lihanghang/NLP-Knowledge-Graph: A Curated Reading Map, Not a Library
自然语言处理、知识图谱、对话系统，大模型等技术研究与应用。
At a glance
- What is it?
- This MIT-licensed Python repository is a Chinese-language index of knowledge graph and KBQA material, not an installable toolkit. It is useful for orientation and reading order, and it will not answer a question or extract a triple for you.
- Who is it for?
- Adopt this repository as a reading list if you are entering Chinese-language KG or KBQA work and want a starting order for papers, mind maps and conference names. Do not adopt it if you need code you can import.
- Can I use it commercially?
- Yes. MIT is a permissive licence: you can use, modify and sell software built on it, as long as you keep its copyright and licence notices.
- Is it still maintained?
- Yes. The repository last received commits 13 days ago.
- What is it written in?
- Mainly Python, according to GitHub's language statistics.
Answers come from the project's GitHub data, last synced on September 15, 2026, and from our analysis. They are not legal advice.
DEEP OPEN-SOURCE ANALYSIS
What This Repository Actually Is
The README opens with a table of contents that reads like a syllabus: Deep learning for Knowledge-Graph, 资源外链, 知识图谱的前世今生, 研究总结列表, 比赛, 思考, Star History, 说明, Contact. Under 研究总结列表 sit subsections for 专题研究, KG&QA理论解析, NLP 论文解析, 中文金融文档结构化研究, 知识图谱相关会议, 事理图谱, 对话系统技术, commercialization notes, open source QA and dialogue systems, semantic platforms, text preprocessing tools, graph storage and query tools, visualization tools, and Chinese and English KG datasets. That is an index. The create time is given as 2019-08-24, and the repository description lists NLP, knowledge graph, dialogue systems and large model research and applications.
The stated direction is worth quoting because it frames the selection: 探索认知智能系列, with two trends named, data fusing with knowledge, and All in LLM. The topics list on the repository includes agent, agentic-ai, bert, ernie, event-driven, kbqa, knowledge-graph, llm, ner, reflection and transformers. In other words, the maintainer is tracking a moving area and updating the reading list as it moves. The last push timestamp is 2026-09-03, which is consistent with an actively maintained index rather than a frozen one.
What this is not: there is no package on PyPI mentioned, no setup.py or pyproject.toml described, no CLI, no config file, no import path. The primary language field says Python, and the commented-out section in the README hints at directory names such as Algorithm-code, datasets, SmartInteraction and 知识存储, but those lines are inside an HTML comment and the README does not document their contents. Treat the Python label as a description of the author's working language, not as a promise that cloning the repository gives you runnable modules.
The Problem It Solves: Reading Order in a Crowded Field
Knowledge graph work sits at an awkward intersection. You need entity and relation extraction, graph storage and query, question answering over the graph, and increasingly an LLM layer. Each of those has its own literature and its own tooling, and the entry points are scattered across conference proceedings, blog posts and vendor platforms. A person starting in Chinese-language KG work has no obvious first paper, second paper, third paper.
This repository answers that specific problem. It gives an ordered list. The KG&QA理论解析 section runs from a 知识图谱综述 mind map through 知识图谱的挑战, 深度学习与知识图谱, CN-DBpedia, KBQA, and a piece on AI self-awareness, ending with an ACL 2019 knowledge graph summary. The NLP 论文解析 section starts with The Illustrated Transformer, then an attention survey, then BERT, then two ERNIE papers (one on knowledge integration, one on informative entities), then T5. Read in that order, the sequence moves from architecture to pretraining to knowledge injection. That ordering is the product.
It is for the reader who already knows they need to build something with a knowledge graph and wants to know what the field considers foundational. It is also for the reader who wants to see which Chinese semantic platforms exist (腾讯文智, 科大讯飞开放语义平台, 玻森, 哈工大语言云) before committing to one. It is not for someone who wants a library to install.
How the Material Is Organized and Where It Points
The structure is flat and link-driven. Most entries are a numbered item with a title and a URL. A large share of the URLs point to naotu.baidu.com, Baidu's mind-mapping service, with a file token in the query string. Others point to mp.weixin.qq.com articles, to nlpprogress.com and paperswithcode.com as external trackers, and to a personal Read the Docs site.
That choice has consequences. A Baidu Naotu link is a shared mind map, not a document in the repository. If the share token is revoked or the service changes its sharing rules, the entry becomes a dead reference and the repository has no local copy. The same applies to the WeChat article links, which are commonly blocked outside mainland China and sometimes rate-limited even inside it. The repository does carry some local files: the commercialization section links to PDFs under 知识图谱基础/CCKS_2013-2018/, organized by year, including a 2017 Xiaomi talk on natural language processing in voice interaction, a 2016 精益知识图谱方法论 talk from 文因互联, a 2016 小i机器人 talk on its Chinese semantic open platform, and a 2014 iFlytek talk on the value and challenges of knowledge graphs from an application perspective. Those are in-tree and survive independently of any third-party host.
