GEM/opusparcus 数据集卡片

链接到主数据卡片

您可以在 GEM Website 找到主数据卡片。

数据集摘要

Opusparcus 是一个包含六种欧洲语言的近义词语料库：德语、英语、芬兰语、法语、俄语和瑞典语。这些近义词由电影和电视剧的字幕组成。

您可以通过以下方式加载数据集：

import datasets
data = datasets.load_dataset('GEM/opusparcus')

数据加载器可以在 here 找到。

数据集概述

数据及其文档的获取方式

网页

Website

下载

Website

论文

LREC

BibTex

@InProceedings{creutz:lrec2018,
  title = {Open Subtitles Paraphrase Corpus for Six Languages},
  author={Mathias Creutz},
  booktitle={Proceedings of the 11th edition of the Language Resources and Evaluation Conference (LREC 2018)},
  year={2018},
  month = {May 7-12},
  address = {Miyazaki, Japan},
  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},
  publisher = {European Language Resources Association (ELRA)},
  isbn = {979-10-95546-00-9},
  language = {english},
  url={http://www.lrec-conf.org/proceedings/lrec2018/pdf/131.pdf}

联系人姓名

Mathias Creutz

联系人电子邮件

firstname dot lastname at helsinki dot fi

有排行榜吗？

否

语言和预期用途

国际化？

是

覆盖的语言

德语、英语、芬兰语、法语、俄语、瑞典语

谁的语言?

Opusparcus 是一个针对六种欧洲语言的近义词语料库：德语、英语、芬兰语、法语、俄语和瑞典语。这些近义词由电影和电视剧的字幕组成。

Opusparcus 的数据是从 OpenSubtitles2016 提取的，而该数据是基于 OpenSubtitles 的数据。

许可证

cc-by-nc-4.0：知识共享署名-非商业性使用4.0国际

预期用途

Opusparcus 是一个用于多种语言的句子近义词语料库，包含口语。

主要任务

近义词

交际目标

可以训练模型，例如近义词检测和生成，即确定两个给定的句子是否意味着相同的事情或为给定的句子生成新的近义词。

信用

谁将数据集添加到 GEM 中？

Mathias Creutz（赫尔辛基大学）

数据集结构

数据字段

• sent1: 一个标记化的句子
• sent2: 另一个标记化的句子，可能是 sent1 的近义词。
• annot_score: 一个介于1.0和4.0之间的值，表示 sent1 和 sent2 是近义词的程度。（对于训练集，该值为0.0，表示未进行人工标注。）
• lang: 数据集的语言
• gem_id: 此条目的唯一标识符

所有字段都是字符串，除了 annot_score 是一个浮点数。

结构原因

对于每种目标语言，Opusparcus 数据被分成三种类型的数据集：训练集、验证集和测试集。训练集很大，包含数百万个句子对，是自动编译的，借助概率排序函数的帮助。开发和测试集包含已被手动标注的句子对；每个集合包含约1000个已经由两位独立标注者确认为可接受的近义词对的句子对。

当您下载 Opusparcus 数据集时，必须始终指定您想要获取的语言，例如：

data = load_dataset("GEM/opusparcus", lang="de")

上述命令将下载德语的验证集和测试集。如果需要获取训练数据，还需要指定所需质量级别，例如“法语，质量为90％的训练数据”：

data = load_dataset("GEM/opusparcus", lang="fr", quality=90)

训练集中的条目已由排名功能自动排序，按可能是近义词的可能性从高到低排列。质量参数指示训练集中可能是真正近义词的比例的估计值（以百分比表示）。允许的质量值范围在 60 到 100 之间，增量为 5（60、65、70，...，100）。60 的值表示训练集中有 60％ 的句子对被估计为真正的近义词（剩下的 40％ 不是）。更高的值会生成更小但更干净的集合。较小的集合是较大集合的子集， quality=95 集是 quality=90 的子集，依此类推。

如果省略， quality 参数的默认值是 100。这根本不匹配任何训练数据，这可能很方便，如果您只对验证和测试集感兴趣，这些集合要小得多，但手动注释。

请注意，与显式输入参数值相反，您可以使用配置名称。下面的命令与上面的命令是等效的：

data = load_dataset("GEM/opusparcus", "de.100")
data = load_dataset("GEM/opusparcus", "fr.90")

