YTCommentVerse: Multilingual YouTube Comment Corpus
- YTCommentVerse is a large-scale, multilingual dataset of YouTube comments linking over 32M comments to video and channel metadata across 15 content categories.
- It employs custom web scrapers leveraging SocialBlade rankings to collect comments and metadata, facilitating analyses on discourse, engagement, and toxicity.
- The corpus serves as foundational infrastructure for research despite limitations like non-temporal design and minimal multimodal context.
YTCommentVerse is a large-scale multilingual, multi-category corpus of YouTube comments introduced as a dataset paper in 2025. It contains 32,236,173 comments from 178,027 videos written by 20,568,637 commenters, spans 15 YouTube content categories, and covers over 50 languages. Each record links comment text to platform-native identifiers and metadata, including videoID, commentID, commenter fields, originalChannelID, upvotes, and category labels. The resource is framed as infrastructure for research on discourse, toxicity, engagement, multilingual NLP, and platform-specific language modeling on YouTube, a platform the paper describes as comparatively underrepresented in public research datasets (Dutta et al., 14 Sep 2025).
1. Position within YouTube data infrastructure
YTCommentVerse is presented as a response to the mismatch between YouTube’s scale and the relative scarcity of large, publicly reusable YouTube comment corpora. The paper argues that YouTube has been harder to study than Twitter/X, Reddit, or Facebook because of data-collection and curation challenges, and it contrasts YTCommentVerse with earlier public datasets that were smaller, narrower in topic, or thinner in metadata (Dutta et al., 14 Sep 2025).
| Dataset | Reported scale | Reported fields |
|---|---|---|
| YouTube18K | ~18K comments | comment, sentiment |
| COVID-YT | ~24K | URL, title, channel, comment text |
| YTSpam | ~1.9K | commentID, commenterName, commentDate, comment, videoName |
| YouTube-180K | ~180K | comment |
| YTCommentVerse | 30M+ | comment, videoID, commentID, commenterName, commenterChannelID, upvotes, originalChannelID, category |
The corpus also sits within a broader lineage of YouTube-comment resources. An earlier large-scale dataset paper, YT-30M, reported the same core scale and a nearly identical comment-centric schema while releasing YT-100K as a 108,694-comment sample; this suggests a direct lineage between the two resources (Dutta, 2024). By contrast, "YouNiverse" released a very large commenter-video incidence table and extensive channel/video metadata, but explicitly stripped comment text from the released comment table (Ribeiro et al., 2020). Earlier domain-specific corpora such as "YouTube AV 50K" focused on narrower topical slices, in that case autonomous-vehicle videos, rather than a cross-category platform-wide comment layer (Li et al., 2018).
2. Collection frame, channel selection, and privacy model
The dataset construction begins from SocialBlade rankings. The authors use SocialBlade to obtain top YouTube channels by subscriber count within each category, then collect all videos posted by those channels and the associated comments and comment metadata using custom web scrapers (Dutta et al., 14 Sep 2025). The process is described in two stages: (1) create a list of videos posted by a YouTube channel and (2) collect video metadata and comments, along with comment metadata, for each video.
The 15 categories used by the paper are: Autos & Vehicles, Comedy, Education, Entertainment, Film, Gaming, Science & Technology, Howto & Style, Music, News & Politics, Nonprofit & Activism, People & Blogs, Pets & Animals, Sports, and Travel (Dutta et al., 14 Sep 2025). Category labels appear to derive from the SocialBlade/YouTube channel category under which channels were selected; the paper does not describe a separate relabeling or validation stage.
A notable feature of the collection description is what it does not specify. The paper does not provide a collection date range or explicit temporal span of the videos and comments, and it explicitly characterizes the resulting resource as non-temporal. It also does not describe a deduplication procedure, spam filtering, bot filtering, deleted-comment handling, minimum language-confidence thresholds, or a detailed cleaning and normalization pipeline (Dutta et al., 14 Sep 2025).
Privacy treatment is explicit. The paper states that all Personally Identifiable Information (PII) has been redacted/anonymized, that the data were collected only from publicly accessible YouTube videos and comment sections, and that individual consent was not sought because contacting millions of users would be impractical (Dutta et al., 14 Sep 2025). The sample records show opaque identifiers such as "b65..." and "2f1...", indicating anonymized commenter fields rather than raw public names.
