SciTopic: Automated Scientific Concept Pages

• SciTopic is a concept inventory that generates pages with concise definitions, related concepts, and key excerpts from scholarly sources.
• It employs a linear pipeline that combines dictionary matching, supervised definition extraction, and unsupervised snippet ranking to assemble each page.
• The system supports over 360,000 pages across 20 domains, enhancing navigation and understanding for 23 million monthly unique visitors.

SciTopic, presented in the Topic Pages system, is an automatically generated inventory of scientific concept pages derived from scientific books and journals. Each page corresponds to a scientific concept and contains three elements extracted from peer-reviewed publications: a definition, related concepts, and the most relevant snippets. The production collection contains over 360,000 Topic Pages across 20 different scientific domains and receives an average of 23 million unique visits per month, positioning it as a large-scale infrastructure for concept-level navigation in scholarly reading environments (Azarbonyad et al., 2023).

1. Concept and scope

SciTopic addresses a specific knowledge-access problem in scholarly communication: readers routinely encounter unfamiliar scientific concepts while reading domain literature, yet the relevant explanatory material is distributed across articles and books. Topic Pages operationalize a concept-centric solution by assembling concise, publication-grounded summaries around individual concepts rather than documents, authors, or venues (Azarbonyad et al., 2023).

The system is anchored in an inventory of scientific terminology and produces one page per concept. In production, the concept source is the Omniscience taxonomy, with approximately 700 K concepts spanning 20 domains. The resulting Topic Pages are hyperlinked back into ScienceDirect articles, so that one click on an annotated concept navigates to its corresponding page. This coupling of concept annotation and page generation makes SciTopic both a retrieval layer and a reading aid (Azarbonyad et al., 2023).

Within the broader literature, SciTopic occupies a distinct position. It is not primarily a topic-modeling system, an ontology-construction framework, or a trend-forecasting tool, although it touches adjacent problems. For example, the Science & Technology Ontology (S&TO) extracts 5,153 topics and 13,155 semantic relations from 393,991 recent articles using BERTopic (Kumar et al., 2023); ScienceWISE formulates article-to-topic assignment through concept–article networks and overlapping clustering (Martini et al., 2016); and SciTrends predicts topic popularity five years ahead across 125 topics using publication histories, patents, and language-model embeddings (Ofer et al., 2023). By contrast, SciTopic is organized around concept explanation and concept-page generation from full-text scholarly content (Azarbonyad et al., 2023).

2. End-to-end pipeline architecture

SciTopic is implemented as a linear pipeline with four major stages, beginning from raw XML documents and ending in published Topic Pages. The source corpus comprises 18 million full-text XML articles and book chapters from 2,700+ journals and 43,000 books. Documents are split into sections, referred to as snippets, and then into sentences (Azarbonyad et al., 2023).

The annotation module performs terminology matching against the Omniscience taxonomy. Concept detection uses exact and alias dictionary lookup, with abbreviation detection via Schwartz & Hearst’s algorithm. When annotations overlap, the system resolves them by selecting the longest matching concept span. For each concept mention, the enclosing sentence and snippet are passed to downstream components (Azarbonyad et al., 2023).

The page-generation workflow then branches into three extraction tasks. First, definition ranking scores every sentence containing a concept and selects the top-scoring sentence as the single-sentence definition. Second, snippet ranking gathers all snippets containing the concept and chooses the top 10 by a location-aware term-frequency measure. Third, related-concept extraction collects other concepts that co-occur with the target concept in those snippets and ranks them by co-occurrence count, returning the top 5 as “Related Concepts.” The selected definition sentence, related-concept list, and top snippets are then assembled into a web page template (Azarbonyad et al., 2023).

A notable property of the architecture is its mixture of supervised and lexical components. Definition extraction uses trained classifiers, while snippet ranking and related-concept extraction remain unsupervised and lexical in production. This division reflects an explicit performance trade-off: simple co-occurrence and lexical ranking methods are extremely fast and low-memory, enabling real-time page generation, but at the cost of deeper semantic relatedness modeling (Azarbonyad et al., 2023).

3. Definition extraction as sentence ranking

Definition extraction is formulated as a ranking problem over candidate sentences that contain the target concept $c$. For each candidate sentence $s_i$, the system computes a score $p_i = \mathrm{Model}(c, s_i)$ and selects the sentence with maximal score as the definition (Azarbonyad et al., 2023).

