Ontology-to-Tools Compilation — Supplementary Information (p.28-47)

7 Supplementary Information

7.1.1 LLM-based agents and ReAct-style reasoning

Large language model (LLM)–based agents couple a generative model with an explicit action space, where each step conditions on a textual context (for example, a user request, intermediate results, and tool outputs) and selects the next action such as calling a tool, querying a knowledge graph, or producing a natural-language response. [48, 53, 56, 60] External capabilities are typically exposed through structured interfaces (for example, function signatures with typed arguments), and tool responses are returned as additional context to enable multi-step interaction with external systems. [48, 53]

ReAct-style agents organise this interaction as an explicit loop of reasoning and acting, alternating between intermediate reasoning traces and tool actions, and incorporating tool observations into subsequent decisions. [60] In information extraction, this framing supports decomposing long inputs, validating intermediate structures, and iteratively refining partial outputs through repeated reasoning–acting cycles.

7.1.2 Tool-calling interfaces and the Model Context Protocol

Earlier tool-calling interfaces for LLMs typically expose application-specific functions through proprietary schemas or prompt templates, and integration logic is often implemented separately for each deployment. [8, 27, 31] The Model Context Protocol (MCP) instead defines an open client–server protocol for connecting LLM applications to external tools and data sources. In MCP, servers register tools, resources, and prompts with machine-readable schemas; clients discover these capabilities and request tool invocations through a standardised messaging layer. [35, 38] MCP supports structured tool definitions, streaming of results, and standardised error reporting across heterogeneous hosts and models, enabling reuse of the same tool servers across different LLM applications.

7.1.3 The World Avatar and its chemistry ontologies

The World Avatar (TWA) is a cross-domain digital ecosystem that represents entities, processes, and their interdependencies using ontologies and RDF knowledge graphs, accessed and modified by software agents at runtime. [4] The reticular-chemistry stack of TWA is organised around two domain-generic core ontologies and one application ontology. OntoSpecies represents chemical species, identifiers, and characterisation data (with extensions for IR and elemental analysis). OntoSyn (OntoSynthesis) represents synthesis procedures as sequences of SynthesisStep operations transforming input Species into ChemicalOutput products. OntoMOPs acts as an application ontology for metal– organic polyhedra (MOPs), modelling MOP instances and their building units and linking OntoSyn procedures and OntoSpecies reactants into the MOP design space. [4, 24, 40, 45] Previous work implemented a MOP synthesis pipeline in TWA that used LLMs guided by

ontology-aligned JSON schemas to populate these chemistry knowledge graphs. [45]

7.1.4 LLM-based knowledge instantiation and ontology grounding

Large language models are increasingly used to construct and maintain knowledge graphs, by inducing schemas, populating instances, or interacting with graph backends. [28, 44] From the perspective of this work, two lines are most relevant: schema-level approaches that define or refine extraction targets, and instance-level approaches that extract and validate concrete assertions.

Schema-level (T-Box). Schema-level pipelines treat the extraction schema as a contract that defines what types, fields, and constraints outputs should satisfy. PARSE refines JSON Schemas as first-class artefacts for extraction, but optimises primarily syntactic constraints encoded at the JSON-schema level rather than ontology axioms. [50] LLMs4SchemaDiscovery mines candidate scientific schemas from text with human feedback, again focusing on schema discovery rather than using a fixed ontology T-Box as the contract. [47] SCHEMA-MINERpro adds ontology grounding by mapping discovered schemas to reference ontologies, but grounding remains a stage separate from instance creation. [46] AutoSchemaKG induces schemas and triples at scale without a predefined domain T-Box, enabling cross-domain graphs but decoupling extraction from ontologyencoded constraints such as units and identification rules. [3]

Instance-level (A-Box). Instance-level pipelines focus on extracting entities, relations, and events as concrete assertions. KGGen uses strict JSON specifications to extract and refine triples, improving structural consistency, but semantic validity with respect to a concrete ontology is enforced via post-hoc checks rather than constraint-aware instance construction. [33] Other pipelines align extracted relations to ontology predicates or use ontologies for filtering and validation, but typically keep ontology enforcement separate from the tool interfaces that govern how instances are created. [22, 23, 36]

