{"id":"59b96c43-40a5-40ae-b53d-ecc1e3d5fa71","arxiv_id":"2504.19536","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"TeleScope is a large multilingual Telegram dataset with 120M messages and unique cross-channel forwarding flows, designed to support social media research.","lead":"TeleScope is a new public dataset covering about 500,000 Telegram channels and 120 million messages from 71,000 public channels, with added forwarding flows and language tags. It gives researchers a large-scale view of Telegram, a platform that is increasingly important for studying misinformation and political communication.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The 'largest of its kind' claim is not established: cited comparators (Pushshift with 317M messages, TGDataset with 120K channels) may each outrank TeleScope on a natural size dimension, and the paper gives no direct comparison.","rationale":"The dataset is a real, functional resource with a DOI, public seed list, crawler code, and a substantial crawl; the paper deserves credit for that. The reader's CONDITIONAL verdict is reasonable. I part ways with the reader's single weakest-assumption framing: the seed-bias concern, while valid, is partially acknowledged in Section 6.1 and does not undermine the literal 'largest' claim. The more load-bearing issue is that the paper's own related-work numbers appear to undercut the abstract's superlative without any explicit comparison metric. This is directly testable. If the comparison shows TeleScope is not largest in any well-defined sense, the abstract and introduction need qualification; if it is largest on a specified dimension, the claim can be sharpened. In either case the central contribution of a large, longitudinal, forwarding-aware Telegram corpus survives, so no verdict change is needed beyond the existing conditional acceptance.","tokens_in":13555,"tokens_out":14680,"duration_ms":156202,"concrete_test":"Build a comparison table from the cited datasets and, if needed, from their public releases: for Pushshift and TGDataset, record number of channels with message metadata, total number of messages, temporal span and crawl continuity, number of languages, and current public availability (direct download or API). Then compare each figure against TeleScope's 71,048 message-bearing channels and 120,024,020 messages. If Pushshift has more messages or TGDataset has more message-bearing channels in comparable units, replace 'largest' with a metric-specific claim such as 'largest collection of channel metadata' or 'one of the largest message-level corpora'.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The abstract's central claim is that TeleScope is 'the largest publicly available, multilingual, longitudinal, and continuous collection from Telegram...'. Section 2 lists two comparators: Pushshift Telegram, with 27K channels and 317M messages, and TGDataset, with 120K channels. TeleScope's own Table 1 reports 71,048 public channels with downloaded message metadata and 120,024,020 messages, plus channel metadata for 534,137 channels. The paper never states the metric on which 'largest' is judged. If the metric is message count, Pushshift is 2.6x larger; if it is channels with message-level data, TGDataset (as cited) is larger than TeleScope's 71K; if it is discovered channel metadata, TeleScope's 534K may win, but that is not the same as downloaded message coverage. Since the superlative is the first contribution claimed in the abstract, this ambiguity is load-bearing: a reader cannot verify the headline claim from the paper's own cited evidence. The sampling bias identified by the reader is a separate, acknowledged limitation (Section 6.1); it does not by itself falsify the size claim, but it would matter for downstream generalization.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces TeleScope, a Telegram dataset suite containing metadata for 534,137 channels, downloaded message metadata for 71,048 public channels (120,024,020 messages), and derived artefacts including a channel-to-channel forwarding graph, message forwarding flows, and aggregated user interaction statistics. The collection procedure starts from 251 seed channels selected from TGStat's top-100 lists by subscribers, citations, and reach, and expands through snowballing via Telegram's message-forwarding metadata. The dataset is released through GESIS with a DOI, and the authors provide enrichment scripts and crawler information on GitHub. The paper claims that TeleScope is the largest publicly available, multilingual, longitudinal, and continuous Telegram collection to date, and it presents exploratory statistics on language distribution, channel properties, active periods, hashtags, and user engagement.","tokens_in":13767,"tokens_out":7127,"duration_ms":71437,"significance":"If validated, TeleScope would be a substantial reusable resource for social media research, particularly for studies of information propagation, multilingual discourse, and low-resource language communities. The paper's strengths include a publicly shared seed list for provenance, a clearly described collection pipeline, reproducible enrichment code, FAIR-aligned hosting with a DOI, and the release of derived propagation and interaction data that are genuinely useful additions over raw Telegram metadata. However, the two most prominent claims of the paper are not yet