The language layer is analytically central but methodologically underspecified. The paper reports language distributions using ISO 639-1 codes, including und for undetermined language, but does not explain the language-identification algorithm, confidence thresholds, treatment of mixed-language comments, or whether language is included as a released field (Dutta et al., 14 Sep 2025).
3. Corpus structure, schema, and linguistic composition
At release scale, YTCommentVerse comprises 32,236,173 comments, 178,027 videos, 20,568,637 commenters, 15 categories, and over 50 languages (Dutta et al., 14 Sep 2025). The paper defines each dataset entry as one comment.
| Field | Meaning |
|---|---|
videoID |
YouTube video ID |
commentID |
Comment ID |
commentorName |
Commenter name |
commentorChannelID |
Commenter’s channel ID |
comment |
Comment text |
upvotes |
Upvotes received by the comment |
originalChannelID |
Channel ID of the original video uploader |
category |
YouTube video category |
The schema text uses upvotes, while the sample JSON uses votes; the paper treats these as the same metadata dimension (Dutta et al., 14 Sep 2025). Just as important are the omissions. The paper does not list timestamp fields, video publication time, language as a released field, sentiment labels, toxicity labels, reply-thread structure, dislikes, number of replies, or country/region labels. The released resource is therefore comment-centric and metadata-rich, but not conversationally structured.
The paper’s language distribution table places English (en) first with 8.9M comments (22.0%), followed by Spanish (es) 4.4M (10.8%), Korean (ko) 3.8M (9.4%), Portuguese (pt) 3.7M (9.0%), Undetermined (und) 2.3M (5.6%), Thai (th) 2.3M (5.6%), Arabic (ar) 1.2M (3.1%), Indonesian (id) 1.2M (2.9%), Somali (so) 1.1M (2.8%), and German (de) 1.1M (2.7%) (Dutta et al., 14 Sep 2025). The language table itself lists approximately 57 entries, counting und, although the paper uses the more conservative summary “over 50 languages.”
The resource’s structure makes it simultaneously textual and relational. Each comment is anchored to a video, uploader channel, commenter identity, category, and engagement signal. This suggests a corpus suitable not only for text analysis but also for commenter-level, channel-level, and cross-category aggregation.
4. Reported descriptive analyses
YTCommentVerse is primarily a dataset-description paper, but it includes several descriptive analyses intended to demonstrate the corpus’s analytical range (Dutta et al., 14 Sep 2025).
First, the paper uses upvotes as a proxy for audience engagement and reports that most comments receive little to no engagement, often fewer than five upvotes. It also notes that upvotes are a partial engagement signal because the dataset does not include dislikes, replies, or richer interaction metadata. This makes engagement modeling possible, but incomplete.
Second, the paper analyzes comment length by category. It reports that Nonprofit has the highest average comment length at 111.7 characters, while Comedy has the highest variability in comment length with . The associated interpretation is that nonprofit discussions tend to be more detailed, whereas comedy supports a much broader range of reaction styles.
Third, the paper presents sentiment results at category level. It reports that Music is the most positively received category with mean sentiment = 0.142, and that Music also has the highest sentiment variability with . It further states that most comments have neutral sentiment and that both positive and negative comments can receive high upvotes, leading to the conclusion that sentiment alone is not a strong predictor of popularity (Dutta et al., 14 Sep 2025).
Fourth, the paper includes a semantic-structure analysis based on a uniform random sample of 100,000 comments, embedded and projected with UMAP, with DBSCAN used to isolate the main dense cluster and remove outliers. In that space, Music, Gaming, and Education form relatively tight clusters, while Comedy, Entertainment, and Tech show more overlap. The paper interprets this as evidence that some categories support more semantically focused audience discourse than others.
Fifth, the paper runs Perspective API analyses using six dimensions: toxicity, severe toxicity, obscene, threat, insult, and identity. The highest toxicity levels appear in Entertainment, Gaming, and Sports, while Education, Animals, and Travel exhibit lower toxicity. Identity-related toxicity is reported as generally low overall, but somewhat more present in Entertainment and Gaming (Dutta et al., 14 Sep 2025).
These descriptive modules are accompanied by a set of intended use cases. The paper explicitly discusses cross-cultural discourse analysis, content moderation, fine-tuning LLMs, and bias and fairness audits as prospective research directions for the corpus (Dutta et al., 14 Sep 2025).
