{"id":"43c12b5b-fd06-4751-92e9-dd675134cffb","arxiv_id":"2505.08359","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"A pipeline that maps German Twitter news sharing in a two-dimensional political space and at story, outlet, and topic levels reveals within-outlet and within-topic partitioning that one-dimensional outlet-level analyses miss.","lead":"This paper introduces a research pipeline that maps news sharing on Twitter in a two-dimensional political space, combining follower-network embeddings, full-text topic modeling, and story-level analysis. Its proof of concept on German Twitter data shows that outlet-level and one-dimensional analyses miss meaningful heterogeneity in how different political groups share news.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Confirmatory hypotheses rest on selected visual exemplars; no systematic statistic across outlets or topics establishes the claimed 'systematic insights'.","rationale":"I keep the reader's CONDITIONAL verdict, but for a different primary reason. The reader's weakest assumption is homophily in the follower network; that assumption is explicit, shared with prior embedding methods, and has independent empirical support (Barberá 2015; Conover et al. 2012). The more fragile link in this paper's argument is internal: the hypotheses are stated as general claims but evaluated by visual selection. The paper itself acknowledges it is a proof of concept, and a proof of concept does not require a full survey; but then the conclusions should be phrased as existence proofs or illustrations, not as confirmed hypotheses. The CHES rotation circularity noted by the reader is a real but secondary concern; the statistical gap would remain even if the axes were perfectly valid. This does not invalidate the pipeline; it conditions its acceptance on either systematic quantitative evaluation or softened claims. Hence the verdict stays CONDITIONAL.","tokens_in":841,"tokens_out":889,"duration_ms":41816,"concrete_test":"Compute for every outlet the standard deviation of per-story mean user positions and the overlap between the story-mean distribution and the outlet-level distribution, with bootstrap confidence intervals. Then locate focus.de and nachdenkseiten.de in the resulting distribution over all 26 outlets and test whether a random selection of two outlets would show the same qualitative contrast as Fig. 3 (e.g., via a permutation test on the difference in story-mean dispersion). If the contrast is not exceptional, or if most outlets show story-mean dispersions nearly as wide as their outlet-level dispersion, H2a's claim that the story level reveals systematic insights is not supported.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central claim is that the pipeline reveals systematic patterns, and it asserts H1, H2a, H2b, and H3 are confirmed (Secs. 4.1, 4.2.1, 4.2.2, 4.3). Each confirmation is illustrated with hand-picked examples: three outlets in Fig. 2, focus.de and nachdenkseiten.de in Fig. 3, two outlets in Fig. 4, and four topics from one metatopic in Fig. 5. No summary statistic, null model, or uncertainty estimate is given for any hypothesis, and no criterion is defined for what would count as 'systematic' versus anecdotal. Because 26 outlets, 220 topics, and 12 metatopics are searched, the selected cases may simply be extreme draws; the confirmatory language overstates what the figures establish. This matters directly for the methodological contribution: a pipeline that can produce illustrative examples is not shown to provide a generalizable mapping unless the claimed heterogeneity is quantified across the full set. The homophily assumption in Sec. 3.1 is shared with established embedding methods and is not the main barrier here; even granting the coordinates, the headline inferential step is missing.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces a methodological pipeline that combines correspondence analysis (CA) of German MPs' Twitter follower networks with a CHES-based rotation to construct a two-dimensional political space, and then maps news sharing of 26 outlets using tweets from March 2023, article full texts, and a structural topic model with 220 topics grouped into 12 metatopics. The authors claim that this pipeline reveals systematic insights unobservable in one-dimensional, outlet-only designs, and they report confirmations of four hypotheses (H1, H2a, H2b, H3) on the basis of selected visual examples: three outlets in Fig. 2, two articles and two outlets in Fig. 3, two outlets across two metatopics in Fig. 4, and four topics within one metatopic in Fig. 5. The paper frames itself as a proof-of-concept methodological contribution, with an accompanying repository for the CA and rotation code.","tokens_in":21844,"tokens_out":4566,"duration_ms":45310,"significance":"If fully substantiated, the pipeline is a valuable advance: it offers a multi-dimensional, multi-level approach to studying news circulation, extends network-embedding ideology estimation beyond a single left-right axis, and foregrounds story- and topic-level heterogeneity in a way that most prior work has missed. The manuscript is transparent about the embedding and rotation steps, provides code, and includes detailed appendices on data collection, topic model validation, and metatopic coding. The main weakness is that the confirmatory evidence for all four hypotheses is anecdotal: selected cases are used where systematic, population-level statistics or null-model comparisons across all outlets, stories, and topics are needed. As a proof of concept the paper has merit, but its central claim of 'systematic insights' currently rests on hand-picked examples rather than on a quantitative demonstration that the pipeline generalizes.","major_comments":[{"comment":"The confirmation of H1 is based on the visual comparison of only three outlets (sueddeutsche.de, faz.net, welt.de), with no summary statistic over the full set of 26 outlets. To establish 'systematic insights' rather than anecdotal differences, the authors should quantify the divergence between outlet sharing distributions in the one- and two-dimensional spaces across all outlets (e.g., compute an overlap or distance metric in each space and compare the distributions of those metrics, with uncertainty estimates). Without such an analysis, the claim that the two-dimensional space reveals systematic patterns is not supported by the evidence presented.","section":"Section 4.1, Fig. 2"},{"comment":"H2a, H2b, and H3 are each confirmed on the basis of a small number of hand-picked examples: two articles and two outlets for H2a, two outlets and two metatopics for H2b, and four topics within a single metatopic for H3. Because the data comprise 26 outlets, 220 topics, and 12 metatopics, the selected cases could be extreme draws, and no criterion is given for what would count as 'systematic'. The authors should provide population-level evidence, such as (i) the distribution of story-level mean sharing positions for all outlets along with a heterogeneity measure (e.g., the share of stories whose mean position falls outside a central interval), and (ii) a test, for all topics or metatopics, of whether topic-level sharing distributions differ significantly from the outlet-level distribution (e.g., via permutation tests or a hierarchical model). Until such quantification is supplied, the confirmatory language in Sections 4.2 and 4.3 overstates what the figures establish.","section":"Sections 4.2.1, 4.2.2, and 4.3 (Figs. 3–5)"},{"comment":"Two data-dependent choices shape all subsequent results: the exclusion of CA dimension 2 (justified by a Spearman correlation of 0.79 with MP follower counts) and the CHES rotation procedure (a correlation cutoff of 0.8, with a second axis built by averaging over four rotation angles within a 22-degree window). These choices are plausible but are presented without sensitivity analyses. To make the pipeline a robust methodological contribution, the authors should demonstrate that the main findings persist (i) when dimension 2 is included or a popularity-adjusted embedding is used, (ii) when the CHES cutoff is varied around 0.8, and (iii) when the second axis is derived from a single best-fitting CHES issue rather than the average of four. In addition, because the rotation is optimized to maximize correlation with the CHES issues, the reported correlations of about 0.9 should be framed as the outcome of an optimization, not as independent validation; an out-of-sample or held-out validation would strengthen the interpretation of the EESP axis.","section":"Section 3.1 and Appendix C"}],"minor_comments":[{"comment":"There are several typographical errors in this section, including 'frequeny' (should be 'frequency'), 'contries' (should be 'countries'), 'expample' (should be 'example'), and 'spce' (should be 'space').","section":"Section 2.1"},{"comment":"The caption contains a typo: 'othogonal' should be 'orthogonal'.","section":"Fig. 1 caption"},{"comment":"The overview states 'we use two Principal Components to construct a two-dimensional space', but the procedure later explains that the authors operate with PCs 1 and 3, excluding PC 2. The overview should be clarified to avoid implying that PCs 1 and 2 are used.","section":"Section 3.1"},{"comment":"The text says 'nearly 1.2 million users' are embedded, while the sample description says '1.2 million followers being included in our pipeline.' Clarify whether this refers to users or followers, and whether the 60% coverage of shares refers to unique users or shares.","section":"Section 3.1 and Section 3.2"},{"comment":"The sentence 'right-leaning focus.de was generally shared often by right to right-wing users' is awkwardly phrased; consider revising for clarity.","section":"Section 5, Discussion"}],"recommendation":"major_revision","confidential_remarks":"The manuscript is a methodologically interesting proof-of-concept, but the confirmatory claims for H1–H3 currently rest on selected examples. I recommend a major revision that adds systematic quantitative evidence across all outlets, stories, and topics, along with sensitivity analyses for the CA dimension selection and the CHES rotation parameters. The authors should be encouraged to deposit additional data (e.g., story-level sharing positions and topic assignments) where platform terms permit, to support reproducibility of the pipeline."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Here's my take on arXiv:2505.08359. The paper is a methods contribution, and the pipeline itself is worth something: combining two-dimensional ideology estimation from follower networks with story-level link data and topic models is a genuinely useful integration, and the German proof of concept shows that outlet-level or one-dimensional designs can miss real structure. The clearest example is the focus.de/nachdenkseiten.de contrast at story level versus outlet level. Credit where due: the methods are described carefully, the code for the embedding and rotation is provided, and the authors are upfront about the homophily assumption and data coverage issues.\n\nThe soft spot is exactly where the stress-test lands. The paper frames H1, H2a, H2b, and H3 as confirmed, but the evidence is a series of hand-picked examples: three outlets in Fig. 2, two outlets in Fig. 3, two in Fig. 4, and four topics from one metatopic in Fig. 5. There is no summary statistic across all 26 outlets or 220 topics, no null model, and no uncertainty quantification. With search over that many units, selected cases can be extreme draws. So the claim that the pipeline reveals 'systematic insights' is not yet supported. This is fixable: the authors need to quantify heterogeneity across the full outlet and topic set, or at least provide baselines showing how often the patterns they showcase actually occur.