Main task almost finished

task/.gitattributes

@@ -1 +1 @@
-main/outputs/eng-nld/logs/main_eng-nld_t3_unsupervised_logs.csv filter=lfs diff=lfs merge=lfs -text
+main/outputs/eng-nld/logs/main_eng-nld_unsupervised_t3_logs.csv filter=lfs diff=lfs merge=lfs -text

task/example/README.md

@@ -11,10 +11,11 @@ example/
 │       ├── example_eng-ita_doc1_input.txt
 │       └── ... # GroTE input files with tags and ||| source-target separator
 ├── outputs/
-│   ├── logs/
-│   │   └── example_eng-ita_t0_logs.csv # GroTE logs
-│   ├── example_eng-ita_doc1_t0_output.txt
-│   └── ... # GroTE output files (one edited target per line)
+│   └── eng-ita/
+│       ├── logs/
+│       │   └── example_eng-ita_t0_logs.csv # GroTE logs
+│       ├── example_eng-ita_doc1_t0_output.txt
+│       └── ... # GroTE output files (one edited target per line)
 ├── doc_id_map.json # Source and doc name maps
 └── README.md
 ```
@@ -40,7 +41,7 @@ Files in `outputs/{{TRANSLATION_DIRECTION}}` contain post-edited outputs (one pe
 The contents of `outputs/{{TRANSLATION_DIRECTION}}/logs` can be used to analyze editing behavior at a granular scale. Each log file has format:
 
 ```python
-"{{TASK_ID}}_{{TRANSLATION_DIRECTION}}_{{TRANSLATOR_ID}}_logs.txt"
+"{{TASK_ID}}_{{TRANSLATION_DIRECTION}}_{{TRANSLATOR_ID}}_logs.csv"
 ```
 
 Refer to [GroTE](https://github.com/gsarti/grote) documentation for more information about the logs.

task/main/README.md

@@ -0,0 +1,74 @@
+# QE4PE Main Task
+
+The goal of the `main` task is to evaluate the speed and quality of translators using different word-level highlight modalities, with the goal of estimating the impact of highlights over the editing process. Refer to the [translators' guidelines](./main_guidelines.pdf) for additional details about the task.
+
+## Folder Structure
+
+```shell
+main/
+├── inputs/
+│   ├── eng-ita/
+│   │   ├── main_eng-ita_doc1_no_highlight_input.txt
+│   │   ├── main_eng-ita_doc1_oracle_input.txt
+│   │   └── ... # GroTE input files with tags and ||| source-target separator
+│   └── eng-nld/
+│   │   ├── main_eng-nld_doc1_no_highlight_input.txt
+│   │   ├── main_eng-nld_doc1_oracle_input.txt
+│   │   └── ... # GroTE input files with tags and ||| source-target separator
+├── outputs/
+│   ├── eng-ita/
+│   │   ├── logs/
+│   │   │   ├── main_eng-ita_no_highlight_t1_logs.csv 
+│   │   │   └── ... # GroTE logs for every translator (e.g. no_highlight_t1)
+│   │   ├── main_eng-ita_doc1_no_highlight_t1_output.txt
+│   │   └── ... # GroTE output files (one edited target per line)
+│   └── eng-nld/
+│       ├── logs/
+│       │   ├── main_eng-nld_no_highlight_t1_logs.csv 
+│       │   └── ... # GroTE logs for every translator (e.g. no_highlight_t1)
+│       ├── example_eng-nld_doc1_no_highlight_t1_output.txt
+│       └── ... # GroTE output files (one edited target per line)
+├── doc_id_map.json # Source and doc name maps
+├── main_guidelines.pdf # Task guidelines for translators
+└── README.md
+```
+
+## Inputs
+
+The `main` task uses 50 documents containing between 6-10 contiguous sentences (tot. ) from the `wmt23` collection that were filtered ad-hoc to ensure sufficient but not perfect quality. Moreover, 15 critical errors were manually introduced in some translations across the two target languages before applying highlight modalities (more info in [manual_errors](../../setup/processed/manual_errors/)).
+
+>The document `doc28` corresponding to the original WMT23 file `aaronsgiles.110351251990540992` was also erroneously included in the main task, but it was also used for the `pretask` and was hence excluded from the main analysis to account for preview effects of editing times.
+
+The four modalities of highlighting tested are `no_highlight`, `oracle`, `supervised` and `unsupervised`:
+
+- `no_highlight`: MT outputs requiring post-editing are presented without any highlighting.
+
+- `oracle`: Highlights are inferred from majority-voting from a previous set of 3 professional post-edits per translation direction. We mark spans edited by 2/3 editors as potential `minor` errors, and 3/3 as potential `major` errors. Refer to the [oracle folder](../../setup/highlights/oracle) for an overview of the highlights in this modality.
+
+- `supervised`: Highlights are produced by XCOMET-XXL as it was done for the `pretask`. We use `minor` errors and merge `major` and `critical` errors into a single `major` potential error category to match levels of other highlighting modalities. Confidence scores were ignored, since they were found to not affect the highlights agreement with human edits in an important way (see paper).
+
+- `unsupervised`: Highlights are produced from the original MT model NLLB 3.3B using Monte Carlo Dropount Variance of probability scores. Based on the proportion of `minor` and `major` potential error spans for the `supervised` modality, we set two threshold values to ensure a matching amount of spans for each of the two categories.
+
+For all methods, word-level error spans are extended to match closest word boundaries whenever necessary. Input files for the task have names using the format:
+
+```python
+"{{TASK_ID}}_{{TRANSLATION_DIRECTION}}_{{DOC_ID}}_{{HIGHLIGHT_MODALITY}}_input.txt"
+```
+
+## Outputs
+
+Files in `outputs/eng-ita` and `outputs/eng-nld` contain post-edited outputs (one per line, matching inputs) using format:
+
+```python
+"{{TASK_ID}}_{{TRANSLATION_DIRECTION}}_{{DOC_ID}}_{{TRANSLATOR_ID}}_output.txt"
+```
+
+> **Note**: `{{TRANSLATOR_ID}}` includes the `{{HIGHLIGHT_MODALITY}}` information, and reflects the speed of the translator on the pretask (e.g. The translator `no_highlight_t1` belongs to Group 1, i.e. the faster one, from the [`pretask` assignments](../pretask/main_task_assignments.json))
+
+The contents of `outputs/{{TRANSLATION_DIRECTION}}/logs` can be used to analyze editing behavior at a granular scale. Each log file has format:
+
+```python
+"{{TASK_ID}}_{{TRANSLATION_DIRECTION}}_{{TRANSLATOR_ID}}_logs.csv"
+```
+
+Refer to [GroTE](https://github.com/gsarti/grote) documentation for more information about the logs.