Names are not good enough: Reasoning over taxonomic change in the Andropogon complex¹

5.1.Taxonomic particulars

The history of the Andro-UC is reviewed in [35,93,94]. The 11 input classifications T₁, …, T₁₁ are each reproduced according to the source publications (Fig. 2). All input articulations were provided by the user in tabular format (Fig. 3), which readily facilitates translation into RCC-5 relations. Strictly speaking, the Andro-UC concerns the “A. virginicus-A. glomeratus complex” as circumscribed in [94]. The use case is thus much narrower in scope than the entire genus-level concept Andropogon sec. Clayton et al. (2013) [19], which includes more than 100 species-level concepts worldwide.

Fig. 3.

Hierarchical, multi-level representations of the 11 input classifications of the Andro-UC (see also Appendix A). Taxonomic name and concept identities (numbered from 1–100) as in Fig. 2.

The classifications of the Andro-UC include, in chronological sequence (Figs 2 and 3): Hackel (1889) [49], Small (1933) [81], Blomquist (1948) [7], Hitchcock & Chase (1950) [50], Radford et al. (1968) [71], abbreviated as “RAB (1968)”, Godfrey & Wooten (1979) [45], Campbell (1983) [11], Campbell (2003) [12], Weakley (2006) [93], Kartesz (2014) [56], referred to as “BONAP (2014)”, and Weakley (2015) [94].

The tabular representation of Fig. 2 encodes taxonomic congruence as a function of occupying the same row (width). For instance, A. capillipes var. capillipes sec. Weakley (2015) (concept 90) == A. virginicus var. glaucus subvar. glaucus sec. Hackel (1889) (concept 4). Hence articulations between concept pairs pertaining to different classifications can be ‘read off’ the table by tracing their cell positions and vertical extensions – encoded with unique shadings – across the respective columns. The consistent column depth across all classifications indicates that the taxonomic boundaries of the ‘complex’ are congruent throughout. This means that all taxonomic incongruences of the Andro-UC are due to differential subdivisions of entities recognized by various authors within jointly agreed-upon outer taxonomic confines.

Another noteworthy aspect of the input representation are higher-ranked entities (compare Figs 2 and 3). These entities are not depicted in Fig. 2, because the table provided by Weakley emphasizes congruence among the narrowest concepts recognized in each classification. However, these higher-level entities are implied by conventions that guide the source taxonomies, and are usually made explicit therein. For instance, the acceptance of two variety-level concepts A. glomeratus var. hirsutior sec. Weakley (2006) (concept 76) and A. glomeratus var. glomeratus sec. Weakley (2006) (concept 77) (Fig. 2) implies recognition of the species-level concept A. glomeratus sec. Weakley (2006) (concept 75) (Fig. 3).

Our representations fully account for the implied higher-level taxonomic concepts, yielding comprehensive alignments with up to four levels (Fig. 3). Where necessary, we have added nominal (type) taxonomic names and concepts to represent comparable ranked entities at all levels; e.g., Andropogon virginicus var. viridis sec. Hackel (1889) was added (concept 6) and is comparable to Andropogon virginicus var. glaucus sec. Hackel (1889) (concept 3) of the same rank and source classification.

5.2.Input configuration, workflow execution, and reproducibility

The Euler/X toolkit is open source and available at [61]. The software can be cloned and then deployed on a desktop using the command-line interface. An overview of the toolkit’s reasoning and visualization options is available through the “help” command. Additional software dependencies include Python, the Answer Set Programming reasoners DLV [25] and Potassco (Gringo, claspD) [39], and GraphViz [38].

The input conventions for labeling concepts and representing parent/child (is_a) relationships and articulations are in accordance with [32–34,36]. They are exemplified in Fig. 1A for the 1948/1933 alignment. We limit our study to showing pairwise taxonomy alignments (see also Discussion), and therefore show outcomes for the following ten input configurations (Figs 4–6): 1933/1889 (Fig. 4A), 1948/1933 (Fig. 4B), 1950/1948 (Fig. 4C), 1968/1950 (Fig. 4D), 1979/1968 (Fig. 4E), 1983/1979 (Fig. 5B), 2003/1983 (Fig. 5C), 2006/2003 (Fig. 5D), 2014/2006 (Fig. 6A), and 2015/2014 (Fig. 6B). This strict chronological sequence is supplemented with two alignments; i.e., (1) 1979/1950 (Fig. 5A), which overcomes the gap in resolution generated by the intermediate, coarse RAB (1968) classification that contains only one species-level concept (Fig. 3); and (2) 2015/1889 (Fig. 6C), representing the largest possible time interval.

