Paper-Reading-ConvAI: a star scale with no legend, and two hosts for the same ACL paper

Two entries at the root, no licence, and an invitation to contribute

The entire repository is a README and a `.vscode/` directory. There is no source directory, no test suite, no build file, no continuous integration configuration, and no license file, and the recorded primary language is unknown, which is what you would expect from a tree with no code in it. There is also no homepage and no GitHub release, so nothing outside the README describes the project. Against that, the opening line invites contributions. That invitation is the only governance statement in the repository, and it arrives with no license beside it and no contributing guide, so the terms for reusing the work are exactly as permissive as the absence of a file suggests, which is to say undefined. The `.vscode/` directory is the one other thing present, an editor configuration checked into a project with nothing to edit. The last recorded push is dated 2026-05-05 and the default branch is master, so for a list whose value is currency, roughly five months is the gap between the newest entry anyone has verified and now. The navigation is hand built as well: the header links to three internal anchors, one per top-level section, and each section that has been written so far closes with a back to top link whose fragment is the project title with its hyphens removed.

The opening sentence names two areas and the contents list has three

The scope is stated in one sentence: a paper reading list in Conversational AI, mainly including dialogue systems, or agents, and natural language generation. Two areas. The table of contents that follows has three top-level entries, because Deep Learning in NLP comes first, ahead of Dialogue Systems (Agents), and Natural Language Generation comes last. So the section a reader meets first is the one the introduction does not mention, and the introduction's own word, mainly, is doing quiet work: the Deep Learning section is a foundation layer rather than part of the conversational remit, but it is presented at the same heading level as the two areas that are. The contents list then descends unevenly. Dialogue Systems carries sixteen named subsections, and four of them, Multimodal Dialogue, Proactive Dialogue, Personalized Dialogue and Emotional Dialogue, each carry their own sub-subsections, eleven in total. Natural Language Generation is laid out flat, with seven subsections and no third level, so the two halves of the list are organised to different depths.

The star markers have no legend and reward 2016 over 2019

Some entries carry a run of star markers and most carry none, and nothing anywhere states what a star means. In the Deep Learning in NLP section, nine entries are marked and nine are not, and the counts split between four and five rather than spanning a range, which suggests a binary judgement recorded as emphasis. What the markers track is not age. Among the five-star entries sit Prompting from 2021, NLP World Scope from 2020, the Transformer from 2017, the Survey on Attention from 2018, the Copy Mechanism with pointer generation from 2017, the Word2Vec Tutorial from 2016 and Gradient Descent from 2016. Among the unmarked entries sit ELMo from 2018, Multi-task Learning from 2017, the VAE introduction from 2019, Transformer-XL from 2019, Additive Attention from 2015 and GloVe from 2014. So a 2016 explainer outranks a 2019 one, and a 2014 paper with a code link is unmarked. Four stars, applied to iNLP from 2023 and to the deep learning based dialogue survey from 2021, sits between those two groups. Whatever the scale was meant to capture, age is not it.

Two hosts serve ACL and EMNLP papers and arXiv gets three URL shapes

The links are inconsistent in ways that matter if you are automating anything against this file. ACL and EMNLP papers are served from two different hosts: Transformer-XL, Additive Attention's successor Multiplicative Attention, both Copy Mechanism entries, ELMo and GloVe all point at `aclweb.org`, while NLP World Scope points at `aclanthology.org`, which is the anthology that replaced the older host. Within the aclweb links the prefix is not even consistent, since the pointer-generator entry omits the `www.` that the other five carry. Three of those links also end in a direct `.pdf` while the rest name a bare anthology identifier. arXiv is worse, arriving in three shapes: the modern `arxiv.org/abs/` form for iNLP and the dialogue surveys, an `arxiv.org/pdf/` form for the VAE introduction, the attention survey, multi-task learning and gradient descent, and a legacy `arxiv.org/paper/` form for Additive Attention. The one entry that carries three separate implementations is the Transformer, labelled official, tf and py. The pattern is consistent enough to be worth checking rather than assuming, since each target is a public identifier that either resolves to the intended paper or does not.

