Connectivity platform for sustainability-oriented themes and ESG factors

Abstract

A computer platform provides an end user with an interactive, color-coded user interface that allows the end user to understand and conceptualize, in an easily and readily understood and efficient and effective manner, how different sustainability themes and ESG factors will impact financial and accounting metrics for investable companies by a grouping, such as an industry or sector grouping. The user interface can also show connectivity logic as to how each sustainability theme will impact the financial and accounting metrics. The user interface can link to the most relevant excerpts of corporate disclosure documents for companies in the grouping to show how those companies are dealing with, or explaining the impact of, the sustainability theme. The relevant excerpts can be curated by NLP processing and / or scoring systems trained through machine learning.

Description

PATENTDocket No. 242206PCTIN THE UNITED STATES RECEIVING OFFICEPatent Cooperation Treaty (PCT) Application For:CONNECTIVITY PLATFORM FOR SUSTAINABILITY-ORIENTED THEMES AND ESG FACTORSInventors: Martin Jarzebowski (Pittsburgh, PA); Luke Fleisch (Pittsburgh, PA); Daniel Clymer (Pittsburgh, PA); and Frank Amato (Wexford, PA)PRIORITY CLAIM

[0001] The present application claims priority to United States provisional patent application Serial No. 63 / 727,800, filed December 4, 2024, titled “Connectivity Platform for Sustainability-Oriented Themes and ESG Factors.”FIELD

[0002] This disclosure relates generally to the field of investment management and more particularly relates to sustainability in relation to financial analysis and / or accounting metrics.BACKGROUND

[0003] Sustainability themes and environmental, social, and governance (ESG) criteria can be used as an informational input when researching and evaluating a corporate issuer. These factors can help a company, its investors, regulators and various stakeholders better understand potential risks and opportunities. The environmental criteria consider how a company may be exposed to different physical risks or may possibly benefit from new product innovations. The social criteria examine how a company manages relationships with suppliers, customers, and employees. The governance criteria deal with the practices and internal controls established by a company’s leadership.

[0004] Incorporating thematic and ESG criteria into investment decisions can be technically challenging due to several factors. There are data availability and quality issues. For example, not all companies disclose sustainability information, and there can be significant gaps, especially for smaller companies or those in emerging markets. Also, ESG data are often not standardized, leading to inconsistencies in how information is reported and measured across companies, industries, and jurisdictions. Also, limited historical sustainability data can make it difficult to analyze trends and assess long-term performance.

[0005] Different third-party ESG rating agencies use different methodologies and1603975172.8Docket No. 242206PCT metrics, resulting in varied and sometimes conflicting ratings for the same company. Still another issue is that many sustainability themes and ESG criteria are qualitative and subjective, making it difficult to measure and evaluate them objectively.

[0006] Another issue is comparability. There are multiple voluntary sustainability reporting frameworks (e.g., the Global Reporting Initiative (GRI), the International Sustainability Standards Board (ISSB), the Sustainability Accounting Standards Board (SASB), and the Task Force on Climate-related Financial Disclosures (TCFD)) that companies can choose from, leading to difficulties in comparing data across different entities. There are also fragmented regulatory and market dynamic issues. ESG-related regulations are continuously evolving across different regions. Keeping up with the latest requirements and standards in different countries can be challenging.

[0007] Additionally, combining sustainability themes and ESG data with traditional financial metrics can be complex and require sophisticated models and tools. Sustainability issues are not a one-size-fits-all approach impacting all economic sectors, industries, and companies in the same manner. Some sustainability themes may be more or less financially relevant for any given industry. For example, a biotechnology company is a different business model when compared with an automobile manufacturer and may be exposed to very different ESG risks and opportunities. Also, investors, stakeholders, and companies may have different time horizons, making it difficult to align all corporate sustainability information with certain methods of financial and valuation analysis performed by investors.

[0008] These and other issues make it technically challenging to incorporate thematic research and ESG criteria into financial assessments, valuation methods, and investment decisions.SUMMARY

[0009] In one general aspect, the present invention is directed to a computer platform, referred to herein sometimes as a financial connectivity platform (FCP), that, through a client-server architecture, provides an end user with an interactive, color-coded user interface that allows the end user to understand and conceptualize, in a easily and readily understood and efficient and effective manner, how different sustainability themes and ESG factors will impact financial and accounting metrics for investable companies by a grouping, such as an industry or sector grouping. The FCP’s user interface can also show economic scenarios and21603975172.8Docket No. 242206PCT connectivity logic as to how, when and where each sustainability theme will impact the financial model and underlying accounting metrics. Still further, the FCP’s user interface can link to the most relevant excerpts of corporate disclosure documents and information for companies in the grouping (e.g., industry) to show how those companies are dealing with, or explaining the impact of, the sustainability theme. The relevant excerpts can be curated by natural language processing (NLP) and / or scoring systems trained through machine learning. The system can quantify, visualize, and qualify how key sustainability themes affect specific financial statement and accounting metrics. The system can also provide company specific insights, industry specific insights, and portfolio specific insights into the where (which industry), how (which accounting metric) and when (time horizon) key sustainability themes (the what) connect to unique financial accounting categories, customized by industry and company. These and other benefits that can be realized through various embodiments of the present invention will be apparent from the description that follows.FIGURES