Two classifiers are described. The first is an LSTM+CNN hybrid. It uses two Bi-LSTMs, one over word embeddings and one over POS-tag embeddings, together with two 1D-CNNs applied separately to word embeddings and POS-tag embeddings. The four representations are concatenated into a vector $h$, followed by a feed-forward layer with sigmoid output:

$\hat{y} = \sigma(W h + b).$

Training uses binary cross-entropy:

$$L = -\frac{1}{N}\sum_{i=1}^N\bigl[y_i\log\hat{y}_i + (1-y_i)\log(1-\hat{y}_i)\bigr].$$

The second classifier is SciBERT-based: the input is formatted as “[CLS] $c$ [SEP] $s$ [SEP]”, the final-layer [CLS] embedding $h_{\mathrm{cls}}$ is fed into a single linear layer with sigmoid output, and the same cross-entropy loss is used (Azarbonyad et al., 2023).

The module relies on standard NLP preprocessing, including sentence tokenization, POS-tagging, and embedding lookup. The LSTM+CNN model uses GloVe embeddings, whereas the SciBERT model uses WordPiece tokenization. The Bi-LSTM component captures sequential dependencies via forward and backward recurrences, and the CNN component captures local $n$-gram patterns through width-$s_i$0 filters. The paper characterizes SciBERT’s transformer layers as jointly encoding semantics and syntax (Azarbonyad et al., 2023).

Training data come from two sources: WCL, consisting of 4,619 Wikipedia sentences labeled as “definition” or “not,” and a proprietary expert-labeled dataset of 43,368 sentence–concept pairs. In evaluation, SciBERT exceeds the hybrid model on the WCL benchmark, with macro-averaged $s_i$1, $s_i$2, and $s_i$3, compared with $s_i$4, $s_i$5, and $s_i$6 for LSTM+CNN. On the proprietary 43 K-sample dataset across eight domains, SciBERT outperforms LSTM+CNN in all but Social Sciences; example per-domain F1 scores are .79 vs .68 in Chemistry, .82 vs .65 in Earth Sciences, and .41 vs .42 in Social Sciences (Azarbonyad et al., 2023).

The domain variation is consequential. SciBERT is pre-trained primarily on biomedical and computer science text, and the reported lower performance on under-represented domains such as Social Sciences indicates a domain-adaptation issue rather than a uniform definition-extraction ceiling. This suggests that the accuracy of page generation depends not only on sentence-ranking architecture but also on the match between pretraining distribution and domain vocabulary (Azarbonyad et al., 2023).

4. Snippet ranking and related-concept extraction

Snippet retrieval operates over all document sections containing at least one occurrence of concept $s_i$7. Each snippet $s_i$8 is scored by a location-aware term-frequency function:

$s_i$9

where $p_i = \mathrm{Model}(c, s_i)$0 is the raw count of $p_i = \mathrm{Model}(c, s_i)$1 in $p_i = \mathrm{Model}(c, s_i)$2, $p_i = \mathrm{Model}(c, s_i)$3 is snippet length in tokens, and $p_i = \mathrm{Model}(c, s_i)$4 is the token index of the concept’s first occurrence. The top 10 snippets by $p_i = \mathrm{Model}(c, s_i)$5 are selected for the Topic Page (Azarbonyad et al., 2023).

This scoring rule is purely lexical and does not use supervised learning in the current production system. It balances three factors already present in the formula: concept repetition, snippet length normalization, and early mention position. A plausible implication is that snippets are favored when they are relatively concise, concept-dense, and concept-frontloaded, which is consistent with the goal of surfacing explanatory passages rather than diffuse contextual mentions (Azarbonyad et al., 2023).

Related concepts are extracted from the same snippet pool. For a target concept $p_i = \mathrm{Model}(c, s_i)$6, the system collects all co-occurring concepts $p_i = \mathrm{Model}(c, s_i)$7 and ranks them by shared-snippet count:

$p_i = \mathrm{Model}(c, s_i)$8

The top 5 co-occurring concepts are displayed as “Related Concepts.” No embeddings are used in the current production system; ranking is purely by co-occurrence count, under the assumption that high co-occurrence correlates with semantic relatedness (Azarbonyad et al., 2023).

This design is computationally economical but methodologically conservative. The paper explicitly notes possible future use of cosine similarity between contextualized embeddings derived from concept definitions, for example with SciBERT-derived vectors and

$p_i = \mathrm{Model}(c, s_i)$9

Similarly, future snippet ranking is expected to exploit click-through data for learning-to-rank objectives such as pairwise hinge or ListNet cross-entropy, but these mechanisms are not yet in production (Azarbonyad et al., 2023).

