Data to be produced and managed for the project include: (1a) Software code written for the Symbiota content management system (primarily written in PHP and with heavy use of JavaScript libraries; and connecting to the open source MariaDB SQL database platform) and (1b) for the Euler/X logic reasoning toolkit (primarily written in Python); (2) specimen occurrence records (with new identifications) managed in the Symbiota-operated SERNEC herbarium portal, and formatted in compliance (where possible; see details below) with the Taxonomic Working Group (TDWG) -endorsed Darwin Core (DwC) and Taxonomic Concept Transfer Schema (TCS) standards (https://github.com/tdwg); and (3) Euler/X toolkit input/output files, presently stored in simple .csv, .gv (GraphViz), .pdf, .txt, and .yaml file formats. We will also (4) author web posts (.html) and instructional videos (.mp4) (see Broader Impacts).

The Symbiota-based SERNEC portal occurrence data are fully Darwin Core-compatible. These data can be bundled through easy-to-use platform functions to yield Darwin Core Archive files for wider sharing. We note, however, that Darwin Core does not presently support all syntactic and semantic conventions of the taxonomic concept approach. In particular, a modularized and flexible management of taxonomic concept labels (TCLs) in conjunction with parent/child relationships and RCC–5 articulations – in some instances under multiple extensional or intensional readings (Section 8.II.1) – is out of scope for DwC. Certain aspects are covered by the TCS. However, this 2005-ratified standard needs revision and expansion, particularly in connection with a fully functional specimen data environment such as Symbiota.

We will adhere to DwC and TCS as much as is conducive to our representation needs. At the same time, this part of the project (Section 8.3.I: taxonomy/occurrence module expansion) is properly viewed as new work required for updating and expanding the TCS ("2.0"). Other services (e.g., GBIF, iDigBio) that 'just' manage DwC syntax and semantics, while not incompatible with our data, will nevertheless be unable to replicate our TCL-based specimen resolution services that critically require RCC–5 integration signals. As a stop-gap solution, we will provide links to alignments on GitHub and/or in DataOne in the "dynamicProperties" field.

At present Euler/X input and output data formats, including the input constraint .txt files and resulting .csv MIR files, are not covered by ratified standards (TDWG or other entities). However, both are ASCII-based, largely translatable into TCS terms and relationships, and easily manageable through standard control version systems (such as Git) that can automatically visualize version differences. The scale of this project – 2,000-3,000 alignments – presents an opportunity to create more formalized input and output data standards. The UIUC team will develop a simple alignment archive format (.aarc). We will also generate an associated and self-contained viewer tool to make taxonomy alignment products (i.e., input, output, and inference rules used to logically connect these products) transparent and reproducible.

Our project operates fully in the Public Domain. The Symbiota software code is published under the GNU General Public License (Version 3, June 2007), whereas the Euler/X code is published with the BSD license (also used by the Open Tree of Life project). All Symbiota-/SERNEC-held data and the new Euler/X alignments are published under the CC0 license (or similar, given certain collections records and image artefacts; see https://www.idigbio.org/content/idigbio-intellectual-property-policy; http://choosealicense.com/licenses/). UIUC's Ludäscher is a member of the DataONE Leadership Team and will work with colleagues in the DataONE Semantics and Provenance Working Group to explore sharing taxonomically (TCL) annotated datasets through DataONE.

Collection- and use case-based data will be published as Darwin Core Archive files. To disseminate DwC–A packages, we will use well-established and separate publication pathways from Symbiota to GBIF (http://www.gbif.org/dataset/) and iDigBio (https://www.idigbio.org/portal/publishers), as preferred by these aggregators. The transformed SERNEC portal will also publish our datasets, as DwC-A files and additionally using the expanded schema (syntax, semantics) for multi-TCL-to-specimen resolution that we will generate. This ensures that our use case results remain accessible and reproducible. Specific data packages authored in relation to the use case publications will be disseminated via means sanctioned by open access (option) journals, using repositories such as Dryad (http://datadryad.org/), figshare (https://figshare.com/), and Zenodo (http://zenodo.org/).

New software code will be published as releases through GitHub or similar openly accessible source code repositories (e.g., http://gitlab.com). SERNEC portal and use case data will be archived through redundant back-ups at ASU, in addition to GBIF and iDigBio. Data persistence will be further assured by establishing a new archival service relationship with DataONE, facilitated by Ludäscher, and specifically through addition of our Project data to the DataONE member node Knowledge Network for Biocomplexity (KNB) Data Respository (https://search.dataone.org/#profile/KNB).

ASU (Franz, Gilbert) assume primary responsibility for project-based managing of data for Symbiota, SERNEC, and the Euler/X alignment repository on GitHub (https://github.com/taxonomic-concept-alignments). All Symbiota code (https://github.com/Symbiota) and contingent software for portal operation are open source. For select code testing purposes, ASU maintains an experimental portal on an institutionally supported VM server (http://hasbrouck.asu.edu/sandbox/). However, all actual SERNEC data are hosted only and directly by the NSF-supported iDigBio infrastructure, which has dedicated Symbiota data servers for multiple hosted data portals. We commit to iDigBio's rules for collaboration, particularly with regards to creating and resolving globally unique specimen identifiers; see https://www.idigbio.org/content/collaborating-idigbio-grant-proposals. UIUC (Ludäscher) is responsible for maintaining the new Euler/X code on GitHub (https://github.com/EulerProject/).

To motivate a complex theme – names, taxa, and concepts – we start with a concrete example. The species epithets "bifaria" (coined by Fernald 1946a, Fernald 1946b) and "divaricata" (Linnaeus 1753) are very good identifiers for some purposes, and poor ones for others (Witteveen 2015, Franz and Sterner 2015, Franz et al. 2016). Each name is rigidly assigned to an individual type specimen (LINN-HL 1059-3; GH00056705), providing "an objective link between the real world of organisms and the world of language (nomenclature)" (Dubois 2005: 381-382). Moreover, in a particular floristic treatment, bifaria and divaricata function as identifiers for endangered orchid species that occur in the southeastern Unites States (Fernald 1950, Radford et al. 1968, Wunderlin and Hansen 2011, Weakley 2015). Depending on the treatment, these epithets may be combined with one of four generic names (Arethusa, Cleistes, Cleistesiopsis, or Pogonia), a varying set of nomenclatural synonyms, a diagnosis or description, and other taxonomic and biological information, such as links to additional specimens or DNA data.

But here we should pause. The phrase "identifiers for species" could imply that we have converged on stable and accurate circumscriptions of two orchid species. It could even imply that we had 'gotten them right' since Linnaeus (1758). Hence the two epithets could reliably stand for the species (Ogden and Richards 1923, Peirce 1998, Franz and Sterner 2015). This turns out to be false. Indeed, with the creation of the epithet bifaria in 1946, a taxonomic subset of what previous authors (e.g., Ames 1922) had placed under divaricata was differentiated from that preceding, more inclusive set (Fig. 1).

Figure 1.  

Taxonomic concept labels and concept-to-concept articulations, represented in a tabular alignment of nine schemata, for the Cleistes use case (sec. A.S. Weakley). The vertical column position and width of taxonomic concept labels indicates taxonomic non-/congruence. The colors approximate taxonomic name lineages, e.g. blue for bifaria and yellow for divaricata.

Until 1946, divaricata had a wide taxonomic referent (= entity for which the name stands), whereas subsequently divaricata started to also stand for a narrower referent. Following Pansarin and Brown (2009), bifaria received yet another less inclusive meaning (Weakley 2015).

