Every issue of a scholarly journal carries a masthead: a page listing the editor-in-chief, the associate editors and the subfields each one handles, the advisory board, and the officers of the sponsoring society. Read across decades, these pages record who controlled peer review in a field, when new subfields got their own editors, and how long gatekeepers stayed in place. Digital libraries index this history but do not read it. They ingest mastheads as malformed "articles", or filter them out as noise.
We propose to recover that record automatically and then use it to answer one historical question: when a research topic surges, does it get a dedicated editor before, during, or after the surge? The extraction is done by a layout-aware sequence labeler, a BiLSTM-CRF. This is a model that reads a page's tokens in both directions and chooses one globally consistent tag sequence under explicit rules about which tag may follow which. We add a check that no single-page extractor has. Consecutive issues of a journal repeat almost the same roster, so disagreements between neighbouring issues expose extraction errors without any extra labels.
Core Claim
The main contribution is the pipeline from pages to career timelines, and its validation method. The key technical claim is that cross-issue consistency is a strong, label-free error signal for masthead extraction. A person who appears in issue $t-1$ and issue $t+1$ but not in issue $t$ is almost always an extraction miss, not a resignation and reappointment. Using this signal to repair and self-train the extractor should raise the accuracy of the final tenure timelines (who held which role, and from when to when) well above what the best single-page extractor achieves.
The claim is stated so that it does not depend on the BiLSTM-CRF beating large generative models page by page. We prefer the conditional random field because it gives per-span probabilities that the cross-issue linker can use, and because its transition rules guarantee well-formed output. Whether it also matches a prompted multimodal model on individual pages is an empirical question the experiments settle.
What Exists and What Does Not
Three lines of prior work supply the method.
- Layout features in a BiLSTM-CRF. Construction of the Literature Graph in Semantic Scholar built ScienceParse, a BiLSTM-CRF that adds font size and bounding-box position to token features to pull titles, authors and references out of PDFs. Its resulting graph contains authors, papers and citations, but no editors.
- A layout benchmark. DocBank supplies token-level layout annotation for 500,000 arXiv pages. Its 12 categories contain nothing resembling an editorial role, and its LaTeX-generated pages are far more uniform than publisher front matter.
- Schema-driven LLM extraction. MOLE and MeXtract extract metadata by prompting or fine-tuning language models against a schema. MeXtract reports that small generative models struggle with deeply nested structures, and nesting is exactly what a masthead has.
None of the three treats front matter as a target. None exploits the repetition of the same document across issues.
Two papers constrain the design:
- Heterogeneous label sets. Marginal Likelihood Training of BiLSTM-CRF for Biomedical Named Entity Recognition from Disjoint Label Sets shows how to train one CRF when different training sets annotate different subsets of the tags. We reuse this to pool mastheads from publishers whose pages expose different fields; some list affiliations and some do not.
- Encoders versus prompting. Few-shot clinical entity recognition in English, French and Spanish found that a RoBERTa-large encoder with a BiLSTM-CRF head beat a prompted Mistral-7B (F1 0.462 versus 0.291 on n2c2 clinical data). That motivates the encoder choice. It is weak evidence against much larger multimodal models, and we do not lean on it beyond that.
The motivating evidence comes from the documents themselves and from bibliometrics.
- The documents. The IEEE publication-information pages (Information Theory, Systems, Man, and Cybernetics, Cybernetics, Human-Machine Systems, Fuzzy Systems) and conference front matter (Front Matter, Front Matter) sit in scholarly indexes as ordinary records. For example, the Information Theory masthead is attributed as an authored paper.
- Bibliometric practice. Revisiting gender bias research in bibliometrics found that 51.4% of the 70 studies it reviewed did no author name disambiguation, and that studies examine bylines, not gatekeepers.
