ESGReveal: Modular Architectures for ESG Data

• ESGReveal is a modular framework that systematically extracts, structures, and integrates ESG data from diverse sources using standards-aligned templates and LLM-driven extraction.
• It employs retrieval-augmented generation and ontology-guided pattern matching to standardize ESG disclosures from corporate reports and news, ensuring traceability and quantitative rigor.
• The framework also incorporates ESG scores into quantitative asset pricing models, enabling optimized portfolio analytics and enhanced risk management in finance.

The ESGReveal Framework refers to a family of technically rigorous, modular architectures for the extraction, structuring, and quantitative incorporation of Environmental, Social, and Governance (ESG) data from disparate, often unstructured, sources. At present, three principal instantiations of the ESGReveal paradigm are established in the literature: (1) LLM- and RAG-driven structured extraction from corporate ESG reports (Zou et al., 2023), (2) semi-automatic ontology-based construction of principle-aligned knowledge graphs from ESG controversies in newsflow (Iwata et al., 13 Sep 2025), and (3) quantitative integration of ESG scores into dynamic asset pricing and portfolio optimization models (Lauria et al., 2022). These frameworks share a core focus on standardizing ESG data representation, aligning extracted data with regulatory or normative guidelines, and enabling downstream analysis or decision-making with explicit support for transparency, traceability, and quantitative rigor.

1. Modular Architecture and Core Components

The ESGReveal implementations are defined by a modular structure that formalizes both data ingestion and processing pipelines. Common architectural features include:

• Metadata Layer: Encodes ESG standards (HKEx, GRI, UN Global Compact) into structured, machine-interpretable templates or query objects (Zou et al., 2023, Iwata et al., 13 Sep 2025).
• Preprocessing and Parsing: Applies document layout analysis (e.g., LayoutLMv3, PubTableTransformer), text segmentation, summarization (e.g., mt5), and dense vector embedding (m3e) to decompose raw PDF/Word documents into indexed knowledge bases. For news sources, LLMs filter and flag candidate articles with ESG-relevance (Zou et al., 2023, Iwata et al., 13 Sep 2025).
• Retrieval or Pattern-Matching Engine:
  • In report-centric ESGReveal (Zou et al., 2023): retrieval-augmented generation (RAG) leverages vector similarity (cosine, coROM) to surface sections relevant to a structured query.
  • In controversy graph variants (Iwata et al., 13 Sep 2025): semi-automatic ontology-driven pattern recognition models (ESGViolationActionPattern) parameterize evidence extraction against pre-defined schema-aligned relational patterns.
• LLM Agent or Extraction Module: LLMs (GPT-4, GPT-4o mini, QWEN, ChatGLM) are prompted with context windows—including structured metadata, expert knowledge, and retrieval results—to output validated JSON or RDF triples (Zou et al., 2023, Iwata et al., 13 Sep 2025).
• Output Data Layer: Extracted indicators are formatted into structured records (JSON for quantitative KPIs; RDF/OWL for event–principle knowledge graphs), supporting downstream querying, analytic tooling, or integration into risk analytics (Zou et al., 2023, Iwata et al., 13 Sep 2025, Lauria et al., 2022).

These modules are orchestrated in a pipeline, enabling multi-modal ingestion (documents, newsflow), semi-automated evidence discovery, and standards-aligned structuring.

2. Methodologies for ESG Data Extraction and Structuring

ESG Disclosure Extraction from Corporate Reports

The canonical ESGReveal configuration for disclosure extraction operates on corporate filings (e.g., Hong Kong Stock Exchange ESG reports) as follows (Zou et al., 2023):

1. ESG Metadata Specification: Each reporting item is formalized as a 4-tuple: ⟨Aspect, KPI, Topic, Quantity⟩, with further <Knowledge> and <SearchTerm> extensions derived from reporting standards.
2. Report Preprocessing: Layout-aware models extract hierarchy and tabular data; text and table snippets are abstracted and vectorized for semantic retrieval.
3. Retrieval-Augmented Generation: Metadata queries generate embeddings for passage retrieval. Cosine similarity and coROM reranking identify top-k relevant passages as context for the LLM.
4. LLM Extraction: Prompts aggregate retrieval context, metadata guidance, and standardized output schemas (for instance, disclosure, KPI, value, unit, target, and key actions). The extraction is temperature=0 deterministic, ensuring repeatability.