established: the 'largest of its kind' statement is not supported by the comparators cited in the paper itself, and the snowball-sampling coverage has only been checked internally, not against an external census or independent channel registry. These issues affect the framing and generalizability of the dataset more than its raw utility.","major_comments":[{"comment":"The headline claim that TeleScope is 'the largest publicly available, multilingual, longitudinal, and continuous collection from Telegram' is not supported by the paper's own cited comparators. Pushshift Telegram (Baumgartner et al. 2020) is reported as containing 317M messages, about 2.6 times TeleScope's 120M messages, and TGDataset (La Morgia et al. 2023) is reported as containing 120K channels, about 1.7 times TeleScope's 71K fully downloaded channels. The paper never states the metric on which 'largest' is judged. I ask the authors to define the comparison metric (e.g., 534K channel metadata entries, or the 261K-node forwarding graph), add a comparison table with existing Telegram datasets, and soften the claim if no dimension is actually the largest.","section":"Abstract and §1, compared with §2"},{"comment":"The convergence among the three seed criteria in Figure 1 is an internal property of the snowball process; it does not establish that the resulting channel set is representative of Telegram's public channel space. The language distribution in Table 3 (82.29% Russian) and the dominance of Russia/Ukraine-related hashtags in Figure 8 indicate strong selection effects from the TGStat seed. If the paper retains the 'comprehensive data coverage' framing in §1 and the general-purpose applicability claims in §7, please add an external validation of coverage—for example, a comparison of the discovered channels against an independent channel registry, or against the channel sets of Pushshift and TGDataset, and an estimate of what fraction of active public channels the 71,048 downloaded channels represent.","section":"§3.1, Figure 1, and §6.1"},{"comment":"The propagation statistics are not well-defined. 'Total number of messages' in Table 2 is 31,227,109, which does not match the 19.6% forwarded share of the 120M messages (about 23.5M forwarded messages), and 'Number of unique messages' (308,147) is not explained. Please define the unit of a forwarding flow, what counts as a message in the flow table, and how unique messages are identified, and report the number of forwarding events and the distribution of flow lengths so that the derived interaction data can be evaluated.","section":"§4.3, Table 2"}],"minor_comments":[{"comment":"There are a few typos: 'dataset suit' in §8 should be 'dataset suite', and 'isFindable' in §5 should be 'is Findable'.","section":"§8 and §5"},{"comment":"The phrase 'metadata for all the channels fully downloaded in our dataset (534,137)' is confusing because only 71,048 channels have downloaded messages; please rephrase to distinguish channel-metadata-only entries from full message downloads.","section":"§4.1"},{"comment":"The active-period analysis aggregates all channels in UTC without per-channel local-time normalization; because the sample is 82% Russian, the observed morning peak may reflect a single timezone. Either normalize to channel-local time or explicitly state this as a limitation.","section":"§6.3 and Figure 7"},{"comment":"The authors state that no Datasheet for the Dataset was created and that one will be provided if the paper is accepted; for a dataset contribution, a datasheet is standard practice and should be added to the final release to document provenance, biases, and intended uses.","section":"Paper Checklist, item 5(g)"}],"recommendation":"major_revision","confidential_remarks":"The paper is a useful and potentially valuable dataset contribution, but the abstract's superlative claim should be checked carefully by the editor: the authors either need to provide a justified comparison or remove 'largest'. The lack of external coverage validation is the second main concern; I would not require a full census of Telegram, but an independent overlap check against Pushshift or TGDataset would substantially strengthen the revision."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"TeleScope is worth knowing about. It ships a real, citable artifact: ~120M message metadata records from 71K public channels, a registry of 534K channels, language and entity enrichments, plus two things that are genuinely new—a reconstruction of cross-channel message forwarding flows and a channel-to-channel forwarding graph with aggregated interaction stats. The data has a DOI, the crawler code is on GitHub, and the authors are transparent about the pipeline. That is reproducible work, and it deserves credit.\n\nWhat is not new is the size claim in the abstract. The paper says TeleScope is 'the largest publicly available, multilingual, longitudinal, and continuous collection from Telegram' but never defines the metric. Pushshift (cited in the paper) has 317M messages. TGDataset (also cited) covers 120K channels. TeleScope has 120M messages and 71K fully downloaded channels. So on message count it is smaller than Pushshift, and on channel coverage it is smaller than TGDataset. The 534K channel metadata registry may be the largest, but that is not the same as downloaded messages. This is not a nitpick—the superlative is the first contribution listed, and a reader cannot verify it from the paper's own cited evidence. The authors need to either state the precise metric and compare directly or drop the claim.