knowledge-graph-learning: a task-organised reading list for knowledge graph work

A list organised by task, not by year

The BrambleXu knowledge-graph-learning repository describes itself as a curated collection of knowledge graph learning resources, including papers, tutorials, datasets, implementations, tools, and communities. It carries the awesome list badge from awesome.re, which is the convention for a curated list of links, and its licence is MIT.

The organising idea is the task, and that is the choice that makes it useful. A reader who wants to build a knowledge graph from text is not looking for the newest paper, they are looking for the reading list that starts with information extraction and open IE and ends with construction, and that is how the Papers and Notes section is laid out.

Each entry is a line with a year, a task name, and a link, and the links point at GitHub issue queries filtered by label. The Information Extraction and Open IE line, the Named Entity Recognition line, and the Relation Extraction line all carry 2023. Named Entity Linking is marked 2020, as are Knowledge Representation Learning and Knowledge Graph Embeddings, and Knowledge Graph Population and Construction, which is annotated as being about constructing knowledge graphs from different sources. Knowledge Base Completion and Knowledge Graph Reasoning is 2019, annotated as entity prediction and link prediction, and Ontology-based Information Extraction and Semantic Role Labeling are 2019 as well. Knowledge-based Recommendation Systems, which links to a specific issue number rather than a label query, is also 2019.

So the year is not the publication year of one paper, and that is the point of the next section. The sections themselves run from Papers and Notes through Tutorials, Surveys, and Slides, Courses and Lectures, Datasets, Projects and Implementations, Libraries and Tools, and Communities, with a Contributing section at the end.

The paper notes live in GitHub issues

The mechanism behind the Papers and Notes section is unusual enough to be worth understanding, and it explains why the repository itself contains almost nothing.

Paper notes are organized with GitHub issue labels. Most existing notes are written in Chinese, but contributions in any language are welcome.

So the repository itself is nearly empty apart from the README and the usual project files, and the notes are in issues, retrieved through label queries. The query for information extraction and open IE is a URL with a label filter encoded in it, and the same pattern repeats for each task. The knowledge-based recommendation entry links to issue 250 directly, which suggests it is a single long note rather than a labelled group.

The design has two clear consequences. The first is that notes can be long and can be discussed, revised, and corrected in the issue tracker, which a README file does not allow. The second is that the list is a front end for a much larger body of content, and the repository is deliberately thin, holding only the index.

That thinness also has a cost for a reader who wants to work offline or to search the notes, because GitHub's issue search is the only index. There is no export, no local copy, and no full text search across the notes other than the label queries the README provides.

The bilingual note is worth stating as a fact rather than an opinion. The README says most of the existing notes are written in Chinese, that contributions in any language are welcome, and separately that resources are available in both English and Chinese, with the language of each resource indicated when it may not be clear from the title. The list of tutorials shows that convention in practice, with Chinese entries marked in parentheses.

The year means the latest meaningful update

Every list has a dating problem, and this one solves it with an explicit convention that is stated up front.

The year shown for each resource is its latest meaningful update: recent repository or collection activity for maintained projects, and the publication or revision year for static papers, datasets, tutorials, and slides. Entries in each section are ordered from newest to oldest.

The two halves of that rule matter because they behave differently. A maintained project keeps its year current indefinitely, so an entry can be much older than the year shown if the repository was touched recently, while a paper is frozen at the year it was published or revised and can never move.

The second half of the rule, ordering each section from newest to oldest, is what makes the list readable as a survey of the field. A section ordered by publication year in ascending order is a history, and a section ordered by latest meaningful update in descending order is a way of answering the question you actually have, which is what is current.

The list also states a maintenance fact plainly. All links were checked on August 26, 2026, and contributions that report broken links or improve the organisation of the list are welcome. The last push to the repository was on the same date, 2026-08-26, and there are no releases, which is the expected shape for a list rather than a package.

The practical consequence is that the list has a shelf life you should measure yourself, not take on trust. A link check on a given day tells you that the links resolved on that day, and nothing about whether the content behind them is still the state of the art.

Tutorials, surveys and slides, with GraphRAG at the top

The tutorials section is the one aimed at a newcomer, and it is where the newest entry appears.

The most recent item is 2026, Awesome GraphRAG, an actively maintained collection of GraphRAG surveys, papers, benchmarks, and open-source projects. It is the only 2026 entry visible in the tutorials, and its presence tells you something about where the field moved: retrieval over a knowledge graph became its own subfield with its own survey literature, and a list that started as a knowledge graph reading list now has to cover it.

The next entry is 2025, the Ontotext Blog, which is a company blog rather than a course, and the list is willing to include that kind of resource. Then 2021, Recent Trends in Entity Linking, a curated collection for one specific task.