Fig. 4.

Visualizations for alignments 1–5 of the Andro-UC, 1889–1979. Representation conventions and annotations as in Fig. 1C. (A) Small (1933)/Hackel 1889 alignment; (B) Blomquist (1948)/Small (1933) alignment; (C) Hitchcock & Chase (1950)/Blomquist (1948) alignment; (D) RAB (1968)/Hitchcock & Chase (1950) alignment; (E) Godfrey & Wooten (1979)/RAB (1968).

Fig. 5.

Visualizations for alignments 6–9 of the Andro-UC, 1950–2006. Representation conventions and annotations as in Fig. 1C. (A) Godfrey & Wooten (1979)/Hitchcock & Chase (1950); (B) Campbell (1983)/Godfrey & Wooten (1979); (C) Campbell (2003)/Campbell (1983); (D) Weakley (2006)/Campbell (2003).

Fig. 6.

Visualizations for alignments 10–12 of the Andro-UC, 1889–2015. Representation conventions and annotations as in Fig. 1C. (A) BONAP (2014)/Weakley (2006); (B) Weakley (2015)/BONAP (2014); (C) Weakley (2015)/Hackel (1889).

In configuring the pairwise alignments, we represent the later (younger) taxonomy as T₂ and the earlier (older) taxonomy as T₁ [33]. Accordingly, the visualizations (Figs 1, 4–6) show concepts unique to T₂ as green rectangles, and concepts unique to T₁ as yellow octagons. Aligned regions with multiple congruent concepts are shown as grey rectangles with rounded corners (Fig. 1C). We use the shorthand of [36] for taxonomic concept labels, where (e.g.) Andropogon virginicus var. decipiens sec. Weakley (2015) becomes “2015.A_virginicus_var_decipiens”. The 12 input files (.txt format) for the Andro-UC are provided in Appendix A.

All alignments were obtained using “polynomial encoding/possible world/reduced containment graph” commands, which show overlapping articulations among input concepts as blue dashed lines in the output visualizations [14,16,61]. The commands generate the set of output MIR (.csv format) and GraphViz-rendered alignment visualizations (.pdf format).

The sets of Maximally Informative Relations (MIR) for each of the 12 alignments are provided in Appendix B. To ensure complete reproducibility, we have also prepared the Andro-UC use case as an experiment at http://recomputation.org/ [41,42].

7.1.Extent and origins of taxonomic incongruence

Each of the 12 input configurations yields a single, consistent, and unambiguously resolved alignment (Figs 4–6). The 12 visualizations clearly illustrate that none of the paired input taxonomies are entirely congruent, instead showing 2–12 unique regions (compare Figs 5B and 5C), and an overall ratio of 56 congruent to 71 non-congruent regions (Table 1). While we cannot examine each alignment in fine detail, we highlight select phenomena that capture the extent and causes of taxonomic incongruence in the Andro-UC. One cause for incongruence is unequal granularity across classifications. For instance, at the lowest taxonomic level, classifications authored from 1933 to 1979 recognize 1–5 concepts, whereas taxonomies published outside of this interval accept 7–9 concepts (Figs 2 and 3). Such differences cause the more finely resolving taxonomy to have one or more non-congruent (properly included) low-level concepts in comparison to its counterpart (e.g., Figs 4A and 5B). For instance, alignments of any taxonomy to that of the most coarse-grained RAB (1968) classification are only congruent with regards to the root-level concepts (Figs 4D and 4E), given that RAB (1968) recognize no additional taxonomic subdivisions within the complex. In the context of its immediate predecessor and successor (Figs 2, 4D, 4E, and 5A), the 1968 classification appears disruptive because the chain of taxonomic resolution between Hitchcock & Chase (1950) and Godfrey & Wooten (1979) is not propagated in RAB (1968).

