View source for Machine learning models/Proposed/Multilingual revert risk

<!--
Welcome to the model card template! Please follow the prompts within comments written below to create a new model card. All arguments marked [required] must be included for a model card to be considered complete. Arguments marked [optional] are not 100% necessary, but are encouraged. Some sections are commented out and marked START [OPTIONAL] SECTION; these are sections of the template considered nonessential. Like arguments marked [optional], they may not be necessary but are encouraged.

Once you believe you are done with this template, please make sure to search the term "[required]" to make sure you've filled in all required fields.

If you want a more stable place with these same questions so that you can copy-paste the answers in, you can make your own duplicate of a Google doc form of this template at the following link: https://docs.google.com/document/d/1Q5aJGGBJB4LN3dXS8_-IjZYi0a3T1MIWtDXDwZeEins/edit?usp=sharing
-->
{{Model-card-draft}}
{{Infobox model card
<!-- [required] Who created this model? Please use wikilinks where possible and the {{List|Person A|Person B|...}} template for multiple people. -->
| creators = {{List|[[User:Trokhymovych|Mykola Trokhymovych]]|Muniza Aslam|Ai-Jou Chou|[[User:Diego_(WMF)|Diego Saez-Trumper]]}}

<!-- [required] Who is currently responsible for this model? Please use wikilinks where possible and the {{List|Person A|Person B|...}} template for multiple people. -->
| owners = {{List|[[User:Diego_(WMF)|Diego Saez-Trumper]]}}

<!-- [required] Where can users go to get predictions from this model? Please use wikilinks if possible. -->
| interface = 

<!-- [optional] Where can users go to see this model's past performance? Please use wikilinks if possible. -->
| past_performance = 

<!-- [optional] Has this model been in any reports or academic papers? Please use wikilinks where possible and the {{List|Paper A|Paper B|...}} template for multiple publications. -->
| publications = 

<!-- [required] Where can users go to find the code for this model and data? Please use wikilinks where possible and the {{List|Link A|Link B|...}} template for multiple repos. -->
| code_repos = {{List|[https://github.com/trokhymovych/KI_multilingual_training training]|[https://gitlab.wikimedia.org/repos/research/knowledge_integrity inference]}}

<!-- [required] Does this model use PII (e.g. IP address, sensitive reader data, etc.) to make predictions? -->
| pii = No

<!-- [required] Is this model in production (i.e. currently hosted by the WMF ML Team)? -->
| production = Yes

<!-- [required] Which project(s) does this model work for? (Possible answers include: 1-2 specific projects, 'many', or 'all') -->
| projects = 

<!-- [required] Single sentence description of the model -->
| schema = This model uses revision content and metadata to predict the risk of being reverted. 
}}

<!-- [required] Introduction:

This is the most important section of your model card, and may be best to write last. In a few sentences (you can take excerpts from other sections), please answer the following prompts:
1. What is the problem your model is trying to address? (Usually a "How can we ____?" question)
How can we help patrollers to detect 
2. Why is the problem difficult to solve without a model?
3. Why should we care about the problem? How would solving this problem be useful?
4. What is the approach your model takes to solving the problem?
5. What data did you use to train your model, and where did it come from?
6. Where can people access your model?
7. What should (and shouldn't) your model be used for?

Write your answers to these questions below.
-->

How can we help editors to identify revisions that need to be “[[en:Wikipedia:Patrolling|patrolled]]”? 

The goal of this model is to detect revisions that might be reverted independently if they were made in good faith or with the intention of creating damage. Wikipedia has a group of dedicated volunteer editors, known as [[en:Wikipedia:Recent_changes_patrol|patrollers]], who work to ensure the accuracy and integrity of the information on the site. These patrollers review and edit articles, monitor for vandalism, and enforce community guidelines. However, their work is not easy, as they have to keep up with the fast pace and language diversity of Wikipedia, where on average, around 16 pages are edited per second in 250+ languages <ref>https://stats.wikimedia.org/</ref>. The aim of this model is to help patrollers quickly identify potential problems, prioritize the work, and revert damaging edits when needed. 

This model is deployed on LiftWing. Right now, it is available for internal usage. This model can be used to detect revisions that might need to be reverted.

== Motivation ==
<!-- [required] What motivates the creation of this model? Which problem are we trying to solve with this model? How would solving this problem positively affect the WMF ecosystem? -->

[[Knowledge Integrity]] is one of the strategic programs of Wikimedia Research with the goal of identifying and addressing threats to content on Wikipedia, increasing the capabilities of patrollers, and providing mechanisms for assessing the reliability of sources<ref name="zia2019"/>. The main goal of the project is to create a new generation of patrolling models, improving accuracy, fairness, and maintainability compared to previous state-of-the-art ORES<ref>https://www.mediawiki.org/wiki/ORES</ref>. 

