Machine learning models/Production/Language agnostic link-based article topic

Model card
This page is an on-wiki machine learning model card.
A model card is a document about a machine learning model that seeks to answer basic questions about the model.
Model Information Hub
Model creator(s)	Isaac Johnson, Martin Gerlach, and Diego Sáez-Trumper
Model owner(s)	WMF Research Team
Model interface	Lift Wing API
Past performance	Previous performance data
Publications	Language-agnostic Topic Classification for Wikipedia
Code	Gitlab repository
Uses PII	No
This model uses links in an article to predict a set of topics that a Wikipedia article may be in.

Predicting what general topic an article is in and doing so consistently across many languages is useful for various analyses of Wikipedia dynamics and filters for edit recommendation systems. However, it is difficult to group a very diverse range of Wikipedia articles into coherent, consistent topics manually across all Wikipedia projects.

This model is a new, language-agnostic approach to predicting which topic an article might be relevant to. It uses the wikilinks in a given Wikipedia article to predict which (0 to many) of a set of 64 topics are relevant to a given article. For example, Mount Everest might reasonably be associated with South Asia, East Asia, Sports, and Earth and the Environment.

The training data for this model was over 30 million Wikipedia articles spanning all languages on Wikipedia. Each article is represented as a collection of its wikilinks as represented by their respective Wikidata items. This data originates from the normal editing activities of Wikipedia and Wikidata editors, and was collected in an automated fashion.

This model is deployed on LiftWing (documentation). Right now, it can also be publicly accessed through a beta testing site. This model may be useful for high-level analyses of Wikipedia dynamics (pageviews, article quality, edit trends), filtering articles, and cross-language analytics. It should not be used for projects outside of Wikipedia, namespaces outside of 0, disambiguations, or redirects.

Motivation

A major challenge for many analyses of Wikipedia dynamics — e.g., imbalances in content quality, geographic differences in what content is popular, what types of articles attract more editor discussion — is grouping the very diverse range of Wikipedia articles into coherent, consistent topics. This problem has been addressed using various approaches based on Wikipedia’s category network, WikiProjects, and external taxonomies. However, these approaches have always been limited in their coverage: typically, only a small subset of articles can be classified, or the method cannot be applied across (the more than 300) languages on Wikipedia.

This language-agnostic approach for classifying articles into a taxonomy of topics, which is based on the links in an article, can be easily applied to (almost) any language and article on Wikipedia. It matches the performance of a language-dependent approach while being simpler and having much greater coverage.

Users and uses

Intended users

• researchers
• bots
• editors
• user scripts and gadgets

Use this model for

• high-level analyses of Wikipedia dynamics such as pageview, article quality, or edit trends — e.g. How are pageview dynamics different between the physics and biology categories?
• filtering to relevant articles — e.g. Filter articles only to those in the music category.
• cross-language comparisons — e.g. How does the quality of articles in the east-asia category differ between French Wikipedia and Japanese Wikipedia?

Don't use this model for

• projects outside of Wikipedia — e.g. Wiktionary, Wikinews, etc.
• namespaces outside of 0, disambiguations, and redirects — the training data for this model explicitly excludes draft pages, talk pages, disambiguations, redirects, and other non-article pages, as they do not have a training label that could be associated with them.

Current uses

This model is in production as of August 2022. Its outputs are incorporated into the Search index where they are used as filters by edit recommender systems such as Newcomer Tasks. For internal Wikimedia Foundation purposes, a snapshot of all model predictions is produced monthly (details).

Ethical considerations, caveats, and recommendations

• The model fits articles into a taxonomy of topics initially developed as a guide for discovering English WikiProjects. While some tweaks have been made to align it better with topic classification, it likely reflects English Wikipedia's interests and distinctions. Other language editions presumably would make different distinctions.
• While 0.5 is the suggested threshold, other thresholds may be more appropriate depending on the language and topic label. Notably, the raw scores from the model are not a measure of topical relevance but a measure of model confidence that a topic is relevant. Thus, a higher score does not mean a topic is more relevant and topics with clearer relevance — e.g., geographic, biographies — will generally have higher scores than topics with more ambiguous relevance — e.g., society, education.
• Gaps in WikiProject coverage are known to lead to biases in recall for certain topics. For example, film labels largely are missing actors from Nollywood (Nigeria) and thus recall is lower for articles about Nigerian films and actors than Hollywood (US) films and actors.