ESG Controversy Alignment and Knowledge Graph Construction

For news-sourced, principle-aligned ESG event structuring (Iwata et al., 13 Sep 2025):

1. Candidate Corpus Selection: Negatively framed ESG news items are filtered through a dual-stage LLM pipeline, yielding a candidate corpus with high controversy density.
2. Lightweight Ontology Schema Construction: An RDFS/OWL-based meta-ontology is defined with explicit class hierarchies (e.g., ESGNewsArticle, ESGViolationActionPattern, UNGCPrincipleTypeEnumeration), supporting reasoning and extensibility.
3. Pattern Template Modeling: Each UNGC principle maps to three formal (Entity A, Action, Entity B) patterns, expressed as both JSON-LD instances (for pattern matching) and promoted to rdfs:Class subclasses (for typing new events).
4. Template Instantiation and Fact Extraction: LLMs extract named entities, perform evidence pattern matching (using "lookFor" and "ignore" lists), and, upon match, output contextually anchored RDF triples asserting principle violations.
5. Knowledge Graph Assembly: Nodes represent articles, named entities, violation events (type-aligned with templates), and UNGC principles; edges assert relationships such as :agent, :patient, and :violatesPrinciple.

Quantitative Integration with Asset Pricing

A distinct branch of ESGReveal applies to financial modeling (Lauria et al., 2022):

• ESG-Valued Returns: An affine combination of normalized ESG score and financial return,

$$\zeta_{i,t}(\lambda) = (1-\lambda) r_{i,t} + \lambda \frac{\varsigma_{i,t}}{c}$$

parameterizes the investor's ESG affinity ($\lambda$).

• Optimization Objective: Portfolio optimization is formulated as

$$\max_{w\in\mathbb{R}^I} \left(w'\mu - \frac{\gamma}{2}w'\Sigma w + \eta(w'\varsigma)\right)$$

where $\gamma$ is risk-aversion, $\eta$ is ESG affinity, and $\varsigma$ the vector of normalized ESG scores.

• Extensions: The same formalism extends to ESG-augmented capital market lines, ESG-valued risk measures, option pricing, and determination of a shadow riskless rate.

3. Formal Schemas, Patterns, and Ontological Alignment

Formalization is central to ESGReveal, enabling rigorous mapping between unstructured input, structured standards, and machine-actionable outputs.

Ontological Structures

• Core Classes: ESGNewsArticle, ESGRegulation, ESGViolationActionPattern, UNGCPrincipleTypeEnumeration, plus template subclasses (e.g., :ForcedLabourPattern).
• Formal Pattern Definitions: Each class is defined in OWL with explicit domain/range and RDFS subclass relations, allowing for logical inference and consistent annotation.
• Pattern Matching Formalism: Given a pattern $p_i$ with lookFor tokens $L_i$ and ignore tokens $I_i$, a sentence $s$ matches $\lambda$0 iff

$\lambda$1

• Alignment with Normative Frameworks: Patterns and templates can be explicitly aligned with third-party taxonomies (e.g., vendor controversy topics), via logic axioms such as

$\lambda$2

Quantitative Schemas

• Report Extraction Schema (JSON):
  • Disclosure: 0/1
  • KPI: string
  • Topic: string
  • Value: numerical
  • Unit: string
  • Target: (string/null)
  • Action: (string/null)
• Reporting Items: 70 total for HKEx Guide; 34 numerical, 36 textual, segmented across environmental, social, and governance dimensions (Zou et al., 2023).

4. Evaluation Protocols and Empirical Results

ESG Report Extraction

Extensive evaluation on ESG reports from HKEx-listed firms (166-company sample, 2,249 reports) employs manually labeled ground truth for both ESG disclosure flags and numeric extractions (Zou et al., 2023):

• Benchmarking LLMs:
  • GPT-4: $\lambda$3 data extraction accuracy ($\lambda$4), $\lambda$5 disclosure detection ($\lambda$6)
  • GPT-3.5, QWEN, ChatGLM: lower performance
• Ablation Study:
  • Enhanced preprocessing and expert metadata both provide significant boosts, confirming dependency on accurate layout parsing and domain-specific prompt guidance.

Ontology-Guided Controversy Extraction

• Manual Evaluation:
  • On 200 LLM-filtered, human-annotated article–triple pairs: ontology-guided extraction achieves $\lambda$7 accuracy and $\lambda$8 precision (vs. $\lambda$9 and $$\max_{w\in\mathbb{R}^I} \left(w'\mu - \frac{\gamma}{2}w'\Sigma w + \eta(w'\varsigma)\right)$$0 for one-shot GPT-4o mini baselines); recall $$\max_{w\in\mathbb{R}^I} \left(w'\mu - \frac{\gamma}{2}w'\Sigma w + \eta(w'\varsigma)\right)$$1 indicating selectivity vs. wide coverage (Iwata et al., 13 Sep 2025).
  • Per-principle precision ranges $$\max_{w\in\mathbb{R}^I} \left(w'\mu - \frac{\gamma}{2}w'\Sigma w + \eta(w'\varsigma)\right)$$2; recall penalty evident for nuanced violations.
• Transparency/Traceability: Each extracted event is linked to both the matching template and source passage, with class-level comments and provenance graph links for auditability.