[0011] Unless specified otherwise, the accompanying drawings illustrate aspects of the innovations described herein. Referring to the drawings, wherein like numerals refer to like parts throughout the several views and this specification, several embodiments of presently disclosed principles are illustrated by way of example, and not by way of limitation. The drawings are not intended to be to scale. A more complete understanding of the disclosure may be realized by reference to the accompanying drawings in which:

[0012] Figure 1 illustrates a computer architecture in which embodiments of the present invention can be activated.

[0013] Figure 2a illustrates a user interface for a financial connectivity platform, according to various embodiments.

[0014] Figure 2b illustrates a user interface with a pop-up window for a financial connectivity platform, according to various embodiments.

[0015] Figure 3 is a diagram of the Al computer system according to various embodiments.

[0016] Figure 4 is a diagram of a process flow performed by the Al computer system according to various embodiments of the present invention.31603975172.8Docket No. 242206PCT

[0017] Figure 5 is a diagram of a taxonomic mapping between a sustainability theme dictionary and connectivity logic according to various embodiments of the present invention.

[0018] Figure 6 illustrates a user interface for the financial connectivity platform, according to an embodiment of the present disclosure.

[0019] Figure 7 illustrates an example of financial connectivity logic for accounting metrics and time horizons, according to an embodiment of the present disclosure.

[0020] Figure 8 illustrates SCORE connectivity framework, according to an embodiment of the present disclosure.

[0021] Figure 9 illustrates a visualization of thematic prevalence, according to an embodiment of the present disclosure.

[0022] Figure 10 illustrates the architecture for creating connectivity logic, according to an embodiment of the present disclosure.

[0023] Figure 11 shows a transformer based LLM model, which is a type of neural network architecture designed for natural language processing (NLP) tasks and which could be used for the LLM.DESCRIPTION

[0024] The present invention relates, in one general, high-level, aspect, to financial relevance and investment connectivity of sustainability-oriented themes and ESG factors in an informational and analytical system which quantifies, visualizes, and qualifies how a broad range of sustainability-oriented themes (sometimes referred to herein simply as “sustainability themes” or “ESG sustainability themes”) directly connect to traditional financial and / or accounting metrics. In one embodiment, a financial connectivity platform (FCP) systematically links sustainability themes to the financial statements of companies customized for different industries.

[0025] Companies can be grouped into industries, sectors, or other relevant groupings based on their primary business activities, as explained further below. For each of the industry groupings, the FCP can make financial linkages between numerous sustainability themes and unique accounting metrics for the companies (categorized by specific types of assets, liabilities, revenues, expenses, and cash flows). For each of the accounting metrics, a41603975172.8Docket No. 242206PCT unique indicator of a time horizon (e.g., short, medium, long) and magnitude (high, medium, low) can be assigned using a methodology. Thus, each sustainability theme can have a particular time horizon and magnitude for each particular accounting metric. The FCP can also provide economic scenarios and logic statements (e.g., connectivity logic) for each of the financial and accounting connections that are fully transparent in a visual platform. In addition, the FCP can provide real-world examples, evidenced by corporate issuer disclosures (e.g., Securities and Exchange Commission (“SEC”) filing excerpts or any other global corporate disclosure repository) that are fully documented in the visual platform. The FCP can identify, store, and visualize the most relevant corporate issuer disclosures for each sustainability theme, for each industry, using natural language processing (NLP) and machine learning.

[0026] In that connection, the FCP can provide a systematic and evidence-based quantification of where (which industry), how (which accounting metric), and when (time horizon) key sustainability themes (the what) connect to unique financial accounting categories, customized by industry and company.

[0027] The FCP can serve, therefore, as a unique assembly and combination of qualitative and quantitative data. It can rely on a taxonomy, ontology, database, and visual user interface to facilitate the understanding of where, how, and when financial connections occur for a given industry and company. This informational and analytical platform allows a user to better understand and contextualize - in an easily usable and understandable manner - the financially connected risks and opportunities presented for a given industry and company.