The current model is able to work on almost any Wikipedia article in any of the 47 chosen languages: <code>['ka', 'lv', 'ta', 'ur', 'eo', 'lt', 'sl', 'hy', 'hr', 'sk', 'eu', 'et', 'ms', 'az', 'da', 'bg', 'sr', 'ro', 'el', 'th', 'bn', 'no', 'hi', 'ca', 'hu', 'ko', 'fi', 'vi', 'uz', 'sv', 'cs', 'he', 'id', 'tr', 'uk', 'nl', 'pl', 'ar', 'fa', 'it', 'zh', 'ru', 'es', 'ja', 'de', 'fr', 'en']</code>

== Users and uses ==

{{ContentGrid|content=
{{Colored box
|title = Use this model for
<!-- [required] What should this model be used for? (use a bullet list for this)

Some general examples:
* high-level analyses of Wikipedia dynamics
* predicting whether or not an edit might be vandalism
...
-->
|content = 
* Define the revert risk of Wikipedia article revision

}}

{{Colored box
|title = Don't use this model for
<!-- [required] What should this model NOT be used for? (use a bullet list for this)

Some general examples:
* making predictions on other language editions of Wikipedia or other Wiki projects (Wiktionary, Wikinews, Wikidata, etc.)
* auto-removing edits that a user makes without another editor in the loop
...
-->
|content = 
* making predictions on language editions of Wikipedia that are not in the listed 47 languages or other Wiki projects (Wiktionary, Wikinews, Wikidata, etc.)
* making predictions on the revisions that are created by bots
* making predictions on the revisions that create a new article (the first revision of a page)
* making predictions on a revision that is the only one for a page
* As any AI/ML model, we recommend to keep humans in the loop, and not consider model's predictions as training data for other ML models. 
}}

{{Colored box
|title = Current uses
<!-- [required] In which contexts is this model currently used? (use a bullet list for this) -->
|content = 

}}
}}

== Ethical considerations, caveats, and recommendations ==
<!-- [required]
- Are there any shortfalls, ethical quandaries, or caveats in your model or data?
- Are there any known gaps in model performance, particularly on sensitive subclasses? (e.g. certain countries, languages, types of editors/readers, gender, race/ethnicity, etc.)
- Does your model rest on assumptions about reader, editor, or wiki behavior?
-->
* This model was developed to improve the performance of it's [[Machine learning models/Proposed/Language-agnostic revert risk|Language Agnostic (RRLA)]] version. The Multilingual version shows a better performance, especially for IP edits. However, it requires more processing power, and might be slower (or given timeouts).

* This models relies on [https://github.com/google-research/bert/blob/master/multilingual.md Multilingual Bert], a [[en:Large Language Model|Large Language Model]], that might contain [https://link.springer.com/article/10.1007/s12559-021-09881-2 certain biases].

== Model ==
The presented model is based on content features extracted using fine-tuned language model mBERT<ref>https://huggingface.co/bert-base-multilingual-cased</ref>, mwedittypes<ref>https://github.com/geohci/edit-types</ref> based features, along with user and page metadata. It is built in a paradigm of having one generalized model for all covered languages, which is currently the 47 most frequently edited languages in Wikipedia. The system includes the following steps:
 
1. Text features preparation:
* Process wikitext and compare with parent revision
* Extract mwedittypes-based features
* Extract texts that were added, removed, and changed
2. Masked Language Models (MLMs) features extraction:
* Pass each of the texts that were added, removed, or changed to the pre-trained classification model
* Apply mean and max pooling to the list of scores of each signal to extract the final unified feature set
3. Final Classification
* Combine all extracted features with user and revision metadata
* Pass the features to the final classifier

[[File:System design inference.png|thumb|left|System design. Inference]]

=== Performance ===
<!-- [required] Please put any tables, graphics, or other measures of model performance on training and testing data here -->



<!-- START [OPTIONAL] SECTION: Are there any specific things to note about how the model performs?
==== Performance Notes ====
END [OPTIONAL] SECTION -->

=== Implementation ===
The presented model is a multistage solution that includes the fine-tuned masked language model (mBERT) for feature extraction and the final classifier (CatBoost) for getting the probability of being reverted based on the extracted features. 


{{ContentGrid|content=
{{Colored box
|title = Model architecture
<!-- [required] What is the architecture of your model? What parameters does it use? It may be useful to use the <syntaxhighlight lang='text'></syntaxhighlight> tags in this section to write them as a code block. -->
|content = 
mBERT models tunning (four models for the title, changes, inserts, and removes):
* Learning rate: 2e-5
* Weight Decay: 0.01
* Epochs: 5
* Maximum input length: 512
* Number of encoder attention layers: 12
* Number of decoder attention layers: 12
* Number of attention heads: 12
* Length of encoder embedding: 768
CatBoost:
* Iterations: 5000
* Learning Rate: 0.01
* Loss: Logloss

