrase corpus for multiple languages containing *colloquial* language (as opposed to news or religious text, for instance).
476
+
477
+ #### Communicative Goal
478
+
479
+ <!-- info: What was the communicative goal? -->
480
+ <!-- scope: periscope -->
481
+ Opusparcus provides labeled examples of pairs of sentences that have similar (or dissimilar) meanings.
482
+
483
+ #### Sourced from Different Sources
484
+
485
+ <!-- info: Is the dataset aggregated from different data sources? -->
486
+ <!-- scope: telescope -->
487
+ no
488
+
489
+
490
+ ### Language Data
491
+
492
+ #### How was Language Data Obtained?
493
+
494
+ <!-- info: How was the language data obtained? -->
495
+ <!-- scope: telescope -->
496
+ `Crowdsourced`
497
+
498
+ #### Where was it crowdsourced?
499
+
500
+ <!-- info: If crowdsourced, where from? -->
501
+ <!-- scope: periscope -->
502
+ `Other crowdworker platform`
503
+
504
+ #### Language Producers
505
+
506
+ <!-- info: What further information do we have on the language producers? -->
507
+ <!-- scope: microscope -->
508
+ The data in Opusparcus has been extracted from [OpenSubtitles2016](http://opus.nlpl.eu/OpenSubtitles2016.php), which is in turn based on data from http://www.opensubtitles.org/.
509
+
510
+ The texts consists of subtitles that have been produced using crowdsourcing.
511
+
512
+ #### Topics Covered
513
+
514
+ <!-- info: Does the language in the dataset focus on specific topics? How would you describe them? -->
515
+ <!-- scope: periscope -->
516
+ The language is representative of movies and TV shows. Domains covered include comedy, drama, relationships, suspense, etc.
517
+
518
+ #### Data Validation
519
+
520
+ <!-- info: Was the text validated by a different worker or a data curator? -->
521
+ <!-- scope: telescope -->
522
+ validated by data curator
523
+
524
+ #### Data Preprocessing
525
+
526
+ <!-- info: How was the text data pre-processed? (Enter N/A if the text was not pre-processed) -->
527
+ <!-- scope: microscope -->
528
+ Sentence and word tokenization was performed.
529
+
530
+ #### Was Data Filtered?
531
+
532
+ <!-- info: Were text instances selected or filtered? -->
533
+ <!-- scope: telescope -->
534
+ algorithmically
535
+
536
+ #### Filter Criteria
537
+
538
+ <!-- info: What were the selection criteria? -->
539
+ <!-- scope: microscope -->
540
+ The sentence pairs in the training sets were ordered automatically based on the estimated likelihood that the sentences were paraphrases, most likely paraphrases on the top, and least likely paraphrases on the bottom.
541
+
542
+ The validation and test sets were checked and annotated manually, but the sentence pairs selected for annotation had to be different enough in terms of minimum edit distance (Levenshtein distance). This ensured that annotators would not spend their time annotating pairs of more or less identical sentences.
543
+
544
+
545
+ ### Structured Annotations
546
+
547
+ #### Additional Annotations?
548
+
549
+ <!-- quick -->
550
+ <!-- info: Does the dataset have additional annotations for each instance? -->
551
+ <!-- scope: telescope -->
552
+ expert created
553
+
554
+ #### Number of Raters
555
+
556
+ <!-- info: What is the number of raters -->
557
+ <!-- scope: telescope -->
558
+ 11<n<50
559
+
560
+ #### Rater Qualifications
561
+
562
+ <!-- info: Describe the qualifications required of an annotator. -->
563
+ <!-- scope: periscope -->
564
+ Students and staff at the University of Helsinki (native or very proficient speakers of the target languages)
565
+
566
+ #### Raters per Training Example
567
+
568
+ <!-- info: How many annotators saw each training example? -->
569
+ <!-- scope: periscope -->
570
+ 0
571
+
572
+ #### Raters per Test Example
573
+
574
+ <!-- info: How many annotators saw each test example? -->
575
+ <!-- scope: periscope -->
576
+ 2
577
+
578
+ #### Annotation Service?
579
+
580
+ <!-- info: Was an annotation service used? -->
581
+ <!-- scope: telescope -->
582
+ no
583
+
584
+ #### Annotation Values
585
+
586
+ <!-- info: Purpose and values for each annotation -->
587
+ <!-- scope: microscope -->
588
+ The development and test sets consist of sentence pairs that have been annotated manually; each set contains approximately 1000 sentence pairs that have been verified to be acceptable paraphrases by two independent annotators.
589
+
590
+ The `annot_score` field reflects the judgments made by the annotators. If the annnotators fully agreed on the category (4.0: dark green, 3.0: light green, 2.0: yellow, 1.0: red), the value of `annot_score` is 4.0, 3.0, 2.0 or 1.0. If the annotators chose adjacent categories, the value in this field will be 3.5, 2.5 or 1.5. For instance, a value of 2.5 means that one annotator gave a score of 3 ("mostly good"), indicating a possible paraphrase pair, whereas the other annotator scored this as a 2 ("mostly bad"), that is, unlikely to be a paraphrase pair. If the annotators disagreed by more than one category, the sentence pair was discarded and won't show up in the datasets.
591
+
592
+ Annotators could also reject a sentence pair as being corrupted data.
593
+
594
+ #### Any Quality Control?
595
+
596
+ <!-- info: Quality control measures? -->
597
+ <!-- scope: telescope -->
598
+ validated by another rater
599
+
600
+ #### Quality Control Details
601
+
602
+ <!-- info: Describe the quality control measures that were taken. -->
603
+ <!-- scope: microscope -->
604
+ If the annotators disagreed by more than one category, the sentence pair was discarded and is not part of the final dataset.
605
+
606
+
607
+ ### Consent
608
+
609
+ #### Any Consent Policy?
610
+
611
+ <!-- info: Was there a consent policy involved when gathering the data? -->
612
+ <!-- scope: telescope -->
613
+ no
614
+
615
+
616
+ ### Private Identifying Information (PII)
617
+
618
+ #### Contains PII?
619
+
620
+ <!-- quick -->
621
+ <!-- info: Does the source language data likely contain Personal Identifying Information about the data creators or subjects? -->
622
+ <!-- scope: telescope -->
623
+ yes/very likely
624
+
625
+ #### Any PII Identification?
626
+
627
+ <!-- info: Did the curators use any automatic/manual method to identify PII in the dataset? -->
628
+ <!-- scope: periscope -->
629
+ no identification
630
+
631
+
632
+ ### Maintenance
633
+
634
+ #### Any Maintenance Plan?
635
+
636
+ <!-- info: Does the original dataset have a maintenance plan? -->
637
+ <!-- scope: telescope -->
638
+ no
639
+
640
+
641
+
642
+ ## Broader Social Context
643
+
644
+ ### Previous Work on the Social Impact of the Dataset
645
+
646
+ #### Usage of Models based on the Data
647
+
648
+ <!-- info: Are you aware of cases where models trained on the task featured in this dataset ore related tasks have been used in automated systems? -->
649
+ <!-- scope: telescope -->
650
+ no
651
+
652
+
653
+ ### Impact on Under-Served Communities
654
+
655
+ #### Addresses needs of underserved Communities?
656
+
657
+ <!-- info: Does this dataset address the needs of communities that are traditionally underserved in language technology, and particularly language generation technology? Communities may be underserved for exemple because their language, language variety, or social or geographical context is underepresented in NLP and NLG resources (datasets and models). -->
658
+ <!-- scope: telescope -->
659
+ no
660
 