Taxonomies produced in 1983 or later show higher levels of congruence between their finest-degree entities (Figs 5C, 5D, and 6). By and large, taxonomists publishing in the past 30 years have adopted Campbell’s (1983) perspective on how finely one should differentiate units within the complex. Incongruences among these recent perspectives are rooted mainly in disagreements on how to name and integrate low-level entities into parent concepts. Interestingly, Hackel (1889) already recognized seven low-level entities, and in that sense his classification is more congruent with contemporary perspectives (Fig. 6C) than with those published in 1933–1979.

In addition to unequal granularity, five alignments show overlapping (><) articulations, most frequently between species-level concepts. These relations are challenging to represent without recourse to RCC-5, because they fracture the hierarchical pattern of inverse/proper concept inclusion from tip- to root-level concepts [33–36]. The resulting alignments become directed acyclical graphs.

The 1950/1948 alignment represents an interesting case of overlap (Fig. 4C). Both Hitchcock & Chase (1950) and Blomquist (1948) recognize three identically named species-level concepts within in the complex, one of which is also taxonomically congruent (1950.A_capillipes == 1948.A_capillipes). The authors’ other two species-level concepts have four reciprocally overlapping articulations.

Of particular note is the articulation 1950.A_glomeratus >< 1948.A_glomeratus. Figure 2 illustrates that the overlap of these two concepts creates three alignment regions: (1) the region congruent with 1950.A_virginicus_ var_hirsutior, (2) the region congruent with 1948.A_virginicus_var_tenuispatheus, and (3) the region that actually constitutes the overlap. This latter region is not uniquely recognized in either input classification, and therefore has no input name assigned to it. Yet other authors, including Hackel (1889), have recognized and named the region (e.g., 1889.A_macrourus_var_abbreviatus; see Fig. 2).

Generalizing the phenomenon exemplified in the 1950/1948 alignment, we observe that overlap of two (or more) concepts creates combined concept regions for which there are no unique names in the respective input taxonomies. Nevertheless, identifiers for these alignment regions are required to express the extent to which the input concepts can be aligned with each other. The toolkit’s “combined concepts” command uniquely labels these regions (see [33]).

Overall, occurrences of differential resolution and overlapping concepts in the Andro-UC result in pairwise alignments with 5–15 regions (Table 1). Taking the 12 alignments in conjunction, 44.1% of the 127 inferred alignment regions are taxonomically congruent (range: 0.0–85.7%), leaving the remaining 55.9% incongruent. This ratio of in-/congruence between paired taxonomies is the semantic basis of the dis-/agreements that taxonomic names are suited to identify and track, though only up to a point, as we analyze in the next section.

7.2.Quantification of name:meaning dissociation

Taxonomic names are reliable identifiers of taxonomic in-/congruence for 77/127 (60.6%) of the regions present in the 12 pairwise alignments of the Andro-UC (range: 38.5–83.3%) (Table 1). The highest ratios are obtained for the 1968/1950 and 1979/1968 alignments. The latter include no congruent regions, since every unique name also symbolizes a unique alignment region (Figs 4D and 4E). The 5:13 ratio (38.5%) for the 2015/1889 alignment (Fig. 6C) is low as expected. In particular, 0/7 congruent concept regions in this 126 year-spanning alignment have reliable names; i.e., each of these regions is labeled by two non-identical names. However, taxonomic names in the Andro-UC do not necessarily perform better over short time intervals, or in alignments whose input taxonomies are closer to the present (2015). One example is the 2006/2003 alignment (Fig. 5D), which has an undesirable 6:9 ratio (40.0%) of reliable:unreliable names.

The 824 output MIR permit finer assessments of name:meaning dissociation (Table 3). Accordingly, among all 60 instances of pairwise taxonomic concept congruence (==) in the MIR, 33 also carry identical names (==:=), whereas 27 pairings have non-identical name strings (==:≠) (ratio 1.22:1). The majority (603/824=78.9%) of the remaining name: meaning relationships are in the |:≠ category. Indeed, any |:≠ combination would constitute erroneous nomenclatural practice, where two reciprocally exclusive concepts have identical names and therefore presumably refer to the same type, which in turn would mean that they are not mutually exclusive (unless homonymy is involved). No such errors are evident in the Andro-UC.

