FinReflectKG

# Dataset Card for FinReflectKG A comprehensive financial knowledge graph dataset extracted from S&P 500 companies' 10-K SEC filings spanning 2014-2024, containing **17.51 million normalized triplets** with full textual context. ## Dataset Details ### Dataset Description **Curated by:** [Domyn](https://www.domyn.com/) **Language(s):** English **License:** CC-BY-NC-4.0 FinReflectKG is a large-scale financial knowledge graph dataset that provides structured representations of financial relationships, entities, and temporal information extracted from regulatory filings. Each triplet represents a structured fact in the format `(entity, relationship, target)` with temporal bounds and rich contextual information. ### Dataset Sources **Paper:** [FinReflectKG: Agentic Construction and Evaluation of Financial Knowledge Graphs](https://arxiv.org/pdf/2508.17906) ## Key Features - **17.51M triplets** across 743 S&P 500 companies - **10+ years** of temporal coverage (2014-2024) - **Full text context** for each triplet - **Normalized entities** and relationships - **Temporal information** with start/end dates - **Rich metadata** including source document information ## Dataset Structure Each row contains a single triplet with the following fields: ### Core Triplet Components - `triplet_id`: Unique identifier for each triplet - `entity`: Named entity (normalized, e.g., "aapl") - `entity_type`: Entity category (ORG, PERSON, GPE, etc.) - `relationship`: Relationship type (normalized, e.g., "discloses", "operates_in") - `target`: Target entity (normalized) - `target_type`: Target entity category - `start_date`: Relationship start date (Month YYYY format) - `end_date`: Relationship end date (Month YYYY or "default_end_timestamp") - `extraction_type`: Extraction method ("default" or "extracted") ### Document Metadata - `ticker`: Company ticker symbol (e.g., "aapl", "msft") - `year`: Filing year - `source_file`: Original PDF filename - `page_id`: PDF page identifier - `chunk_id`: Text chunk identifier ### Context & Features - `chunk_text`: Full text context surrounding the triplet - `triplet_length`: Length of triplet text representation - `chunk_text_length`: Length of context text - `has_context`: Whether contextual text is available ## Temporal Information ### Temporal Fields Explained FinReflectKG employs sophisticated temporal extraction to capture when relationships occur or are valid. The temporal information consists of: #### Start Date & End Date - **Format**: "Month YYYY" (e.g., "January 2024", "March 2023") - **Purpose**: Captures the temporal validity period of each relationship #### Default Timestamps When explicit temporal information cannot be reliably extracted from the text context: - **`default_start_timestamp`**: Used when no explicit start date is mentioned or inferable - **`default_end_timestamp`**: Used when no explicit end date is mentioned or inferable **Examples**: - **Extracted**: "Environmental spending was approximately $21 million in fiscal 2014" → `start_date`: "January 2014", `end_date`: "December 2014" - **Default**: "Cintas depends on multiple suppliers" → `start_date`: "January 2014" (filing year), `end_date`: "default_end_timestamp" ### Extraction Type Classification The `extraction_type` field indicates the reliability and source of temporal information: #### "extracted" - **When**: Both start_date AND end_date are successfully extracted from the text - **Example**: Financial metrics with specific fiscal periods, dated announcements, time-bounded events #### "default" - **When**: Either start_date OR end_date (or both) uses default values - **Example**: General operational relationships, ongoing business activities without specified timeframes ## Example Usage ### Loading the Dataset ```python from datasets import load_dataset # Load the complete dataset dataset = load_dataset("domyn/FinReflectKG") # Access the data triplets = dataset["train"] print(f"Dataset size: {len(triplets):,} triplets") # Example triplet example = triplets[0] print(f"Entity: {example['entity']}") print(f"Relationship: {example['relationship']}") print(f"Target: {example['target']}") print(f"Context: {example['chunk_text'][:200]}...") ``` ### Filtering by Company ```python # Filter for Apple (AAPL) triplets apple_data = dataset["train"].filter(lambda x: x["ticker"] == "aapl") print(f"Apple triplets: {len(apple_data):,}") ``` ### Filtering by