MCP-assisted LLM-based solutions. MCP has emerged as a general mechanism for connecting LLM applications to tools and data sources. [35, 38] Existing MCP-based knowledge-graph servers often expose generic CRUD-style interfaces over graph backends, demonstrating tool-based access but typically leaving ontological correctness to application logic or post-hoc validation rather than compiling T-Box axioms into tool signatures and runtime checks. [18]

TWA agent composition framework Prior to recent LLM-based agent systems, The World Avatar (TWA) explored how semantic descriptions can support the discovery, com- position, and execution of software agents within a knowledge-graph ecosystem. [69] introduced a semantic agent composition framework for TWA, in which agents are described using a lightweight agent ontology and grounded execution metadata to enable automated composition of agent workflows across domains.

This work motivates a view of ontologies not only as data models but also as interfaces to execution, where semantic descriptions govern which agents can be composed and how they may interact. In contrast to agent discovery and composition based on pre-defined interfaces, the present work compiles domain ontologies into executable tool interfaces and run-time constraints that directly regulate generative behaviour.

7.2 Supplementary Methods

This section provides extended implementation and evaluation details referenced in the main Methods, including dataset curation and annotation protocol, evaluation metrics and matching rules (including the CBU formula-only criterion), MCP server/tool specifications, and runtime policies (iterations, error handling, and repair strategies).

Agentcreated MCP tools

Predefined scripts

T/A Boxes

Predefined MCP tools

LLM-powered Agents

Figure 7.6: Ontology-to-tools compilation workflow. A preparation agent takes as input the ontology T-Box and a small set of manually authored, domain-agnostic meta-prompts for extraction and KG construction. It synthesises (i) an ontology-specific MCP server exposing ontology-aware tools with machine- checkable schemas and (ii) domain- and task-specific instantiation prompts (for synthesis steps, reaction chemicals, characterisation entities, and CBUs) that steer the runtime agent. A ReAct-style instantiation agent then executes the generated prompts and invokes the MCP tools to construct knowledge- graph instances from documents.

Figure 7.7: UML workflow of the two-agent framework for ontology-aligned knowledge- graph construction within the World Avatar (TWA) environment. The MCP Creation Agent runs in the TWA context and takes the domain ontology (T- Box; schema) and an MCP Creation Prompt to generate a KG instantiation library. The MCP Integration Agent then takes an MCP Integration Prompt plus the validated library (and its function descriptions) to package it into a deployable MCP server, exposing these functions as callable MCP tools and registering them into a shared MCP tool pool for reuse by downstream LLM agents, including TWA workflows.

Figure 7.8: Class diagram of the structured task specification: a task decomposition model, guided by the Ontology T-Box and a meta-prompt, generates itera- tions that bundle and KG-building steps, file flows (inputs/outputs), optional sub-iterations, and MCP tool configurations.

Figure 7.9: Illustrative example of the interaction between the OM T-Box (top), a sin- gle create_temperature call within the MCP server logic (middle), and the resulting OM instance (bottom). The top box shows the ontology T-Box, which defines schema-level constraints for temperature quantities, such as the relevant classes and properties. The middle box shows an example script call that enforces these constraints by looking up the unit, validating the nu- meric value, and only then creating the corresponding OM individuals for the temperature value and its unit. The bottom box shows the resulting example instance that satisfies the T-Box constraints, linking the created quantity to a valid unit and a compliant numerical value.

7.2.1 Iteration 1: System Meta Prompt

You are a knowledge graph construction prompt expert specializing in

ITERATION 1 KG building prompts.

ITERATION 1 is special: it only creates top-level entity instances from

paper content, WITHOUT creating any related entities (inputs,

outputs, sub-components, or detailed steps).