Among the remaining 161 non-congruent articulations (>, <, ><) in the set of output MIR, 16 refer to concept pairings with identical names (2>:=; 9<:=; and 5><:=), and 145 have non-identical names. These types of name:meaning categorization yield an overall ratio of 781:43 reliable:unreliable names for the entire Andro-UC. The highest occurrences of unreliable names appear in the 1933/1889 (5 MIR), 2006/2003 (7 MIR), and 2015/1889 (10 MIR) alignments.

Quantification of name:meaning cardinality over the 126-year period of the Andro-UC reveals that 18/36 taxonomic names (50.0%) have been used in multiple treatments, whereas the other 18 names are particular to single treatments (Table 4). Cumulatively, the use case entails 88 taxonomic name usages and 46 unique name:meaning combinations (ratio: 1.91:1). Only one name – A. virginicus – is used in every classification. Three additional names – i.e., A. capillipes, A. glomeratus, and A. virginicus var. virginicus – are used in 6–8 of the 11 input classifications. The other (32/36) names occur in less than half of them. Most, though not all, taxonomic names with 2–4 usages appear in temporally consecutive taxonomies (relation: 9/13).

The most frequently used name – A. virginicus – also has the highest number of incongruent taxonomic meanings, with a name:meaning cardinality of 1:6 (Table 4). Six consecutive classifications authored in 1933–1983 all propagate incongruent meanings of “A. virginicus” (Fig. 2). Only three additional names have more than one meaning in the Andro-UC; viz. A. capillipes (name:meaning cardinality = 1:2), A. glomeratus (1:4) and A. virginicus var. virginicus (1:2).

The 88 name usages in the Andro-UC correspond to 21 chains of taxonomically congruent concepts (Table 5). Of these, the chain symbolized by 2015.A_glaucopsis (most recent member) is the longest, with elements appearing in 9/11 classifications and under four non-identical names. Other long chains include 2015.A_virginicus (8 usages/4 non-identical names), 2014.A_capillipes (8/2), 2015.A_hirsutior (7/4), and 2015.A_tenuispatheus (7/4). At the other end of the spectrum, five concepts display globally unique meanings that whose meanings are unique to one classification (two authored in 1979; and one in 1968, 1950, and 1948, respectively).

At the other end of the length spectrum, there are five concepts whose meanings are unique to one classification (two authored in 1979; and one in 1968, 1950, and 1948, respectively).

The least favorable name:meaning cardinality among the 21 chains 4:1; meaning that four non-identical names are used to identify sets of taxonomically congruent concepts. This ratio applies to six concept chains: 2015.A_capillipes, 2015.A_dealbatus, 2015.A_virginicus, 2015.A_glaucopsis, 2015.A_hirsutior, and 2015.A_tenuispatheus. Conversely, a cardinality of 1:1 is obtained in 9/21 chains, of which only four have more than one usage (Table 5).

The information shown in Tables 4 and 5 provides an intuitive sense of how taxonomic names fare in the longer term as identifiers of taxonomy meanings in the Andro-UC. The performance of names should be evaluated in the context of taxonomic stability. High taxonomic stability would be reflected in an abundance of occupied cells in Table 5, because early-authored concepts would have congruent successors – with either identical or non-identical names – in the 1889–2015 time interval. This is not the case: only 85/231 cells (36.8%) have values, and 14/16 chains (87.5%) with multiple elements are ‘interrupted’.

In spite of persistent taxonomic meaning evolution, identifiers could nevertheless (in principle) be designed to achieve a name:meaning cardinality of 1:1. In that case taxonomic names would simultaneously show a score of 1 in the “Meanings” column of Table 4 and a score of 1 in the “Names” column of Table 5. Thirty-two names meet the former condition, and nine names meet the latter condition. However, the intersection of these two sets of includes only one name – A. virginicus var. abbreviatus. This name, used exclusively in Godfrey & Wooten (1979), is the only identifier that requires neither the “sec.” annotation nor an articulation with RCC-5 to reliably identify its associated meaning in the entire Andro-UC. The remaining 35/36 names are syntactically and/or semantically ‘compromised’, showing name:meaning cardinality relations other than 1:1. This outcome is also reflected in the accumulation of numbers greater than 1 across the columns of Tables 4 and 5.