Time Period ```python # Filter for recent years (2022-2024) recent_data = dataset["train"].filter(lambda x: x["year"] >= 2022) print(f"Recent triplets: {len(recent_data):,}") ``` ### Filtering by Relationship Type ```python # Find all "discloses" relationships disclosure_data = dataset["train"].filter(lambda x: x["relationship"] == "discloses") print(f"Disclosure triplets: {len(disclosure_data):,}") ``` ## Entity Types The dataset includes various entity types relevant to financial documents: - **ORG**: Filing Company (The public company that is the subject of the 10-K filing) - **ORG_GOV**: Government bodies (e.g., United States Government) - **ORG_REG**: Domestic or international regulatory bodies (e.g., SEC, Federal Reserve, ECB) - **GPE**: Countries, states, or cities mentioned in geographic operations or risks - **PERSON**: Key individuals (e.g., CEO, CFO) - **COMP**: External companies referenced in the filing, including competitors, suppliers, customers, or partners - **PRODUCT**: Products or services (e.g., iPhone, AWS) - **EVENT**: Material events such as pandemics, natural disasters, M&A events - **SECTOR**: Sectors or industries relevant to the filer ORG or COMP entities (e.g., Technology, Healthcare, Financials) - **ECON_IND**: Quantitative metrics that reflect economic trends or conditions (e.g., Inflation Rate, GDP Growth, Unemployment Rate, Interest Rate, Consumer Confidence Index) - **FIN_INST**: Tradable financial assets or liabilities (e.g., bonds, derivatives, options) - **FIN_MARKET**: Financial indices and market dynamics (e.g., S&P 500, Dow Jones) - **CONCEPT**: Abstract concepts or theories (e.g., Artificial Intelligence, Digital Transformation, Circular Economy) - **RAW_MATERIAL**: Essential raw materials (e.g., Lithium) - **LOGISTICS**: Supply chain and logistics entities (e.g., Ports) - **ACCOUNTING_POLICY**: Key accounting policies (e.g., revenue recognition, lease accounting, goodwill impairment) - **RISK_FACTOR**: Documented risks (e.g., market risk, supply chain risk, cybersecurity risk, geopolitical risk) - **LITIGATION**: Legal disputes or proceedings including lawsuits, regulatory investigations, or settlements disclosed in the 10-K - **SEGMENT**: Internal divisions or business segments of the filer ORG (e.g., Cloud segment, North America retail) - **FIN_METRIC**: Financial metrics or values (e.g., Net Income, EBITDA, Long-Term Debt, CapEx, R&D Expense) - **ESG_TOPIC**: Environmental, Social, and Governance themes (e.g., Carbon Emissions, DEI, Renewable Energy, Climate Risk) - **MACRO_CONDITION**: Qualitative or thematic macroeconomic trends that affect the company or industry (e.g., Recession, Inflationary Pressures, Tightening Monetary Policy, Economic Uncertainty, Labor Shortages) - **REGULATORY_REQUIREMENT**: Specific regulations or legal frameworks (e.g., Basel III, SEC rules, GDPR) - **COMMENTARY**: Statements or disclosures from company management (e.g., outlooks, explanations, guidance) ## Relationship Types The dataset includes comprehensive relationship types for financial knowledge graphs: - **Has_Stake_In**: Indicates full or partial ownership or equity interest (e.g., ORG owns SEGMENT or has stake in COMP) - **Announces**: Publicly discloses or communicates (e.g., ORG announces PRODUCT launch or ESG_TOPIC initiative) - **Operates_In**: Indicates operational geography or market presence (e.g., ORG operates in GPE) - **Introduces**: Rolls out or implements a new product, policy, or segment (e.g., ORG introduces PRODUCT or ACCOUNTING_POLICY) - **Produces**: Manufactures or develops a product or service (e.g., ORG produces PRODUCT) - **Regulates**: Exerts control or regulatory oversight (e.g., ORG_REG regulates ORG) - **Involved_In**: Specifies direct involvement in an event such as a merger, acquisition, or litigation - **Impacted_By**: Indicates that the entity was materially affected by a major event (e.g., Amazon impacted_by COVID-19) - **Impacts**: Specifies the broad influence or effect an entity or event has on financial performance, market trends, or other key outcomes - **Positively_Impacts**: Contributes to positive outcomes (e.g., ESG_TOPIC positively impacts FIN_METRIC) - **Negatively_Impacts**: Contributes to adverse outcomes (e.g., RISK_FACTOR negatively impacts FIN_METRIC) - **Related_To**: Indicates a connection or relationship - **Member_Of**: Indicates formal affiliation or group membership (e.g., COMP is member of FIN_MARKET index) - **Invests_In**: Allocates financial or strategic capital (e.g., ORG invests in COMP or ESG_TOPIC) - **Increases**: Denotes a growth or rise in value or activity (e.g., ORG increases CapEx or debt) - **Decreases**: Denotes a decline in value or activity (e.g., ORG decreases headcount or R&D spend) - **Depends_On**: Requires support or shows reliance on another entity (e.g., ORG depends on RAW_MATERIAL or COMP) - **Causes_Shortage_Of**: Indicates supply constraint driven by an event or condition (e.g., EVENT causes shortage of RAW_MATERIAL) - **Affects_Stock**: Indicates direct influence on stock price or valuation (e.g., RISK_FACTOR affects_stock of ORG) - **Stock_Decline_Due_To**: Specifies factor causing drop in stock price (e.g., MACRO_CONDITION stock_decline_due_to ORG) - **Stock_Rise_Due_To**: Specifies factor causing increase in stock price (e.g., ESG_TOPIC stock_rise_due_to ORG) - **Market_Reacts_To**: Indicates market response to external events or disclosures (e.g., FIN_MARKET market_react_to EVENT) - **Discloses**: Reveals or reports (e.g., ORG discloses FIN_METRIC, ESG_TOPIC, or RISK_FACTOR) - **Faces**: Encounters legal or regulatory challenges (e.g., ORG faces LITIGATION or REGULATORY_REQUIREMENT) - **Guides_On**: Provides management commentary or forecast (e.g., COMMENTARY guides_on FIN_METRIC or MACRO_CONDITION) - **Complies_With**: Meets regulatory or policy requirements (e.g., ORG complies_with REGULATORY_REQUIREMENT) - **Subject_To**: Indicates entity is governed or affected by (e.g., ORG subject_to ACCOUNTING_POLICY or REGULATORY_REQUIREMENT) - **Supplies**: Indicates vendor or supplier relationship (e.g., COMP supplies RAW_MATERIAL to ORG) - **Partners_With**: Indicates formal or strategic collaboration (e.g., ORG partners_with COMP for PRODUCT) ## Data Quality The dataset has undergone extensive cleaning and validation: - **99.08% clean dates** in proper "Month YYYY" format - **Normalized entities** and relationships using lemmatization - **Deduplicated triplets** - **Filtered invalid data** - **Comprehensive validation** of data structure and integrity ## Temporal Coverage **📊 TOTAL TRIPLETS: 17,513,372** **📅 YEAR RANGE: 2014-2024** **🏢 COMPANIES: 743** | Year Range | Triplet Count | Exact Count | Companies | |------------|---------------|-------------|-----------| | 2014-2018 | 7.55M | 7,549,552 | 743 | | 2019-2021 | 5.04M | 5,043,004 | 743 | | 2022-2024 | 4.92M | 4,920,816 | 743 | | **TOTAL** | **17.51M** | **17,513,372** | **743** | ## Applications This dataset is suitable for: ### Research Applications - **Financial NLP**: Named entity recognition, relation extraction - **Knowledge Graph Construction**: Building financial knowledge bases - **Temporal Analysis**: Studying financial relationships over time - **Risk Assessment**: Analyzing risk factors and their evolution - **Compliance Research**: Understanding regulatory relationships ### Industry Applications - **Financial Intelligence**: Automated analysis of SEC filings - **Due Diligence**: Comprehensive company relationship mapping - **ESG Analysis**: Environmental, social, governance insights - **Market Research**: Understanding competitive landscapes - **Regulatory Technology**: Compliance and risk monitoring ## Data Source The dataset is derived from S&P 500 companies' 10-K annual reports filed with the SEC. ## Ethical Considerations - **Public data**: All source data is publicly available SEC filings - **No personal information**: Focus on corporate and financial entities - **Regulatory compliance**: Respects SEC disclosure requirements - **Research use**: Intended for academic and research purposes ## Citation If you use this dataset in your research, please cite: ```bibtex @article{arun2025finreflectkg, title={FinReflectKG: Agentic Construction and Evaluation of Financial Knowledge Graphs}, author={Arun, Abhinav and Dimino, Fabrizio and Agarwal, Tejas Prakash and Sarmah, Bhaskarjit and Pasquali, Stefano}, journal={arXiv preprint arXiv:2508.17906}, year={2025}, url={https://arxiv.org/pdf/2508.17906} } ``` ## Dataset Creation For detailed information about the dataset creation process, methodology, and evaluation, please refer to the attached paper: [FinReflectKG: Agentic Construction and Evaluation of Financial Knowledge Graphs](https://arxiv.org/pdf/2508.17906). ## Dataset Card Contact Reetu Raj Harsh ([e-mail])