[0028] The FCP can incorporate company-specific insights such as qualitative and quantitative company research, company performance metrics, company quantitative scoring, company financial modeling and analysis, company valuation modeling and analysis, company data visualization, company stewardship and engagement record keeping, and company reporting. The financial connectivity platform can also include industry-specific insights such as industry data analytics and visualizations, industry thematic research, industry performance metrics, industry quantitative scoring, industry thematic disclosure aggregation and analysis, industry stewardship and engagement record keeping, and industry reporting. The financial connectivity platform includes portfolio-specific insights such as investment portfolio analysis, portfolio data visualizations, portfolio thematic research, portfolio performance metrics, portfolio stewardship and engagement record keeping, and51603975172.8Docket No. 242206PCT portfolio reporting.

[0029] With reference now to the figures, Figure 1 illustrates a computer architecture for the FCP 8 in which embodiments of the present invention can be activated. The computer architecture includes a back-end computer system 10 in communication with a user at a client device 12 via an electronic data network 14, such as the Internet, a LAN, a WAN, etc. The client device 12 may be any computer device with a browser or app (e.g., mobile app for a mobile client device) for accessing and interacting with web pages served from the back-end computer system. For example, the client device 12 could be a desktop computer, a laptop, or a mobile device (e.g., smartphone or tablet).

[0030] The back-end computer system 10 may comprise a web server 16 for serving web pages, as described herein, to the client device 12. The web server 16 may communicate with the client device 12 via HTTP (HyperText Transfer Protocol) and / or HTTPS (HTTP Secure) protocol, for example. The client device 12 can establish a TCP (Transmission Control Protocol) connection with the web server 16 using the server’s IP address and the port number (typically port 80 for HTTP and port 443 for HTTPS). Once the connection is established, the client device 12 sends an HTTP request to the web server 16. This request can include a request line, which can specify the HTTP method (e.g., GET, POST), the URL, and the HTTP version; headers, which provide additional information about the request (e.g., Host, User-Agent, Accept); and an optional body that contains data to be sent to the web server 16, used with methods like POST. The web server 16 receives the request, processes it, and determines how to respond. This may involve, as explained herein, retrieving static files (HTML, CSS, images); running server-side scripts or applications (e.g., PHP, Python, Node.js); and / or querying a database. To send an HTTP response, the web server 16 can send an HTTP response back to the client device 12. This response can include a status line, which can include the HTTP version, a status code (e.g., 200 OK, 404 Not Found), and a reason phrase; headers, which can include additional information about the response (e.g., Content-Type, Content-Length); and a body, e.g., the requested resource or an error message. When using HTTPS, the communication typically involves additional steps for security. For example, before sending the HTTP request, the client device 12 and the web server 16 can perform an SSL / TLS handshake to establish a secure encrypted connection. This can involve exchanging certificates, generating encryption keys, and agreeing on encryption methods. Once the secure connection is established, all HTTP requests and responses are encrypted,61603975172.8Docket No. 242206PCT ensuring data privacy and integrity.

[0031] Embodiments of the present invention could also be used where the client device 12 is a mobile device with a mobile app for accessing the FCP 8. Data in this context can be exchanged in JSON or XML instead of HTML. The app can interact with the server via API requests, typically using HTTP methods like GET, POST, PUT, DELETE. These requests are structured to retrieve data (e.g., user information, product details) or to send data (e.g., user login credentials, form submissions). The app manages the presentation and navigation of data received from the server within its own user interface.

[0032] The back-end computer system 10 can comprise one or more networked servers, and it can comprise a centralized repository 22, as shown in Figure 1, for storing structured and unstructured data. For example, the centralized repository 22 can store, in electronic format (e.g., PDF, word processing, spreadsheet, XML) source documents for the various companies. The source documents can comprise corporate disclosures and filings for publicly-listed or private companies in any country globally, news, unstructured text, audio, video, and other publicly available information. For example, the source documents could include Securities and Exchange Commission (SEC) filings or similar filings in other countries for a specific corporate issuer, such as annual reports, 10-K filings, 8-K filings or supplemental filings, which can be indexed in the centralized repository 22 by sector, industry, company, year and date. The source documents could also comprise, in various embodiments and / or to the extent available, CSR (corporate sustainability) reports, published patents and patent applications, government reports such as EPA (Environmental Protection Agency) reports, non-governmental organization reports such as CDP (carbon disclosure project) reports, SBTi (Science-Based Targets initiative) reports, company website disclosures, social media, aggregated news reports, transcripts of public statements, and / or copies and summarizations of correspondence. The source documents may include different sources and different modalities.

[0033] In some embodiments, the centralized repository 22 may be included in the data storage 29.

[0034] CSR reports are documents published by companies to communicate their efforts in managing the social, environmental, and economic impacts of their operations. These reports typically cover a range of topics such as environmental management, social equity,71603975172.8Docket No. 242206PCT ethical business practices, community engagement, business risks and economic performance. The main goal is to provide stakeholders, including investors, customers, employees, and regulators, with transparent information on the company’s sustainability initiatives and performance. Reports may include details on green house gas (GHG) emissions, energy efficiency projects, governance practices, and community development programs.