Your task:

Analyze the provided ontology (T-Box) to identify the top entity

type (the main class that represents the overall procedure/process)

Extract all relevant rules, constraints, and identification

guidelines from rdfs:comment annotations

Generate a focused, tool-oriented KG building prompt for ITERATION 1

CRITICAL CONSTRAINTS FOR ITER1:

ONLY create instances of the top-level entity class (one per

procedure described in the paper)

Do NOT create any related entities (inputs, outputs, sub-components,

steps) in this iteration

Link each top-level entity to its source document
Apply strict identification rules from the ontology to determine what

qualifies as a valid instance

Include all cardinality, scope, exclusion, and linking rules from the

ontology

OUTPUT REQUIREMENTS:

Start with “Follow these generic rules for any iteration.” followed

by the global rules

Include MCP tool-specific guidance (error handling, IRI management,

check_existing_* usage)

Include explicit identification section (document identifier handling)
Clearly state the task scope: create top-level entities only, no

related entities

List all constraints from the ontology rdfs:comment for the top-level

entity class

Include clear termination conditions
Be concise and tool-oriented (this prompt is for an MCP agent using

function calls)

Be completely domain-agnostic (no specific compound types, no

domain-specific terminology)

Do NOT include:

Domain-specific examples (\eg specific compound names, specific

synthesis types)

Variable placeholders like {doi} or {paper_content} - these will be

added programmatically

Verbose ontology explanations - focus on actionable rules
Any mention of specific chemical entities, materials, or

domain-specific concepts

7.2.2 Iteration 1: User Meta Prompt

Based on the ontology and available MCP tools below, generate a KG

building prompt for ITERATION 1.

ITERATION 1 SCOPE:

Create ONLY top-level entity instances (the main procedure/process

class)

Do NOT create any related entities (inputs, outputs, sub-components,

steps)

Link to source document

ONTOLOGY (T-Box):

{tbox}

MCP MAIN SCRIPT (Available Tools) [Python]:

{mcp_main_script}

The MCP Main Script shows all available tools with their descriptions,

parameters, and usage guidance. Use this to understand what tools

are available and how they should be called.

REQUIREMENTS:

Extract ALL rules from the top-level entity class rdfs:comment,

especially:

Scope (what qualifies as a valid instance)
Different forms / methods / variations rules
Exclusions (what NOT to create)
Cardinality requirements
Linking requirements
Conservative behavior guidelines
Critical exclusions for extraction

Include global MCP rules:

Tool invocation rules (never call same tool twice with identical

args)

IRI management (must create before passing, use check_existing_*

tools)

Error handling (status codes, already_attached, retryable)
Placeholder policies
Termination conditions (run_status: done)

Include identification section:

Document identifier handling (treat as sole task identifier, reuse

consistently)

Entity focus guidance (for when entity_label/entity_uri provided)

Be concise and actionable - this is for an MCP agent making function

calls

Be completely domain-agnostic:

Do NOT mention specific compound types, materials, or chemical

entities

Use generic terminology that applies to any domain
Adapt wording from the ontology to be domain-neutral where possible

Generate the prompt now (do NOT include variable placeholders - those

will be added programmatically):

7.2.3 Extension Ontology: System Meta Prompt

You are an expert in creating knowledge graph building prompts for

extension ontologies.

Extension ontologies are simpler ontologies that extend a main ontology

with additional specialized information. They use MCP tools to

build A-Boxes that link to the main ontology’s A-Box.

Your task is to analyze a T-Box ontology and MCP tools to create a KG

building prompt that:

Provides a clear task route for building the extension A-Box
Emphasizes using MCP tools to populate the A-Box
Requires comprehensive population (making certain MCP function calls

compulsory)

Emphasizes IRI reuse from the main ontology A-Box
Includes domain-specific requirements extracted from the T-Box

comments

Forbids fabrication - only use information from the paper

CRITICAL RULES:

Read ALL classes, properties, and rdfs:comment fields in the T-Box
Extract domain-specific requirements from rdfs:comment (\eg required

fields, cardinality constraints)

Focus on HOW to build the A-Box, not just WHAT to extract
Emphasize the connection between the extension A-Box and the main

ontology A-Box

Make the prompt actionable with clear steps
Output ONLY the prompt text (no markdown fences, no commentary)

7.2.4 Extension Ontology: User Meta Prompt

Generate a KG building prompt for an extension ontology.