The contrast with later topic-page systems is instructive. TOPICAL, focused on biomedical entities, replaces lexical snippet ranking with a retrieval–clustering–prompting pipeline over PubMed, SPECTER2 PRX embeddings, community detection, and GPT-4 prompting, generating a structured page with a definition statement, main content, future directions, and inline PMIDs (Giorgi et al., 2024). SciTopic instead keeps page assembly extractive at the definition/snippet layer and reserves learned modeling for sentence-level definition detection (Azarbonyad et al., 2023).

5. Evaluation, scale, and operational profile

SciTopic reports both component-level and page-level evaluation. For definition extraction, performance is measured with macro-averaged precision, recall, and F1. On the WCL dataset, the paper compares against Jin et al. (2013), Li et al. (2016) LSTM, and Navigli & Velardi (2010), with the SciBERT model achieving the highest reported F1 at .93 (Azarbonyad et al., 2023).

At the page level, the system is already deployed at substantial scale. The production inventory contains 363,000+ Topic Pages across 20 science domains, with approximately 23 million monthly unique visits. The average missing-definition rate, defined as cases where no good candidate is found, is approximately 17%, corresponding to approximately 63,000 pages (Azarbonyad et al., 2023).

The operational scale is tied directly to the throughput of the XML pipeline. It processes more than 18 M documents while annotating mentions against a 700 K-entry taxonomy. Domain coverage includes Chemistry, Engineering, Medicine, Social Sciences, and other scientific areas. These figures indicate that SciTopic is not a small benchmark system but a high-volume production pipeline designed for continual concept coverage (Azarbonyad et al., 2023).

The reported performance profile reveals a characteristic engineering compromise. The dictionary-driven annotation stage is simple but robust for large terminology inventories; the definition module applies heavier supervised models where precision matters most; and snippets and related concepts use lightweight lexical or co-occurrence statistics. This suggests a design principle in which model complexity is concentrated on the component most visible to readers—the definition sentence—while keeping the remaining stages computationally tractable (Azarbonyad et al., 2023).

6. Position within the scientific-topic infrastructure landscape

SciTopic belongs to a broader family of systems that structure scientific knowledge around topics, concepts, and their dynamics, but its function remains distinct. Ontology-oriented systems such as S&TO construct topic networks and semantic relations, including relatedIdentical, superTopicOf, subTopicOf, and CommonArticles, from recent scientific corpora (Kumar et al., 2023). Sci-OG advances this direction with semi-automated ontology generation based on topic discovery, relationship classification, and ontology construction, and reports an overall F1 of 0.951 for its hybrid relationship classifier on 21,649 annotated semantic triples (Pisu et al., 6 Aug 2025). These systems emphasize structured topic graphs rather than reader-facing concept pages.

Trend-oriented systems pursue a different objective. SciTrends forecasts scientific topic popularity five years in advance from normalized publication counts, review-to-research ratios, patent features, and topic-name embeddings, with CatBoost + embeddings achieving $h$0 and MAE approximately 58.5 for the popularity target $h$1 (Ofer et al., 2023). “Scientific X-ray” models topic evolution through idea trees and Knowledge Entropy, reporting that visible depth does not exceed 6 jumps for 99.9% of sampled topics and that 99.11% of high-KE nodes have $h$2 (Li et al., 2021). More recently, an OpenAlex-based framework quantifies topical prevalence and short-term dynamics across journals, countries, regions, and domains using prevalence measures and five-point OLS slopes (Sorzano, 26 May 2026). These approaches analyze scientific topics as evolving systems rather than as concepts requiring explanatory pages.

A separate adjacent line concerns topic discovery and clustering. ScienceWISE models article–concept relations as bipartite and projected networks, filters generic concepts using entropy-based criteria, and applies Louvain and HiReCS to derive overlapping topics (Martini et al., 2016). A later SciTopic system enhances scientific topic discovery with a textual encoder over title, abstract, and metadata, entropy-based sampling, and LLM-guided triplet fine-tuning, reporting gains over baselines such as BERTopic and FASTopic across NeurIPS, DBLP, arXiv, and PubMed datasets (Li et al., 28 Aug 2025). These methods seek latent topical structure in corpora; the Topic Pages system instead operationalizes concept explanation from explicit concept mentions (Azarbonyad et al., 2023).

The principal limitations of SciTopic are also explicit. Related-concept extraction does not yet use embeddings; snippet ranking does not yet use supervised learning; and SciBERT’s pretraining profile leaves under-represented domains less well served. Planned improvements include domain-specific fine-tuning of SciBERT using in-domain definition pairs, including Wikipedia extractions, and supervised learning for snippets and related concepts using user click logs (Azarbonyad et al., 2023). This suggests a development trajectory in which the current extractive, concept-centric architecture could be incrementally augmented with semantic similarity modeling and interaction-driven ranking without changing its basic page abstraction.