In summary, even though names used in the Andro-UC act as identifiers of meanings with reliability ratios of 56.6% or higher in the local, pairwise alignments (Tables 1–3), their global reliability is such that >97.2% diverge from an ideal name:meaning cardinality of 1:1. This assessment remains adequate even if taxonomic change is taken into account.

8.1.New knowledge products

What aspects of our approach are new and valuable? The Andro-UC illustrates the unique ability of RCC-5 multi-taxonomic alignments to resolve taxonomic meaning evolution at more granular levels than is possible using taxonomic names and nomenclatural relationships [70]. This follows directly from the input information – we can only represent and align the entities shown in Fig. 1 if taxonomic concept labels and RCC-5 articulations are used. Critically, the approach requires an initial set of articulations provided by human (expert) users, and grounded in their assessments of pertinent taxonomic evidence, that satisfy criteria of consistency and (lack of) ambiguity to yield well-specified alignments. Compliance with these criteria is achieved by the interactive toolkit workflow [33].

New knowledge products for the Andro-UC include the output MIR, alignment visualizations, and name:meaning cardinality analyses. Through the reasoning process, the set of 88 user-provided input articulations is logically tested and augmented to yield 824 Maximally Informative Relations (1025 MIR if the root concepts are included). The MIR derived for each alignment can be queried to determine whether any concept pairs (and ancillary biological data) are suitable for integration, or not [36,43,53,85,86]. In particular, articulations of congruence (“yes, integrate”) and exclusion (“no”) between two concepts are reciprocally actionable in this context. Proper inclusion and inverse proper inclusion are least unilaterally actionable without ambiguity (“add data assigned to the less inclusive concept to those of the more inclusive one”). Overlap is the most challenging articulation for the purpose of merging ancillary information. However, in some instances overlap at higher levels in an alignment can be resolved into proper inclusion at lower levels. For instance, the 2014/2006 alignment (Fig. 6A) shows the articulation 2014.A_glomeratus >< 2006.A_glomeratus at the level directly following the root. At the next lower level, the overlap is resolved: both concepts share congruent subelements, and each also has additional subelements that are unique to it as a parent-level concept, but nevertheless congruent with another child-level concept in the alternative classification (and with non-identical names and ranks). Hence biological data identified at this lower level can be fully integrated across the two very disparate taxonomic perspectives. In summary, the reasoner-inferred MIR provide new, human- and machine-interpretable information that can drive the integration of biological data linked to different, aligned concept taxonomies with more precision and reliability than taxonomic names alone [6,29,33,36,43].

The alignment visualizations are logically congruent with the output MIR [14,16,23,87]. Their unique value lies in aiding human users to understand multi-concept relationships through tree-like representations. Visualization tools for multi-taxonomy relationships have advanced significantly over the past 20 years [3,5,16,46,47,98]. Nevertheless, the Euler/X toolkit is the first platform to leverage RCC-5 relationships and logic reasoning to yield comprehensive, tree-like alignment visualizations.

The visualizations communicate uniquely valuable information. For instance, Figs 1–3 all show information related to the 1948/1933 alignment. Figure 2 effectively visualizes the lowest-level entities and articulations of the entire Andro-UC, but is not well suited for input taxonomies nested into three or more ranks (or phylogenetic levels). Such tables are ‘flattened’ into two dimensions. Figure 3, in turn, can shows all nested entities for the individual 1948/1933 taxonomies, but does not provide accurate multi-concept alignment information. Using names to navigate across these trees may lead to erroneous conclusions such as 1948.A_virginicus | 1933.A_glomeratus, when the proper articulation is ><. Although they are commonly used to convey information about taxonomic stability and change [8,82,95], collections of unaligned tree graphs have limited resolution power in absence of RCC-5 alignment information.

