AI evidence extraction

Automated Identification of Research Gaps Using Keyword Clustering and an Embedding Model

Paweł Karol Frankowski, Joanna Wiśniewska, Sebastian Matysik · Journal of the Association for Information Systems · 2026

AI-generated evidence extraction, verified across multiple analytical personas. Not a substitute for the peer-reviewed original.

9/10

Relevance

0/4

Quality (LMQS)

Evidence

Citations

0.00

FWCI

Methodology & findings

Study design

Case study with computational pipeline validation; semi-automated systematic literature review analysis using semantic clustering and gap-detection criteria applied to 14,302 records from Scopus; comparison with alternative methods (VOSviewer, LLM validation across five models)..

Primary method

Design science research (computational artifact development with validation)

Main result

The pipeline identified eight candidate gaps in a gamification-in-marketing case study with 14,302 records. "Compared to VOSviewer co-occurrence maps, our approach shows superior semantic coherence and is less affected by keyword variation. Validated across five LLM shows method effectiveness - nearly twice as many gap classifications and a 24–30% increase in novelty scores."

Research paradigm

Positivist/empiricist (computational/algorithmic validation)

Author conclusions

The authors conclude that their approach "combine[s] semantic keyword clustering with quantitative gap-detection criteria to semi-automate a formerly manual process, improving scalability and reproducibility" and that "Compared to VOSviewer co-occurrence maps, our approach shows superior semantic coherence and is less affected by keyword variation."

Risk of bias

Single domain case study (gamification-in-marketing) may not generalize; Dependence on GPT-4 for cluster labeling introduces LLM bias; Scopus-only querying may miss non-indexed publications; Gap detection criteria rely on user-defined thresholds (95th percentile); No blinded validation or inter-rater reliability reported; Single case study (gamification-in-marketing) may not generalize to other research domains; Reliance on GPT-4 for cluster labeling introduces model-specific bias; Semantic similarity threshold (95th percentile) is data-dependent and may require re-tuning across domains; Publication bias inherent in Scopus-indexed literature; User problem description formulation could influence results via embedding similarity

Open questions raised

Future work could extend validation across additional domains beyond gamification-in-marketing, investigate threshold sensitivity for gap-detection criteria, and explore integration with other LLMs beyond the five tested.
The paper does not explicitly identify future research directions or additional gaps in the provided abstract.

Data: not_statedCode: not_statedExtracted from: pdfAgreement 63%

Explore related topics

Systematic Review Automation Literature Discovery Design Science