The Deep Learning section starts newest-first and then stops

Read the years down the Deep Learning in NLP section and the intent is clear for the first six entries: 2023, 2021, 2021, 2020, 2019, 2017. Then the ordering breaks and never recovers. The VAE introduction is 2019 and sits after the 2017 Transformer. The Copy Mechanism with pointer generation is 2017 and sits after the 2015 Memory Net. The 2016 Copy Mechanism paper is followed by ELMo at 2018, Multi-task Learning at 2017 and Gradient Descent at 2016, in that order, so the last three entries run backwards. Compare the Survey on Dialogue section, which is exact: 2024, 2024, 2023, 2023, 2022, 2021, 2020, 2017, 2017, with no exception in nine entries. So the file contains at least two different curation habits. Within the dialogue surveys the ordering is reliable enough to navigate by date, and within the foundation section it is not, which means the second one has to be read rather than skimmed.

Personalized Dialogue holds three names for one idea

The taxonomy gets thin in three places, and each is a case where a reader has to guess the boundary. Personalized Dialogue splits into Character-based Dialogue, Personality-aware Dialogue and Persona-based Dialogue, three labels for what a working system probably treats as the same substrate with different emphases, and nothing in the file says where one ends and the next begins. Emotional Dialogue splits into Emotional Support Dialogue and Empathetic Dialogue, a distinction that is real in the literature and still needs a sentence to explain. Proactive Dialogue splits into Misc. of Proactive Dialogue, Target-oriented Dialogue and Non-collaborative Dialogue with persuasion and negotiation in the title, so one of the three slots in a category is explicitly a holding pen. And Recommendation Dialogue and CRS uses an abbreviation that the file never expands, in a list where every other category name is spelled out. The anchor links show the same strain, since a title with a full stop in it and a title with parentheses both had to be rewritten by hand to become usable fragments.

Eleven venue labels for one field, and findings papers linked to arXiv

Each entry carries a venue and a year in parentheses, and the labels do not belong to one taxonomy. Reading across the visible entries gives arXiv, ACL, ACL-Findings, EMNLP, NAACL, NeurIPS, ICLR, IJCAI, TOIS, SIGKDD Explorations and ACM Computing Surveys, which mixes four conference families, three journals and two magazine-style survey venues, plus the preprint server standing in for all of them. arXiv is doing quiet work in that list. The 2021 data augmentation survey is labelled ACL-Findings and is linked to arXiv rather than to an anthology, and the 2017 dialogue systems survey is labelled SIGKDD Explorations and is also linked to arXiv, while the 2020 open-domain challenges paper labelled TOIS is linked through the ACM digital library instead. So the label and the host disagree about how authoritative the source is. IJCAI appears once, on the 2023 proactive dialogue survey, and that entry is an arXiv link too, so a flagship conference proceedings paper is represented exactly like a preprint. For a list whose job is to be a reliable map, that mismatch is worth resolving per entry, and there are only a few hundred of them.

Only one entry distinguishes an official implementation from a port

Code availability is recorded as an optional annotation on an entry, and most entries omit it. In the visible portion, six entries carry a code link of some kind, and five of those use the same bare label: Transformer-XL, the pointer-generator copy mechanism, ELMo and GloVe each get one repository. The Transformer is the only entry treated differently, because it gets three links at once, one labelled official and pointing at the TensorFlow tensor2tensor repository, one labelled tf, and one labelled py pointing at a third-party PyTorch implementation. That is the single most useful piece of curation in the file, and it exists for exactly one paper. Everything else, including the attention survey, the VAE introduction, gradient descent and multi-task learning, is text with no pointer to an implementation, even where one is easy to find. The pattern suggests the code links were added opportunistically rather than as a policy, which is the same conclusion the star markers point to.

Editorial conclusion

Treat this as a bibliography rather than as software, and use it the way you would use any hand-maintained list: as a map of where a subfield sits, checked against the papers themselves before you cite anything. It has real virtues. The dialogue taxonomy is finer than most, splitting proactive, personalized and emotional dialogue into named sub-branches instead of treating them as one row, and the survey section is ordered cleanly from 2024 back to 2017, which makes it the fastest entry point in the file. Three things to know before you rely on it. The stars are undocumented, and they do not track recency or venue, so do not read them as a recommendation signal. The link targets are inconsistent, mixing two hosts for ACL and EMNLP papers and three URL shapes for arXiv, so verify a link resolves before you cite it. And the repository carries no licence and no contribution guide, so the invitation to contribute sits next to no stated terms; that gap matters if you intended to reuse the taxonomy itself rather than just the reading list. The last recorded push is dated 2026-05-05 on the master branch.

Frequently asked questions

Official sources

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