[0035] Governmental reports are official documents produced by government agencies or departments at the local, state, federal, or international level. They present findings, data, and analysis on a range of themes and topics and may include policy and legislative evaluations, financial statements, and statistical data. One example of a governmental report is EPA reports; the EPA (a U.S. government agency) publishes reports and data on environmental policies, regulations, and enforcement actions to protect human health and the environment. These reports cover various environmental issues such as air and water quality, hazardous waste management, climate change, and toxic substances control. The EPA’s reports aim to inform the public and policymakers about the state of the environment, regulatory actions, and the effectiveness of environmental programs. Examples include the annual “Inventory of U.S. Greenhouse Gas Emissions and Sinks” and “National Water Quality Inventory Report.”

[0036] CDP is a global non-profit organization that runs a disclosure system for companies, sectors, industries, and regions to manage their environmental impacts. CDP collects data on emissions, climate change, water security, and deforestation. Companies report on their greenhouse gas emissions, climate risks, water usage, and deforestation impacts. The aim is to inform investors, companies, and governments on climate themes by providing high-quality environmental data at the industry and corporate issuer level. Companies report via CDP questionnaires, which are then compiled into reports and scorecards that assess environmental and governance performance and transparency.

[0037] SBTi is a collaboration between CDP, the United Nations Global Compact, World Resources Institute (WRI), and the World Wide Fund for Nature (WWF). It helps companies set greenhouse gas emission reduction targets in line with the latest climate science. SBTi reports detail companies’ commitments and progress towards reducing their emissions in accordance with the Paris Agreement goals, typically targeting a limit of global warming to 1.5°C or well below 2°C above pre-industrial levels. The initiative aims to ensure that corporate targets are science-based and contribute to global efforts to mitigate climate81603975172.8Docket No. 242206PCT change. SBTi reports might include verification of companies’ targets, progress updates, and case studies demonstrating best practices in emissions reduction.

[0038] The centralized repository 22 can be implemented with cloud-based data storage and / or on-premises data storage. Cloud-based data storage could be implemented via Amazon S3, Azure Data Lake Storage and / or Google cloud storage, for example. Onpremises data storage could be implemented with a distributed file system, such as Apache Hadoop HDFS. The back-end computer system 10 can use web-scraping tools and libraries to collect and store the source documents periodically and / or from time to time as they become available. The back-end computer system 20 can execute scripts to navigate to the target web pages for the source documents, extract the necessary data, and handle pagination if needed. A data pipeline to ingest the scraped data into the centralized repository 22 can be used, such as JSON, CSV, or Parquet, for example. Metadata for the source documents can be stored in the centralized repository 22, thereby making it easier to search and retrieve relevant source documents.

[0039] The back-end computer system 10 may also comprise on-premises data storage and / or cloud-based data storage 29 to store other relevant data files, such data files for: the industry sector / group 36 for each investable company covered by the FCP 8 the magnitudes for each sustainability theme for each industry 30; the time horizons for each accounting factor for each sustainability theme for each industry 32; and the connectivity logic for each sustainability theme for each industry 34.

[0040] The back-end computer system 10 may also comprise an Al computer system 40 for, in various embodiments and among other things, determining which paragraphs of a company’s source documents are most relevant to the various sustainability themes. More details about potential embodiments, features and capabilities of the Al computer system 40 are described further below.

[0041] For example, when a user searches for and opens a specific company in the FCP 8 the textual disclosure excerpts pertaining to a given theme are visually displayed in the platform along with filing type, year, and page number. These are determined using the machine learning model to identify and extract company disclosures to provide one dimension of the financially connected context of a sustainability theme.

[0042] For example, a user can create a custom scoring system based on the prevalence91603975172.8Docket No. 242206PCT of a specific theme being disclosed in each industry. The FCP 8 can have a gauge which shows the rate of change (historical time series) in disclosure for a particular theme by a specific company. The FCP 8 will also have a flag which signals if a company is currently a leader or laggard as it pertains to the level of transparency on a given theme relative to its industry peers. In addition, the manner in which the machine learning process was designed, an interactive user prompt will allow the ability to search for a greater range of themes / topics.

[0043] Figure 2a illustrates a user interface 100 that the FCP 8 backend computer system 10 can provide to the client device 12, according to various embodiments. As shown in the example of Figure 2a, the UI 100 quantifies and visually displays, in a readily intuitive and easily usable manner, where, how, and when a sustainability theme and topic connect to various financial accounting categories and specific financial statement line items, each of which may be customized by industry and company. The UI 100 shown in Figure 2a is for a particular company (e.g., “Company X”) that the user entered in a search query, for example, or otherwise activated. Search query and activation can include, but not limited to an entity name or unique identifier (e.g. symbol, ISIN, CUSIP, etc.) The possible corporate issuers that can be entered can be limited to publicly traded companies, private companies, government-sponsored entities, or a sub-category thereof, such as publicly traded companies in a particular geographic region(s) or private companies covered by a particular entity or organization using the FCP 8. The data storage 29 can store a corresponding industry grouping for each company.

