mathiascreutz
commited on
Commit
·
d6a4efb
1
Parent(s):
04ebbb9
Minor modifications
Browse files
README.md
CHANGED
@@ -123,8 +123,8 @@ data = load_dataset("GEM/opusparcus", lang="de")
|
|
123 |
|
124 |
The above command will download the validation and test sets for
|
125 |
German. If additionally, you want to retrieve training data, you need
|
126 |
-
to specify the level of quality you desire, such as "90%
|
127 |
-
|
128 |
|
129 |
```
|
130 |
data = load_dataset("GEM/opusparcus", lang="fr", quality=90)
|
@@ -283,55 +283,85 @@ and largest (`quality=60`) train configuration have been shown.
|
|
283 |
|
284 |
### Curation Rationale
|
285 |
|
286 |
-
|
|
|
|
|
287 |
|
288 |
### Source Data
|
289 |
|
290 |
#### Initial Data Collection and Normalization
|
291 |
|
292 |
-
|
|
|
|
|
|
|
|
|
293 |
|
294 |
#### Who are the source language producers?
|
295 |
|
296 |
-
|
|
|
297 |
|
298 |
### Annotations
|
299 |
|
300 |
#### Annotation process
|
301 |
|
302 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
303 |
|
304 |
#### Who are the annotators?
|
305 |
|
306 |
-
|
|
|
307 |
|
308 |
### Personal and Sensitive Information
|
309 |
|
310 |
-
|
311 |
|
312 |
## Considerations for Using the Data
|
313 |
|
314 |
### Social Impact of Dataset
|
315 |
|
316 |
-
|
317 |
|
318 |
### Discussion of Biases
|
319 |
|
320 |
-
|
|
|
321 |
|
322 |
### Other Known Limitations
|
323 |
|
324 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
325 |
|
326 |
## Additional Information
|
327 |
|
328 |
### Dataset Curators
|
329 |
|
330 |
-
|
331 |
|
332 |
### Licensing Information
|
333 |
|
334 |
-
|
335 |
|
336 |
### Citation Information
|
337 |
|
|
|
123 |
|
124 |
The above command will download the validation and test sets for
|
125 |
German. If additionally, you want to retrieve training data, you need
|
126 |
+
to specify the level of quality you desire, such as "French, with 90%
|
127 |
+
quality of the training data":
|
128 |
|
129 |
```
|
130 |
data = load_dataset("GEM/opusparcus", lang="fr", quality=90)
|
|
|
283 |
|
284 |
### Curation Rationale
|
285 |
|
286 |
+
Opusparcus was created in order to produce a *sentential* paraphrase corpus
|
287 |
+
for multiple languages containing *colloquial* language (as opposed to
|
288 |
+
news or religious text, for instance).
|
289 |
|
290 |
### Source Data
|
291 |
|
292 |
#### Initial Data Collection and Normalization
|
293 |
|
294 |
+
The data in Opusparcus has been extracted from
|
295 |
+
[OpenSubtitles2016](http://opus.nlpl.eu/OpenSubtitles2016.php), which
|
296 |
+
is in turn based on data from http://www.opensubtitles.org/.
|
297 |
+
|
298 |
+
The sentences have been tokenized.
|
299 |
|
300 |
#### Who are the source language producers?
|
301 |
|
302 |
+
The texts consist of subtitles that have been produced using
|
303 |
+
crowdsourcing.
|
304 |
|
305 |
### Annotations
|
306 |
|
307 |
#### Annotation process
|
308 |
|
309 |
+
The development and test sets consist of sentence
|
310 |
+
pairs that have been annotated manually; each set contains
|
311 |
+
approximately 1000 sentence pairs that have been verified to be
|
312 |
+
acceptable paraphrases by two indepedent annotators.
|
313 |
+
|
314 |
+
The `annot_score` field reflects the judgments made by the annotators.
|
315 |
+
If ´the annnotators fully agreed on the category (4.0: dark green,
|
316 |
+
3.0: light green, 2.0: yellow, 1.0: red), the value of
|
317 |
+
`annot_score` is 4.0, 3.0, 2.0 or 1.0. If the two annotators
|
318 |
+
chose adjacent categories, the value in this field will be 3.5, 2.5 or
|
319 |
+
1.5. For instance, a value of 2.5 means that one annotator gave a
|
320 |
+
score of 3 ("mostly good"), indicating a possible paraphrase pair,
|
321 |
+
whereas the other annotator scored this as a 2 ("mostly bad"), that
|
322 |
+
is, unlikely to be a paraphrase pair. If the annotators disagreed by
|
323 |
+
more than one category, the sentence pair was discarded and won't show
|
324 |
+
up in the datasets.
|
325 |
|
326 |
#### Who are the annotators?
|
327 |
|
328 |
+
Students and staff at the University of Helsinki (native or very
|
329 |
+
proficient speakers of the target languages)
|
330 |
|
331 |
### Personal and Sensitive Information
|
332 |
|
333 |
+
The datasets do not contain any personal or sensitive information.
|
334 |
|
335 |
## Considerations for Using the Data
|
336 |
|
337 |
### Social Impact of Dataset
|
338 |
|
339 |
+
The goal of Opusparcus is to promote the support for colloquial language.
|
340 |
|
341 |
### Discussion of Biases
|
342 |
|
343 |
+
The data reflect the biases present in the movies and TV shows that
|
344 |
+
have been subtitled.
|
345 |
|
346 |
### Other Known Limitations
|
347 |
|
348 |
+
The sentence pairs in the validation and test sets have been selected
|
349 |
+
in such a manner that their Levenshtein distance (minimum edit
|
350 |
+
distance) exceeds a certain theshold. This guarantees that the manual
|
351 |
+
annotation effort focuses on "interesting" sentence pairs rather than
|
352 |
+
trivial variations (such as "It is good." vs. "It's good."). The
|
353 |
+
training sets, however, have not been prefiltered in this manner and
|
354 |
+
thus also contain highly similar sentences.
|
355 |
|
356 |
## Additional Information
|
357 |
|
358 |
### Dataset Curators
|
359 |
|
360 |
+
Mathias Creutz, University of Helsinki, Finland
|
361 |
|
362 |
### Licensing Information
|
363 |
|
364 |
+
CC-BY-NC 4.0
|
365 |
|
366 |
### Citation Information
|
367 |
|