If names are potentially ambiguous, then how should we model the evolving relationships between identifiers, meanings, and natural entities? We propose the following definitions (Franz and Sterner 2015). Genus- and species-level names are anchored by rigidly chosen types – either type names or type specimens (Franz et al. 2008, Witteveen 2015, Witteveen 2016). At any given time, taxonomic names (= symbols) also stand for taxa, which we define here as historically, evolutionarily coherent sets of organisms in nature. However, history shows that we need a third element in our model, i.e., taxonomic concepts. We define a taxonomic concept as an empirically informed theory of the identity and definitional boundaries of the perceived taxon, expressed by a particular author at a particular time (Berendsohn 1995, Geoffroy and Berendsohn 2003, Franz and Peet 2009). "Name usage" or "taxonomic circumscription/ meaning" (= reference; see the semiotic triangle of Ogden and Richards 1923) are common synonyms of "taxonomic concept". We say "perceived taxon" because we represent the interaction between taxonomic concepts – the theories proposed by human authors – and natural taxa as one of successive approximation. Instead of saying "Linnaeus (1758) lumped the species (taxon) divaricata", we say "the taxonomic concept divaricata sec. (secundum, according to) Linnaeus (1758) is more inclusive than the taxonomic concept divaricata sec. Pansarin and Brown (2009)". In other words, we are realists about the evolutionary identities of taxa (which are unlikely to have changed much in the past 258 years), but conceptualists about human taxonomic making (which progresses rapidly). Eventually our scientific process is expected to 'get there', within epistemic limits.

It follows that taxonomic names have three roles in our data systems (Franz and Sterner 2015): (1) they identify types, (2) they pragmatically stand for taxa, but with the possibility of imprecision or error, and (3) they identify lineages of one to (very) many taxonomic concepts proposed in the history of human taxonomic making.

The third role is critical for querying non-type specimens. Fig. 1​ shows that the meanings of bifaria and divaricata are not stable, instead becoming less inclusive over time. Thus, if a user queries an aggregated data system for specimens identified to Cleistes bifaria, the system's immediate response should be: "According to any source, or only according to a particular one? Here are your options [showing five or more schemata] – please specify further." And, depending on the user's query refinement, they would receive distinctly labeled and partially overlapping sets of specimens in return. These sets (if georeferenced) would present different signals on visualized distribution maps (Fig. 2).

Figure 2.  

"Where do these endangered orchid species occur?" – visualizing the taxonomic variable for aggregated herbarium data. Mappings for the same 250 SERNEC specimens (not all resolved at this geographic scale) according to four distinct taxonomies. (A) sec. SERNEC (2016) 'consensus', (B) sec. Radford et al. (1968), (C) sec. Kartesz (2010), (D) sec. Weakley (2015). In A, C, and D, unequal sets of specimens labeled as bifaria are red; those labeled as divaricata are blue. In B, all specimens are identified as divaricata, and hence a query for bifaria would not return any specimens.

Likely, one or another specification would lead the user to make distinct biological inferences based on these derivative maps. This is how the user can assess the robustness of their hypotheses vis-à-vis the taxonomic variable.

What we describe is hard to do (Koperski et al. 2000, Boyle et al. 2013, Lepage et al. 2014, Franz et al. 2016a, Franz et al. 2016b). Few if any biodiversity data providers excel at resolving specimen information so granularly. The vast majority are name-based, and not yet concept-based. Creating a "which concept?" query-/counter-query service that reaches to the herbarium specimen level is our overarching objective.

To begin building a solution, we need a new term for the identifier "Cleistesiopsis divaricata [name author, year] sec. Pansarin and Brown (2009)". We call these taxonomic concept labels (TCLs) (Franz et al. 2015, Jansen and Franz 2015, Franz et al. 2016a). The cardinality relationship between TCLs and taxonomic concepts is one-to-one (Remsen 2016), i.e., they select exactly one taxonomic concept out of a system that may represent many potentially conflicting concepts. We say that the taxonomic name participates in the TCL. But the name, by itself, is not suited to pick out a particular taxonomic concept, instead selecting – via nomenclatural (type) identity – the corresponding lineage of concepts (Franz and Sterner 2015). Often the cardinality relationship between taxonomic names and taxonomic concepts is one-to-many, where the "many" will differ in terms of their meanings (Geoffroy and Berendsohn 2003, Kennedy et al. 2005, Remsen 2016). In that sense bifaria is a poor identifier: it identifies a lineage of non-congruent meanings. We can also have the inverse situation, where multiple nomenclatural synonyms have a congruent taxonomic reference (Remsen 2016). However, we model these data not as 'one and the same concept', but instead as multiple taxonomic concepts – each with a unique TCL – that happen to have congruent meanings (Franz and Peet 2009). So, we would not say "Cleistes divaricata var. bifaria sec. Fernald (1950) and Cleistesiopsis bifaria sec. Kartesz (2010) is the 'same' concept". Instead we manage two TCLs whose meanings are taxonomically congruent. The two names involved also happen to have a nomenclatural synonymy relationship, at least according to recent treatments. However, it is not deductively sound to equate synonymy with congruence in our sense, because the former is type-based or may only be "pro parte" (Geoffroy and Berendsohn 2003, Franz et al. 2016a, Franz et al. 2016b). Arethusa divaricata is a synonym of Cleistesiopsis divaricata (FNA 2015), whereas Arethusa divaricata sec. Linnaeus (1758) is congruent with the sum of three species-level concepts sec. Weakley (2015), including Cleistesiopsis divaricata sec. (Fig. 1).

Thus, in addition to modeling TCLs, we need a new language to express concept-to-concept relations (Koperski et al. 2000, Geoffroy and Berendsohn 2003, Franz and Peet 2009). These relationships are of two kinds: (1) between concepts authored in a single treatment, or (2) between concepts authored in multiple treatments. The former are parent/child relationships. Example:

• Cleistesiopsis sec. Weakley (2015) is a parent of Cleistesiopsis oricamporum sec. Weakley (2015)

Such parent/child relationships are explicit in the hierarchy asserted by the particular treatment. And the latter, between-hierarchies relationships are RCC–5 articulations, where "RCC" stands for Region Connection Calculus (Randell et al. 1992, Thau and Ludäscher 2007, Thau et al. 2008). RCC–5 is a generic language for expressing to what extent two regions overlap. The five types of pairwise articulations are: congruence (A == B), proper inclusion (A > B), inverse proper inclusion (A < B), overlap (A >< B), exclusion (A {| or !} B). This allows us to say (compare with Fig. 1).

• Cleistesiopsis bifaria sec. Weakley (2015) < Cleistesiopsis bifaria sec. Kartesz (2010)
• Cleistesiopsis divaricata sec. Weakley (2015) ! Cleistes divaricata var. bifaria sec. Fernald (1950)
• Cleistesiopsis oricamporum sec. Weakley (2015) == Cleistes bifaria "Coastal populations" sec. Smith et al. (2004)

Armed with the new syntax (TCLs) and semantics (parent/child relationships and RCC–5 articulations), we are much closer to responding to the counter-query "Please specify your preferred name usage".

In Fig. 2 we approximate the impact of representing and linking taxonomic concepts in an aggregated system. The specimen data are from the SouthEast Regional Network of Expertise and Collections portal (SERNEC 2016), as of August, 2016. SERNEC is supported by Symbiota (Gries et al. 2014) and is the focus of our proposed work. The 250 records were obtained by querying the portal for "Cleistes, Cleistesiopsis". Georeferencing was added to ensure that all specimens are Google-mapped. The dataset was then imported four times independently into a local Symbiota sandbox. One of the four imports (2A) was left unchanged; with specimen identifications as currently represented "sec. SERNEC (2016)". For each of the other imports, specimens were re-identified according to one of the following three floristic treatments: (2B) Radford et al. (1968), (2C) Kartesz (2010), and (2D) Weakley (2015). For the latter two treatments, this step was not done very thoroughly yet, because we are visualizing an argument, not a final product. We estimate 5-10% error of identifications, especially along eastern coastal regions. We then specified queries in such a way that all specimens will be mapped according to the most granular TCL(s) present in each import.

