README.md

---
annotations_creators:
- expert-generated
language:
- en
license: cc-by-sa-4.0
multilinguality:
- monolingual
size_categories:
- 1K<n<10K
source_datasets:
- original
task_categories:
- feature-extraction
- summarization
- question-answering
- text-classification
- fill-mask
- zero-shot-classification
task_ids:
- named-entity-recognition
- topic-classification
- news-articles-summarization
pretty_name: ISAFpressreleases
dataset_info:
  features:
  - name: name
    dtype: string
  - name: eventrefnumber
    dtype: string
  - name: text
    dtype: string
  - name: StartDate
    dtype: timestamp[s]
  - name: eventtype
    dtype: string
  - name: province
    dtype: string
  - name: citydistrict
    dtype: string
  - name: village
    dtype: string
  - name: targetgroup
    dtype: string
  - name: commander
    dtype: string
  - name: position
    dtype: string
  - name: minkilled
    dtype: string
  - name: mincaptured
    dtype: string
  - name: capturedcharacterisation
    dtype: string
  - name: killedcharacterisation
    dtype: string
  - name: killq
    dtype: string
  - name: captureq
    dtype: string
  - name: killcaptureraid
    dtype: string
  - name: airstrike
    dtype: string
  - name: noshotsfired
    dtype: string
  - name: dataprocessed
    dtype: string
  - name: flagged
    dtype: string
  - name: glossarymeta
    dtype: string
  - name: minleaderskilled
    dtype: string
  - name: minfacilitatorskilled
    dtype: string
  - name: minleaderscaptured
    dtype: string
  - name: minfacilitatorscaptured
    dtype: string
  - name: leaderq
    dtype: string
  splits:
  - name: train
    num_bytes: 4351888.182082124
    num_examples: 4098
  - name: test
    num_bytes: 768854.8179178764
    num_examples: 724
  download_size: 1781814
  dataset_size: 5120743.0
configs:
- config_name: default
  data_files:
  - split: train
    path: data/train-*
  - split: test
    path: data/test-*
---

## ISAF Press Releases Dataset Description

- **Homepage:** [N/A]
- **Repository:** [N/A]
- **Paper:** [A Knock on the Door: 22 Months of ISAF Press Releases](https://www.afghanistan-analysts.org/en/special-reports/a-knock-on-the-door-22-months-of-isaf-press-releases/)
- **Point of Contact:** Alex Strick van Linschoten ([@strickvl](https://huggingface.co/strickvl))

### Dataset Summary

The ISAF Press Releases dataset contains data used as the basis for the research
paper "A Knock on the Door: 22 Months of ISAF Press Releases". The dataset
provides a comprehensive collection of press releases issued by the
International Security Assistance Force (ISAF) in Afghanistan from December 1,
2009, to February 21, 2013. The press releases were collected, processed, and
annotated to extract information about kill-capture missions carried out by ISAF
during this period. The dataset offers valuable insights into the nature and
extent of these operations, providing a historical record of ISAF's activities
in Afghanistan. It consists of 4822 press release reports, each labelled with 
information about the event, including the date, location, target group, and
the number of people killed or captured (as represented in the data). The
dataset is provided as a single Parquet file without predefined splits.

### Supported Tasks and Leaderboards

- `named-entity-recognition`: The dataset can be used to train a model for Named Entity Recognition, which involves identifying mentions of people killed or captured in the press releases. Success on this task is typically measured by achieving a *high* [F1 score](https://huggingface.co/metrics/f1). The baseline metric would be to compare against the labels provided in the dataset.

- `event-extraction`: The dataset supports Event Extraction, which involves identifying occurrences of kills or captures mentioned in the press releases. Success on this task is typically measured by achieving a *high* [F1 score](https://huggingface.co/metrics/f1). The baseline metric would be to compare against the labels provided in the dataset.

- `text-classification`: The dataset can be used for Document Classification, where the goal is to classify press releases as mentioning kills, captures, or other categories. Success on this task is typically measured by achieving a *high* [Accuracy](https://huggingface.co/metrics/accuracy). The baseline metric would be to compare against the labels provided in the dataset.

### Languages

The press releases in the dataset are entirely in English. They contain military jargon and Afghanistan-specific places and context, which are explained in the associated research paper.