 
661
 
662
  ### Discussion of Biases
663
 
664
+ #### Any Documented Social Biases?
 
665
 
666
+ <!-- info: Are there documented social biases in the dataset? Biases in this context are variations in the ways members of different social categories are represented that can have harmful downstream consequences for members of the more disadvantaged group. -->
667
+ <!-- scope: telescope -->
668
+ no
669
 
670
+ #### Are the Language Producers Representative of the Language?
 
 
 
 
 
 
671
 
672
+ <!-- info: Does the distribution of language producers in the dataset accurately represent the full distribution of speakers of the language world-wide? If not, how does it differ? -->
673
+ <!-- scope: periscope -->
674
+ What social bias there may be in the subtitles in this dataset has not been studied.
675
 
 
676
 
 
677
 
678
+ ## Considerations for Using the Data
679
 
680
+ ### PII Risks and Liability
681
 
682
+ #### Potential PII Risk
683
 
684
+ <!-- info: Considering your answers to the PII part of the Data Curation Section, describe any potential privacy to the data subjects and creators risks when using the dataset. -->
685
+ <!-- scope: microscope -->
686
+ The data only contains subtitles of publicly available movies and TV shows.
687
+
688
+
689
+ ### Licenses
690
+
691
+ #### Copyright Restrictions on the Dataset
692
+
693
+ <!-- info: Based on your answers in the Intended Use part of the Data Overview Section, which of the following best describe the copyright and licensing status of the dataset? -->
694
+ <!-- scope: periscope -->
695
+ `non-commercial use only`
696
+
697
+ #### Copyright Restrictions on the Language Data
698
+
699
+ <!-- info: Based on your answers in the Language part of the Data Curation Section, which of the following best describe the copyright and licensing status of the underlying language data? -->
700
+ <!-- scope: periscope -->
701
+ `non-commercial use only`
702
+
703
+
704
+ ### Known Technical Limitations
705
+
706
+ #### Technical Limitations
707
+
708
+ <!-- info: Describe any known technical limitations, such as spurrious correlations, train/test overlap, annotation biases, or mis-annotations, and cite the works that first identified these limitations when possible. -->
709
+ <!-- scope: microscope -->
710
+ Some subtitles contain typos that are caused by inaccurate OCR.
711
+
712
+ #### Unsuited Applications
713
+
714
+ <!-- info: When using a model trained on this dataset in a setting where users or the public may interact with its predictions, what are some pitfalls to look out for? In particular, describe some applications of the general task featured in this dataset that its curation or properties make it less suitable for. -->
715
+ <!-- scope: microscope -->
716
+ The models might memorize individual subtitles of existing movies and TV shows, but there is no context across sentence boundaries in the data.
717
+
718
+ #### Discouraged Use Cases
719
+
720
+ <!-- info: What are some discouraged use cases of a model trained to maximize the proposed metrics on this dataset? In particular, think about settings where decisions made by a model that performs reasonably well on the metric my still have strong negative consequences for user or members of the public. -->
721
+ <!-- scope: microscope -->
722
+ A general issue with paraphrasing is that very small modifications in the surface form might produce valid paraphrases, which are however rather uninteresting. It is more valuable to produce paraphrases with clearly different surface realizations (e.g., measured using minimum edit distance).
723
 
 
724