[0044] The UI includes several fields that are described further below. The UI 100 includes a sustainability theme field or pane 110, which lists the applicable sustainability themes. A magnitude pane 102 identifies an industry-specific magnitude (e.g., high, medium, low) for each of the sustainability themes. The magnitudes for each industry / grouping can be stored in the data file 30 and in the data storage 29. An accounting metrics pane 104 can include a grid that shows, for each theme, and for the company’s particular industry, the time horizon (short, medium, long) for various company accounting / financial metrics. In various embodiment, a short time horizon means less than 1 year; a medium time horizon means greater than 1 year and less than 5 years; and a long time horizon means more than 5 years. Five accounting / financial metrics are shown in the example of Figure 2a: Assets, Liabilities, Revenue, Expenses and Cash Flow. In other embodiments, different accounting / financial metrics are used. For example, the UI can expand to showcase the time horizon and financial101603975172.8Docket No. 242206PCT connectivity for multiple types of underlying Assets, Liabilities, Revenues, Expenses and Cash Flows. In this example, the GHG emissions theme has a high magnitude and a medium time horizon for a company’s underlying asset line items. Underlying asset categories may include tangible assets, intangible assets and financial assets. This means that GHG emissions will have a high financial magnitude for companies in Company X’s industry in 1- 5 years (the medium time horizon). As another example, the Digital Rights theme has a medium magnitude rating and a short time horizon for Liabilities, which indicates that digital rights will have medium magnitude on specific underlying liabilities for companies in Company X’s industry in the next year.

[0045] The industry-specific magnitudes, time horizons, and connectivity logic for each sustainability theme stored in the data files 30, 32, 34, 36 can be determined, for example, by subject matter experts, rule-based systems, and / or machine learning about factors that companies in each industry face. Financial magnitude is determined through a multidimensional process that incorporates disclosed corporate metrics that align with elements of the thematic taxonomy, thematically aligned factor performance analysis, corporate disclosure aggregation and analysis, and subject matter expertise. Through this systematic process, a broad global universe of corporate issuers is separated into industry classifications and subjected to an analysis consisting of some or all of the previously listed components. This analysis results in a theme level scoring system that allows for sorting of underlying industries into high, medium, and low magnitude groupings.

[0046] Impact timing is assigned to each underlying accounting metric that has an established financial connection. To assess the timing of the metric impact, the applicable SCORE 86 stakeholder scenario logic 600 attached to the metric is assessed. Since SCORE 600 is formulated on specific scenarios, it is those scenarios that are isolated and assigned a timing of short, medium, or long. The timing assessment is based on historically observed timing impacts for like scenarios (e.g. a known financial penalty is generally charged to the income statement as a non-operating expense within 12 months of it being handed down - a short time horizon; investment in a new production facility in the form of capital expenditures and physical property, plant and equipment assets may take a longer time to materialize, 1-5 years - a medium time horizon). These financial scenarios, connectivity logic, and time horizons are fully documented, stored 29 and part of the user visual interface 100.111603975172.8Docket No. 242206PCT

[0047] In an example grid style interface, the sustainability themes can be listed on the left side of the UI 100, and the corresponding magnitude and time horizons for each sustainability theme can be listed in the corresponding row for each sustainability theme. Also, the grid can be color coded. High, medium and low in the magnitude field 102 can each have a different color or hue or accent (such as different shades of blue for each of high, medium, and low). Similarly, the short, medium, and long time horizon indicators in the accounting metrics pane can have correspondingly different colors, hues or accents (such as blue for high, orange for medium, and green for short, in one example). Note that the example of Figure 2a shows only six different sustainability themes. There could be many more applicable sustainability themes, such as the 36 themes as shown below in Table 2 in some embodiments, and a user of the FCP 8 could scroll up and down the list by, for example, clicking and dragging a scroll bar 113, using a mouse wheel, touch interactions, etc.