## Dataset Structure

### Data Instances

Here is an example instance from the dataset:

```json
{
 "name": "Afghan-Led Clearing Operation Kills more than 20 Insurgents in Paktiya",
 "eventrefnumber": "2010-08-CA-135",
 "text": "Afghan-Led Clearing Operation Kills more than 20 Insurgents in Paktiya\n ISAF Joint Command - Afghanistan 2010-08-CA-135  For Immediate Release Download PDF\n\n KABUL, Afghanistan (Aug. 12) - An Afghan and coalition security force killed more than 20 armed insurgents during an on-going clearing operation aimed at disrupting the Haqqani Network's freedom of movement in Dzadran district of Paktiya province.   This area is known Haqqani Network safe haven and used to stage attacks into Kabul and the Khost-Gardez pass. When the security force arrived, dozens of insurgents occupied entrenched fighting positions throughout the mountainous area, preparing to engage the security force with machine guns, rocket-propelled grenades and automatic weapons. An air weapons team suppressed the enemy resulting in more than 20 insurgents killed so far.   \"The Haqqani Netrwork is a prevalent insurgent threat in Afghanistan right now. Afghan and coalition forces are focused on smothering their influence and power,\" said U.S. Army Col. Rafael Torres, International Security Assistance Force Joint Command Combined Joint Operation Center director. \"Every day we are taking more insurgents off the battlefield so that Afghans can live in peace.\"   This clearing operation is still ongoing. ISAF will provide updates as they become available.",
 "StartDate": "2010-08-12",
 "eventtype": "insurgentskilled",
 "province": "Paktya",
 "citydistrict": "Zadran",
 "village": "",
 "targetgroup": "Haqqani",
 "commander": "",
 "position": "",
 "minkilled": "21",
 "mincaptured": "0",
 "capturedcharacterisation": "",
 "killedcharacterisation": "more than 20",
 "killq": "true",
 "captureq": "false",
 "killcaptureraid": "true",
 "airstrike": "true",
 "noshotsfired": "false",
 "dataprocessed": "true",
 "flagged": "false",
 "glossarymeta": "false",
 "minleaderskilled": "0",
 "minfacilitatorskilled": "0",
 "minleaderscaptured": "0",
 "minfacilitatorscaptured": "0",
 "leaderq": "false"
}
```

### Data Fields
- `name`: The name or title of the press release.
- `eventrefnumber`: The reference number for the event, usually mentioned in the press release itself.
- `text`: The full text of the press release.
- `StartDate`: The date of the event itself, not the press release.
- `eventtype`: The type of event (e.g., "insurgentskilled").
- `province`: The province where the event occurred.
- `citydistrict`: The city or district where the event occurred.
- `village`: The village where the event occurred.
- `targetgroup`: The target group (e.g., "Haqqani") mentioned in the press release.
- `commander`: A boolean indicating whether a "commander" was part of the event.
- `position`: The position of the targeted individual(s) mentioned in the press release.
- `minkilled`: The minimum number of people killed based on the text.
- `mincaptured`: The minimum number of people captured based on the text.
- `capturedcharacterisation`: The language used to characterize the number of people captured.
- `killedcharacterisation`: The language used to characterize the number of people killed.
- `killq`: A boolean indicating whether someone was killed.
- `captureq`: A boolean indicating whether someone was captured.
- `killcaptureraid`: A boolean indicating whether the event was a kill-capture raid.
- `airstrike`: A boolean indicating whether an airstrike was involved.
- `noshotsfired`: A boolean indicating whether the press release mentioned that the raid was carried out with no shots fired.
- `dataprocessed`: An internal field indicating whether the entry was processed.
- `flagged`: An internal field indicating whether the entry was flagged for being interesting.
- `glossarymeta`: A boolean indicating whether the entry contains interesting characterizations for compiling a glossary.
- `minleaderskilled`: The minimum number of "leaders" killed.
- `minfacilitatorskilled`: The minimum number of "facilitators" killed.
- `minleaderscaptured`: The minimum number of "leaders" captured.
- `minfacilitatorscaptured`: The minimum number of "facilitators" captured.
- `leaderq`: A boolean indicating whether the press release claimed that a "leader" was captured or killed.

### Data Splits
The dataset is provided as a single Parquet file without predefined splits.
Original JSON files are available at `original_json_data/` in this repository,
as is the script (`scripts/export_script.py`) that was used to convert the
individual JSON files to the single `.parquet` file.

## Dataset Creation
### Curation Rationale
The dataset was created to provide a comprehensive and publicly accessible record of ISAF's kill-capture missions in Afghanistan, as reported in their press releases. The motivation was to enable scholars, legal teams, and others to analyze and understand the nature and extent of these operations, as the original ISAF website no longer exists in the same form. The dataset serves as an important historical artifact for Afghan history and provides a means to reverse-engineer the minimum numbers of people killed and captured by ISAF during the specified period.

