digital library teams

Abstract: Scientific document embeddings such as SPECTER2 are optimized for retrieval over large, heterogeneous corpora. On a small specialized corpus their similarity distributions can become highly compressed, which limits their usefulness for exploratory literature mapping. We study whether weakly supervised in-domain adaptation, using only metadata already available in open citation databases, changes this geometry. On ArtCon, a curated corpus of 704 publications on artificial and machine consciousness, we adapt the SPECTER2 proximity model with a lightweight LoRA adapter trained on positive pairs derived from bibliographic coupling in OpenAlex, using no manually annotated similarity labels. We compare the adapted embeddings with lexical (TF-IDF), citation-based (bibliographic coupling) and generic semantic (SPECTER2) representations under thresholded, fixed-density and symmetric k-nearest-neighbour graph constructions. Adaptation substantially decompresses the similarity distribution (mean pairwise cosine similarity 0.9104 to 0.8663, standard deviation 0.0259 to 0.0594) and yields graphs with markedly higher Louvain modularity and better connectivity than the base encoder under matched edge and neighbourhood budgets. The effect is stable across all three graph constructions and is therefore not an artefact of a single sparsification choice. We report this as a case study on one corpus, and we are explicit about what it does not establish: the clustering indices used here are computed against community labels derived from each representation's own similarity graph, so they measure internal geometric self-consistency rather than agreement with an external topical ground truth, and adaptation and evaluation use the same documents. Evaluation against independent labels, on held-out documents, and against geometric controls is ongoing work.

Abstract: Science maps describe relationships between papers. Text embedding models build these relationships by turning a paper's text into a list of numbers that gives it a position. Would another model give the same picture? I compare ten models using titles and abstracts of 500,000 publications from the OpenAlex bibliographic catalogue, published in 2000-2024 across 26 disciplines, including Medicine and Physics. Representing each discipline by its papers' average position produces similar arrangements across models, including when other papers are selected. Yet within a discipline and period, two models share, on average, only eight of the 25 papers closest to a given paper. Supplying titles alone also changes much of that list, with effects that vary between models. Models can also disagree about which discipline's papers are more spread out. One measure of spread compares angles between lines drawn from a common point to each paper's position. Moving that point can reverse which discipline appears more spread out, even when selecting other papers leaves the answer unchanged. The broad arrangement can remain similar while answers about nearby papers and differences between disciplines depend on the model, the text it receives and how its output is processed.

Abstract: Transforming scholarly PDFs into machine-readable fulltext remains a bottleneck for large-scale information systems. Recent vision-based parsers improve accuracy, but need GPUs and may introduce noise into the extracted text. GROBID, a modular font-stream parser running on CPU, is the de-facto standard for structuring scientific articles and underpins several of the largest open scholarly corpora. We pair it with a lightweight CPU detector localising figure, table, and paratext (header, footer, page number) regions, encoded as typed-area masks whose tokens are routed to GROBID's specialised models or discarded. On two PMC corpora, Bioinformatics (1,926 articles) and Materials Science (2,595), scored against JATS with a section-aware structural protocol, our extension improves over plain GROBID on most metrics (NS $+0.025$/$+0.013$; $+0.086$ paragraph recall on Materials Science, $d_z{=}1.08$), and caption-linked figure recovery improves on both corpora. On the external Table-BRGM benchmark, table detection recovers F1 $0.16 \to 0.94$ and table structure follows (GriTS-Top $0.27 \to 0.78$, below the strongest GPU system). On body text, against four vision-based systems (Docling, MinerU, olmOCR, dots.ocr), it has the best paragraph precision on both corpora, the best section detection on Materials Science, and a character error rate within 0.004 of the best GPU parser. End-to-end on CPU, it costs $2.7$--$3.2\times$ less than the cheapest GPU system (Docling) and $10$--$14\times$ less than generative parsers.

Abstract: Long-context models read a novel the way a person reads a printout: one token after another, in narrative order, with the whole history competing for a fixed budget of attention. A detective does not work that way. They sort what happened when, and they keep a map of who relates to whom, so a clue from chapter one can meet a question asked at the end of the book. We test whether a frozen knowledge graph can give a small local model that same freedom. Thirty detective novels and 234 multiple-choice questions are answered by one fixed qwen3.5:9b reader under nine conditions: five graph routes, a recent-window baseline, whole-book compression, ordinary vector retrieval, and a question-only control. The strongest graph route reaches 53.85% (126/234) against 46.15% for the recent window, 51.28% for compression, 51.71% for vector retrieval and 40.17% for question-only. On the subset that no model can answer without the book, the graph route reaches 42.86%. None of the fifteen graph-baseline contrasts survives Holm correction, so we present the result as exploratory evidence about a design. Two structural findings survive scrutiny better than the headline number: annotated evidence concentrates in the topological core of these graphs (2.35x enrichment, pooled), and the two graph-building pipelines differ so much in annotation coverage (16% versus 73% of clue paragraphs) that pooled accuracy alone would hide which bottleneck is being measured.

Abstract: Dense retrieval systems increasingly rely on multi-billion-parameter language models, whose memory and computational requirements make large-scale indexing, frequent corpus updates, and low-latency serving costly. We present a three-stage training pipeline for developing compact and efficient retrievers that remain competitive with substantially larger models. The pipeline combines cross-lingual alignment, relational knowledge distillation, and contrastive fine-tuning. It requires no original ground-truth relevance labels, relying exclusively on supervision generated by strong embedding models and rerankers utilised as teachers. Using this pipeline, we develop PolDense and EuroDense, both supporting contexts of up to 8,192 tokens. PolDense is a family of six Polish retrievers ranging from 17M to 1B parameters. EuroDense is a 435M-parameter retriever supporting nine European languages. We conduct an extensive evaluation covering 41 Polish and 150 multilingual retrieval tasks. The results demonstrate strong quality-efficiency trade-offs. PolDense-1B outperforms the evaluated retrievers with up to 9B parameters, while the PolDense family forms the Pareto frontier across model sizes. Among the evaluated models below 1B parameters, EuroDense ranks first in both task-averaged and language-averaged performance and leads in seven of nine languages. We release all models publicly.