[0048] The user interface 100 preferably includes interactive UI components that allow a user of the user interface 100 to access additional information pertaining to what is shown in the user interface 100. The interactive UI components could be, for example, hyperlinks, buttons (e.g., clickable elements that trigger specific actions), drop-down menus, accordions (expandable sections that reveal or hide content when clicked), collapsible sections (similar to accordions, collapsible sections can be expanded or collapsed to show or hide additional information), tabs (e.g., clickable elements that switch between different content panels or views within a single page), tab panels (the content sections displayed based on the selected tab), tooltips (small, temporary text boxes that appear when a user hovers over an element, providing additional context or information), popovers (similar to tooltips, but often larger and more detailed, they can contain images, links, or other interactive elements), pop-up window (similar to popover, but can be triggered by various actions, such as clicking a button or link; they may remain visible until closed by the user), and / or modals (e.g., overlay windows that appear on top of the current page). In the example of Figure 2a, the interactive UI components include the “view more” hyperlinks in the connectivity logic pane 108 and the disclosure examples pane 112. When the user activates any of these interactive UI components, additional information is provided to the user corresponding to what the user selected. For example, when the user activates one of the “view more” interactive UI components in the connectivity logic window 108, a popover or pop-up window 114 can be displayed on the user interface as shown in the example of Figure 2b, where the popover / pop- up window 114 displays the connectivity logic that the user requested by activating the121603975172.8Docket No. 242206PCT interactive UI component. A similar popover or pop-up window (or other means of displaying additional information) may appear when the user activates an interactive UI components in the disclosure examples pane 112. Other types of interactive UI components may be used in other embodiments of the present invention.

[0049] In some embodiments, the UI has drop-down capability for a Frequently Asked Questions (FAQ) tab. The FAQ tab may include information regarding how to use the UI 100. The FAQ tab may list commonly asked questions about the user interface and the information provided in the list.

[0050] In some embodiments, a drop-down menu may be shown on the UI 100 for the top themes. The top themes may be customized to showcase themes which are of the highest financial magnitude for a given industry, company, time horizon, accounting category, connectivity scenario, or other dimension. The top themes may be n-number of top themes, where n is any number less than the number of possible themes stored in data storage 29 or the centralized repository 22. The user may be able to create report of the top n-themes by pressing a button on the UI 100.

[0051] The financial connectivity platform 8 may also enable the ability to automatically associate and contextualize the magnitude, time horizon, and specific accounting impact of each theme for a given industry and / or company.

[0052] The financial connectivity platform 8 may permit users to transition between sustainability, thematic, accounting, industry-level and company-level data, research, and disclosure information, thereby streamlining access and usability. The financial connectivity platform 8 may systematically filter unstructured data into financially connected categorizations which are configured in a visual display to simplify research, reporting, financial analysis, valuation analysis, and improve decision-useful information.

[0053] The financial connectivity platform 8 may display information for a company within an industry group or other business classification. The corporate issuer may be placed within a sector or industry group. The financial connectivity platform 8 may determine which sector or industry group a given company falls within.

[0054] Any suitable groupings for the companies can be used, such as standard or proprietary sector or industry groupings. For example, the Global Industry Classification131603975172.8Docket No. 242206PCTStandard (GICS) industries may aid in forming industry categories with similar business model dynamics. For example, certain industries retain a high degree of physical and tangible assets (e.g., manufacturing hardware) versus intangible assets (e.g., services or software). Different economic and sustainability themes are more or less relevant for tangible or intangible businesses. In other embodiments, the internal or 3rdparty determinations of business line exposure weighted by applicable revenue, market value or other financial activity metric could be used to group the companies. In yet other embodiments, modifications to any of these known classification systems could be used, and / or proprietary groupings could be used.

[0055] The FCP 8 comprises a database (e.g., centralized repository 22) and data visualizations (e.g., Figs. 2a-2b) fully customized for each industry. For example, the FCP 8 can group the companies into the 27 industries listed in the left column in Table 1 below. Table 1 shows in the right column an associated GICS sector for each of the industries listed in the left column of Table 1. Thus, in various embodiments, more refined groupings can be used than GICS sectors. Any company which is searched for in the FCP 8 is automatically mapped to one of the industries, with the mappings stored in the industry sector / group data file 36 in the data storage 29, for example. The company to industry group / sector mappings may be determined by subject matter experts, rule-based systems, and / or machine learning. The FCP 8 may retain, and be used to cover, thousands of investable companies, corporate issuers and government-sponsored entities from around the world and continue to grow in sample size.141603975172.8Docket No. 242206PCTTable 1

[0056] The financial connectivity platform 8 has a user interface for data visualization that depicts the themes 110 (what), accounting metrics 104 (where), magnitudes 102 (how), and time horizon 106 (when) exclusively for each of the industries or companies. For example, the underlying methodology, unique combinations, and data visualizations for a pharmaceutical company’s output are completely different relative to an automobile company’s output.