In contrast, the alignment visualizations (Fig. 1) simultaneously communicate information about nomenclatural identity, multi-level tree hierarchy, and multi-tree in-/congruence. Their interpretation is intuitive; for instance, the proportion and position of grey squares versus green rectangles or yellow octagons communicate the extent and localization of taxonomic in-/congruence in an alignment (compare, e.g., Figs 5B and 5C). The relative occurrences of (=, ≠, +, −) annotations show the degree to which taxonomic names can reliably integrate taxonomic meanings.

8.2.Building better identifiers for biodiversity data

How relevant is our representation approach to the broader, semantics-facilitated biodiversity data realm? Multiple reviewers raised this important question. We think that it is too early to attempt a comprehensive answer. Technical, scientific, cognitive-evolutionary, and socio-cultural constraints affect how identifier granularity is managed in the biological domain. Predicting how the RCC-5 alignment approach will fare in light of these trade-offs is beyond the scope of this analysis. We can, however, assess the particulars and generalities of the Andro-UC, and what it teaches us about identifying and linking taxonomically identified information in open-ended biodiversity data environments.

The scale of the Andro-UC is small. Weakley’s (2015) classification recognizes seven species-level concepts – more than any other taxonomy. The taxonomic history is evidently complex, but no more complicated than that of many other continuously revised groups [1,20,32,35,43,68,77,90]. The poor performance of names as identifiers of divergent taxonomic meanings is not exceptional for the field (herein broadly defined to include phylogenetics).

The problem of name:meaning dissociation in biological taxonomy is systemic. It is rooted in Code-mandated principles that promote stability and change in naming (largely) as a function of nomenclatural type identity and priority. To some degree the inadequacies are manageable through social processes, including conservative re-/naming practices or ‘standardized’ taxonomies [6,44,55,79,91]. In practice, the long-term drawbacks of using taxonomic names as concept identifiers are frequently mitigated by the ability of well-trained human scientists to contextualize name usages and thereby infer the intended meanings [30,32,69]. However, no counteracting human practice can alter the insight that taxonomic names and nomenclatural relationships are fundamentally not designed to track granular similarities and differences in taxonomic meaning of the sort exemplified in the Andro-UC. Computer algorithms in particular struggle with inferring what “Andropogon virginicus” means ‘in the currently intended context’ [62], i.e., when the relevant context of the name usage is not made explicit. Something beyond the type-anchored name identity is needed if taxonomic perspectives are to be translated into entities fit for use in open-ended, semantics-enabled information environments.

Specifying the referential extension of taxonomic names for reliable reuse requires more than ostension (the act of pointing) to exemplars (types). Ostensive definitions of taxonomic meanings are bound to under-specify the intended meanings in many applied contexts, such as those of the Andro-UC. Instead it is more appropriate to model the name-to-(currently-perceived-)taxon linkage as a matter of theory construction [31,75]. The challenge of integrating biodiversity data then becomes one of aligning multiple taxonomic theories, which can be modeled with the RCC-5 approach.

The aforementioned insufficiencies are most apparent in cases of multi-concept overlap. Such cases are frequent in taxonomy, and they cannot be reduced to the differences in degree of resolution [26,65,68,77]. As an example, the 1950/1948 classifications of the Andro-UC concur that there are three identically named species-level concepts entailed in the complex (Fig. 4C). They also concur that 1950/1948.A_virginicus has three variety-level child concepts. However, they disagree on the extent to which the available, type-anchored names reach out to perceived, and necessarily more inclusive, taxa presumed (more precisely: theorized) to constitute natural, evolutionary entities [9]. As a consequence of this differential inference of ‘extra-typical’ taxonomic boundaries, the four 1950/1948 species-level concepts overlap in complex ways (Fig. 4C). Such multi-theory overlap is more frequent at higher taxonomic levels, where the performance of names as identifiers of taxonomic meanings becomes increasingly poor [33,35,36].

The herein demonstrated alignment approach paves the way for building better taxonomic concept identifiers and multi-taxonomy resolution services.