\n\nOne more substantive issue: the axis validation is partly circular. The rotation is optimized to maximize correlation with selected CHES issues, and then the high CHES correlation is presented as confirmation. That's not a valid independent validation, though it's not fatal because the news-sharing data are external to the embedding. The second-axis interpretation is also a judgment call; excluding CA dimension 2 because it correlates with follower counts is reasonable, but it should be tested with robustness checks.\n\nThe homophily assumption is standard for this embedding literature and the authors cite the relevant validation work. I don't treat it as the main barrier here.\n\nWho benefits: computational social scientists and political communication researchers who want a ready-made multi-level news-sharing mapping. It deserves serious peer review—as a methods paper, it's a step forward—but the confirmatory language needs to be toned down and the systematic evidence added. I'd engage with it if I were working in this area, and I'd cite it for the pipeline, not for the empirical claims as they stand.","headline":"A useful integrated pipeline, but the proof-of-concept evidence is more illustrative than systematic; the confirmatory hypotheses outrun the selected examples.","tokens_in":22362,"tokens_out":3989,"would_cite":true,"duration_ms":35265,"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":"Two political dimensions reveal news-sharing splits hidden from a left-right-only view.","keywords":["news sharing","network embeddings","topic models","digital trace data","social media","Twitter","two-dimensional political space","political cartography"],"falsifier":"A validation study that matches a sample of embedded users to survey responses on left-right and EU/protectionism attitudes and checks whether their embedding coordinates predict reported attitudes would settle the question; if the agreement is weak or the second axis fails to discriminate, the pipeline's political map is not measuring what it claims.","tokens_in":21385,"feed_emoji":"🗺️","tokens_out":7367,"duration_ms":69582,"temperature":0.7,"pith_summary":"This paper argues that studying who shares which news on social media needs three things at once: a political space with more than one dimension, a view of individual articles rather than only outlets, and a content taxonomy built from the texts themselves. It builds a pipeline that places roughly 1.2 million German Twitter users in a two-dimensional political space inferred from the MPs they follow, then maps about 500,000 March 2023 shares of articles from 26 outlets onto that space, with article full texts grouped into 220 topics and 12 metatopics. The proof of concept shows patterns invisible in a one-dimensional outlet-only design: outlets like focus.de and nachdenkseiten.de look alike at outlet level but differ sharply at story level, and the same outlet can be shared by different camps for different topics. If the pipeline holds up, it gives researchers a generalizable way to measure news circulation by source and content, not just by source.","feed_headline":"News sharing splits along a second political axis","feed_subtitle":"Story- and topic-level Twitter data show the same outlets feeding different camps different articles.","key_machinery":"The load-bearing object is the CHES-validated two-dimensional political space. It starts from the follower network of 433 German MPs: a correspondence analysis embeds 1.2 million users who follow at least three MPs, and the axes are rotated until they align with party positions from the Chapel Hill Expert Survey, yielding a left-right x-axis and a y-axis combining elite-skepticism, EU-skepticism, internal-market position, and protectionism (the EESP axis). Around this space, the pipeline layers link-level sharing data, roughly 500,000 tweets from March 2023 across 26 outlets, full texts scraped with newspaper3k, and a 220-topic Structural Topic Model aggregated into 12 metatopics. The key analytic move is comparing distributions at three granularities: users per outlet, users per individual story, and users per topic or metatopic, which is what lets the paper separate \"shared by both sides\" from \"different sides share different items from the same outlet.\"","core_discovery":"The central claim is that a two-dimensional political space with story- and topic-level resolution reveals systematic news circulation patterns that one-dimensional, outlet-level analyses miss. The paper confirms this with a German proof of concept: H1, H2a, H2b, and H3 are all answered positively. Concretely, the AfD occupies its own region high on the elite-/EU-skeptical/protectionist axis; outlets that look similar on the left-right axis (faz.net versus welt.de) separate on the second axis; focus.de and nachdenkseiten.de have similar outlet-level sharing distributions but different story-level structures, with focus.de's stories rarely bridging the EESP cleavage and nachdenkseiten.de's stories often shared across it; and climate-change news circulates across the spectrum while right-wing EESP users share stories about climate institutions while left-of-centre users share extreme-weather stories. The paper presents these as proof that the pipeline can uncover systematic insights in news sharing.","pith_inferences":["A natural next test is to run the same pipeline on a second month or another country; if the EESP axis does not reproduce, the German result may be specific to the timing or the particular MP list rather than to the method.","The 60 percent embedding coverage means the untraced 40 percent of shares could come from users who follow fewer politicians and are systematically different in age, tenure, or engagement; comparing