[0057] The financial connectivity platform 8 comprises a visualization of the sustainability themes 110. The themes may be categorized under environmental, social, governance, or strategy. Each theme is stored with unique definitions, keywords, and ontology. As shown in the Table 2, there may be 36 themes. In other embodiments, there may be more or less than 36 themes, and different themes could be used in other embodiments. The FCP 8 may visually display on the user interface 100 the themes 110 with their corresponding financial connections 104, magnitude 102, time horizon 106, connectivity logic 108, and corporate disclosure examples 112.151603975172.8Docket No. 242206PCTTABLE 2

[0058] The unique definitions, keywords, and ontology of each theme can be used to curate specific textual analysis and extraction from corporate disclosure data and other sources stored in the centralized repository 22 by the Al computer system 40 as explained herein. Relevant corporate disclosure extracts corresponding with a given theme (in a financially connected context) can be identified by the Al computer system for incorporation into a side bar display of the interface 100 for the given company.

[0059] As shown in Figure 2b, a pop-up window 114 may be shown on the UI when the user clicks (or otherwise activates) one of the interactive UI components in the connectivity logic pane 108. As shown in the example of Figure 2b, the pop-up window can include a logical explanation of how the corresponding sustainability theme (GHG emissions in this example) affects the financial and accounting metrics of a given industry or company. In the illustrated example, reducing internal combustion engine (ICE) emissions will affect the161603975172.8Docket No. 242206PCT assets, non-operating expenses, costs of goods sold (COGS), operating expenses and capital expenditures (CapEx) of companies in the specific industry; a changing product mix will affect the assets, non-operating expenses, COGS, operating expenses and capital expenditures and revenues; violations of vehicle efficiency standards will affect liabilities and nonoperating expenses; and changes to input costs via supply chain emissions reduction will affect COGS. The connectivity logic for each theme and for each industry can be determined by subject matter experts, rule-based systems, and / or machine learning. Preferably, the connectivity logic maps issues pertaining to the theme to specific accounting and financial metrics (e.g., assets, liabilities, non-operating expenses, COGS, operating expenses, CapEx, etc.). Preferably, only issues for a theme that have a direct and material effect on an accounting / financial metric are shown in the connectivity logic. To ensure comprehensiveness, the connectivity logic for the themes can be considered from multiple stakeholder perspectives, such as suppliers, customers, owners, regulators and employees of companies in the relevant sector / industry group (e.g., a so-called, acronymic “SCORE” framework).

[0060] The SCORE 600 framework is a proprietary framework utilized to identify, design, and catalogue connectivity logic for the themes in the taxonomy 82 and the respective industry groups (e.g. why and how themes impact the financial statements of enterprises in a given industry). SCORE 600 incorporates a multi-step process in which, first, the external “forces” 604 incentivizing for driving leadership decisions and business outcomes are identified and documented in understandable, textual format. Forces 604 are broken up into four distinct categories: customers, Regulators & Gatekeepers, Geophysical Disruptions, and Ownership. For each theme / industry combination, SCORE 86 identifies and documents the manifestation of customer decision making and impact drivers, regulatory / oversight limits and incentives, physical environmental impact drivers, and ownership’s own incentives to enhance the profitability of the enterprise. The formulation of the driver characteristics for customers, regulators, geophysical disruptions, and ownership is executed by way of structured criteria that considers underlying industry constituent business models. This driver characteristic assessment can be separately stored for use in additional business model and corporate sustainability evaluation processes. From each of these forces stems “responses” from the enterprise or firm 606 and the marketplace 608. Responses from the enterprise 606 are broken into the distinct categories of operational responses, supplier related responses, remedial responses, and employee related responses. In effect, enterprise related responses171603975172.8Docket No. 242206PCT606 capture how the cost structures, investments, purchasing, hiring, and non-recurring expenses of the enterprise react to previously identified forces. Separately, market related responses 608 are broken into two categories: revenue opportunities and revenue reductions. These categories capture how the markets for an industry’s constituent enterprise may react, negatively or positively, to the identified forces. All of the forces and responses “logic statements” are catalogued 86 distinctly in textual format within a data table. Lastly, the logic statements can be analyzed versus empirical and forward-looking evidence to determine which specific accounting categories may be impacted within the context of the theme / industry combination. Each accounting category is distinctly linked via a labeling system to one or multiple responses from the enterprise 606 or the marketplace 608. This creates a clear, evidence-based lineage between financial outcomes and themes.

[0061] Definitions for the exemplar sustainability themes, according to one embodiment of the present invention, are provided in Table 3 below. Subject matter experts, rule-based systems, and / or machine learning can apply these definitions when determining the magnitudes, time horizons, and / or connectivity logic for the sustainability themes, e.g., they can serve as a starting point to establish unique combinations of financial connectivity logic for each of the industries.