On the historical side, Will the Technological Singularity Come Soon? fits multi-logistic growth waves to AI publication counts. We reuse its curve-fitting method to date topic surges. We then ask how editorial structure moves relative to those surges, a question that growth-curve paper cannot address from counts alone.
Research Program
Benchmark. The first artifact is a masthead benchmark. We will collect every available issue's front-matter page for roughly forty journals from IEEE, ACM, Elsevier and Springer, plus conference front matter in the style of Past IJCAI Conferences. We will hand-annotate about 1,500 pages, stratified by publisher and decade. The annotation guideline must be fixed and double-annotated from the start, with agreement reported. Five Years of Argument Mining documents how inconsistent guidelines capped progress in that field.
Tag set. The tags are:
- role headers, such as "Editor-in-Chief" or "Associate Editors";
- area subheaders, such as "Coding Theory";
- person names;
- affiliations;
- locations;
- other text.
Each uses the begin/inside/outside scheme.
Handling nesting. A linear-chain CRF cannot represent nesting directly. We therefore split the work in two:
- The CRF's transition rules encode the local grammar. For example, an affiliation cannot start without a name earlier in the same visual block.
- A deterministic stack assembler turns the tag sequence into a tree of tier, area, person and affiliation, where each header governs everything until the next header of the same or higher level.
Extractor. Reading order and block boundaries come from token coordinates, following ScienceParse. The encoder is a pretrained transformer feeding a BiLSTM-CRF.
Linking across issues. Extracted entries for one journal are aligned issue by issue. The linker matches persons across issues by name similarity, affiliation and role, using the CRF's span probabilities as match weights. It then applies a stability prior: a single-issue gap in an otherwise continuous tenure is treated as a likely extraction error. Each flagged gap sends the page back for constrained re-decoding, with the expected span pinned. Pages whose repaired output becomes consistent are added as pseudo-labels for self-training.
Outputs. The output is a set of tenure timelines per journal. Each record is a person, a role, an area, and start and end issues. Cross-journal person identity is resolved with a disambiguation step whose error rate we report in the format recommended by the SoDA-card paper.
Historical analysis. For each associate-editor area title, we map it to a topic and fit a multi-logistic curve to that topic's publication counts, following Ni et al. We then measure the lag between the area's first appearance on a masthead and the curve's inflection point. This produces a distribution of lags, not an anecdote.
Baselines
Naive reference: layout-aware rules. This is an OCR-plus-rules parser that splits on bold or capitalized lines matching a lexicon of role titles, then assigns the following lines as names and affiliations using punctuation patterns. It is what a digital library would build quickly. MOLE describes legacy rule-based systems such as CERMINE and ParsCit as brittle across publisher formats. It runs on the same pages. Beating it should be easy on modern IEEE pages and is only meaningful on older and non-IEEE pages, so we report results by decade and publisher.
Strongest existing method: a multimodal LLM prompted with a JSON schema. A large multimodal model receives the page image plus the extracted text and a schema, in the style of MOLE. We also run a small fine-tuned model in the MeXtract style, trained on the same labeled pages as our extractor. No reported result was found on Lacuna for masthead or front-matter extraction by any method. The nearest reported numbers come from different tasks and are not comparable:
- ScienceParse, author extraction: F1 92.1%.
- LayoutLM-Large on DocBank: macro F1 0.9350.
- Gemini 1.5 Pro on MOLE: average F1 75.23.
Fairness conditions. All baselines see the same test pages and the same 1,500 labels, split identically. Prompt engineering for the LLM gets the same development-set tuning budget, measured in development evaluations, as hyperparameter search for the CRF. Two outcomes count as beating the LLM:
- matching its span F1 within two points while producing zero structurally invalid trees; or
- after cross-issue repair, exceeding its tenure-timeline accuracy.
A per-page F1 win is not required.
Ablations. Three ablations isolate the key ingredient:
- The same BiLSTM-CRF with cross-issue linking but without repair and self-training.
