ICE-Map Visualization a demo of icemap
The ICE-Map Visualization, as introduced by Eckert, Stuckenschmidt and Pfeffer (2007) and described comprehensively by Eckert (2012), depicts a concept hierarchy as a treemap and colours it by the information content of the concepts in a set of indexed documents. This demo applies the web implementation icemap to two knowledge organization systems that Eckert (2012, chapter 3) used to evaluate subject indexing, the Medical Subject Headings and the STW Thesaurus for Economics, with current open data.
31,128 MeSH descriptors and tree categories, and samples of MEDLINE citations indexed with them. Since April 2022, the U.S. National Library of Medicine indexes MEDLINE automatically, and reviews part of the results; each citation records whether it was indexed by people, automatically, or automatically and then reviewed.
6,697 descriptors and thesaurus categories of the STW, and samples of the records of ECONIS, the catalogue of the ZBW – Leibniz Information Centre for Economics, from two periods of publication.
Loading the hierarchy …
What the map shows
Indexing against the structure. As for the intellectual indexing of MEDLINE in Eckert (2012, Figs. 3.9 and 3.10), a few branches are used far more often than their size lets one expect. Most of them hold check tags, descriptors that are considered for every citation: the age groups (Adult, Middle Aged, Aged) make Named Groups [M] the reddest category; Humans (80 % of the citations) and Animals lie in Organisms [B]; Female and Male have no tree number at all. The chemicals and drugs [D], a third of the thesaurus, are used about as often as their size suggests.
Automated against people, 2021. In 2021, NLM already indexed part of MEDLINE automatically. The two sets therefore differ first in what was indexed: environmental science went to the automated indexing (Water Pollutants, Chemical, Soil Pollutants, Microplastics), experimental biology to people. Within this division, some descriptors are assigned by people and hardly ever automatically, such as Tumor Cells, Cultured (111 citations against none), Dose-Response Relationship, Drug or Structure-Activity Relationship. People assigned 12.4 descriptors per citation, the automated indexing of 2021 7.0, that of 2025 10.6.
Reviewed against automated, 2025. Part of the automated indexing is reviewed by people. The reviewed citations carry descriptors that the automated indexing alone rarely assigns: the type of study (Review Literature as Topic, Systematic Reviews as Topic, Double-Blind Method) and the population studied (Population Groups, e.g. Asian People, Black or African American, Hispanic or Latino). The reviewed citations are not a random selection: laboratory science (Mice, Inbred C57BL, Limit of Detection) is blue because it is rarely among them.
Data
- Hierarchy
- MeSH 2026 (descriptors), nested by their tree numbers; the first tree number of a descriptor gives its place in the map, the others are listed under “Also under”. The 16 tree categories form the top level.
- 2021
- a sample of the MEDLINE citations that entered PubMed in 2021, 150 per day: 12,963 indexed by people (IndexingMethod “Manual”), 34,909 automatically (“Automated”)
- 2025
- a sample of 2025, 80 per day: 22,538 indexed automatically
- Reviewed
- 5,662 citations of the 2025 sample (“Curated”: indexed automatically, then reviewed by people)
- Counting
- every citation once per area, however many of its descriptors lie below; qualifiers are not counted
What the map shows
Indexing against the structure. As in the evaluation of intellectual indexing with the STW by Eckert (2012, Figs. 3.6–3.8), the usage of the thesaurus departs from its structure in characteristic ways. The geographic names (G) are given to almost half of the records, and the general descriptors (A) to a fifth, mostly Theory, which marks theoretical works. The commodities (P), a large part of the thesaurus, are rarely used: paper, wood products and clothing not at all in the sample.
2023–2025 against 2008–2010. The comparison of the two periods shows how the literature, and with it the indexing, has moved. Red are descriptors that were hardly or never used in 2008–2010: Coronavirus, Digitization, Machine learning, Virtual currency, Blockchain and the Sustainable Development Goals. Blue are topics of the earlier period, such as the subprime financial crisis, RFID or East Germany. Records of the later period also carry more descriptors (5.7 against 4.7).
Data
- Hierarchy
- STW Thesaurus for Economics 9.20 without deprecated concepts: the thesaurus categories above, the descriptors below them. A descriptor is placed under a broader descriptor of the same category where there is one, else under its first category; further broader concepts are listed under “Also under”.
- ECONIS 2008–2010
- 15,896 of 20,000 records published 2008–2010 have subjects that match STW descriptors; 4.7 descriptors per record
- ECONIS 2023–2025
- 17,385 of 20,000 records published 2023–2025; 5.7 descriptors per record
- Sampling
- 20 pages of 1,000 records at evenly spaced positions of the result list of the EconBiz API. The subjects of a record (labels, German and English, mixed with free keywords) are matched to the preferred and alternative labels of the descriptors, ignoring case.
- Counting
- every record once per area, however many of its descriptors lie below
How to read the map
The size of an area is determined by the structure alone: it is the number of concepts in its subtree. The colour shows the information content (IC) of a concept or the difference between two ICs, computed as in Eckert (2012, Eqs. 2.4–2.8). A concept counts all documents indexed with it or with one of its narrower concepts, so deviations deep in the hierarchy also show in the areas above them.
- Structure (IIC): the intrinsic information content, derived from the hierarchy alone. The fewer concepts below a concept, the higher its IIC and the paler its area. This is what the map would show if the documents were spread over the hierarchy according to its structure.
- IC − IIC: the information content in a set of indexed documents minus the intrinsic one. Red: the concept is used more often than the structure lets one expect; blue: less often.
- IC − IC: two sets of indexed documents compared. Red: the concept is used more often in the first set than in the second, the reference.
Double-click an area to zoom in; the path above the map, Escape or Backspace lead back out. Click an area or a line of the hierarchy beside the map to see its numbers. Contrast narrows the colour range, so smaller differences show; balance shifts the neutral point.
Citation
If you use the ICE-Map Visualization or this implementation, please cite:
Eckert, K. (2012). Usage-driven Maintenance of Knowledge Organization Systems (Doctoral dissertation). Universität Mannheim. https://madoc.bib.uni-mannheim.de/31754/
The ICE-Map Visualization was first presented at the 4th International Conference on Knowledge Capture (K-CAP 2007), where the paper received the best paper award. Further publications are listed in the repository.
References
- Eckert, K. (2012). Usage-driven Maintenance of Knowledge Organization Systems (Doctoral dissertation). Universität Mannheim. madoc.bib.uni-mannheim.de/31754
- Eckert, K., Stuckenschmidt, H., & Pfeffer, M. (2007). Interactive thesaurus assessment for automatic document annotation. In Proceedings of the 4th International Conference on Knowledge Capture (K-CAP 2007) (pp. 103–110). ACM. doi:10.1145/1298406.1298426
- U.S. National Library of Medicine. Frequently asked questions about indexing for MEDLINE. nlm.nih.gov/bsd/indexfaq.html