those users' aggregate sharing with embedded users' sharing would bound this bias.","Because the second axis rests on excluding CA dimension 2 on the grounds that it tracks MP follower counts, re-running the pipeline with that dimension included, or with popularity partialled out, would show whether the EESP findings survive.","Story-level analysis could be converted into a quantitative \"unbundling score\": the fraction of an outlet's stories whose mean sharing position falls on one side of a cleavage, a metric that could be compared across platforms and over time."],"forward_implications":["If the pipeline is adopted beyond Germany, any country covered by CHES party-position data can get the same two-dimensional treatment, and the method also allows more than two dimensions in principle.","Story-level distributions become a direct measure of news unbundling: an outlet whose story means straddle a cleavage is supplying different camps with different items, while an outlet whose story means sit between poles is producing genuinely cross-cutting content.","Topic-by-outlet comparisons expose selective exposure at the level of beats, for example right-wing EESP users ignoring focus.de's climate coverage while sharing welt.de's, which outlet-level averages flatten.","The metatopic-versus-topic distinction provides a bottom-up way to detect subtler forms of biased information environments: the same broad issue is shared by all camps, but systematically different aspects circulate in each camp.","Combining follower-based ideology with text-based content lets researchers study circulation without pre-labelling outlets as partisan or non-partisan and without pre-judging articles as true or false."],"supporting_citations":[{"why":"Supplies the principle that following politicians on Twitter reveals ideological positions, the foundation of the follower-network embedding.","marker":"Barberá (2015)"},{"why":"Extends the revealed-preference assumption to Facebook page likes, strengthening the homophily premise.","marker":"Bond and Messing (2015)"},{"why":"Provides the Correspondence Analysis approach used to embed users and MPs in the low-dimensional space.","marker":"Bonica (2014)"},{"why":"Introduces the expert-survey rotation that turns abstract embedding axes into interpretable CHES dimensions.","marker":"Ramaciotti Morales et al. (2022)"},{"why":"Contributes the Chapel Hill Expert Survey party positions used to name the x- and y-axes.","marker":"Jolly et al. (2022)"},{"why":"Shows that US news sharing is more segregated at story level than at outlet level, the prior the paper builds on for H2a.","marker":"Green et al. (2025)"},{"why":"Supplies the Structural Topic Model used to infer 220 topics from article full texts.","marker":"Roberts et al. (2019)"},{"why":"Provides the newspaper3k scraper through which roughly 75 percent of shared links received full texts.","marker":"Ou-Yang (2020)"}],"fun_headline_variants":["Story-level data expose a second axis in news sharing","Outlet-level analysis misses a hidden news-splitting dimension","AfD stands apart: news sharing reveals a second political axis","Beyond left-right: news sharing splits on a second dimension","Twitter news sharing maps a two-dimensional political space"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"Everything rests on the assumption that users follow politicians they agree with; if following on Twitter is driven by fame, novelty, or social reasons, the political coordinates of every user, outlet, story, and topic shift.","fun_headline_variants_meta":{"raw":{"variants":["Story-level data expose a second axis in news sharing","Outlet-level analysis misses a hidden news-splitting dimension","AfD stands apart: news sharing reveals a second political axis","Beyond left-right: news sharing splits on a second dimension","Twitter news sharing maps a two-dimensional political space"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001292,"raw_usage":{"total_tokens":5247,"prompt_tokens":888,"completion_tokens":4359,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":504,"completion_tokens_details":{"reasoning_tokens":4280}},"tokens_in":504,"tokens_out":4359,"duration_ms":32341,"temperature":1.0,"reasoning_tokens":4280,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T21:56:26.849845+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"A validation study that matches a sample of embedded users to survey responses on left-right and EU/protectionism attitudes and checks whether their embedding coordinates predict reported attitudes would settle the question; if the agreement is weak or the second axis fails to discriminate, the pipeline's political map is not measuring what it claims.","supporting_citations":[{"cited_title":"and Messing, S","cited_arxiv_id":null,"evidence_quote":"Extends the revealed-preference assumption to Facebook page likes, strengthening the homophily premise."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the Correspondence Analysis approach used to embed users and MPs in the low-dimensional space."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Introduces the expert-survey rotation that turns abstract embedding axes into interpretable CHES dimensions."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Contributes the Chapel Hill Expert Survey party positions used to name the x- and y-axes."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Shows that US news sharing is more segregated at story level than at outlet level, the prior the paper builds on for H2a."},{"cited_title":"E., Stewart, B","cited_arxiv_id":null,"evidence_quote":"Supplies the Structural Topic Model used to infer 220 topics from article full texts."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the newspaper3k scraper through which roughly 75 percent of shared links received full texts."}],"review_version":1}