We need to be cautious in interpreting these 'mostly real' data visualizations that make SERNEC (2A) look dismal. The 250 specimens are housed in 33 different herbaria. They were vouchered over the period of 1869 to 2011, which likely means that they were variously re-/identified using any/all relevant treatments starting with Brown (1813) (Fig. 1). Many identifications are evidently not recent, because only four specimens are identified to the genus-level name Cleistesiopsis. This name was not coined until Pansarin and Barros (2008), Pansarin and Brown (2009), and in these treatments Cleistes was initially placed into synonymy of Cleistesiopsis.

But this proposal is as much about the design of aggregating systems (trust) as it is about promoting more, and more accurate, identifications (quality). If we look at the SERNEC (2016) Taxonomic Thesaurus in Fig. 3, we note that its syntax and semantics are systemically misdesigned to represent the taxonomic concept label and articulation information shown Fig. 1.

Figure 3.  

SERNEC (2016) Taxonomic Thesaurus for the genus-level name Cleistesiopsis, as of August 2016. [] = synonym.

In particular, the system will not permit users to submit specimen queries in accordance with a particular taxonomic perspective, other than 'the portal consensus'. And we note that 'the consensus' is actually an evolving body of data, yet without adequate version tracking through time (Scoble 2004, Berendsohn and Geoffroy 2007, Cheney et al. 2007, Franz and Thau 2010, Midford et al. 2013. At some point, someone with access to the Thesaurus may add "Cleistesiopsis oricamporum". There will be no special flag, however, that we now have a new 'consensus' version. No ability to switch back and forth between versions, or to show the differential. This is commonplace (Page and Valiente 2005, Leonelli 2013, O’Malley 2013, Winsberg et al. 2014, Hinchliff et al. 2015): often when we hear terms like consensus or synthesis in the context of aggregation, what we actually have is a poorly sourced and connected set of unitary, taxonomically incongruent classification snapshots. Each snapshot fails to represent the actual diversity of taxonomic perspectives in use at the time, and obscures the localized practice of identifying herbarium specimens to particular (preferred) sets of taxonomic concepts.

What needs to change? The prevailing name-based designs of aggregating systems improperly conflate two semi-independent processes. One might say with reasonable accuracy that, given a particular taxonomic perspective, the application of valid names and nomenclatural relationships is an undemocratic, logically contingent process. However, adherence to this or that perspective is democratic. At present, herbaria networked in SERNEC (2016) may follow one or more of five variously authored and endorsed perspectives that relate incongruently to the orchid specimens in question (Figs 1, 2). Because the business of aligning our concepts with taxa is not finished, a consensus perspective should not be 'dictated by design' (Fig. 3). Instead, the system should explicitly model the alternative views, resolve incongruences as much as possible, and allow users to empirically assess the impact of varied outcomes.

This proposal has conceptual, technical, social, and hence trust-related implications for biodiversity data science. The difference between the four visualizations (Fig. 2) is not just that (2A) resolves semantically ambiguous taxonomic concept lineages, whereas (2B–2D) resolve unique sets of TCLs (with increasing granularity). Surely it matters in some hypothesis-testing contexts whether we apply one perspective or another (Flanagan et al. 2006). A more fundamental difference is that, qua the process of data aggregation, the individual, time-stamped, consenting or dissenting 'voice' of the taxonomic expert(s) has been designed away in (2A). The aggregation design has disenfranchised the expert authors, whose unique views are no longer accessible to configure specimen queries. But SERNEC (2016) herbaria use at least five incongruent taxonomic schemata to identify the specimens in question! In contrast, noone in particular owns the 'consensus', which may not align with any treatment used in a herbarium to perform identifications. This is not just a matter of data quality, i.e., of updating all records once to 'the latest consensus'. Instead it is a matter of poor data system design that affects trust in aggregated data. The design must improve first to enable a better data culture.

While our development focuses on SERNEC 2016, a shift to concept-based specimen resolution potentially affects biodiversity informatics at a global scale. We believe that this field is, and should be, part of the big data movement that many sciences are experiencing. But "big data" is a malleable term, and challenges encountered while transitioning to the next levels of aggregation are specific to particular domains (Leonelli 2013, O’Malley 2013, Winsberg et al. 2014). Our challenge is this: taxonomic names and nomenclatural relationships are by design imprecise identifiers and connectors of conflicting/evolving taxonomic perspectives in aggregating systems. Most systems nevertheless legislate one view at a time, through algorithmic and social processes (Page and Valiente 2005, Jones et al. 2011, Döring 2013, Hinchliff et al. 2015, Döring 2016b). As a result, they create unitary syntheses that noone in particular owns or subscribes to (Figs 2, 3). But few biodiversity scientists would want a future where big data is equated with the disenfranchisement of individual taxonomic expertise, or where poor identifiers support only a subset of reliable inferences because of the way in which the data were (mis-)aggregated via so-called synthetic classifications.

The service envisioned in Fig. 2 can be built now for the SERNEC (2016) portal. Doing so is both a milestone for the taxonomic concept approach (Franz et al. 2008) and a powerful demonstration of its impact on specimen-based inferences. Why? – Because specimens are special. Many biologists understand the limitations of name-based aggregating systems 'in principle' (Kennedy et al. 2005, Patterson et al. 2016, Remsen 2016). But experiencing a herbarium portal solution developed from A to Z, where users are shown the effect of multiple incongruent taxonomic representations of their specimens on a map of their focal region, is not just novel but forceful. For many good reasons including reproducibility (Vink et al. 2012), we care greatly about specimen-based data. Once the options displayed in Figure 2 become reality, we predict that it will be hard to return to 'just names'.

Below we describe why SERNEC (2016) is ideally suited for this purpose. Thereafter we present our specific ABI Development objectives, implementation and management plans, and broader impacts.

A constellation of prior efforts in four different areas uniquely identifies SERNEC as the target for developing a concept-based system.

1. SERNEC is an active, NSF-supported Thematic Collections Network (TCN), with currently 60 member herbaria (goal: 200) and nearly 1.97 million herbarium records (goal: 3–4 million). More than 85% of SERNEC's specimens are from the Mid-Atlantic and Southeastern United States region. Led by Zack Murrell (Appalachian State University), the community is primed to shift to taxonomic concept resolution.
2. The SERNEC (2016) portal and underlying data are sustained by the open source Symbiota software platform (Gries et al. 2014), hosted by the Integrated Digitized Biocollections project (iDigBio; see https://www.idigbio.org). Symbiota is the go-to solution for the majority of NSF-supported TCNs (August, 2016: 11/17 TCNs; ~ 25 portals; > 750 collections; > 32 million records; > 9 million images; > 7,300 users including 2,875 active contributors). It is the most impactful mid-level aggregator for millions of distributed, Darwin Core-compliant records (Wieczorek et al. 2012) that nevertheless allows full data access and authorship by individual collections and expert contributors. Co-PI (PI = Principal Investigator) Edward Gilbert (Arizona State University; ASU) is the lead Symbiota developer, and is singularly qualified to transform Symbiota as part of his developing biodiversity informatics research program.
3. Co-PI Bertram Ludäscher (University of Illinois Urbana-Champaign; UIUC) and PI Franz (ASU) are co-leaders of the Euler/X toolkit (Chen et al. 2014b, Dang et al. 2015, Ludäscher et al. 2016) for achieving logically consistent multi-taxonomy alignments (Section 6). Application of this toolkit will allow us to provide RCC–5 articulations at scale. Lastly,
4. Co-PI Alan Weakley (University of North Carolina; UNC) is the author of the Weakley Flora (Weakley 2015; see http://www.herbarium.unc.edu/flora.htm and Fig. 4). This 1320-page treatment covers nearly the entire SERNEC region, except southern Florida (expansion in progress). Weakley specifies as valid 7,000 taxonomic concepts at the sub-/specific level. In addition, he provides RCC–5 articulations to taxonomic concepts authored in multiple relevant preceding treatments (Fig. 4; http://tinyurl.com/wf-rcc5). These logically actionable concept-to-concept reconciliation data are unique and invaluable to the project. We have on hand 75,621 RCC–5 articulations that reconcile each of Weakley's 7,000 concepts (on average nearly 11 times) with alternative concepts stemming from 465 sources, starting as early as Coulter and Rose (1900) and
5. and reaching the present time (Schilling et al. 2015). The 20 most comprehensive treatments – e.g. Small (1933), Wunderlin and Hansen (2011)​, or FNA (2015) – have 1,402–5,543 articulations, for a total of 66,996 or 88.6% of all articulations.