标签选择的方法是什么？

标注者在测试和验证集中使用以下评分来标记句子对：

4：近义词的好例子（注释工具中的深绿色按钮）：这两个句子可以在相同的情况下使用，并且基本上“意思相同”。

3：近义词的大部分好例子（注释工具中的浅绿色按钮）：可以接受地认为这两个句子指的是同一件事，尽管一个句子可能更具体，或者在样式上存在差异，例如礼貌形式与熟悉形式之间的区别。

2：近义词的大部分不好例子（注释工具中的黄色按钮）：这两个句子之间存在一些联系，可以解释它们一起出现的原因，但实际上不认为它们的意思相同。

1：近义词的不好例子（注释工具中的红色按钮）：没有明显的联系。句子有不同的含义。

如果两个标注者对于某个类别完全一致，则 annot_score 字段中的值为 4.0、3.0、2.0 或 1.0。如果两个标注者选择了相邻的类别，则此字段中的值将为 3.5、2.5 或 1.5。例如，值为 2.5 表示一个标注者给出了 3 的分数（表示可能的近义词对），即可能是近义词对的可能性，而另一个标注者将此分数评为 2（“大部分不好”），即不太可能是近义词对。如果标注者之间的意见相差超过一个类别，则将丢弃句子对，并且不会出现在数据集中。

训练集没有手动标注。这在 annot_score 字段中的值为 0.0 表示。

有关标注者间的一致性评估，请参阅Aulamo等人的文章（2019）。 Annotation of subtitle paraphrases using a new web tool. In Proceedings of the Digital Humanities in the Nordic Countries 4th Conference, Copenhagen, Denmark。

实例示例

{'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 ."}

数据拆分

数据分为训练集、验证集和测试集。验证集和测试集有两个版本，常规验证集和测试集以及完整集，称为 validation.full 和 test.full。完整集包含由标注者成功标注的所有句子对，包括被拒绝作为近义词的句子对。完整集的注释分数范围从 1.0 到 4.0。常规验证集和测试集仅包含由标注者评分介于 3.0 和 4.0 之间的句子对，合格作为近义词。

数据拆分中句子对的数量如下，为每种语言显示了最小（quality=95）和最大（quality=60）的训练配置范围。

train	valid	test	valid.full	test.full
de	0.59M .. 13M	1013	1047	1582	1586
en	1.0M .. 35M	1015	982	1455	1445
fi	0.48M .. 8.9M	963	958	1760	1749
fr	0.94M .. 22M	997	1007	1630	1674
ru	0.15M .. 15M	1020	1068	1854	1855
sv	0.24M .. 4.5M	984	947	1887	1901

具体来说，加载英语数据，并请求训练拆分质量为 95％，将产生以下结果：

>>> data = load_dataset("GEM/opusparcus", lang="en", quality=95)

>>> data
DatasetDict({
    test: Dataset({
        features: ['lang', 'sent1', 'sent2', 'annot_score', 'gem_id'],
        num_rows: 982
    })
    validation: Dataset({
        features: ['lang', 'sent1', 'sent2', 'annot_score', 'gem_id'],
        num_rows: 1015
    })
    test.full: Dataset({
        features: ['lang', 'sent1', 'sent2', 'annot_score', 'gem_id'],
        num_rows: 1445
    })
    validation.full: Dataset({
        features: ['lang', 'sent1', 'sent2', 'annot_score', 'gem_id'],
        num_rows: 1455
    })
    train: Dataset({
        features: ['lang', 'sent1', 'sent2', 'annot_score', 'gem_id'],
        num_rows: 1000000
    })
})

>>> data["test"][0]
{'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 ."}

>>> data["validation"][2]
{'annot_score': 3.0, 'gem_id': 'gem-opusparcus-validation-1586', 'lang': 'en', 'sent1': 'No promises , okay ?', 'sent2': "I 'm not promising anything ."}

>>> data["train"][1000]
{'annot_score': 0.0, 'gem_id': 'gem-opusparcus-train-12501001', 'lang': 'en', 'sent1': 'Am I beautiful ?', 'sent2': 'Am I pretty ?'}

#### Splitting Criteria

<!-- 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. -->
<!-- scope: microscope -->
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.)

Please note that previous work suggests that a larger and noisier training set is better than a
smaller and clean set. See 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*, 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*.




