Paper-Reading-ConvAI: A Curated Reading List for Dialogue Systems and NLG

What Paper-Reading-ConvAI actually is

The repository has two top-level entries: .vscode/ and README.md. There is no source directory, no package manifest, no test folder and no build configuration. That is the whole project. It is a paper list, and the README says so in its first line: a reading list in conversational AI, "mainly including dialogue systems (agents) and natural language generation."

The intended audience is narrow and identifiable. Someone starting a literature review on dialogue systems, or a graduate student who needs to know which survey to read before which specific paper, is the person this list serves. It is not for someone looking for a library to import, a benchmark harness to run, or a dataset to download. The README links to papers on arXiv, ACL Anthology, ACM Digital Library and NeurIPS proceedings; the reading happens elsewhere.

A practical consequence: the repository has no releases, and the license is not stated in the README. If you plan to copy the taxonomy into your own notes or an internal wiki, the absence of a declared license is a real question you have to resolve, not a formality.

The taxonomy: two pillars and how they are split

The README's table of contents is the actual product. It splits into Deep Learning in NLP, Dialogue Systems (Agents), and Natural Language Generation.

The dialogue pillar is the largest and is subdivided by research problem rather than by model family: Survey on Dialogue, Conversational LLMs and Benchmarks, LLM-based Dialogue Agents, Multi-turn Interaction, Long-term Dialogue, Multimodal Dialogue, Proactive Dialogue, Personalized Dialogue, Emotional Dialogue, Tutoring Dialogue, Recommendation Dialogue and CRS, Knowledge-grounded Dialogue, Task-oriented Dialogue, Open-domain Dialogue, Dialogue Evaluation, and Dialogue Misc. Several of these have second-level splits. Multimodal Dialogue breaks into Situated and Embodied Dialogue, Visually-grounded Dialogue, and Spoken Dialogue. Proactive Dialogue breaks into Misc. of Proactive Dialogue, Target-oriented Dialogue, and Non-collaborative Dialogue (Persuasion and Negotiation). Personalized Dialogue breaks into Character-based, Personality-aware and Persona-based.

The NLG pillar is organized differently, partly by theory and partly by method: Survey on NLG, NLG Theories and Techniques, Diffusion Models for NLG, Controllable Generation, Text Planning, Decoding Algorithms, and NLG Evaluation.

That inconsistency is worth noticing. Dialogue is indexed by task and application context; NLG is indexed by technique. If you are looking for work that crosses both, such as controllable generation for task-oriented dialogue, no single heading holds it, and you will need to check two branches.

How entries are formatted and what the stars mean

Each entry follows the same pattern: a bolded short name, a quoted paper title, a venue and year, then one or more bracketed links such as [[paper]] or [[code]], and sometimes a star rating. For example, the Transformer entry reads "Attention is All you Need". NeurIPS(2017), with links to the paper, an official TensorFlow implementation, and two third-party implementations, followed by five stars.

The star rating is the part to treat with caution. The README does not define it. Nothing in the README states whether stars reflect citation count, the maintainer's personal recommendation, or reading priority. The Prompting survey, the NLP World Scope paper, the attention survey, the pointer-generator paper and the word2vec tutorial all carry five stars, which suggests a difficulty or importance signal, but that is inference, not documentation. Use the stars as a rough ordering hint, not as a quality measure.

Link coverage is also uneven. Some entries offer code links, others only a paper link, and a few, such as the VAE entry, point to an arXiv URL. The list does not indicate whether any linked code has been maintained.

Reading the list: no install step, just a local copy

There is no installation procedure because there is nothing to install. The README gives no setup instructions, no package to add and no command to run. You get the content by cloning the repository or reading it on the web.

If you want a local copy to search and annotate, clone it:

git clone https://github.com/iwangjian/Paper-Reading-ConvAI.git

After that, the only file that matters is README.md. You can search it directly for a topic keyword, which is faster than scrolling the table of contents. Any text search tool works; the point is that the file is plain Markdown with no index beside it. There is no search interface, no tagging system and no metadata file, so any filtering you want beyond a text search is work you do yourself.

Where the list stops being useful

The first limitation is structural: a README is a flat document. The table of contents gives you anchors, but a paper that belongs to two categories is filed under one, and there is no cross-reference mechanism. If you are working on emotional support dialogue that also involves long-term memory, you have to check Emotional Dialogue, Long-term Dialogue and possibly Multi-turn Interaction separately, and reconcile the overlap by hand.

The second limitation is that a reading list ages in a specific way. Individual entries do not become wrong, but the balance of a field shifts and the list does not rebalance itself. The last push was on 2026-05-05, so anything published after that date is not present. The README does not state an update cadence or a contribution review process beyond an open invitation for contributions.

The third limitation is the one that matters most for adoption decisions: there is no code, no data and no tooling. If your task is to reproduce a dialogue model, evaluate a benchmark or build a retrieval index over papers, this repository gives you none of that. It is the wrong tool for anything that needs to execute.

How it compares with Papers with Code and Semantic Scholar

The obvious alternative is Papers with Code, which indexes papers alongside linked implementations, datasets and leaderboards, and lets you filter by task and by whether code exists. The difference in approach is that Papers with Code is generated and maintained at scale from multiple sources, while Paper-Reading-ConvAI is a hand-curated list with a fixed taxonomy chosen by its maintainers. The trade-off runs both ways: automated indexes cover far more papers but give you no editorial ordering, while a curated list gives you a suggested reading path at the cost of coverage and freshness.

Semantic Scholar and the ACL Anthology are the other comparison points. Both are search and citation systems rather than reading lists. They answer "find papers about X" well; they do not answer "what should I read first in this subfield." Paper-Reading-ConvAI is aimed at the second question. If you already know the subfield and just need to find a specific paper, the list adds little over a search engine.

Maintenance cost and licensing

The cost of using this repository is close to zero: clone it, read it, and re-pull when you want newer entries. The cost of depending on it is the cost of any unversioned document, which is that the structure can change between pulls and there is no changelog to tell you what moved.

Contributions are explicitly invited by the README, so if you find a missing paper the expected path is a pull request against README.md rather than an issue requesting a feature. That is a low-friction way to keep the list useful for your own subfield.

On licensing: the README does not state a license for the repository. The papers linked from the README belong to their respective publishers and venues, and their terms are separate from anything in this repository. If you intend to redistribute the list or embed its taxonomy in a product, confirm the repository's licensing terms first; that is a factual gap in the project, not a legal opinion.

Editorial conclusion

Adopt Paper-Reading-ConvAI if you need a starting map of dialogue-system subfields and want survey papers grouped by topic before you commit to a deep reading plan. Skip it if you need runnable code, dataset download links or a maintained package; the repository contains only .vscode/ and README.md at the top level, and the last push was on 2026-05-05. Before relying on it, open the README and check whether the subcategory you care about, such as multi-turn interaction or dialogue evaluation, actually lists the papers you already know, because that is the fastest way to judge whether the curation matches your field.

Frequently asked questions

Official sources

1. Issues
2. iwangjian/Paper-Reading-ConvAI on GitHub
3. README