T-Box (analyze to understand the ontology structure and requirements):

{tbox}

MCP Main Script (understand available tools, their parameters, and

calling sequences):

{mcp_main_script}

Your prompt MUST:

State the task - Extend the main ontology A-Box with the extension A-Box

Provide a task route - Give a recommended sequence of steps for

building the KG

Emphasize MCP tools - Instruct to use the MCP server to populate the

A-Box

Require IRI reuse - Emphasize reusing existing IRIs from the main

ontology A-Box

Extract T-Box requirements - Include any compulsory requirements from

rdfs:comment (\eg required fields, minimum instances)

Forbid fabrication - Only use information from the paper content

Include tool-specific notes - If the T-Box or domain requires special

handling (\eg external database integration, data transformations),

include those notes

Structure your output as:

Your task is to extend the provided A-Box of [MainOntology] with the


[extension ontology] A-Box, according to the paper content.


You should use the provided MCP server to populate the [extension


ontology] A-Box.


Here is the recommended route of task:


[Step-by-step guidance based on T-Box structure]


36


Requirements:


[List of requirements based on T-Box rdfs:comment and MCP tool


constraints]


Special note:


[Any domain-specific notes based on T-Box comments]


Here is the DOI for this run (normalized and pipeline forms):


- DOI: {{doi_slash}}

- Pipeline DOI: {{doi_underscore}}


Here is the [MainOntology] A-Box:


{{main_ontology_a_box}}


Here is the paper content:


{{paper_content}}

CRITICAL:

Extract domain-specific requirements from the T-Box rdfs:comment

fields. Do NOT invent requirements. ALL requirements must be

justified by the T-Box or MCP tool constraints.

Output EXACTLY the structure shown above. Do NOT add any additional

sections after {{paper_content}}. This is the END of the prompt.

Generate the prompt now:

7.2.5 Illustrative trace of constraint-triggered repair

Listing 1 shows a representative repair cycle triggered by an ontology constraint on measurement units. The ontology restricts temperature units to a controlled vocabulary that includes degree Celsius . When the instantiation agent first instantiates a temperature condition using the verbatim unit token C from text, the ontology-aware MCP tool rejects the input and returns a structured validation error listing allowed values. The agent then normalises the unit to an allowed entry and retries, producing a valid instance.

Listing 1: Representative MCP trace: ontology-constrained unit repair (C → degree Cel- sius).

Input evidence (paper text):

”… heated at 120 C for 12 h …”

Attempt 1 (verbatim unit token):

Tool: mop.create_temperature_condition

Input: { “context”: “reaction_step_3”,

“value”: 120,

“unit”: “C” }

Tool response (constraint violation):

{ “ok”: false,

“error_type”: “OntologyConstraintViolation”,

“field”: “unit”,

“message”: “Unit value ’C’ is not permitted by the ontology.”,

“allowed_values”: [“degree Celsius”, “kelvin”]

}

Repair (from tool feedback):

Map shorthand unit token “C” to the allowed vocabulary entry

“degree Celsius”.

Retry tool call with corrected unit.

Attempt 2 (normalised unit):

Tool: mop.create_temperature_condition

Input:

{ “context”: “reaction_step_3”,

“value”: 120,

“unit”: “degree Celsius”

}

Tool response (success):

{

“ok”: true,

“instance_iri”: “twa:TemperatureCondition_0f3a…”,

“validated”: true

}

7.3 Supplementary Evaluation Results

Tables 7.4–7.6 report per-paper extraction performance against the full ground truth for the three evaluated outputs: synthesis steps (Table 7.4), chemical building units under a formula-only matching criterion (Table 7.5), and characterisation entities (Table 7.6). For each paper (indexed by DOI), we provide true positives (TP), false positives (FP), false negatives (FN), and the derived precision, recall, and F1 scores, followed by an overall aggregate row.

Table 7.1: Benchmark profile (30 papers). Ground-truth positives are computed as TP + FN from Table 7.2.

Category Papers Ground-truth positives CBU 30 164 Characterisation 30 926 Steps 30 4940 Chemicals 30 675 Total 30 6705

7.3.1 Detailed instance visualisation for a UMC-1 synthesis

The Supplementary Information includes a detailed visualisation of one instantiated ontology subgraph corresponding to the synthesis of the metal–organic polyhedron UMC-1 (Supplementary Fig. 7.10). The figure shows the complete set of instantiated individuals and relations for this synthesis recipe under the OntoSyn and OntoMOPs ontologies, including typed synthesis steps, chemical inputs, and the resulting product entity, without the abstraction and condensation used in the main-text example (Fig. 1).