Quantitative Finance

• Portfolio Experiments:
  • DJIA stocks, Refinitiv ESG scores: increasing ESG affinity ($$\max_{w\in\mathbb{R}^I} \left(w'\mu - \frac{\gamma}{2}w'\Sigma w + \eta(w'\varsigma)\right)$$3) steers optimal portfolios toward high-ESG constituents; efficient frontier shifts toward higher aggregate ESG with moderate reduction in pure financial return (Lauria et al., 2022).
  • ESG-valued riskless rates and derivatives can be defined using modified drift terms, supporting full integration into classical asset pricing theory.

5. Alignment with Standards, Explainability, and Extensibility

A central design goal is alignment of structured outputs to global sustainability standards (UN Global Compact, SDGs) and existing reporting frameworks (HKEx ESG Guide, GRI, SASB).

• Framework Flexibility: Patterns and templates are parameterized so that new normative requirements (e.g., CSRD, SFDR) can be added without major re-engineering (Iwata et al., 13 Sep 2025).
• Explainability and Ontological Anchoring: Use of RDFS class hierarchies, rdfs:comments (often LLM-generated), and fine-grained provenance links support downstream audit, interpretability, and integration into reasoning-based systems.
• Interoperability: Mapping facilities allow bridging between proprietary vendor datasets (e.g., RepRisk taxonomy) and open standard schemas.
• Scalability: Pattern templates can be expanded for new principles or regulatory motifs; modularity enables adoption in diverse linguistic and document contexts (future directions: multi-language support, VLM-based figure/table extraction) (Zou et al., 2023, Iwata et al., 13 Sep 2025).

6. Limitations, Open Challenges, and Future Directions

• Recall Limitations: Indirect, implicit, or highly nuanced ESG events are less well captured by existing pattern-based, LLM-guided extractions.
• Coverage Gaps: Current ESGReveal pipelines do not support pictorial or embedded graphics extraction—planned work seeks to integrate vision-LLMs for figure/table parsing (Zou et al., 2023).
• LLM Dependence: Extraction accuracy is sensitive to LLM selection and instruction prompt quality; model behavior varies across vendors and versions.
• Evolution of Standards: The modular pattern and template approach is designed for extensibility, but frequent updates to standards or emergence of new regulatory requirements may necessitate ongoing prompt and schema maintenance.
• Automatic Pattern Discovery: Future work may leverage LLM fine-tuning or bootstrapped semi-supervised techniques to generate new violation patterns or template classes automatically, reducing human pre-specification (Iwata et al., 13 Sep 2025).
• Fine-grained, Industry-Specialized Datasets: The development of more granular evaluation sets and benchmarking across industries will improve model robustness and reduce generalization error in diverse reporting contexts (Zou et al., 2023).

7. Impact, Applications, and Significance

The ESGReveal family delivers explicit, machine-checkable representations of ESG data for both compliance/risk monitoring (via ontological controversy graphs) and investment/portfolio analytics (via quantitative ESG scoring and asset pricing). Immediate use cases include:

• Automated ESG Compliance Validation: Principle-aligned extraction of negative events or controversies from newsflow with traceable provenance (Iwata et al., 13 Sep 2025).
• High-throughput Disclosure Analytics: Extraction and normalization of ESG indicators across large corpora of corporate reports, enabling cross-firm and cross-industry comparison, monitoring of disclosure rates, and identification of sectoral gaps (Zou et al., 2023).
• Risk/Return Optimization in ESG-Weighted Finance: Portfolio design, risk evaluation, and derivative valuation accounting for ESG affinity in a transparent, parameterized manner (Lauria et al., 2022).
• Standardization and Auditability: Explicit linkage to international guidelines (UNGC, HKEx) and support for semantic SPARQL/OWL query paradigms increase the trustworthiness and comprehensibility of ESG data pipelines.

A plausible implication is that as regulatory pressure and investor scrutiny around ESG intensify, frameworks of the ESGReveal class—which combine alignment, interpretability, and scalable extraction—are likely to become reference implementations both in academic research contexts and real-world auditable disclosure and compliance architectures.