The 2018 entries are the most substantial part of the section, and they are the conference tutorial type. There is a KDD tutorial on knowledge extraction and inference from text, split into a part one and a part two as separate PDF links. There is a WSDM tutorial on mining knowledge graphs from text, hosted on its own site. There are Neo4j slides on the power of graph-based search. And there are two Chinese entries, one a Zhihu collection of knowledge graph papers and one a knowledge graph introduction series, both marked as Chinese.

The older entries go back further. A 2017 IBM tutorial on enterprise knowledge graphs for large scale analytics, a 2017 metaphacts slide deck called Getting Started with Knowledge Graphs, and a 2016 Yahoo slide deck on knowledge integration in practice.

The pattern in that list is instructive on its own. The generic introductory resources stop around 2016 to 2018, which is roughly when the field's vocabulary settled, and the task-specific resources keep appearing afterwards. If you are starting from scratch, the 2016 to 2018 resources are a reasonable syllabus, but you should take the current resources from the task sections rather than from the general tutorials.

The related NLP tasks, and the model families

Two expandable sections sit under the knowledge graph tasks, and they are the reason the list is useful beyond a narrow reading list.

The first is a block of related NLP and machine learning tasks, with a Tasks subsection and a Models and methods subsection. The tasks are Recommendation from 2024, Question Answering and Machine Comprehension and Relational Reasoning from 2023, Annotation and Domain Adaptation and Domain-specific Learning from 2021, Coreference Resolution from 2020, and a 2019 cluster covering Data Augmentation, Dependency Parsing, Natural Language Understanding, Neural Machine Translation, Slot Filling, Summarization, and Text Classification.

The model families follow the same shape. BERT, Embeddings and Pre-trained Models, and Multi-task and Joint Learning are 2023, while End-to-end Models, Graph Neural Networks, and Transformer-based Models are 2019.

The reason this matters is that knowledge graph work is rarely a single task. Building a knowledge graph from text means entity extraction, relation extraction, entity linking, and then embeddings or a language model, and a practitioner who only reads the knowledge graph section ends up with a gap in the middle of the pipeline. The expanded sections make those adjacencies explicit, and they point at the same label-based issue queries, so the notes follow the same pattern as the main tasks.

The 2019 dates on the model families are also a reminder about how fast this part of the field moves. A 2019 entry for graph neural networks or transformer-based models is a pointer to a set of notes, and the notes themselves are where the currency lives, because the year attached to the entry only tells you when the note was last meaningfully updated.

The remaining sections of the list, for courses and lectures, datasets, projects and implementations, libraries and tools, and communities, follow the same labelling approach, with a courses section that has its own entries and a datasets section that is the one most likely to matter if you are planning to run an experiment.

What a curated list can and cannot give you

It is worth being clear about what this repository is, because the awesome list format invites the wrong expectations.

It is an index with a curation policy, not a library, and not a course. There is no code to install, no dataset to download, and no API. The repository itself contains only a README and the usual project files, and everything it points to lives elsewhere, most of it in issues in the same repository and the rest in third party sites.

The strengths follow from that. Because the curation is a human reading the field, the list compresses a lot of search into an ordered structure, and the task organisation means you can enter at the stage you care about rather than at the beginning. The bilingual coverage is a genuine advantage for a field with a large Chinese research community, and the issue-based notes mean that a correction is a pull request on a discussion rather than a pull request on a file, which makes them easier to keep right. The link checking policy and the explicit dating convention are both things that most reading lists do not do, and both make the list more trustworthy than a bare list of links.

The limitations are the mirror image. A link that resolved on a given date may lead to a moved page, a dead repository, or a paywall, and you will find out at the worst moment. The years attached to entries are the latest meaningful update, not the publication date, so you cannot use them to judge how old a technique is. The list has no stated policy for removal, so a resource that has been superseded may still be there, and the ordering by recency within a section can push an older but more complete survey down the page.

The right way to use it is as a map and a community rather than a reference. Read the task section for your problem, follow the entries into the notes, check the underlying paper or repository, and treat the list as a starting point rather than a decision.

Editorial conclusion

knowledge-graph-learning fits a student or engineer entering knowledge graph work who needs a map of the field by task, with a route from a survey to a tutorial to an implementation, and who is comfortable reading Chinese notes alongside English papers. It does not fit a team that needs a production library, a dataset with a documented schema, or a benchmark it can trust without checking. Verify first that the section for your task has been updated recently, since the list is maintained by hand and the years attached to each entry are the latest meaningful update rather than the publication date, that the dataset you pick has a maintained upstream, and that the code you plan to reuse is still compatible with the library versions current today.

Frequently asked questions

Official sources

1. BrambleXu/knowledge-graph-learning on GitHub
2. Issues
3. License: MIT
4. README