The conference table is plain data: ACL, CVPR, ICML, IJCAI, EMNLP, CIKM, AAAI, SIGKDD, TKDE and SIGIR, each with a grade (A类 or B类), a type (学术会议 or 学术期刊) and a field. It is a quick way to know where KG work gets published. It is not a ranking of the work itself.
Getting It Running: Clone, Read, and Nothing Else
There is no installation step documented. The only concrete action the README supports is fetching the repository and reading it. The commands that follow from the repository layout are the ordinary ones:
git clone https://github.com/lihanghang/NLP-Knowledge-Graph.git cd NLP-Knowledge-Graph
After that, the README offers no entry point. No pip install line, no python -m invocation, no environment file, no config keys. The commented-out block lists candidate directories (Algorithm-code, datasets, my summary, NLP&KG基础, SmartInteraction, 认知科学, 事理图谱, 文本相似度, 知识存储, 知识构建, 中文金融文档抽取) but it is commented out, so the README itself does not assert that these directories exist or what they contain. If you clone the repository, inspect the tree before assuming any of those names are present.
One link in the README points to a GitHub repository at lihanghang/Knowledge-Graph, described as a 技术集成项目 under the name 小智v1.0. That is a different repository from the one reviewed here, and it is inside an HTML comment as well. The README also states that contributors can contact the maintainer, which is the only contribution channel it names. There is no CONTRIBUTING file described, no issue template described, and no release has been retrieved for this repository, so there is no versioned artifact to pin to.
The Real Limitation: Links Decay and Scope Drifts
The failure mode here is link rot, and it is structural rather than incidental. A curated index whose entries live on a mind-mapping service and a social publishing platform inherits both platforms' retention policies. The repository cannot guarantee that a 2019 Naotu token still resolves. The README gives no archived snapshots, no DOIs, and no local copies for those entries, so a reader who finds a dead link has no fallback inside the repository.
Scope drift is the second issue. The description covers NLP, knowledge graphs, dialogue systems and large models. The topics list adds agent, agentic-ai, event-driven and reflection. Those are broad, and the README's own TOC shows the breadth: a section on 事理图谱 (event-centric graphs), a section on dialogue system technology, a section on open source QA and dialogue systems, a section on semantic platforms, and a section on commercialization. A reader who wants depth on KBQA specifically will find the KBQA entry is a single mind-map link, not a treatment. The repository points; it does not teach.
A third constraint is language. The README, the section names and the annotations are in Chinese. The linked papers are mostly in English, but the framing around them is not. A reader without Chinese will get the URLs and lose the curation, which is the actual value here. That is not a defect in the project, but it is a hard filter on who benefits.
Where It Sits Against Papers with Code and NLP-Progress
The README itself links to two alternatives, and they are genuinely different in approach. paperswithcode.com attaches code implementations and reported results to papers, so its unit is the paper plus its reproduction. NLP-Progress tracks datasets and current state of the art for common NLP tasks, so its unit is the task leaderboard. Both are maintained by teams, both are web-first, and both are oriented toward what currently performs best.
This repository's unit is the reading sequence. It does not report numbers, it does not link to implementations for most entries, and it does not maintain leaderboards. What it does that the other two do not is group material by a Chinese-language practitioner's mental model: 前世今生, 理论解析, 事理图谱, 对话系统技术, 商业化. It also carries the CCKS 2013-2018 PDF set locally, which neither external tracker does. If you want the best current score on a task, go to NLP-Progress. If you want to know which papers to read first and how Chinese industry talks about knowledge graphs, this repository is the better starting point. Choosing between them is not a quality judgement; they answer different questions.
Maintenance Cost and the MIT License
The maintenance burden for a user is near zero in the operational sense: nothing to upgrade, no dependency to patch, no runtime to monitor. The cost is in verification. Because the content is links, the useful maintenance work is checking that links still resolve, and that is work the reader ends up doing. The last push timestamp of 2026-09-03 indicates the maintainer is still touching the repository, but the README does not describe a link-checking process, and no releases have been retrieved, so there is no changelog telling you what changed between visits.
The license is MIT. That permits use, copying, modification and redistribution with the license and copyright notice preserved. It says nothing about the linked third-party content, which carries its own terms: the Baidu Naotu maps, the WeChat articles and the CCKS presentation PDFs are not covered by this repository's MIT grant unless the repository states otherwise, and the README does not state otherwise. The commercialization section is explicitly marked 仅供交流学习, which is a scope note from the maintainer rather than a license term. If you plan to reuse any of those PDFs in a product or a training corpus, check the original source's terms. This is a description of what the license text covers, not legal advice.
Editorial conclusion
Adopt this repository as a reading list if you are entering Chinese-language KG or KBQA work and want a starting order for papers, mind maps and conference names. Do not adopt it if you need code you can import. Before relying on it, open the linked Baidu Naotu and WeChat URLs yourself, because several entries are third-party pages whose availability this repository does not control, and the README carries no installation or API surface to verify.
Community notes