The graph was initially rendered from the RDF serialisation using an RDF-to-Graphviz visualisation tool [4] and then manually edited to adjust the layout for presentation and page fitting.

rdf2dot

Table 7.2: Overall evaluation summary by category (system output recovered from in- stantiated graphs vs. full ground truth). CBU denotes grounded/derived CBUs (Methods).

Category TP FP FN Precision Recall F1 CBU 128 40 36 0.762 0.780 0.771 Characterisation 669 102 257 0.868 0.722 0.788 Steps 4225 857 715 0.831 0.855 0.843 Chemicals 393 0 282 1.000 0.582 0.736 Micro-aggregate 5415 999 1290 0.844 0.808 0.826 Macro-average - - - 0.865 0.735 0.785

Table 7.3: Ablation study. CBU denotes grounded/derived CBUs (Methods).

Setting CBU F1 Char. F1 Steps F1 Chem. F1 Full system 0.771 0.788 0.843 0.736

  - external tools 0.000 0.610 0.800 0.732

  - constraint feedback 0.768 0.788 0.572 0.724

Table 7.4: Per-paper extraction results for synthesis steps against the full ground truth.

DOI TP FP FN Precision Recall F1 10.1002.anie.201811027 [11] 185 31 19 0.856 0.907 0.881 10.1002.anie.202010824 [12] 305 10 19 0.968 0.941 0.955 10.1002.chem.201604264 [30] 245 28 30 0.897 0.891 0.894 10.1002.chem.201700798 [20] 47 10 8 0.825 0.855 0.839 10.1002.chem.201700848 [10] 111 20 30 0.847 0.787 0.816 10.1021.acs.cgd.6b00306 [43] 35 10 4 0.778 0.897 0.833 10.1021.acs.chemmater.8b01667 [6] 105 35 24 0.750 0.814 0.781 10.1021.acs.inorgchem.4c02394 [54] 216 38 44 0.850 0.831 0.840 10.1021.acs.inorgchem.8b01130 [14] 207 74 70 0.737 0.747 0.742 10.1021.acsami.7b09339 [39] 117 35 40 0.770 0.745 0.757 10.1021.acsami.7b18836 [26] 80 24 6 0.769 0.930 0.842 10.1021.acsami.8b02015 [5] 118 28 37 0.808 0.761 0.784 10.1021.cg4018322 [41] 50 4 10 0.926 0.833 0.877 10.1021.ic050460z [21] 139 63 41 0.688 0.772 0.728 10.1021.ic402428m [29] 180 51 36 0.779 0.833 0.805 10.1021.ic501012e [52] 147 7 3 0.955 0.980 0.967 10.1021.ic802382p [42] 84 14 5 0.857 0.944 0.898 10.1021.ja042802q [51] 389 135 80 0.742 0.829 0.783 10.1021.ja105986b [67] 63 23 25 0.733 0.716 0.724 10.1021.jacs.7b00037 [62] 128 4 7 0.970 0.948 0.959 10.1021.jacs.8b10866 [63] 214 51 28 0.808 0.884 0.844 10.1039.C2CC34265K [9] 166 22 33 0.883 0.834 0.858 10.1039.C5CC05913E [2] 55 13 4 0.809 0.932 0.866 10.1039.C5DT04764A [66] 161 0 3 1.000 0.982 0.991 10.1039.C5RA26357C [7] 50 14 14 0.781 0.781 0.781 10.1039.C6CC04583A [65] 206 64 52 0.763 0.798 0.780 10.1039.C6DT02764D [64] 103 15 15 0.873 0.873 0.873 10.1039.C7CC01208J [61] 92 10 10 0.902 0.902 0.902 10.1039.C8DT02580K [13] 119 5 8 0.960 0.937 0.948 10.1039.D3QI01501G [59] 106 20 12 0.841 0.898 0.869 Overall 4223 858 717 0.831 0.855 0.843

Table 7.5: Per-paper formula-only extraction results for chemical building units (CBUs) against the full ground truth.