Figure 4.  

Species concept entry Weakley's Flora (Weakley 2015: 355), with 15 RCC–5 articulations to name usages endorsed in incongruent, preceding treatments (such as Radford et al. 1968 ["RAB"]).

What do Weakley's RCC–5 articulations signal? Weakley's articulations measure the performance of taxonomic names as identifiers of taxonomic meanings (Table 1; Geoffroy and Berendsohn 2003, Franz et al. 2016a, Franz et al. 2016b). The signal is good for the majority of names: 57.1% of the pairwise RCC–5 relationships involving the same name(s) are also taxonomically congruent (==). For the remaining cases (42.9%), either the meanings are incongruent (2.9%), or the names are different (18.3%), or both (21.7%). We note, however, that these are pairwise articulations, in each case directed from Weakley (2015) to one preceding flora. The percentage of reliable names decreases if longer concept lineages are measured (Franz et al. 2016a).

Furthermore, Weakley's work focuses on providing one lowest level, closest matching articulation to a concept in another treatment. This has numerous implications. (1) Weakley's articulations do not directly address the genus level, although often species-level incongruences will propagate up (Fig. 1), affecting genus-level SERNEC queries (Fig. 2). The proportion of taxonomic conflict increases at this level, and hence names perform worse (Franz et al. 2016b). (2) Weakley does not directly provide RCC–5 articulations between concepts authored in two non-Weakley treatments. However, such articulations are needed for the service envisioned in Fig. 2. In the great majority of cases the articulations are logically implied, due to transitivity constraints (e.g., if A == B and B > C, then A > C; (Geoffroy and Güntsch 2003, Thau et al. 2009, Thau et al. 2010, Chen et al. 2014a, Chen et al. 2015). Euler/X is designed precisely to remove the two deficiencies. (3) Weakley's articulations are effectively a 'heat map' of taxonomic conflict. If we filter the RCC–5 data for "><" (overlap) or "?" (uncertainty), we get exactly that set of 1,002 articulations where controlling for the taxonomic variable will have the most impact. Such a 'trouble' filter has two invaluable functions: it concentrates our work on a feasible subset of all records, and it provides an empirically preselected set of use cases where the positive effects of transitioning to concept-based resolution of specimen data will be maximal. For these reasons, Weakley's Flora and RCC–5 articulations, and hence Euler/X, SERNEC, and Symbiota – are uniquely indicated to carry concept taxonomy to the specimen level, at scale.

Application of Euler/X to data explicit and implicit in Weakley's Flora will yield 1.5–2.5 billion additional RCC–5 articulations. Here is how. The toolkit ingests two or more taxonomies (T₁,T₂,…,T_N) at a time (Franz et al. 2015). These are read off the respective floras, manually if needed. The concepts entailed in each have parent/child (is_a) relationships. Further input includes the between-taxonomy RCC–5 articulations (A) by Weakley, and finally a set of covering constraints (C) applicable to taxonomic hierarchies (e.g., no parent is childless; siblings are reciprocally disjoint) (Thau and Ludäscher 2007). Euler/X is a powerful, specialized, logic constraint solver (Chen et al. 2014b, Franz et al. 2016a, Franz et al. 2016b). The toolkit represents the entirety of the input (T_1–N,A,C) as a set of constraints which, if consistent (not self-contradictory), will generate one or more output alignments of the input taxonomies, as dictated by the RCC–5 articulations and all logically implied relations. The toolkit's main functionality is to interactively infer – driven by user inputs – output alignments that are both consistent and maximally precise (Chen et al. 2014a, Chen et al. 2015, Franz et al. 2015).

The input constraints and derived alignments can be visualized (Fig. 5:A–C). The alignment visualizations (Fig. 5:C) are directly interpretable by users as multi-taxonomy integration maps, i.e., they are logic-vetted analogs of Fig. 1 that communicate how the input taxonomies are congruent or not, and therefore can or cannot be integrated. Just as critical is the reasoner-inferred set of Maximally Informative Relations (MIR), output either as a simple spreadsheet (Fig. 5:D) or via a scalable matrix visualization tool (Dang et al. 2015). For any concept pair, there are 32 possible RCC–5 articulations (2⁵) if we allow for uncertainty to be expressed by disjunctions (A {== or >} B; etc.). The MIR are exactly that set of relations from which the truth or falseness of any relation in the "R₃₂ lattice" for all concept pairs can be deduced. The formula for the number of MIR inferred via reasoning is simple: If T₁ has m concepts and T₂ has n concepts, then we obtain m * n MIR. This means, e.g., that aligning Weakley (2015) with the Flora of North America (FNA 2015) generates 7000 * 5408 = 37,856,000 MIR logically implied by Weakley's input. Thus if we took the 20 most articulated non-Weakley treatments and produced a matrix of reciprocal pairwise alignments for all of them, we would obtain 2,586,545,944 or nearly 2.6 billion MIR.

Figure 5.  

Euler/X toolkit products. (A) Part of the complex "Andropogon use case", with 1948/1950/1968 input sec. Weakley (2015) (see also Fig. 1). (B) Euler/X visualization of input constraints. Each concept taxonomy is uniquely colored. Ten RCC–5 articulations are provided. (C) Euler/X alignment visualization, showing congruent (grey, rounded) and incongruent (colored) alignment regions. Arrows represent proper inclusion; blue dashed lines show overlap. (D) Subset of 48 inferred, Maximally Informative Relations (MIR). The "*" represents the 1948.Androgon_var_tenuispatheus concept lineage throughout (A) and (C).

Application of Euler/X will generate vast numbers of concept-to-concept relations that speak directly to the query: "To what extent can these two concepts be integrated?" The toolkit will create a comprehensive corpus of RCC–5 signals that will newly drive name-based integration for SERNEC specimen data.

We target the ABI Development level because our innovations consist primarily of key increments to well-established service components, and in making the newly integrated infrastructure work in conjunction with SERNEC's specimen data.

1. Through annual workshops and continuous research on specific use cases, we will foster community engagement to transform SERNEC into the first concept-based herbarium specimen data culture.
2. We will apply the Euler/X reasoning toolkit to generate comprehensive genus- to variety-level concept alignments for at least 10 major treatments with relevance to SERNEC. Both visual and MIR-based toolkit products (> 10⁹ RCC–5 articulations) will be stored in a GitHub repository for user access from the portal, and the RCC–5 articulations will drive the future reconciliation of specimen data.
3. We will expand and optimize the Symbiota taxonomy and occurrences modules, and critical related user interfaces, to represent both name- and concept-based information. We will build modules to manage multiple concept taxonomies, utilize Euler/X reasoning products for specimen queries and visualizations, and perform semi-automated upgrades of identifications to entirely transition to TCLs.
4. Using Symbiota's new identification module and the expertise of the UNC team and SERNEC collaborators, we will augment minimally 80% of the in-region specimen identifications (up to 1.6 million) from the current name-level to the most granular available TCL. Where needed, the identifications (and hence specimens) will be flagged with an "uncertainty signal" (need for study).
5. We will research and publish at least 12 use cases "on the impact of controlling for the taxonomic variable" that showcase the diverse scientific (evolution, ecology) and societal (global change, conservation) significance of further developing a more robust, concepts-to-specimens data culture.