\n\nThe second soft spot is sampling coverage. The snowball starts from the TGStat top 100 in three criteria. The convergence argument in Figure 1 shows that the three seeds end up overlapping, which is interesting but does not show representativeness relative to Telegram as a whole. There is no external census check. The language distribution confirms the concern: 82% Russian. That is fine as a description of the dataset, but the paper's 'multilingual' framing needs real caveats. To the authors' credit, they do acknowledge the private-channel gap and the missing-source-channel limitation in the forwarding flows—those are honest and correct.\n\nThe core artifact—forwarding flows, the graph, the interactions—holds up. The methodology is straightforward and the limitations are mostly stated. I would send this to peer review, but I would require the authors to fix the size claim and add a coverage-limitation paragraph before acceptance. This is a resource paper; its value to the community is real, but its headline claims are currently overbuilt.","headline":"Useful and genuinely novel Telegram dataset with forwarding flows, but the 'largest' claim is not supported by the paper's own numbers.","tokens_in":14263,"tokens_out":2687,"would_cite":false,"duration_ms":24360,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"This paper introduces TeleScope, a public dataset of about 120 million Telegram messages from 71,000 public channels, with reconstructed forwarding paths that let researchers trace how content spreads.","keywords":["Telegram","dataset","message forwarding","channel-to-channel graph","information propagation","longitudinal data","multilingual corpus","social media analysis"],"falsifier":"A reader could draw an independent random sample of Telegram public channels—say, through Telegram's own search or an unrelated directory—and check whether TeleScope's language distribution and forwarding graph match that sample; a large mismatch, such as most non-Russian channels being absent, would falsify the representativeness implied by the snowball convergence.","tokens_in":13379,"feed_emoji":"📡","tokens_out":6980,"duration_ms":67553,"temperature":0.7,"pith_summary":"TeleScope is a public dataset suite that the authors say is the largest of its kind for Telegram: metadata for about 534,000 channels, complete message metadata for 71,048 public channels, and roughly 120 million messages crawled between February and October 2024. The dataset's distinctive feature is that it does not stop at raw messages: it reconstructs forwarding flows, builds a channel-to-channel graph, and aggregates views, forwards, and reactions per message across the crawled network. The authors' aim is to give social scientists the kind of fine-grained material for Telegram that Twitter once provided for studying information spread, communities, virality, and misinformation. If the resource works as advertised, it removes a major access barrier to one of the world's largest and least transparent messaging platforms.","feed_headline":"Largest open Telegram dataset maps 120M message flows","feed_subtitle":"Public metadata and forwarding graphs let researchers trace how content spreads across channels.","key_machinery":"The mechanism that carries the argument is Telegram's own message-forwarding feature, used twice: once as a discovery engine and once as a tracing tool. The crawler starts from 251 seed channels—selected from a public registry's top-100 lists ranked by subscribers, citations, and reach—and whenever it sees a forwarded message, it adds the source channel and recurses. This snowball sampling both expands the channel list to over 1.2 million discovered channels and produces the backward traces needed to construct message forwarding flows and the channel-to-channel graph. The graph is the load-bearing derived object: it turns a flat collection of messages into a network along which information spread can be measured.","core_discovery":"The central claim is that a large, longitudinal, multilingual Telegram corpus can be assembled from public data and enriched so that message propagation is visible instead of opaque. The authors present TeleScope as the largest publicly available collection of its kind: 534,137 discovered channels, 71,048 fully downloaded public channels, and 120,024,020 messages, spanning February 1 to October 29, 2024. The novel part is the derived layer—message forwarding flows that trace each forwarded message back to its source channel, a directed channel-to-channel graph with 261,171 nodes and 2,733,720 edges, and aggregated user interaction counts computed across all crawled copies of a message. On top of this, TeleScope adds language labels, per-channel hourly activity, and extracted Telegram entities such as hashtags, mentions, URLs, and formatting spans. The authors argue that these pieces together make Telegram research possible at the level of detail previously limited to Twitter-style platforms.","pith_inferences":["The convergence of the three seed criteria shown in the paper is evidence that snowballing saturates around the seed region, not that it covers all Telegram; a random-sample validation would strengthen any cross-platform