DOI TP FP FN Precision Recall F1 10.1021.acsami.7b18836 [26] 3 1 1 0.750 0.750 0.750 10.1039.C8DT02580K [13] 4 0 0 1.000 1.000 1.000 10.1002.chem.201700848 [10] 4 2 2 0.667 0.667 0.667 10.1021.acs.inorgchem.4c02394 [54] 8 2 2 0.800 0.800 0.800 10.1021.ic402428m [29] 8 0 0 1.000 1.000 1.000 10.1039.C5RA26357C [7] 2 0 0 1.000 1.000 1.000 10.1039.C6DT02764D [64] 2 2 2 0.500 0.500 0.500 10.1021.ic501012e [52] 4 0 0 1.000 1.000 1.000 10.1002.chem.201604264 [30] 8 2 2 0.800 0.800 0.800 10.1021.jacs.7b00037 [62] 1 3 3 0.250 0.250 0.250 10.1039.C7CC01208J [61] 2 2 2 0.500 0.500 0.500 10.1021.acschemmater.8b01667 [6] 6 2 0 0.750 1.000 0.857 10.1039.C5DT04764A [66] 3 3 3 0.500 0.500 0.500 10.1039.C6CC04583A [65] 10 0 0 1.000 1.000 1.000 10.1021.ic802382p [42] 2 0 0 1.000 1.000 1.000 10.1002.anie.202010824 [12] 7 3 3 0.700 0.700 0.700 10.1021.acs.inorgchem.8b01130 [14] 4 2 2 0.667 0.667 0.667 10.1039.C2CC34265K [9] 3 3 3 0.500 0.500 0.500 10.1021.acsami.7b09339 [39] 8 0 0 1.000 1.000 1.000 10.1002.anie.201811027 [11] 4 4 0 0.500 1.000 0.667 10.1021.acs.cgd.6b00306 [43] 0 2 2 0.000 0.000 0.000 10.1039.D3QI01501G [59] 2 0 2 1.000 0.500 0.667 10.1021.cg4018322 [41] 2 0 0 1.000 1.000 1.000 10.1039.C5CC05913E [2] 2 0 0 1.000 1.000 1.000 10.1021.ja105986b [67] 2 0 0 1.000 1.000 1.000 10.1021.ic050460z [21] 6 0 0 1.000 1.000 1.000 10.1021.jacs.8b10866 [63] 6 2 2 0.750 0.750 0.750 10.1021.ja042802q [51] 7 5 3 0.583 0.700 0.636 10.1021.acsami.8b02015 [5] 6 0 2 1.000 0.750 0.857 10.1002.chem.201700798 [20] 2 0 0 1.000 1.000 1.000 Overall 128 40 36 0.762 0.780 0.771

Table 7.6: Per-paper extraction results for characterisation entities against the full ground truth.

DOI TP FP FN Precision Recall F1 10.1002.anie.201811027 [11] 26 0 4 1.000 0.867 0.929 10.1002.anie.202010824 [12] 36 11 16 0.766 0.692 0.727 10.1002.chem.201604264 [30] 10 7 16 0.588 0.385 0.465 10.1002.chem.201700798 [20] 5 2 5 0.714 0.500 0.588 10.1002.chem.201700848 [10] 21 4 15 0.840 0.583 0.689 10.1021.acs.cgd.6b00306 [43] 8 1 1 0.889 0.889 0.889 10.1021.acs.chemmater.8b01667 [6] 36 3 10 0.923 0.783 0.847 10.1021.acs.inorgchem.4c02394 [54] 43 9 16 0.827 0.729 0.775 10.1021.acs.inorgchem.8b01130 [14] 38 1 11 0.974 0.776 0.864 10.1021.acsami.7b09339 [39] 29 1 11 0.967 0.725 0.829 10.1021.acsami.7b18836 [26] 17 2 2 0.895 0.895 0.895 10.1021.acsami.8b02015 [5] 19 2 8 0.905 0.704 0.792 10.1021.cg4018322 [41] 13 1 17 0.929 0.433 0.591 10.1021.ic050460z [21] 31 1 7 0.969 0.816 0.886 10.1021.ic402428m [29] 20 0 24 1.000 0.455 0.625 10.1021.ic501012e [52] 14 4 6 0.778 0.700 0.737 10.1021.ic802382p [42] 7 2 5 0.778 0.583 0.667 10.1021.ja042802q [51] 51 2 6 0.962 0.895 0.927 10.1021.ja105986b [67] 8 2 5 0.800 0.615 0.696 10.1021.jacs.7b00037 [62] 16 6 4 0.727 0.800 0.762 10.1021.jacs.8b10866 [63] 41 8 9 0.837 0.820 0.828 10.1039.C2CC34265K [9] 23 8 13 0.742 0.639 0.687 10.1039.C5CC05913E [2] 7 2 4 0.778 0.636 0.700 10.1039.C5DT04764A [66] 24 2 6 0.923 0.800 0.857 10.1039.C5RA26357C [7] 8 2 5 0.800 0.615 0.696 10.1039.C6CC04583A [65] 43 10 15 0.811 0.741 0.775 10.1039.C6DT02764D [64] 18 2 4 0.900 0.818 0.857 10.1039.C7CC01208J [61] 18 2 4 0.900 0.818 0.857 10.1039.C8DT02580K [13] 16 4 4 0.800 0.800 0.800 10.1039.D3QI01501G [59] 22 4 11 0.846 0.667 0.746 Overall 668 105 264 0.864 0.717 0.784