8.1. Community engagement. Community engagement is absolutely critical because we aim to build a new data culture by example. Working with the SERNEC (2016) community is advantageous because this community is highly active and coherent, with ASU (Franz, Gilbert) presently acting as a TCN grant subawardee in charge of all Symbiota-related development. SERNEC botanists are deeply familiar with Weakley's Flora, and his concept mapping efforts are well known, valued, and used (Weakley 2009, Weakley 2014).

To further deepen the engagement, we have identified a core of 12 use cases that will be taken up from the planning to the publication stage by leaders within the SERNEC community (Section 8.5). To directly engage SERNEC scientists, we will hold annual workshops (2^nd quarter of each project year) with as many as 10 non-local invitees plus 10-20 local participants at UNC (year 1), the University of Florida and iDigBio (year 2), and Appalachian State University (year 3). Workshop goals will evolve with the advancement of use cases. Each workshop will run for two full days, plus travel. During the interim periods, we will communicate virtually with use case groups (e.g. via iDigBio's Adobe Connect) and through monthly updates to the SERNEC (2016) community and the iDigBio/Symbiota Working Group (https://www.idigbio.org/wiki/index.php/Symbiota_Working_Group). We budget funds to cover presentation costs at inter-/national conferences, and publication costs for peer-reviewed publications of use case outcomes in impact-maximizing open access (option) journals.

8.2. Euler/X concept alignments. ASU and UIUC will concentrate on this task, with Franz mentoring undergraduate students at ASU to create and publish the alignments, and Ludäscher mentoring a graduate (Ph.D.) student at UIUC to develop new toolkit capabilities for special reasoning and visualization challenges of the SERNEC use case. Weakley's group (UNC) will provide expert input as needed.

I. Scope. We will produce comprehensive – all with all – alignments for the 6-12 most abundantly applied treatments for SERNEC, given the taxonomic subgroup (see Weakley 2015: 7–10). The alignments will be partitioned by family rank sec. Weakley (2015), yielding 413 high-level partitions. This means that input data for, e.g., Orchidaceae sec. Weakley (2015) will include 33 genus-level concepts and their respective children – to be reconciled with orchid concepts authored in one or more (up to 12) additional treatments. However, we will represent each taxonomy as a multi-rooted hierarchy with (in this example 33) unconnected genus-level parents. This is because we will not infer family-level congruence with RCC–5; at that level, name identity 'takes over'. Instead we will model concepts exclusively at the genus- to variety-level, in two complementary ways called "extensional" and "intensional" (Franz and Peet 2009, Franz et al. 2015, Franz et al. 2016b). Roughly, extensional alignments compute genus-level congruence only from the joint signal provided by the children. Sampling of children matters greatly then: if one treatment lists one additional species-level concept – perhaps an introduced entity or one that occurs within the treatment's greater geographic range – that logically means non-congruence at the genus level. Intensional alignments, in turn, respond to the question: "regional sampling issues aside, are these two treatments in obvious conflict in terms of feature-based diagnostics, or not?" As an example, Arizona and Florida have different sets of oak species (non-congruence, extensional), but two concepts of Quercus sec. the respective flora treatments (Arizona, Florida) likely congruently entail "all plants that have acorns and no plants that lack acorns" (congruence, intensional). Both types of alignments are valuable for different purposes, either (1) for bringing out fine differences among parent-level taxonomic concepts or (2) for maximizing concept integration at the generic level across differentially sampled floras. We have optimized Euler/X translations of either kind (Franz et al. 2015).

II. Feasibility. The task of producing two types of 413 family-level alignments that are 6-12 taxonomies deep and reciprocally comprehensive may appear daunting. We are certain that it is not, given prior experience, efforts, and project resources allocated to this task. The reasoning capacity is already there (Franz et al. 2016b). At present, Euler/X can complete on a laptop (> 13 hours) pairwise alignments of taxonomies that are 7-11 ranks deep, have 3,200 input concepts, and yield 2.46 million MIR. In comparison, the most speciose family Asteraceae sec. Weakley (2015) has a total of 1,120 concepts. Most other treatments are less granular, but even doubled (2,240 concepts) these alignments are already well within range (< 2 hours) if done pairwise. In fact they will be faster because we only represent three ranks (genus to variety), which greatly diminishes the reasoning burden. In cases nearing the scale of the Asteraceae, we will be limited to aligning 2-3 treatments at a time. But for many less diverse cases we can align 5-10 taxonomies at once with current capacity, which we expect to increase with further development at UIUC.

III. Approach. We have run thousands of successful alignments with Euler/X, including larger sub-alignments (all Gymnospermae, all Rosaceae) of Weakley (2015) versus Radford et al. (1968). Based on these experiences, and given (1) Weakley's comprehensive RCC–5 articulations and (2) the predominance of congruent (==) signals therein (75.4%; see Table 1), we are confident that a majority of consistent, well-specified alignments are achievable using simple translation scripts. The RCC–5 spreadsheet can be directly reconfigured (e.g., with Python scripts) to extract, hierarchically arrange, and pre-articulate all concept sets as Euler/X input (Franz et al. 2015), leaving only minimal cleaning and completion work for us to infer the alignments. We will generate these scripts (UIUC, UNC), build up efficient workflows, and scale them from less to more complex alignments. The majority of the hands-on work will take place at ASU, under Franz' lead, carried out by a rotating group of undergraduate students who will be trained in systematic theory, toolkit use, and the interpretation/resolution of conflicting treatments. We aim to generate 2,000-3,000 alignments – on average 6-10 per family sec. Weakley (2015) – and more than 10⁹ logically inferred RCC–5 articulations. The latter will drive specimen data integration in the transformed SERNEC portal. These products will be openly accessible – also via links from portal interfaces – in a structured GitHub repository (https://github.com/taxonomic-concept-alignments/SERNEC).

IV. New Euler/X development. SERNEC 2016 poses special challenges for reasoning and visualization. We identify three critical needs. (1) In 2016 Ludäscher's UIUC team released the first toolkit version with a custom RCC–5 reasoner, on which the 3000+ input concepts per alignment capacity is based. However, much of the process of generating alignments relies on constraint specifications, logic-based diagnostics, and visualization functions that remain operationally linked to off-the-shelf "Answer Set Programming" reasoners (Chen et al. 2014a, Chen et al. 2014b, Chen et al. 2015). One example is the representation of hybrid concepts (Fig. 3; Cleistesiopsis x ochlockneensis), achieved by locally relaxing the sibling disjointness constraint (Thau and Ludäscher 2007). To fully deploy all needed Euler/X functions at increasingly large scales, we will require further toolkit development in connection with custom RCC–5 reasoning. (2) The SERNEC alignments have few ranks per taxonomy (range: 2-3), but are deep in terms of numbers of taxonomies (range: 2-12). Moreover, the RCC–5 input articulations are "star-like", i.e., always directed from Weakley 2015 to one alternative treatment. We need to generate matrices of all possible pairs. This will require new logic development and application of RCC–5 transitive reasoning. We will review and advance how each element of the R₃₂ lattice (see Section 6) can be propagated along a chain of three or more concepts. This problem was initially explored by Geoffroy and Güntsch 2003 under the theme of "navigating the potential taxon graph". But the theoretical work remains under-developed, and solutions have not yet been implemented (Thau 2010, Thau et al. 2010). The subject is therefore highly suitable for a Ph.D. thesis in computer science, to be mentored at UIUC by Ludäscher. The SERNEC use case is ideal for optimizing the task of RCC–5 propagation beyond paired taxonomies, which addresses general and specific needs to develop comprehensive RCC–5 coverage for all SERNEC taxonomies. (3) Lastly, we need to visualize alignments of concept-rich but rank-poor taxonomies. The task entails special color/shape assignment and spatial efficiency constraints (Graham and Kennedy 2010, Dang et al. 2015). Indeed, if we limit ourselves to the most granular level, Weakley's tabular views are most effective (Figs 1, 5:A). To find the best pragmatic solutions for 2-3 levels and 2-12 taxonomies, we will create custom stylesheets that translate the toolkit products (Fig. 5) into easy-to-interpret GraphViz (Gansner and North 2000) and PDF outputs for users.