- The LLM baseline given the identical cross-issue repair, feeding its outputs through the linker with uniform confidence. This separates the value of consistency from the value of CRF probabilities.
- The CRF without its transition rules.
Why not just…?
Why not just prompt a multimodal LLM on every page? This is the real threat, and cost is not a decisive answer. Forty journals over several decades amounts to tens of thousands of pages, which is affordable with an API. The answer rests on three properties that the experiments measure directly:
- Structural validity: how many outputs have names without roles, or broken nesting.
- Reproducibility across reruns.
- Usefulness of confidence estimates for linking.
The clinical comparison favouring encoder-plus-CRF models (Naguib et al.) used a 7B model and does not settle this. If the LLM wins on pages and its outputs repair equally well under the cross-issue method, we will adopt it as the extractor. The contribution then becomes the benchmark, the consistency-based repair, and the timelines. The second ablation above is designed to reveal exactly this outcome.
Why not just write rules per publisher? Rules work for a stable template and fail when the template changes. Template changes happen over the decades that make longitudinal history interesting. The per-decade breakdown will show where rule accuracy collapses. If it does not collapse, that is a useful finding in its own right and makes the whole corpus cheaper.
Why not just query OpenAlex, Crossref or Semantic Scholar? They do not hold the data. The Semantic Scholar literature graph has no editor nodes (Groeneveld et al.), and indexes list mastheads as authored papers (the Information Theory page is one example). They are useful only for cross-journal person disambiguation, and we use them there.
Why not hand-curate a few journals? For a single journal's history, hand curation is better, and we will do it for five journals as our gold standard. The lag analysis needs dozens of journals and hundreds of area titles, and the manual route cannot reach that scale.
Proposed Experiments
Page-level extraction. We report span F1 per tag, and triple F1 over (person, role, area) triples. Results are stratified by publisher and decade and compared across all baselines.
Timeline accuracy (decisive check). For five journals, we hand-verify full editorial timelines. We then measure:
- the fraction of tenures recovered;
- the error in start and end issue;
- the rate of spurious one-issue gaps and phantom appointments.
The hypothesis fails if cross-issue repair does not reduce timeline errors relative to the unrepaired extractor. It also fails if the reduction is no larger for the CRF than for the LLM given the same repair; in that case CRF probabilities add nothing.
Label efficiency. Self-training on repaired pages should let the CRF trained on 300 labeled pages approach the one trained on 1,500. We will plot the learning curve.
Historical result. We report the distribution of lags between each area's editorial appearance and its topic's growth inflection, with bootstrap intervals over topics. A lag distribution centred clearly after inflection would say journals formalize subfields only once the literature has already surged. A distribution centred before inflection would suggest editorial structure helps seed them. Either answer comes from the extracted data, not from retrospective narrative.
References
- Construction of the Literature Graph in Semantic Scholar
- DocBank: A Benchmark Dataset for Document Layout Analysis
- MOLE: Metadata Extraction and Validation in Scientific Papers Using LLMs
- MeXtract: Light-Weight Metadata Extraction from Scientific Papers
- Marginal Likelihood Training of BiLSTM-CRF for Biomedical Named Entity Recognition from Disjoint Label Sets
- Few-shot clinical entity recognition in English, French and Spanish: masked language models outperform generative model prompting
- IEEE Transactions on Information Theory publication information
- IEEE Transactions on Systems, Man, and Cybernetics publication information
- IEEE Transactions on Cybernetics publication information
- IEEE Transactions on Human-Machine Systems publication information
- IEEE Transactions on Fuzzy Systems publication information
- Front Matter
- Front Matter
- Revisiting gender bias research in bibliometrics: Standardizing methodological variability using Scholarly Data Analysis (SoDA) Cards
- Will the technological singularity come soon? Modeling the dynamics of artificial intelligence development via multi-logistic growth process
- Past IJCAI Conferences
- Five Years of Argument Mining: a Data-driven Analysis