Model

Performance

As of the most recent model training in November 2022, the model had the following test statistics:

Overall Model Performance (November 2022, threshold=0.5 for all topics)

Precision Percentage of articles classified as a topic that were actually the topic	Recall Percentage of articles that actually were a topic that were classified as the topic	F1 Harmonic mean of precision and recall	Average Precision Mean precision across all thresholds
0.881 (micro); 0.841 (macro)	0.795 (micro); 0.676 (macro)	0.833 (micro); 0.744 (macro)	0.893 (micro); 0.796 (macro)

Detailed Model Performance (November 2022, threshold=0.5 for all topics)
Topic n Number of articles in the test set with the label (8% of all data) FPR Percentage of articles not in a topic classified as being in the topic Precision Percentage of articles classified as a topic that were actually the topic Recall Percentage of articles that actually were a topic that were classified as the topic F1 Harmonic mean of precision and recall Average Precision Mean precision across all thresholds europe 889127 5.6% 88.3% 87.3% 0.878 0.953 biography 772873 3.0% 92.5% 92.1% 0.923 0.971 stem 541929 1.2% 94.8% 89.5% 0.921 0.972 asia 418330 1.7% 89.9% 84.3% 0.870 0.933 sports 390101 0.5% 96.9% 94.6% 0.957 0.980 media 383948 1.5% 90.8% 88.0% 0.894 0.951 north-america 318785 1.9% 84.3% 76.5% 0.802 0.895 western-europe 317806 1.3% 89.8% 84.0% 0.868 0.944 biology 287075 0.3% 97.4% 93.9% 0.956 0.986 geographical 252242 1.6% 81.7% 70.1% 0.755 0.834 eastern-europe 209804 0.7% 91.5% 85.0% 0.881 0.944 southern-europe 196306 0.9% 87.7% 80.4% 0.839 0.917 northern-europe 183064 0.9% 85.4% 73.7% 0.791 0.872 history 157208 0.9% 79.7% 58.3% 0.673 0.751 music 148941 0.3% 94.2% 86.6% 0.902 0.942 films 139596 0.7% 87.5% 84.7% 0.861 0.925 women 138163 1.0% 69.6% 43.5% 0.535 0.622 military-and-warfare 112782 0.6% 80.9% 62.9% 0.707 0.783 east-asia 114672 0.4% 89.2% 81.1% 0.850 0.906 politics-and-government 110200 0.7% 76.7% 55.9% 0.647 0.718 west-asia 105185 0.4% 89.9% 81.0% 0.852 0.915 visual-arts 102622 0.6% 78.6% 59.9% 0.680 0.749 philosophy-and-religion 99928 0.6% 78.9% 59.2% 0.676 0.726 transportation 91054 0.2% 92.6% 84.3% 0.882 0.921 literature 85679 0.5% 78.3% 54.3% 0.641 0.718 south-asia 83060 0.2% 91.9% 84.0% 0.878 0.918 africa 84073 0.4% 84.9% 69.0% 0.761 0.829 north-asia 74433 0.3% 87.3% 75.6% 0.810 0.882 south-america 70327 0.3% 85.8% 75.6% 0.804 0.876 oceania 67774 0.2% 90.7% 75.8% 0.826 0.876 business-and-economics 59146 0.4% 73.1% 45.7% 0.562 0.606 technology 51349 0.4% 77.0% 61.0% 0.681 0.733 architecture 48185 0.3% 75.6% 52.6% 0.620 0.681 engineering 48126 0.2% 86.7% 66.1% 0.750 0.795 medicine-and-health 43245 0.2% 82.9% 67.0% 0.741 0.798 society 46268 0.3% 64.3% 27.9% 0.389 0.421 southeast-asia 43758 0.1% 89.3% 73.9% 0.809 0.859 television 44118 0.2% 85.1% 66.6% 0.747 0.802 earth-and-environment 45095 0.2% 84.3% 61.8% 0.713 0.759 space 39521 0.1% 95.4% 89.9% 0.926 0.961 linguistics 36010 0.1% 89.8% 64.8% 0.753 0.787 computing 28099 0.1% 83.1% 68.4% 0.750 0.824 internet-culture 27017 0.1% 89.9% 77.3% 0.831 0.881 entertainment 26086 0.2% 69.9% 40.8% 0.516 0.566 central-america 28158 0.1% 83.7% 59.2% 0.694 0.738 education 26072 0.1% 65.3% 25.5% 0.367 0.386 northern-africa 25855 0.1% 80.2% 57.3% 0.669 0.719 chemistry 24297 0.1% 86.2% 76.9% 0.812 0.880 food-and-drink 23315 0.1% 83.5% 61.5% 0.708 0.757 performing-arts 20426 0.1% 77.2% 43.6% 0.557 0.597 books 18727 0.1% 78.2% 51.8% 0.623 0.657 eastern-africa 19151 0.1% 85.1% 64.2% 0.732 0.789 mathematics 17261 0.1% 84.9% 70.7% 0.772 0.837 video-games 18180 0.0% 95.9% 87.9% 0.918 0.944 physics 18626 0.1% 78.9% 62.1% 0.695 0.756 comics-and-anime 17743 0.0% 91.8% 77.5% 0.840 0.878 western-africa 17738 0.1% 85.6% 66.6% 0.749 0.811 software 15889 0.1% 75.0% 67.0% 0.708 0.744 southern-africa 10577 0.0% 85.4% 61.2% 0.713 0.748 central-asia 10589 0.1% 78.2% 56.8% 0.658 0.702 central-africa 8296 0.0% 80.9% 54.7% 0.653 0.695 fashion 7320 0.0% 73.6% 42.2% 0.537 0.545 radio 5268 0.0% 85.7% 51.3% 0.642 0.632 libraries-and-information 4169 0.0% 72.8% 36.1% 0.482 0.440 Topic n FPR Precision Recall F1 Average Precision