generalization.","Because the language distribution is 82% Russian, cross-lingual studies built on this dataset will need reweighting or stratified subsampling to avoid conflating 'Telegram' with 'Russian-speaking Telegram'.","The message forwarding flows could be used to estimate cascade lengths and branching factors, but any such estimate is a lower bound, since forwards that leave the 71K crawled channels are invisible.","A natural next step, not demonstrated in the paper, is to join the channel-to-channel graph with extracted URLs and hashtags to study topic-level diffusion across communities."],"forward_implications":["Researchers can study message propagation at scale by using the forwarding flows instead of scraping each channel separately.","Retweet-style analyses—virality prediction, information diffusion, and community detection—can be replicated on Telegram using the channel graph and aggregated interaction scores.","The 47 detected languages, including low-resource ones, open a path for multilingual NLP and cross-community comparisons not feasible with earlier Telegram datasets.","Public channel metadata for 534K channels, including creation dates and flag statuses, supports longitudinal studies of platform growth and content moderation.","Planned annual snapshots would let later work track how channels, forwarding networks, and communities evolve over time."],"supporting_citations":[{"why":"Provides the Pushshift Telegram dataset with 27K channels and 317M messages, the baseline against which TeleScope positions its enriched coverage.","marker":"Baumgartner et al. 2020"},{"why":"Introduces TGDataset with 120K channels, the earlier large-scale Telegram corpus that TeleScope compares against.","marker":"La Morgia, Mei, and Mongardini 2023"},{"why":"Uses Telegram forwarding to characterize information propagation in fringe communities, motivating the snowball expansion and flow construction.","marker":"Hoseini et al. 2024"},{"why":"Demonstrates prior use of the same public channel registry for Telegram research, supporting the choice of registry-based seeding.","marker":"Alvisi, Tardelli, and Tesconi 2024"},{"why":"Provides earlier Telegram community detection work built on the same registry, reinforcing the seed-selection approach.","marker":"Tikhomirova and Makarov 2021"},{"why":"Presents BelElect, a specialized Telegram dataset for bias research, representing the targeted datasets that TeleScope aims to generalize beyond.","marker":"Höhn, Mauw, and Asher 2022"},{"why":"Offers a Telegram corpus for hate speech and online harm, another specialized dataset whose scope TeleScope broadens.","marker":"Solopova, Scheffler, and Popa-Wyatt 2021"}],"fun_headline_variants":["TeleScope: 120M Telegram messages with forwarding maps for spread tracing","Largest Telegram dataset reveals 120M messages and 2.7M forwarding links","Trace Telegram discourse: 500K channels, 120M messages, forwarding graphs","TeleScope: Open corpus of 120M Telegram messages with propagation paths","Follow Telegram info flow: 120M messages and 2.7M forwarding edges"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The entire dataset starts from 251 channels picked from the top-100 lists of a public registry, and the paper assumes that snowballing from these seeds eventually gives a representative view of Telegram's public channels without comparing against a full census of Telegram.","fun_headline_variants_meta":{"raw":{"variants":["TeleScope: 120M Telegram messages with forwarding maps for spread tracing","Largest Telegram dataset reveals 120M messages and 2.7M forwarding links","Trace Telegram discourse: 500K channels, 120M messages, forwarding graphs","TeleScope: Open corpus of 120M Telegram messages with propagation paths","Follow Telegram info flow: 120M messages and 2.7M forwarding edges"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000886,"raw_usage":{"total_tokens":3815,"prompt_tokens":922,"completion_tokens":2893,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":538,"completion_tokens_details":{"reasoning_tokens":2789}},"tokens_in":538,"tokens_out":2893,"duration_ms":18294,"temperature":1.0,"reasoning_tokens":2789,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-16T05:49:47.059522+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"A reader could draw an independent random sample of Telegram public channels—say, through Telegram's own search or an unrelated directory—and check whether TeleScope's language distribution and forwarding graph match that sample; a large mismatch, such as most non-Russian channels being absent, would falsify the representativeness implied by the snowball convergence.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the Pushshift Telegram dataset with 27K channels and 317M messages, the baseline against which TeleScope positions its enriched coverage."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Uses Telegram forwarding to characterize information propagation in fringe communities, motivating the snowball expansion and flow construction."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides earlier Telegram community detection work built on the same registry, reinforcing the seed-selection approach."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Offers a Telegram corpus for hate speech and online harm, another specialized dataset whose scope TeleScope broadens."}],"review_version":1}