QE4PE Pretask

The goal of the pretask is to familiarize translators with the GroTE interface, indentify potential issue and inform the modality assignment for the main task to ensure a uniform distribution of editing times across modalities. Refer to the translators' guidelines for additional details about the task.

Folder Structure

pretask/
├── inputs/
│   ├── eng-ita/
│   │   ├── pretask_eng-ita_doc1_input.txt
│   │   ├── pretask_eng-ita_doc2_input.txt
│   │   └── ... # GroTE input files with tags and ||| source-target separator
│   └── eng-nld/
│   │   ├── pretask_eng-nld_doc1_input.txt
│   │   ├── pretask_eng-nld_doc2_input.txt
│   │   └── ... # GroTE input files with tags and ||| source-target separator
├── outputs/
│   ├── eng-ita/
│   │   ├── logs/
│   │   │   ├── pretask_eng-ita_t1_logs.csv 
│   │   │   └── ... # GroTE logs for every translator (e.g. t1)
│   │   ├── metrics/
│   │   │   ├── pretask_eng-ita_t1_metrics.csv
│   │   │   └── ... # Metrics for every translator (e.g. t1)
│   │   ├── pretask_eng-ita_doc1_t1_output.txt
│   │   └── ... # GroTE output files (one edited target per line)
│   └── eng-nld/
│       ├── logs/
│       │   ├── pretask_eng-nld_t1_logs.csv 
│       │   └── ... # GroTE logs for every translator (e.g. t1)
│       ├── metrics/
│       │   ├── pretask_eng-nld_t1_metrics.csv
│       │   └── ... # Metrics for every translator (e.g. t1)
│       ├── example_eng-nld_doc1_t1_output.txt
│       └── ... # GroTE output files (one post-edited segment per line)
├── doc_id_map.json # Source and doc name maps
├── main_task_assignments.json # Translator assignments to main task modalities
├── pretask_eng-ita_guidelines.pdf # Task guidelines for translators
├── pretask_eng-nld_guidelines.pdf # Task guidelines for translators
└── README.md

Inputs

The pretask uses six documents containing between 6-10 contiguous segments from the same wmt23 collection of the main task that were matching all the main task requirements, but were not selected for the main collection. doc_id_map.json shows the document assignments from the original collection. We use the supervised highlights produced by XCOMET-XXL to familiarize translators with editing highlighted texts with highlights being assumed to be of a good but not "perfect" quality (as opposed to oracle). Word-level error spans are extended to match closest word boundaries whenever necessary. Input files for the task have names using the format:

"{{TASK_ID}}_{{TRANSLATION_DIRECTION}}_{{DOC_ID}}_input.txt"

Main Task Assignments

From the pretask results, we create 3 groups per translation direction representing faster (1), average (2) and slower (3) translators by splitting the time-ranked list of translators in three blocks of equal size. In every block, we assign a modality to each translator randomly (random seed = 42). All gaps >5 minutes in the logs are omitted from the calculation to account for accidental AFK time during logging.

The final assignments (also in main_task_assignments.json for machine-readable format) are:

English - Italian

Name	Time	Modality	Alias
Group 1
t8	37m	supervised	supervised_t1
t11	45m	oracle	oracle_t1
t10	51m	no_highlight	no_highlight_t1
t7	62m	unsupervised	unsupervised_t1
Group 2
t1	68m	oracle	oracle_t2
t5	70m	unsupervised	unsupervised_t2
t6	74m	supervised	supervised_t2
t12	100m	no_highlight	no_highlight_t2
Group 3
t9	106m	no_highlight	no_highlight_t3
t3	122m	oracle	oracle_t3
t2	164m	unsupervised	unsupervised_t3
t4	185m	supervised	supervised_t3

English - Dutch

Name	Time	Modality	Alias
Group 1
t4	23m	oracle	oracle_t1
t12	27m	supervised	supervised_t1
t1	31m	no_highlight	no_highlight_t1
t7	36m	unsupervised	unsupervised_t1
Group 2
t8	44m	supervised	supervised_t2
t2	44m	unsupervised	unsupervised_t2
t11	65m	oracle	oracle_t2
t5	66m	no_highlight	no_highlight_t2
Group 3
t9	68m	no_highlight	no_highlight_t3
t6	76m	supervised	supervised_t3
t3	103m	unsupervised	unsupervised_t3
t10	152m	oracle	oracle_t3

t13 was added for Dutch as a replacement for t9, which compromised a part of the task by not following the guidelines. t13 was assigned to no_highlight using no_highlight_t4 as an Alias. The pre-task editing time for t13 was 19 minutes.

Outputs

Files in outputs/eng-ita and outputs/eng-nld contain post-edited outputs (one per line, matching inputs) using format:

"{{TASK_ID}}_{{TRANSLATION_DIRECTION}}_{{DOC_ID}}_{{TRANSLATOR_ID}}_output.txt"

The contents of outputs/{{TRANSLATION_DIRECTION}}/logs can be used to analyze editing behavior at a granular scale. Each log file has format:

"{{TASK_ID}}_{{TRANSLATION_DIRECTION}}_{{TRANSLATOR_ID}}_logs.csv"

Refer to GroTE documentation for more information about the logs.

The contents of outputs/{{TRANSLATION_DIRECTION}}/metrics contain aggregated metrics for each translator, with name format:

"{{TASK_ID}}_{{TRANSLATION_DIRECTION}}_{{TRANSLATOR_ID}}_metrics.csv"

Metrics include:

max/min/mean/std of BLEU, chrF, TER and COMET scores computed between the MT or the PE texts and all (other, in case of PE) post-edited variants of the same segment for {{TRANSLATION_DIRECTION}}.
XCOMET-XXL segment-level QE scores and errors for the MT and PE texts.