}}
{{Colored box
|title = Output schema
<!-- [required] What output schema does your model return? It may be useful to use the <syntaxhighlight lang='text'></syntaxhighlight> tags in this section to write them as a code block. -->
|content =<syntaxhighlight lang='text' >
{
  lang: <language code string>,
  rev_id: <revision_id string>,
  score: {
     prediction: <boolean decision result>
     probability: {
        true: <probability of being reverted>,
        false: <probability of being NOT reverted>
  }
}
</syntaxhighlight>
}}

{{Colored box
|title = Example input and output
<!-- [required] Please provide an example input to and output from the model. It may be useful to use the <syntaxhighlight lang='text'></syntaxhighlight> tags in this section to write them as a code block.  -->
|content = Input
<syntaxhighlight lang='text' >
curl https://api.wikimedia.org/service/lw/inference/v1/models/revertrisk-multilingual:predict -X POST -d '{"rev_id": 123855516, "lang":"ru"}'
</syntaxhighlight>
Output
<syntaxhighlight lang='text' >
{
  lang: "ru",
  rev_id: 123855516,
  score: {
     prediction: true
     probability: {
        true: 0.9392203688621521,
        false: 0.0607796311378479
  }
}
</syntaxhighlight>

}}
}}

== Data ==
The model was trained on a dataset collected using the two tables from the [[wikitech:Analytics/Data_Lake|Wikimedia Data Lake]]. We used the [[wikitech:Analytics/Data_Lake/Edits/MediaWiki_history|MediaWiki History]] table, and the [[wikitech:Analytics/Data_Lake/Content/Mediawiki_wikitext_history|Wikitext History]] one. Snapshot dated 2022-07 was used with the observation period from 2022-01-01 to 2022-07-01 (6 months) for training and the following week for testing. We also filtered out revisions related to [[en:Wikipedia:Edit war|edit wars]] and revisions created by [[en:Wikipedia:Bot|bots]]. 



{{ContentGrid|content=
{{Colored box
|title = Data Pipeline
<!-- [required] Broadly speaking, how was the data for this model collected, labeled, preprocessed, and fed into the model? -->
|content = The data was collected using [[wikitech:Analytics/Data_Lake|Wikimedia Data Lake]] and [[wikitech:Analytics/Cluster|Wikimedia Analytics cluster]].
For each language, we collected revisions data. Then we merged the wikitext data and extracted the required features from the content using udf functions. Data collection pipeline for one language can be found in [https://github.com/trokhymovych/KI_multilingual_training/blob/main/modules/data_collection.py data collection script]
}}

{{Colored box
|title = Training data
<!-- [required] How was the training data split off from the data writ large? -->
|content = 
* Data period: 6 months
* Number of revisions: 8,586,362
* [[en:Wikipedia:IP users|IP users]] edits rate: 0.17
* Revert rate: 0.08
* Random sample of up to 300,000 revisions per language

}}

{{Colored box
|title = Test data
<!-- [required] How was the test data split off from the data writ large? -->
|content = 
* Data period: 1 week
* Number of revisions: 1,079,265
* [[en:Wikipedia:IP users|IP users]] edits rate: 0.19
* Revert rate: 0.07

}}
}}
<!-- START [OPTIONAL] SECTION: If you marked PII as being used in this model, please describe the kind of PII that is used and how it is used.
=== PII ===
END [OPTIONAL] SECTION -->
<!-- START [OPTIONAL] SECTION: If you would like, you can use the ContentGrid/Colored box template below to depict an example data pipeline.
{{collapse top |title=Example data pipeline}}
{{ContentGrid|content=
{{Colored box
|title = 
|content = 
}}
{{Colored box
|title = 
|content = 
}}
}}
{{collapse bottom}}
END [OPTIONAL] SECTION -->

== Licenses ==
<!-- [required] What are the licenses for your model code and model outputs? These must be free and open-source.

Typically we use the following licenses:

* Code: [[:en:MIT_License|MIT License]]
* Model: [[:en:Creative_Commons_license|CC0 License]]
 -->

* Code: [[:en:Apache_License|Apache 2.0 License]]
* Model: [[:en:Apache_License|Apache 2.0 License]] 

<!-- [optional] Default LaTeX citation of this model. If the model has been featured in a publication, you can put alternate citation information here. -->
== Citation ==
Cite this model as: 
<syntaxhighlight lang='bibtex' >
@inproceedings{trokhymovych2023fair,
  title={Fair multilingual vandalism detection system for Wikipedia},
  author={Trokhymovych, Mykola and Aslam, Muniza and Chou, Ai-Jou and Baeza-Yates, Ricardo and Saez-Trumper, Diego},
  booktitle={Proceedings of the 29th ACM SIGKDD Conference on Knowledge Discovery and Data Mining},
  pages={4981--4990},
  year={2023}
}
</syntaxhighlight>

== References ==
<references>
<ref name="zia2019">Zia, Leila and Johnson, Isaac and Mansurov, Bahodir and Morgan, Jonathan and Redi, Miriam and Saez-Trumper, Diego and Taraborelli, Dario. 2019. Knowledge Integrity. https://doi.org/10.6084/m9.figshare.7704626</ref>
</references>


[[Category:Machine learning models]]