8.3.Scalability of the RCC-5 alignment approach

How widely applicable (or scalable) is the RCC-5 multi-taxonomy alignment approach within the field of biological taxonomy? Generally speaking, reasoning about taxonomies with RCC-5 remains in its infancy [16,33,36,43,87]. At present, the Euler/X toolkit can effectively process consistent, well-specified, pairwise input taxonomies with up to 200–400 concepts each [14,16,61]. While this scale is sufficient for small- to medium-sized alignment use cases, future toolkit development should concentrate on modularizing the reasoning process, specifically by using a divide-and-conquer approach that better leverages the hierarchical structure of the input constraints and dynamic user/reasoner interactions. Demonstrating the practical value of the approach requires making the toolkit accessible to larger use cases and biodiversity data environment where taxonomy evolution is an important variable to identify and control.

The analysis of the Andro-UC demonstrates the potential of reasoning about taxonomies and at the same time leaves much room for further work. In particular, the 11 input classifications allow for 55 pairwise comparisons, of which only 12 are presented here. This omission is deliberate. New toolkit releases will have the ability to align more than two input taxonomies simultaneously (but remain in development). Such an approach entails new reasoning challenges and products. For instance, we could ask to what extent 12 alignments produced in the current study are sufficient for recovering the full set of 55 pairwise alignments, based on transitive reasoning. Solutions to such challenges are relevant to the issue of scalability, and can inform the users’ practice of engaging with the toolkit.

Pathways to broader implementation should focus on directly integrating the use of taxonomic concept labels, parent/child relationships, RCC-5 articulations, and reasoning and visualization services into prominent biodiversity data platforms [2,34,35,48,57,59]. We envision information environments where identifications of organismal occurrence records are augmented to the level of carrying taxonomic concept labels [53]. The circumscriptions of the respective concepts are also managed in the platform, and consistent, well-specified RCC-5 alignments are provided. Building such an infrastructure would permit biologically significant queries of the following types. (1) Return all records identified to the name Andropogon virginicus (optionally, with synonyms or algorithmically matched alternative spellings). This query type corresponds to the current capability of many environments [10,70,73]. (2) Return all records identified to the taxonomic concept label Andropogon virginicus sec. Weakley (2015) and, alternatively, Andropogon virginicus sec. RAB (1968). Show the corresponding occurrence-based distribution maps (less inclusive concept [2015] – small set of records; more inclusive concept [1968] – large set of records). (3) Translate all occurrence records originally identified to Weakley (2015)-endorsed concept labels into the corresponding BONAP (2014)-endorsed concept labels (see Fig. 6B). (4) Highlight ‘problem records’ identifiable to multiple non-congruent concepts in the set of aligned classification standards used for identifications. (5) Show records in this target region as identified according to the most, or least, granular concept-level taxonomy. (6) For any set of records (and related biological data) identified to any pair of taxonomic concept labels, assess whether the records and data can be retrieved and integrated based on the reasoner-inferred MIR.

The above queries (2)–(6) are biologically significant and depend on utilizing the RCC-5 alignment approach to achieve the desired degree of resolution. Such logic-enabled integration services are urgently needed in our assessment to build open-ended biodiversity data environments that can manage the complexities of evolving taxonomic knowledge. Strengths of the RCC-5 approach in this context include explicitness, consistency, machine-interpretability, and flexibility in processing diverse forms of taxonomic concept input ranging from minimally structured lists of taxonomic concept labels to phylogenies and monographic revisions [8,33,35,36,53].

8.4.Non-taxonomic alignment challenges

The RCC-5 alignment approach has so far been limited to use cases in biological taxonomy. Explorations of the toolkit’s performance in relation to other integration challenges is generally recommendable if the new focal domain shares several of the toolkit’s critical (taxonomic) input/output constraints [87]. This means that other semantic integration challenges that need to consistently align and visualize multiple, hierarchically structured sets of concepts with coverage and/or disjoint siblings constraints may benefit from exploring the RCC-5 alignment approach.

Our approach can be complemented by Semantic Web methods that reason over concept similarity and drift by leveraging Natural Language Processing techniques and relationships defined in OWL-DL ontologies [13,22,27,37,73,80,92]. Such complementary analyses of concept identity and semantic evolution are now possible.