Languages: en de zh
Multilinguality: translation
Size Categories: 1K<n<10K
Language Creators: found
Source Datasets: extended|wikipedia

# Dataset Card Creation Guide

### Dataset Summary

From the homepage: This shared task (part of WMT20) will build on its previous editions to further examine automatic methods for estimating the quality of neural machine translation output at run-time, without relying on reference translations. As in previous years, we cover estimation at various levels. Important elements introduced this year include: a new task where sentences are annotated with Direct Assessment (DA) scores instead of labels based on post-editing; a new multilingual sentence-level dataset mainly from Wikipedia articles, where the source articles can be retrieved for document-wide context; the availability of NMT models to explore system-internal information for the task.

Task 1 evaluates the application of QE for post-editing purposes. It consists of predicting:

• Word-level tags. This is done both on source side (to detect which words caused errors) and target side (to detect mistranslated or missing words).
• Target. Each token is tagged as either OK or BAD. Additionally, each gap between two words is tagged as BAD if one or more missing words should have been there, and OK otherwise. Note that number of tags for each target sentence is 2N+1, where N is the number of tokens in the sentence.*
• Source. Tokens are tagged as OK if they were correctly translated, and BAD otherwise. Gaps are not tagged.
• Sentence-level HTER scores. HTER (Human Translation Error Rate) is the ratio between the number of edits (insertions/deletions/replacements) needed and the reference translation length.

From the homepage:

For sentence-level QE, submissions are evaluated in terms of the Pearson's correlation metric for the sentence-level HTER prediction. For word-level QE, they will be evaluated in terms of MCC (Matthews correlation coefficient). These are the official evaluation scripts.

### Languages

There are two language pairs in this dataset:

• English - German (en - de)
• German - Chinese (en - zh)

## Dataset Structure

### Data Instances

An example looks like this:

{
'translation': {
'en': 'favorite fish include cod , salmon , winter flounder , haddock , striped bass , pollock , hake , bluefish , and , in southern New England , Tautog .',
'de': 'zu den Lieblingsfischen gehören Kabeljau , Lachs , Winterflounder , Schellfisch , gestreifter Bass , Pollock , Seehecht , Rotbarsch und in Südengland Tautog .',
}
'src_tags': [1, 1, 0, 1, 1, 1, 1, 1, 1, 1, 0, 1, 1, 0, 1, 0, 1, 1, 1, 1, 1, 1, 0, 0, 0, 0, 0, 0, 1, 1],
'mt_tags': [1, 1, 1, 1, 1, 1, 1, 0, 1, 1, 1, 1, 1, 1, 1, 1, 1, 0, 1, 1, 1, 1, 1, 1, 1, 0, 1, 0, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 0, 1, 1, 0, 0, 1, 0, 1, 1, 1, 1, 1],
'pe': 'zu den Lieblingsfischen zählen Kabeljau , Lachs , Winterflunder , Schellfisch , Wolfsbarsch , Pollock , Seehecht , Bluefish und im Süden Neuenglands Tautog .',
'hter': 0.3199999928474426,
'alignments': [[2, 0], [2, 1], [2, 3], [3, 2], [3, 4], [4, 5], [5, 6], [6, 5], [7, 6], [8, 6], [9, 7], [10, 8], [10, 10], [11, 9], [12, 12], [13, 13], [14, 11], [15, 12], [15, 15], [16, 14], [17, 17], [19, 16], [20, 16], [21, 20], [22, 18], [23, 19], [23, 21], [24, 22], [25, 21], [26, 22], [27, 22], [28, 23], [29, 24]],
}


### Data Fields

• translation: Dictionary with pairs (source,target).
• src_lg: sequence of text in source language.
• tgt_lg: sequence of text in target language.
• src_tags: source word-level tags. 0=BAD, 1=OK. [] if N/A (only for test).
• mt_tags: target word-level tags. 0=BAD, 1=OK. [] if N/A (only for test).
• pe: post-edited version of NMT output. "" if N/A (only for test).
• hter: human translation error rate. -10_000 if N/A (only for test).
• alignments: Word aligments. List of pairs of integers.

### Data Splits

There are 2 configurations in this dataset (one for each available language pair). Each configuration is composed of 7K examples for training, 1K for validation and 1K for (blind) test.

## Dataset Creation

### Curation Rationale

The original text is extracted from Wikipedia.

From the homepage:

Word-level labels have been obtained by using the alignments provided by the TER tool (settings: tokenised, case insensitive, exact matching only, disabling shifts by using the -d 0 option) between machine translations and their post-edited versions. Shifts (word order errors) were not annotated as such (but rather as deletions + insertions) to avoid introducing noise in the annotation.

HTER values are obtained deterministically from word-level tags. However, when computing HTER, we allow shifts in TER.

The baseline system is a neural predictor-estimator approach implemented in OpenKiwi (Kepler at al., 2019), where the predictor model will be trained on the parallel data used to train the NMT model.

## Considerations for Using the Data

Unknown

### Citation Information

Not available.


### Contributions

Thanks to @VictorSanh for adding this dataset.

None yet