Table 7.7: Per-paper extraction results for chemicals against the full ground truth.

DOI TP FP FN Precision Recall F1 10.1002.anie.201811027 [11] 8 0 6 1.000 0.571 0.727 10.1002.anie.202010824 [12] 24 0 6 1.000 0.800 0.889 10.1002.chem.201604264 [30] 22 0 13 1.000 0.629 0.772 10.1002.chem.201700798 [20] 3 0 6 1.000 0.333 0.500 10.1002.chem.201700848 [10] 12 0 7 1.000 0.632 0.774 10.1021.acs.cgd.6b00306 [43] 5 0 3 1.000 0.625 0.769 10.1021.acs.chemmater.8b01667 [6] 7 0 1 1.000 0.875 0.933 10.1021.acs.inorgchem.4c02394 [54] 24 0 14 1.000 0.632 0.774 10.1021.acs.inorgchem.8b01130 [14] 7 0 25 1.000 0.219 0.359 10.1021.acsami.7b09339 [39] 5 0 1 1.000 0.833 0.909 10.1021.acsami.7b18836 [26] 11 0 3 1.000 0.786 0.880 10.1021.acsami.8b02015 [5] 7 0 4 1.000 0.636 0.778 10.1021.cg4018322 [41] 5 0 14 1.000 0.263 0.417 10.1021.ic050460z [21] 12 0 3 1.000 0.800 0.889 10.1021.ic402428m [29] 28 0 5 1.000 0.848 0.918 10.1021.ic501012e [52] 18 0 12 1.000 0.600 0.750 10.1021.ic802382p [42] 7 0 7 1.000 0.500 0.667 10.1021.ja042802q [51] 31 0 14 1.000 0.689 0.816 10.1021.ja105986b [67] 6 0 5 1.000 0.545 0.706 10.1021.jacs.7b00037 [62] 14 0 13 1.000 0.519 0.683 10.1021.jacs.8b10866 [63] 30 0 29 1.000 0.508 0.674 10.1039.C2CC34265K [9] 17 0 15 1.000 0.531 0.694 10.1039.C5CC05913E [2] 6 0 6 1.000 0.500 0.667 10.1039.C5DT04764A [66] 12 0 26 1.000 0.316 0.480 10.1039.C5RA26357C [7] 7 0 9 1.000 0.438 0.609 10.1039.C6CC04583A [65] 23 0 12 1.000 0.657 0.793 10.1039.C6DT02764D [64] 11 0 5 1.000 0.688 0.815 10.1039.C7CC01208J [61] 11 0 9 1.000 0.550 0.710 10.1039.C8DT02580K [13] 11 0 3 1.000 0.786 0.880 10.1039.D3QI01501G [59] 9 0 6 1.000 0.600 0.750 Overall 393 0 282 1.000 0.582 0.736

Figure 7.10: Detailed ontology instance subgraph for a UMC-1 synthesis.

Ontology-to-Tools Compilation — Supplementary Information (p.28-47)

7 Supplementary Information

7.1 Supplementary Note 1: Background and related work

7.2 Supplementary Methods

7.3 Supplementary Evaluation Results