## Source Data
### Initial Data Collection and Normalization
The initial data collection involved manually copying the text of press releases from the ISAF website (http://www.isaf.nato.int/article/isaf-releases/) into Tinderbox software. The press releases were collected from December 1, 2009, to February 21, 2013, covering a period of over 21 months. All available press releases during this period were included in the dataset.

The collected press releases were then processed to split them into individual incident reports. If a press release mentioned multiple incidents, they were separated into distinct entries. The text of the press releases was not modified and remains in its original form.

### Who are the source language producers?
The press releases were written by the press office and media relations team of ISAF (International Security Assistance Force) and later NATO (North Atlantic Treaty Organization) during the specified period. They were created by human writers as official communications from ISAF/NATO.

## Annotations
### Annotation process
The annotation process involved reading each press release and evaluating it against a set of variables. The annotations were performed using Tinderbox software. The variables included:

- Basic data: Incident name, reference number, date of the incident.
- Location data: Province, district, village name (if provided).
- Target data: Target group, claimed capture of a "leader" or someone in a leadership position, specific position of the target.
- Numerics: Indication of someone being killed or detained, minimum number
  killed or detained, exact terms used to refer to those detained or killed,
  numbers of "leaders" and "facilitators" claimed to be killed or captured,
  classification as a "capture-or-kill" raid, involvement of an airstrike. The
  annotator used a fixed list of interpretations for certain terms when
  estimating the minimum numbers of people killed or detained. Detailed
  explanations of the annotation process and variable definitions are provided
  in the associated research paper.

### Who are the annotators?
The sole annotator of the dataset was Alex Strick van Linschoten, the lead researcher on the project, with support from Felix Kuehn.

### Personal and Sensitive Information
The dataset contains names and locations mentioned in the original press releases, which were publicly issued by ISAF. No additional personal or sensitive information was included, and no steps were taken to anonymize or protect the identities of individuals mentioned in the press releases.

## Considerations for Using the Data
### Social Impact of Dataset
The dataset provides an important historical record of ISAF's activities in Afghanistan during the specified period. It enables scholars, legal teams, and others to access and analyze this information, contributing to a better understanding of the events and their impact. The availability of this data is crucial for Afghan history and ensures that the information remains accessible even though the original ISAF website no longer exists in its previous form.

However, it is important to consider the potential biases and limitations of the dataset, as discussed in the following sections.

### Discussion of Biases
The dataset reflects the inherent biases and limitations of ISAF's presence and understanding of Afghanistan. The press releases were created by ISAF/NATO and may not provide a complete or unbiased account of the events. The categorization of those detained or killed may not be standardized across the entire period, and certain incidents may not have been published.

It is important to recognize that the dataset represents the perspective and narrative of ISAF/NATO and may not necessarily reflect the experiences or viewpoints of other stakeholders, particularly the Afghan population.

### Other Known Limitations
The dataset has several known limitations:

- Incomplete coverage: The dataset does not provide a complete picture of all ISAF's capture-or-kill operations in Afghanistan during the specified period. It is limited to the incidents reported in the press releases.
- Underrepresentation of air strikes: Air strikes are underrepresented in the dataset due to the vague descriptions of "air assets" involved in incidents. It is often impossible to determine the type of weapons or aircraft used.
- Potential inconsistencies: The categorization of those detained or killed may not be standardized across the entire period, and there may be inconsistencies in the reporting of incidents.
- Users of the dataset should be aware of these limitations and consider them when drawing conclusions or making interpretations based on the data.

## Additional Information
### Dataset Curators
The dataset was created by Alex Strick van Linschoten and Felix Kuehn as part of their research project.

### Licensing Information

This dataset is licensed under the Creative Commons Attribution-ShareAlike 4.0 International (CC BY-SA 4.0) license. For more information, see [https://creativecommons.org/licenses/by-sa/4.0/](https://creativecommons.org/licenses/by-sa/4.0/).

### Citation Information

```
@article{strick_van_linschoten_kuehn_2011,
  author    = {Alex Strick van Linschoten and Felix Kuehn},
  title     = {A Knock on the Door: 22 Months of ISAF Press Releases},
  journal   = {Afghanistan Analysts Network},
  year      = {2011},
  month     = {October},
  day       = {12},
  url       = {https://www.afghanistan-analysts.org/en/special-reports/a-knock-on-the-door-22-months-of-isaf-press-releases/}
}
```

## Contributions

Many thanks to the [Afghanistan Analysts Network](https://www.afghanistan-analysts.org/en) for funding the research.