8.3. Adding taxonomic concept representation to Symbiota and SERNEC. This objective requires a large part of the project's resources for new Symbiota development at ASU. Three major task domains are involved. (1) Symbiota's underlying taxonomic and occurrence schemas must be expanded to support TCLs, source-specific parent/child relationships, and RCC–5 articulations. (2) A subset of Symbiota's graphic user interfaces will be changed accordingly, and new interfaces will be created to manage multiple taxonomies and efficiently upgrade specimen identifications to TCLs. (3) A name-to-concept transition plan for the SERNEC portal will be executed such that (a) existing named-based data are not functionally compromised and (b) new concept-based data become the portal norm – most immediately to support our use cases. Below we describe the sequence of actions that will achieve this transition.

I. Schema expansion. The expansion of Symbiota's taxonomic and occurrence modules will be guided by the remarkable example of Avibase (Lepage et al. 2014; http://avibase.bsc-eoc.org) which manages more than 1.5 million taxonomic concepts. Symbiota's current taxonomic module (http://symbiota.org/docs/taxonomic-schema-2) is a well-developed branch of the data structure that already permits source-specific views on valid versus synonymous names. As in the case of Avibase, we will transition to maintain a single, specific higher-level classification for higher taxonomic levels (kingdom to family; Weakley 2015, Stevens 2016), but represent taxonomic concepts at lower levels where the impact on SERNEC specimen searches is most significant. We will complete and release a new "references" module, and add new tables to explicitly model TCLs, parent/child relationships, and RCC–5 articulations. We will also 'elevate' the representation of synonymy relationships to the concept level, meaning that nomenclatural relationships between valid and synonymous names (each rendered as TCLs) will be displayed as a specific function of the treatment that recognizes them (unlike Fig. 3). Once transitioned, we will propagate default settings for occurrences that require using TCLs from the expanded taxonomic module.

II. User interfaces. We will upgrade a critical subset of Symbiota's interfaces to enable concept use. In particular, we will modify the taxonomy viewing and editing interfaces to only accept the new syntax and semantics. New concept taxonomies can be uploaded piece-meal or through batch functions. We will also create simple formatting and loading tools to ingest multiple taxonomies into the Euler/X alignment toolkit and re-integrate the outcomes (MIR) into the RCC–5 table. Based on the latter, we will generate a new "incongruence alert" table that entails precisely those taxonomic names that, when searched by users, require additional specification of a TCL to identify a consistent circumscription (Figs 1, 2).

Symbiota already has an effective visualization interface for single taxonomies. Rather than building a new multi-tree visualization interface – which is both difficult and redundant (Graham and Kennedy 2010, Dang et al. 2015, Hinchliff et al. 2015) – we will (1) provide drop-down menus for users to select and view a taxonomy at a time, and (2) links for each genus-level concept to the corresponding set of Euler/X alignments with other treatments stored in the GitHub repository. This approach is more modular and feasible than dynamically integrating Euler/X alignment visualizations within Symbiota, while allowing users to explore SERNEC's classifications interactively.

We will reconfigure the occurrence identification interface to interact with the new taxonomy module. Again, this will include drop-down options to select preferred sources, view alignments, and populate a TCL. Very substantive upgrades will be made to the add batch determinations interface, which presently permits selecting names or individual specimens. In collaboration with SERNEC experts, we will expand this interface to represent the subset of Darwin Core fields most decisive for filtering occurrences so that batch updates can follow. Target fields will include (e.g.) the source collection, collecting locality and date, collector/identifier information (who/when?), and references (where available). Combinations of these fields will facilitate smart queries of the kind: "show me all specimens in this region, collected in this time period, and identified to this name by members of this herbarium community". These queries, combined with expert knowledge and specimen images, will facilitate upgrades of many identifications to TCLs at once. A second, innovative map-based determinations function will be developed as an extension of Symbiota's Map Search module. It will allow experts to gather specimen sets for batch determinations directly through the map interface (http://tinyurl.com/sernec-mapint), by using area shape selectors. Because granular taxonomic concepts are often geographically separated (Fig. 2:D), the function will allow finding all specimens in need of TCL identification conveniently, and for any given region.

Lastly, we will transform the primary search and display specimens interfaces. Key goals are to promote TCL-based specimen queries and mappings, with the option to relabel an initially queried set according to an alternative treatment (Fig. 2). Superficially, the specimen query interface will not change much. However, if for instance the user enters the name "Cleistes divaricata", this name is pre-identified in the "alert" table to have incongruent meanings (see above). Hence the user will be prompted with the counter-query "please specify further". They will see a drop-down list of treatments, and links to alignments on GitHub that will explain the concept resolution issues. Once a TCL is chosen (which remains optional – we will never prohibit name searches), the output is rendered accordingly (Fig. 2). At this point users have two additional options to assess robustness. (1) They can rerun the query with alternative TCLs and compare maps, which may show different specimen sets; or (2) they can rerun the query on the same specimen set but with different TCLs. We will highlight options to view and save TCL-based query and relabeling outcomes. Symbiota is known for the relative ease of moving data; typically only 2-3 clicks are needed to extract datasets in various formats, including Darwin Core Archive (GBIF 2010). This ensures that explorations of the taxonomic variable are fully documented and reproducible.

III. Transition plan. Realizing the above changes requires a sound transition plan. It is critical not to break existing services while building new ones for transition. We also recall that 57.1% of the RCC–5 articulations identify reliable name usages that (at present) need no additional specification (Table 1).

Once the taxonomy module is expanded, it is necessary to 'reify' the SERNEC consensus taxonomy (Fig. 3) at a specific date to become static. This means, for instance, that a specimen identified only to the name "Cleistes divaricata" will now be automatically updated to the TCL "Cleistes divaricata sec. SERNEC consensus, March 1^st, 2018", where the latter is an immutable classification. Multiple (6-12) rigorously sourced and aligned classifications will be added under the new conventions. Submitting 'just names' or new TCLs without RCC–5 articulations is no longer admissible. We thereby shift to the new culture. However, for reasons explained in Sections 1–3, the reified 'TCLs' of the SERNEC consensus cannot be precisely aligned with TCLs of specific treatments. Initially, we will have created a syntax change without actually improving the core semantics (!). Yet this gives us an excellent success metric: the degree to which we will have transitioned from name- to TCL-based specimen resolution is directly proportional to the number of "sec. SERNEC 2018" TCLs that are re-identified to a specific source ("sec. Weakley 2015", etc.). In other words, the syntactic shift to TCLs for SERNEC specimens is instantaneous (March 1^st, 2018) and will affect all specimens identified only to a name at that time. In the following, gaining semantic resolution becomes a function of gradually replacing 'consensus' TCLs with (typically the most granular) externally sourced TCLs. The use cases will facilitate the beginnings of this transition.

8.4. Augmenting SERNEC specimen identifications to TCLs. Using the new tools, our goal is to upgrade minimally 80% of all specimen identifications from "sec. SERNEC 2016" to either (1) a specific, maximally granular treatment or (2) an added "uncertainty signal". For the use cases, which – as shown in Table 2 – compromise 35,422 specimens, we will aim for close to 100% granular TCL identifications.