## Dataset in GEM

### Rationale for Inclusion in GEM

#### Why is the Dataset in GEM?

<!-- info: What does this dataset contribute toward better generation evaluation and why is it part of GEM? -->
<!-- scope: microscope -->
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.

#### Similar Datasets

<!-- info: Do other datasets for the high level task exist? -->
<!-- scope: telescope -->
yes

#### Unique Language Coverage

<!-- info: Does this dataset cover other languages than other datasets for the same task? -->
<!-- scope: periscope -->
yes

#### Difference from other GEM datasets

<!-- info: What else sets this dataset apart from other similar datasets in GEM? -->
<!-- scope: microscope -->
There is another data set containing manually labeled Finnish paraphrases.

#### Ability that the Dataset measures

<!-- info: What aspect of model ability can be measured with this dataset? -->
<!-- scope: periscope -->
Sentence meaning


### GEM-Specific Curation

#### Modificatied for GEM?

<!-- info: Has the GEM version of the dataset been modified in any way (data, processing, splits) from the original curated data? -->
<!-- scope: telescope -->
yes

#### GEM Modifications

<!-- info: What changes have been made to he original dataset? -->
<!-- scope: periscope -->
`other`

#### Modification Details

<!-- info: For each of these changes, described them in more details and provided the intended purpose of the modification -->
<!-- scope: microscope -->
Training sets have been prepared for each the "quality levels" 60% – 95%.

In the original release, this task was left to the user of the data.

#### Additional Splits?

<!-- info: Does GEM provide additional splits to the dataset? -->
<!-- scope: telescope -->
yes

#### Split Information

<!-- info: Describe how the new splits were created -->
<!-- scope: periscope -->
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.

#### Split Motivation

<!-- info: What aspects of the model's generation capacities were the splits created to test? -->
<!-- scope: periscope -->
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. 


### Getting Started with the Task

#### Pointers to Resources

<!-- 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. -->
<!-- scope: microscope -->
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).

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.

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.

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). 

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.




## Previous Results

### Previous Results

#### Measured Model Abilities

<!-- info: What aspect of model ability can be measured with this dataset? -->
<!-- scope: telescope -->
Sentence meaning

In a scenario of paraphrase detection, the model determines whether two given sentences carry approximately the same meaning.

In a scenario of paraphrase generation, the model generates a potential paraphrase of a given sentence.

#### Metrics

<!-- info: What metrics are typically used for this task? -->
<!-- scope: periscope -->
`BLEU`, `BERT-Score`, `Other: Other Metrics`

#### Other Metrics

<!-- info: Definitions of other metrics -->
<!-- scope: periscope -->
PINC

#### Proposed Evaluation

<!-- info: List and describe the purpose of the metrics and evaluation methodology (including human evaluation) that the dataset creators used when introducing this task. -->
<!-- scope: microscope -->
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.

#### Previous results available?

<!-- info: Are previous results available? -->
<!-- scope: telescope -->
yes

#### Other Evaluation Approaches

<!-- info: What evaluation approaches have others used? -->
<!-- scope: periscope -->
See publications on using Opusparcus

#### Relevant Previous Results

<!-- info: What are the most relevant previous results for this task/dataset? -->
<!-- scope: microscope -->
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).



## Dataset Curation

### Original Curation

#### Original Curation Rationale

<!-- info: Original curation rationale -->
<!-- scope: telescope -->
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).

#### Communicative Goal

<!-- info: What was the communicative goal? -->
<!-- scope: periscope -->
Opusparcus provides labeled examples of pairs of sentences that have similar (or dissimilar) meanings.

#### Sourced from Different Sources

<!-- info: Is the dataset aggregated from different data sources? -->
<!-- scope: telescope -->
no


### Language Data

#### How was Language Data Obtained?

<!-- info: How was the language data obtained? -->
<!-- scope: telescope -->
`Crowdsourced`

#### Where was it crowdsourced?

<!-- info: If crowdsourced, where from? -->
<!-- scope: periscope -->
`Other crowdworker platform`

#### Language Producers

<!-- info: What further information do we have on the language producers? -->
<!-- scope: microscope -->
The data in Opusparcus has been extracted from [OpenSubtitles2016](http://opus.nlpl.eu/OpenSubtitles2016.php), which is in turn based on data from [OpenSubtitles.org](http://www.opensubtitles.org/).

The texts consists of subtitles that have been produced using crowdsourcing.