5. Relation to established research agendas on YouTube comments
YTCommentVerse enters an already diverse literature on YouTube comments, but it does so with substantially broader scale and category coverage than most earlier resources. This broader scale is especially relevant for research that requires commenter linkage, channel linkage, and cross-video structure.
One established line of work models comments as networks rather than isolated text. "Commenter Behavior Characterization on YouTube Channels" constructed per-channel co-commenter networks for 20 channels, 7,782 videos, 294,199 commenters, and 596,982 comments in a misinformation-related military domain, using graph features, PCA, clustering, clique analysis, and cross-channel overlap to detect commenter mobs (Shajari et al., 2023). "Detecting Suspicious Commenter Mob Behaviors on YouTube Using Graph2Vec" studies the same basic problem with a graph-embedding perspective (Shajari et al., 2023). Because YTCommentVerse releases comment text together with commenter identifiers, videoID, originalChannelID, and category labels, this suggests direct compatibility with large-scale co-commenting, coordination, and audience-overlap analyses.
Another major line of work studies YouTube comments as political discourse. "Cross-Partisan Discussions on YouTube" analyzed 973 channels, 274,241 political videos, and 134M comments to show substantial but asymmetric cross-partisan participation, with comment ranking making cross-partisan comments modestly less visible (Wu et al., 2021). "Asymmetric Dynamics of Partisan Warriors in YouTube Comments" examined 1,854,320 comments around the 2024 U.S. second presidential debate and shifted the focus from generic incivility to target-specific hostile cross-cutting behavior and channel-level ecological effects (Lee et al., 23 Mar 2026). YTCommentVerse does not itself annotate ideology or attack targets, but its multilingual, category-linked scale makes such downstream political analyses plausible.
A third line concerns safety and moderation. "Hate, Obscenity, and Insults: Measuring the Exposure of Children to Inappropriate Comments in YouTube" collected approximately four million records around children’s videos and reported 11% toxic comments in that setting (Alshamrani et al., 2021). Here the relevance of YTCommentVerse is infrastructural rather than label-complete: it provides the raw comment layer, but not native moderation labels. A similar pattern appears in instructional-video QA: "YTCommentQA" extracted 2,332 questions from 2,004 videos from naturally occurring YouTube comments and showed that answerability depends on whether evidence is present in the video and on which modality carries that evidence (Yang et al., 2024). These adjacent studies indicate that YTCommentVerse can function as a base corpus for moderation, discourse, retrieval, QA, and platform-analysis tasks, even though it is not itself a benchmark suite for any one of them.
6. Access, limitations, and scholarly significance
The release model is tiered. The paper states that a smaller version, YT-100K, is available on Hugging Face, while the full YTCommentVerse dataset was available by directly emailing the authors and is also said to be uploaded to Zenodo; the release format is described as an SQLite database (Dutta et al., 14 Sep 2025). The paper also invokes FAIR principles, tying findability to Hugging Face, accessibility to dataset identifiers, interoperability to a Hugging Face datasets-style structure, and reusability to documentation and dataset cards.
The paper is explicit about several limitations. It identifies public visibility bias, because only publicly visible comments are included; non-temporal design, because the dataset is not suitable for many temporal modeling tasks; lack of multimodal context, because no video, audio, image, or thumbnail information is included; and a limited engagement signal, because only upvotes are released, not dislikes, replies, or richer interaction metadata (Dutta et al., 14 Sep 2025). It also leaves several important methodological issues unspecified: the exact language-identification method, the collection time frame, filtering and preprocessing details, and formal annotation or validation procedures.
These constraints place YTCommentVerse closer to a large-scale YouTube comment infrastructure dataset than to a heavily annotated benchmark corpus. Its central value lies in breadth: 32.2M comments, 178K videos, 20.6M users, 15 categories, and 50+ languages, combined with identifiers that permit comment-level, commenter-level, video-level, and channel-level aggregation (Dutta et al., 14 Sep 2025). Its central limitation is the absence of temporal, conversational, and multimodal fields. Within the YouTube data landscape, it is therefore best understood as a foundational corpus for large-scale empirical work on comment text and platform-linked metadata, rather than as a complete representation of the full interactional ecology of YouTube comments.