Improving identifications will be facilitated by technology, but is only feasible because of the direct involvement of experts. Some 15-30% of the SERNEC records have partial identification-related information recorded (expert, year, taxonomic reference used). We will utilize these data to identify the best-fitting TCL. Collective experience strongly indicates that, even for problematic cases, a remotely working expert can confidently assert TCL identifications by drawing on their sophisticated background knowledge of spatially/temporally localized identification practices. For instance, a very large number of non-Floridian SERNEC herbaria have treated Radford et al. 1968 as "the bible" during the past four decades. Weakley's concepts tend to be more granular, but in geographically and/or ecologically structured patterns (Fig. 2:D). Thus, all specimens identified as "Cleistes divaricata" but documented from Gulf Coast savannas can be confidently updated to "Cleistesiopsis oricamporum sec. Weakley 2015". The scope of the SERNEC TCN grant is critical here: all newly digitized herbarium sheets are scanned/uploaded as high-resolution images (http://tinyurl.com/sernec-at-idigbio). Presently, this includes 1.25 million or 63.6% of all records – a proportion that will only increase with time. On average, our use case experts will have these images available to make maximally granular identifications in 2/3 of all ambiguous cases.

Thousands of herbarium sheets will nevertheless remain "uncertain" with regards to the narrowest TCL. Removing uncertainty may require direct study of morphological/molecular data, likely in the context of new revisions. Although hypotheses are weakened in such cases, we regard this as a positive contribution to explicitly identify 'problem specimens'. In analogy to the "alert" table for incongruent name usages, we will create a special flag for uncertain TCL identifications. Flagged specimens will be retrievable by query, and uniquely colored on maps, with an option to show only non-ambiguous specimens.

8.5. Use case selection, approach, and impact. We have enlisted five SERNEC botanists (plus the UNC postdoc; Section 9) to lead 12 use cases (Table 2​). The selection rationale is as follows. Use cases (1) must be feasible in terms of taxonomic and specimen volume; (2) entail significant conflicts in taxonomic concepts; (3) have availability of taxonomic experts; and (4) differential concept resolution should affect the robustness of basic and applied biological inferences. We are confident that several hundreds of such use cases exit within the SERNEC domain. Hence the current selection and teams are subject to much expansion, but have emerged as a strategy to best demonstrate impact. We briefly describe each.

I. Use case particulars. The Andropogon glomeratus-virginicus "complex" is notorious for its taxonomic instability (Campbell 1983, Weakley et al. 2011, Weakley 2015). An Euler/X analysis of 11 classifications (Franz et al. 2016a) revealed a name:meaning cardinality ranging from 1:6 to 4:1, with only 1/36 names being reliable in the sense of Table 1. Different taxonomic schemata dramatically alter the recognition of species- and variety-level diversity, ranges, ecological preferences (hydrology), and hybridization hypotheses for these "bluestem" grasses. Weakley will guide this paradigmatic concept use case. Asarum and Hexastylis (birthwort family) have longstanding 'porous' generic limits (Kelly 1998), disagreements regarding species-level diversity, and hence also distribution patterns (Weakley 2015). Murrell, who has extensively studied "concept creep" in this group using electrophoretic and morphological methods (Murrell et al. 1998), will lead. Cleistes and Cleistesiopsis sec. auctorum entail concepts for rare, endangered "Pogonia" orchids (Pansarin and Barros 2008, Pansarin and Brown 2009, Smith et al. 2004). Alternative views have disparate implications for distribution ranges (Figs 1, 2) and thus for conservation management decisions (Gregg and Kéry 2006). This use case, headed by Weakley, is particularly suited to replicate the seminal (but pre-Darwin Core) Peterson and Navarro-Sigüenza 1999 analysis of species concepts altering conservation priorities – at the specimen level. The composite Coreopsis is diverse and poorly delimited from Bidens (Kim et al. 1999, Crawford and Mort 2005, Sorrie et al. 2013). Our hypotheses of narrow ecological specialization, the evolution of polyploidy, and historical biogeography in the Southeast are contingent on the taxonomic variable. Weakley and postdoc will lead. The use case for Cornus (dogwood family) will be coordinated by specialist Murrell (Murrell 1993, Xiang et al. 2006). Several complex subgroups and possible hybrids are involved; with implications for understanding Tertiary continental migration patterns.

Euphorbia (spurge family) is the most speciose complex, including recent introductions not yet keyed out by Weakley 2015. Many species-level groups are geographically limited and require specialized edaphic conditions, with implications for biogeography, population genetics, conservation, and even medicinal discovery (Park 1998, Dorsey et al. 2013, Ernst et al. 2016). Mac Alford (University of Southern Mississippi) will coordinate. Galactia is a legume complex (Duncan 1979, Isely 1998, Ward and Hall 2004) and "one of the worst genera ever" (Alan Franck, University of South Florida, pers. comm.), with incongruences in species delimitation and distribution that will only increase after an upcoming revision co-authored by Franck who will take on this use case. Gonolobus and Matelea (dogbane family) are variously 'joined' or not and entail several contentious subgroups, with differential outcomes for our understanding of their distributions, reproductive biology, and breeding system evolution (Lipow and Wyatt 1998, Krings 2008). Mark Fishbein (Oklahoma State University) will concentrate on this complex. Lantana is a small use case, with Franck leading, where incongruent concepts involve ecological interactions of endemic/native and invasive/exotic groups (Sanders 2006, Sanders 2012) – an important theme for the flora of Florida (Lee et al. 2009). The composite Liatris (Mayfield 2002, FNA 2015, Weakley 2015) is the second most speciose use case, with Alford as lead, where several inconsistently recognized groups are threatened in eastern prairie habitats. The group is also a model for understanding density-dependent foraging (Figlar and Nooteboom 2004). For Magnolia, 12 incongruent schemata have been published (Figlar and Nooteboom 2004, Sima and Lu 2012, Weakley 2015) – a situation that requires concept resolution for these abundant, cherished, and often cultivated and then naturalized groups. Weakley has aligned all views and will head the use case. Alternative schemata will affect deep- and shallow-time inferences of disjunctions and dispersal processes between temperate and tropical regions (Azuma et al. 2001, Azuma et al. 2011). Lastly, Monotropis (heath family) is the smallest use case, though very complex due to different family- to species-level assignments. Members have specialized heteromycotrophic interactions with fungi and their (oak or pine) hosts, and these are variously captured in taxonomic perspectives (Klooster and Culley 2009, Rose and Freudenstein 2014). Weakley will analyze.

II. Research approach. We expect that use case leaders will engage many additional SERNEC members. Although the TCL identification efforts (Section 8.4) will be similar in each case, the ultimate research goals will vary greatly. Some may take on the form of a review – though rooted in specimen-level data and visualizations – of taxonomic inconsistencies affecting our basic understanding of biodiversity and distribution. Others may reassess the specimen-level evidence and inter-taxonomic robustness of very specific ecological or evolutionary hypotheses. Still others may characterize to what extent conservation and global change assessments are contingent on a specific taxonomic commitment (Peterson and Navarro-Sigüenza 1999, Rosenberg 2014, Rylands and Mittermeier 2014, Borsch et al. 2015). This variety of motivations is deliberate: we will thereby demonstrate that the need to control for the effect of the taxonomic variable on integrating specimen-level data is potentially as broad as comparative biology.