#### Topics Covered

<!-- info: Does the language in the dataset focus on specific topics? How would you describe them? -->
<!-- scope: periscope -->
The language is representative of movies and TV shows. Domains covered include comedy, drama, relationships, suspense, etc.

#### Data Validation

<!-- info: Was the text validated by a different worker or a data curator? -->
<!-- scope: telescope -->
validated by data curator

#### Data Preprocessing

<!-- info: How was the text data pre-processed? (Enter N/A if the text was not pre-processed) -->
<!-- scope: microscope -->
Sentence and word tokenization was performed.

#### Was Data Filtered?

<!-- info: Were text instances selected or filtered? -->
<!-- scope: telescope -->
algorithmically

#### Filter Criteria

<!-- info: What were the selection criteria? -->
<!-- scope: microscope -->
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.

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.


### Structured Annotations

#### Additional Annotations?

<!-- quick -->
<!-- info: Does the dataset have additional annotations for each instance? -->
<!-- scope: telescope -->
expert created

#### Number of Raters

<!-- info: What is the number of raters -->
<!-- scope: telescope -->
11<n<50

#### Rater Qualifications

<!-- info: Describe the qualifications required of an annotator. -->
<!-- scope: periscope -->
Students and staff at the University of Helsinki (native or very proficient speakers of the target languages)

#### Raters per Training Example

<!-- info: How many annotators saw each training example? -->
<!-- scope: periscope -->
0

#### Raters per Test Example

<!-- info: How many annotators saw each test example? -->
<!-- scope: periscope -->
2

#### Annotation Service?

<!-- info: Was an annotation service used? -->
<!-- scope: telescope -->
no

#### Annotation Values

<!-- info: Purpose and values for each annotation -->
<!-- scope: microscope -->
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.

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.

Annotators could also reject a sentence pair as being corrupted data.

#### Any Quality Control?

<!-- info: Quality control measures? -->
<!-- scope: telescope -->
validated by another rater

#### Quality Control Details

<!-- info: Describe the quality control measures that were taken. -->
<!-- scope: microscope -->
If the annotators disagreed by more than one category, the sentence pair was discarded and is not part of the final dataset.


### Consent

#### Any Consent Policy?

<!-- info: Was there a consent policy involved when gathering the data? -->
<!-- scope: telescope -->
no


### Private Identifying Information (PII)

#### Contains PII?

<!-- quick -->
<!-- info: Does the source language data likely contain Personal Identifying Information about the data creators or subjects? -->
<!-- scope: telescope -->
yes/very likely

#### Any PII Identification?

<!-- info: Did the curators use any automatic/manual method to identify PII in the dataset? -->
<!-- scope: periscope -->
no identification


### Maintenance

#### Any Maintenance Plan?

<!-- info: Does the original dataset have a maintenance plan? -->
<!-- scope: telescope -->
no



## Broader Social Context

### Previous Work on the Social Impact of the Dataset

#### Usage of Models based on the Data

<!-- 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? -->
<!-- scope: telescope -->
no


### Impact on Under-Served Communities

#### Addresses needs of underserved Communities?

<!-- 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). -->
<!-- scope: telescope -->
no


### Discussion of Biases

#### Any Documented Social Biases?

<!-- 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. -->
<!-- scope: telescope -->
no

#### Are the Language Producers Representative of the Language?

<!-- 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? -->
<!-- scope: periscope -->
What social bias there may be in the subtitles in this dataset has not been studied.



## Considerations for Using the Data

### PII Risks and Liability

#### Potential PII Risk

<!-- 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. -->
<!-- scope: microscope -->
The data only contains subtitles of publicly available movies and TV shows.


### Licenses

#### Copyright Restrictions on the Dataset

<!-- 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? -->
<!-- scope: periscope -->
`non-commercial use only`

#### Copyright Restrictions on the Language Data

<!-- 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? -->
<!-- scope: periscope -->
`non-commercial use only`


### Known Technical Limitations

#### Technical Limitations

<!-- 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. -->
<!-- scope: microscope -->
Some subtitles contain typos that are caused by inaccurate OCR.

#### Unsuited Applications

<!-- 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. -->
<!-- scope: microscope -->
The models might memorize individual subtitles of existing movies and TV shows, but there is no context across sentence boundaries in the data.

#### Discouraged Use Cases

<!-- 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. -->
<!-- scope: microscope -->
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).