Performance Notes

• Performance will suffer in articles with few outlinks, though generally this means precision remains high but recall drops. Atypical linking practices might also lead to inexplicable results, though this is difficult to define. In practice, the number of outlinks in a given Wikipedia article does vary by language edition, region of the world, and age of article.
• The groundtruth data for this model (WikiProject tags) are quite comprehensive but also certainly missing many legitimate tags – this means that false positive and precision rates are likely better than they look but also that false negative and recall rates are likely conservative and so not as good as they seem.
• Evaluation factors: number of outlinks, topic

Implementation

Model architecture

Overview at: https://fasttext.cc/docs/en/supervised-tutorial.html

• Epochs: 2
• Learning rate: 0.1
• Window size: 20
• Min count (under which QID is not retained in vocab): 20
• No pre-trained embeddings used
• Embeddings dimension: 50
• Total number of model params: 3200 (50 x 64)
• Vocab size: 4,535,915
• Total number of embeddings params: 226,795,750 (vocab size * embeddings dimension)
• Model size on disk: 944 MB
• Decision thresholds: 0.5 for all labels

Output schema

{
  article: <url string>,
  results: [
	{score: <score (0-1)>, topic: <string>},
	... (up to 64 topics)
	{score: <score (0-1)>, topic: <string>}
  ]
}

Example input and output

Input

$ curl https://api.wikimedia.org/service/lw/inference/v1/models/outlink-topic-model:predict -X POST -d '{"page_title": "Frida_Kahlo", "lang": "en", "threshold": 0.1}' -H "Content-type: application/json"

Output

{"prediction": {
  "article": "https://en.wikipedia.org/wiki/Frida_Kahlo",
  "results": [
    {"score": 0.863, "topic": "Culture.Biography.Biography*"},
    {"score": 0.516, "topic": "Geography.Regions.Americas.North_America"},
    {"score": 0.477, "topic": "Culture.Visual_arts.Visual_arts*"},
    {"score": 0.275, "topic": "Culture.Biography.Women"}
    ]
  }
}

Data

The training data for this model was over 30 million Wikipedia articles spanning all languages on Wikipedia. Each article was represented as the list of wikidata items associated with its outlinks. This data originated from the editing activities of Wikipedia and Wikidata editors, and was collected in an automated fashion.

Data pipeline

Wikilinks in a Wikipedia article (to other namespace 0 articles in that wiki) at the most current revision are selected from the pagelinks table and mapped to their corresponding Wikidata IDs at a set date. If there is no Wikidata ID or their Wikidata ID is not within the vocabulary, that data is dropped. The resulting bag-of-Wikidata-IDs is fed into the model, which maps each ID to a 50-dimensional embedding, averages them together, and then uses multinomial logistic regression to predict labels.

Training data

• 90% sample of every language in Wikipedia
• In practice, this meant that English Wikipedia provides 17.9% of the data and then French (4.4%), German (3.7%), Italian (3.4%), Spanish (3.3%), Egyptian Arabic (3.1%) and all other languages are below 3%
• Sampling is done by Wikidata ID, so all the language versions of a given article either appear in training, validation, or test but not across multiple splits.

Test data

• Same data pipeline and general approach as training data
• 8% sample of every language in Wikipedia
• 2% retained for validation

Example data pipeline
1) Initial article Magdalena Carmen Frida Kahlo y Calderón (6 July 1907 – 13 July 1954) was a Mexican painter known for her many portraits, self-portraits, and works inspired by the nature and artifacts of Mexico. Inspired by the country's popular culture, she employed a naïve folk art style to explore questions of identity, postcolonialism, gender, class, and race in Mexican society... 2) Isolate links Frida Kahlo: [ [[Self-portrait]], [[Mexico]], [[Culture of Mexico]], [[Naïve art]], [[Folk art]], [[Postcolonialism]], ... ] 3) Convert to list of Wikidata items Q5588: [ # Frida Kahlo Q192110, # self-portrait Q96, # Mexico Q2317008, # Culture of Mexico Q281108, # Naïve art Q1153484, # Folk art Q265425, # Postcolonialism ... ] 4) Train model This set of Wikidata IDs is then fed into the model, which maps each ID to an embedding, averages them together, and then uses a multinomial logistic regression to predict labels.

Licenses

• Code: MIT License
• Model: CC0 License

Citation

Cite this model as:

@article{johnson2021classification,
   author={Johnson, Isaac and Gerlach, Martin and Sáez-Trumper, Diego},
   title={Language-agnostic Topic Classification for Wikipedia},
   journal={WWW '21: Companion Proceedings of the Web Conference 2021},
   month=April,
   year=2021,
   pages={594–601}
}