III. Innovative impact. Rather than specifying each worthwhile research question, we exemplify the kinds of questions that our development will facilitate, and why this matters. Accordingly, in the case of Andropogon, users can query (often sequentially):

1. "Show me all specimens identified to the taxonomic name Andropogon virginicus in the Carolinas" [returns many records, resolved only with the ambiguity of an incongruent taxonomic concept lineage].
2. "Now show me all specimens identified to the TCL Andropogon virginicus sec. Weakley 2015" [returns a subset of these records, reflecting a choice for this particular, granular TCL resolution].
3. "Now me all specimens of Andropogon virginicus sec. Radford et al. 1968, yet translated into the TCLs sec. FNA 2015" [returns (again) many records, but specifically represents and contrasts two treatments, as opposed to providing the ambiguous lineage view of (1)].
4. "Show me all specimens whose identifications are ambiguous with regards to FNA 2015 versus Weakley 2015" [using the "uncertainty" flag, points to specimens needing further study].
5. "For the Carolinas and for this inclusive specimen set, show me the composite least versus most granular taxonomic perspective(s) available" [returns a potentially multi-taxonomy 'composite' that represents opposite extremes in resolution granularity].
6. "Save and output all results in Darwin Core Archive format".

This is what we mean by "controlling the taxonomic variable". The services will be basic, as dictated by realism, and the control offered to users is not explicitly of a statistical kind. Yet we are confident that queries (2-6) – which are not supported by any existing herbarium specimen network – will yield impactful outcomes when applied to the aforementioned use cases and research goals. We will work to carry each of these to publication in international journals with open access options that variously expand the reach of our approach, such as Biodiversity Data Journal, Conservation Biology, Global Ecology and Biogeography, PeerJ, PLoS ONE, Systematic Biology, Taxon, and Trends in Ecology and Evolution.

The project lead personnel – Franz, Gilbert, Ludäscher, and Weakley – is introduced in Section 5. Franz will mentor 15 undergraduate students at ASU to achieve the Euler/X alignments (Section 8.2.I–III). At UIUC, Ludäscher will mentor a Ph.D. student who will concentrate on the reasoning and visualization challenges (Section 8.2.IV). At UNC, Weakley will mentor a postdoc (year 1), and utilize applications analyst Michael Lee to provide critical support at the interface of Symbiota module development, data population, and service optimization for the use cases. Weakley and the postdoc will play an immense role in overseeing the rapid integration of the floristic legacy, new tools, and expert contributions. Gilbert and Lee will translate new conceptual and TCL identification-related functions and data from various sources into the transformed SERNEC portal (Sections 8.3 & 8.4). We invest significant resources to support expert co-leadership of the use cases (Sections 8.1 & 8.5).

Fig. 6 lists the development milestones and deliverables corresponding to the five main objectives and for the three allocated project years. Symbiota's structure is modular, and the development release approach is such that all new code is immediately functional in portals that need the new functions. Select modular functions will go live in SERNEC within weeks of the project start, although the complete service as described in Section 8.4.III will take more time to be built and fully sustain the use cases.

Figure 6.  

Project management time line; see also Sections 5 & 8. G = graduate student; UG = undergraduate students.

I. Scientific. The intellectual case was presented in Sections 1–4. Our project is focused on just one herbarium network, but the potential impact is much wider. Space does not permit reviewing the many aggregators that concede, in one form or another, that taxonomic concept resolution is needed but seemingly out of reach. The list includes (e.g.) Catalogue of Life, GenBank, Global Biodiversity Information Facility, Global Names Architecture, iDigBio, Open Tree of Life, and the Taxonomic Name Resolution Service (Jones et al. 2011, Federhen 2012, Boyle et al. 2013, Döring 2013, Hinchliff et al. 2015, Döring 2016b, Patterson et al. 2016). Meanwhile, research communities are viewing the inherent ambiguity of taxonomic names in networked environments as a perpetual Achilles' heel for aggregation – our weakest semantic link in the entire system (Por 2007, Bortolus 2008, Peet et al. 2012, Hjarding et al. 2015, Maldonado et al. 2015, Franklin et al. 2016, Wiser 2016).

This project is designed to advance a global agenda, by demonstrating that conceptual and technical challenges can be addressed at scale if communities are willing to engage in concept taxonomy. Trust in data is also a design feature of allowing expert access and resolving multiple conflicting views that reflect the realities of ongoing taxonomic research. If only 1% of SERNEC's data display the issues shown in Fig. 2, trust in all data suffers. We believe that better science will follow from our demonstrations, first reaching other Symbiota portals and eventually large-scale aggregators. In light of vast investments into vouchered data mobilization, our project is a highly cost-effective way to improve data quality and trust. Taking well over 1 million SERNEC records from name- to TCL-identification will show that "big" specimen data can pass the credibility threshold needed to validate the mobilization investment.

II. Educational. We will directly train one postdoctoral researcher (UNC), two Ph.D. students (ASU, UIUC), and at least 15 undergraduate students (ASU). Each of our workshops will capacitate 10-15 SERNEC experts, who in turn can inform and recruit colleagues and students at their home herbaria. Project members Alford, Fishbein, Franck, Franz, Gilbert, Murell, Soltis, and Weakley regularly offer plant/biodiversity courses to undergraduate students at their respective institutions, reaching an estimated 300-500 students per year, with ca. 10-40% minority students (range: Oklahoma – Mississippi). Each has committed to integrating our project's theme and use cases as new sections into their future biodiversity teaching plans. At ASU, this will include two new three-hour sections of the undergraduate-focused biodiversity informatics course "Discovering Biodiversity – Field to Database", offered in the spring of 2017 and 2019 to 25 students. At each institution, project members will make a sustained, systematic effort to recruit new students from underrepresented groups, working through institutional (e.g., sponsored STEM minority mentor programs) and local student organizations to advertise project opportunities and thereby proactively broaden participation.

Murrell's leadership of SERNEC will promote our advances with nearly 200 herbarium scientists in the region. Alford's involvement in the Magnolia grandiFLORA project (http://www.mississippiplants.org/), which has an educational component for K–12 teachers, will add exposure. At ASU, Franz and Gilbert will promote the project through virtual and personal outreach, aided by their leadership of the Biodiversity Knowledge Integration Center (BioKIC; https://biokic.asu.edu/). We will publish a BioKIC monthly blog post with project updates, to be shared with the iDigBio/Symbiota Working Group. Conference presentations will mainly target the global TDWG community (http://www.tdwg.org/).

We budget funds for two additional forms of outreach. The first will be a feature story "Where do plant species occur?" (see Fig. 2) – with accessible web content, learning activities, and videos – for ASU's popular "Ask A Biologist" learning resource, visited 22,000 times daily (https://askabiologist.asu.edu/). The second is a comprehensive series of "How-To" videos for creating multi-taxonomy alignments. Connecting the primary taxonomic literature with the RCC–5 logic is not trivial. Applying the concepts and tools is a bottleneck for further adoption. We will produce 10-15 well-structured and -narrated videos that exemplify steps the entire project workflow, including efficient use of the Euler/X toolkit. The video projects will be presented by ASU undergraduate students, with input from project experts. The series will be published and promoted through iDigBio's Vimeo channel (https://vimeo.com/idigbio).

Mid-term prospects (~ 5-15 years) for our development and data innovations are very strong. Our project operates inside an upward-trending information culture (Gries et al. 2014, Jones et al. 2014, Nelson et al. 2015). The SERNEC community is 233 herbaria strong, active since 2005, and recently organized as the SouthEast Chapter of the Society of Herbarium Curators, creating a region-wide research/training/proposal engine. In the past year the portal was accessed ca. 86x daily by a total of 13,410 users (Google Analytics). SEINet's metrics (http://swbiodiversity.org/seinet/) are 10x higher – a realistic SERNEC target. The data, portal, and software are sustained by standing commitments from SERNEC, ASU, and iDigBio (University of Florida). Odds of sustaining the concept taxonomy innovations are less clear; thus we prioritize community engagement and use case-driven development.