Cache the programmes
Session pages for 2020–2027 were collected gently and cached locally, with parsers for the different historical page layouts.
Transparent by design
How the matcher
was built.
The system turns historical EGU programmes into a rigorously evaluated retrieval model, then packages only the small model and current programme as static browser assets.
For one new abstract
What happens in your browser
Every inference step stays on the device. The server provides static files but never receives the submitted title or abstract.
- 01
Read several views of the text
The title, complete title-plus-abstract, and tokenizer-aware specialist passages are read separately. Long abstracts use overlapping chunks instead of silently losing the ending.
- 02
Create semantic vectors
A quantized MiniLM transformer uses attention layers and mean pooling to create normalized 384-dimensional representations. A compact learned query transformation adapts them to EGU retrieval.
- 03
Compare every scientific session
The query views are compared with precomputed title and description vectors for all 1,259 typical EGU27 sessions. Special events such as medals, short courses, geo panels, town halls, and administrative sessions are excluded.
- 04
Combine five relevance signals
The scorer combines joint text similarity, title-to-title, title-to-description, passage-to-title, and passage-to-description matches using weights learned on past assignments.
- 05
Adapt the number of suggestions
A validation-calibrated score-gap rule returns a short set for focused abstracts and a broader set for ambiguous ones. You can still reveal the next ten sessions through the complete ranking.
Ranking, not classification
The score is comparative
Each cosine term measures the angle between two normalized vectors. The learned non-negative weights determine how much each scientific view contributes.
The value is useful for ordering sessions, but it is not the probability that a session is “correct.” Close leading scores indicate genuine ambiguity.
Historical construction
From programme pages to training pairs
The preparation pipeline is deliberately separate from the public site. It can be computationally expensive; the exported matcher remains compact.
Recover abstracts
Crossref records with Copernicus EGU DOI prefixes provide titles and abstracts. Original source-file and line locations remain traceable.
Join accepted assignments
Abstract identifiers from programme pages link each abstract to its recorded session. Multiple valid assignments are preserved instead of forcing one label.
Prevent leakage
Session representations contain only official code, division, title, and description—never abstracts already accepted into that session.
Evaluation through time
Past years keep their roles
EGU27 has no assignments yet, so the final model learns from every labeled year and is calibrated on the most recent one. The same recipe was first backtested one year earlier: trained on 2022–2024, calibrated on 2025, and scored on EGU26, which it never saw. Sparse 2020–2021 public assignments are retained only as auxiliary catalogues.
Learn retrieval transformations and ranking weights from 68,000 assigned abstracts.
Choose the stopping point and calibrate the 80% prediction set on the most recent year.
The same recipe, shifted one year back, reached 77.5% Recall@10 on an unseen year.
Rank the newly published EGU27 programme.
79.5%EGU26 backtest coverageunseen year · median set 9
77.5%EGU26 backtest Recall@10unseen year · MRR 0.469
80.4%EGU26 calibration coveragefinal model · median set 9
Coverage means at least one historically recorded assignment appears in the calibrated set. It does not mean that assignment was uniquely correct; authors and conveners can reasonably choose different interdisciplinary homes.
Static deployment
Expensive preparation.
Lightweight use.
The selected model is exported to quantized ONNX. Session vectors are prepared once, stored compactly as float16, and verified by checksum when loaded.
- Web Workerkeeps the interface responsive
- WebGPUpreferred local acceleration
- WebAssemblycompatible CPU fallback
- Static cachesupports later offline inference
No API endpoint receives the abstract. The only network requests fetch immutable application, model, and catalogue assets.
Use with scientific judgment
What the matcher cannot decide
Historical assignments are imperfect labels: some abstracts could fit several sessions, and some may have been submitted strategically or incorrectly.
Similarity does not account for conveners’ unpublished preferences, late programme changes, invitation status, or author eligibility.
Always read the official description and submission rules. The tool supports navigation; it does not replace the programme or make a submission decision.
Try the current catalogue