The exploration layer before screening
Every topic is built from keyword pairs that passed a hypergeometric test with false-discovery-rate correction. Every paper is assigned by a second test. You get the map — and the methods paragraph to cite it.
Legend
Node size degree
Edge width φ
Selected term
resilience
Co-occurring
Documents (23)
Building resilient supply networks under disruption
Open DOI ↗
Recovery strategies after systemic shocks
Open DOI ↗
VOSviewer shows you clusters. It can’t tell you whether they’re real.
Bibliometric mapping tools cluster keywords by raw co-occurrence and stop there. aiSlr runs a significance test on every link and every assignment, so each cluster comes with a p-value, a correction procedure and a threshold you can report. That is the whole difference, and it is a narrow one: they draw the map; we also test it.
How it works
No black-box model decides what your field is about. Every link, every topic and every assignment comes with a test statistic, a correction procedure and a threshold you can put in the methods section.
Scopus or Web of Science CSV, RIS, BibTeX, PubMed MEDLINE, OpenAlex JSON, and CNKI / Wanfang exports. Formats are detected, columns mapped, duplicates removed — and you see the parsed records before anything is spent.
For each candidate keyword pair, a one-sided hypergeometric test asks whether the pair co-occurs more than chance allows. p-values are corrected with the Benjamini–Hochberg procedure and only surviving pairs enter the network.
Louvain community detection groups the validated network into topics with a fixed random seed, so the same data gives the same map. One threshold controls how strict the links must be.
Each document is assigned to topics by a second hypergeometric test with Bonferroni correction, so every cluster is backed by a list of papers, not a hunch.
Every run writes a methods paragraph with the exact tests, corrections, thresholds and seed, plus a statistical appendix and the parameters needed to regenerate it.
What you take to the manuscript
Every run produces a paragraph like this one, with the exact tests, corrections, thresholds and random seed used — plus a statistical appendix and the parameters needed to regenerate the result. Copy it, cite it, or download it as Markdown.
Generated by aiSlr 0.1.0 from a 600-record arXiv export (machine learning in healthcare, 2014–2026) at ρ > 0.3. Reproduced verbatim.
Topic structure was derived from the titles and abstracts of 600 records, of which 600 contained abstracts of at least 40 words and were retained for analysis. Text was Unicode-normalised (NFKC), lower-cased and tokenised; adjacent word pairs occurring at least 2 times with a pointwise mutual information of at least 3.0 were merged into phrase tokens. Tokens were reduced to stems with the Porter algorithm (PorterStemmer, NLTK) and filtered against the NLTK English stop-word list (n = 198), a built-in list of academic boilerplate stems (n = 51) and 0 project-specific stop words; tokens shorter than 3 characters were discarded.
Terms occurring in at least 2 documents were retained as candidate nodes. For each of the 965877 candidate term pairs, the null hypothesis of independent occurrence was tested with a one-sided (right-tail) hypergeometric test, and p-values were corrected for multiple comparisons with the Benjamini–Hochberg false discovery rate procedure at α = 0.05. Pairs surviving correction with a binary Pearson correlation above ρ = 0.3 were retained, yielding a validated co-occurrence network of 1337 terms and 1762 edges.
Topics were identified by Louvain community detection on the validated network, with edges weighted by ρ and a fixed random seed of 42, producing 385 communities (modularity Q = 0.944).
Documents were assigned to topics by a second one-sided hypergeometric test of the overlap between each document's validated terms and each topic's term set, corrected with the Bonferroni procedure at α = 0.01; a document may be assigned to several topics or to none. 211 documents received at least one assignment and 389 received none. A further Bonferroni-corrected test at the same α identified 14 documents whose vocabulary drew significantly on the validated term set as a whole; these are reported separately as general documents.
Where this fits
aiSlr sits between your search and your screening tool. It is not a Covidence or Rayyan competitor — it is the step before them.
Run your query in the databases you already use.
Scopus · Web of Science · PubMed · OpenAlex · CNKI
Import the export, get a statistically validated topic map, the methods paragraph and the timeline.
aiSlr
Title/abstract screening, dual review, conflict resolution.
Rayyan · Covidence · DistillerSR
Paste the methods paragraph, cite the appendix, export the figure.
Your manuscript
What it doesn’t do
Stating the limits plainly is the point. If you need these, you need a screening platform — and you can still start here.
One workspace
Zoom from the whole corpus down to a single sentence in a single paper, and back, without losing the thread.
Topic overview
Open a finished analysis and the first thing you see is the whole field: one circle per topic, sized by how many papers it holds.
Hide the small ones, hover for detail, click to go in. You always know which level you are on.
Co-occurrence network
Drill into a topic and the keywords that define it lay themselves out as a network. Thicker lines are stronger correlations — and every line passed the test.
Click a term to light up its neighbours, see the exact correlation values, and filter the papers down to the ones that use it.
Global keywords
Step back from topics and look at the vocabulary of the entire review: every keyword ranked by importance, with the documents behind it.
Raise the minimum document count to cut noise, or search to jump straight to a term on the map.
Evidence panel
Nothing on the map is decorative. Select anything and the panel lists the papers that support it, with abstracts a click away and DOIs that open.
Built for the moment a reviewer asks where a theme came from.