[0062] The definitions, collectively forming a dictionary, can also be stored within the centralized repository 22 and / or the data storage 29. A keyword dictionary and machine learning ontology can also be stored for each sustainability theme. In various embodiments, each theme’s definition, words, future words in this general nature, and keyword dictionary can be used to create a unique ontology for machine learning and prompt engineering. For example, SEC filings (e.g., Form 10-K) or other regulatory filings and data sources in various countries globally may be systematically ingested into the centralized repository 22 and analyzed to identify corporate disclosure language extracts which pertain to each theme in a financially connected context for a given company. Links to the extracts can be included in the field 112 of the UI (see Figs. 2a-2b). When a user clicks on a disclosure example in field 112, one or more relevant excerpts can appear in a presentation mechanism such as a pop-up or slide-in window, for example, showing how companies in the industry describe the business risk, economic opportunity and / or financial and accounting impact of the sustainability theme. The relevant excerpts can be determined by the Al computer system 40 based on the vector embeddings, as described herein.181603975172.8Docket No. 242206PCT

[0063] According to various embodiments, the definitions set forth in Table 3 below, and terms in this general nature, can be used for the sustainability themes:191603975172.8Docket No. 242206PCT201603975172.8Docket No. 242206PCT211603975172.8Docket No. 242206PCTTable 3

[0064] The sustainability themes, whether the ones identified above in Table 2 or some other set of themes, including their keywords, phrases and definitions (such as the example definitions in Table 3 above) preferably constitute a taxonomy of themes, e.g., a classification and organization of the themes into groups or categories based on shared characteristics. Moreover, as shown in the example of Figure 5, the taxonomy of sustainability themes 80 may be mapped, by a taxonomic mappings 82, to connectivity logic 84 for the themes based on the perspectives of various stakeholders 86 in the companies, such as suppliers, customers, owners, regulators and employees of companies in the relevant sector / industry group (e.g., a so-called “SCORE” stakeholder logic). That is, the taxonomic mappings 82 preferably map, for each industry, the various sustainability themes to the relevant, affected financial and accounting metrics for the connectivity logic 84, with relevant connectivity logic for a themeindustry pairing being displayed, for example, in the connectivity logic pop-up window 114 of Fig. 2b. As such, the taxonomic mappings can be tuned and / or particularized for different industries, countries and regions. For example, countries and regions may have unique laws and regulations which govern the companies who do business inside their borders. The taxonomic mapping 82 retains country and regional flexibility. In various embodiments, one or more subject matter experts, rule-based systems, and / or machine learning can generate the taxonomic mappings 82. In other embodiments, the taxonomic mappings 82 may be generated or updated via computer-implemented techniques, such as via natural language processing, machine learning (e.g., supervised or unsupervised learning), or knowledge graphs with automated link prediction and / or semantic reasoning. The taxonomic mapping 82 may be stored digitally, for example, in the data storage 29 in a spreadsheet or in a221603975172.8Docket No. 242206PCT database, or in some other suitable digital format.

[0065] Referring back to Figs. 2a and 2b, the magnitude pane 102 on the UI displays magnitude values (e.g., high, medium, low, or numerical values in other contexts) for each sustainability theme. The magnitude value for each theme may be stored within the magnitude database 30 shown in Figure 1. The magnitude values are theme-dependent (e.g., different sustainability themes can have different magnitude value), with the magnitude values being determined by a subject matter expert, rule-based systems, and / or machine learning, for example.

[0066] The magnitude values displayed in the magnitude pane 102 preferably signify which sustainability themes retain a relatively higher or lower financial accounting impact for a given industry. For example, a “high” magnitude value means that the theme has a high financial accounting impact and relevance in the industry and correspondingly a “low” magnitude value means that the theme has a low financial accounting impact and relevance in the industry. High, medium, and low are relative terms, where low has the lowest financial accounting impact, medium has a higher financial accounting impact than low, and high has a higher financial accounting impact than low and medium. In other embodiments, different gradations or levels can be used for the magnitude values, such as numerical values (e.g., with greater numbers indicating greater magnitude).

[0067] A user may benefit from seeing which themes 110 are connected to which financial accounting category and the relative financial importance of the theme 110 to a given industry. Some themes may retain a high financial magnitude for a given industry, whereas other themes may be more relevant to environmental or societal impact. Providing this type of clarity and context does not currently exist in conventional systems. The potential to fill this practitioner void helps in various applications including legal, regulatory and compliance. For example, financial services frequently create, market and distribute products and services in a diverse set of jurisdictions, each with their own specific regulations governing pecuniary and non-pecuniary factors.

[0068] In order to determine which theme 110 would be categorized as high, medium or low, for a given industry, instead of solely relying on subject matter experts, the financial connectivity platform 8 may establish a systematic, computer-based system to determine the financial importance to the industry. In yet another embodiment, a machine learning231603975172.8Docket No. 242206PCT classifier, implemented by the Al computer system 40, could be used to classify the magnitude for each theme for each industry.

[0069] For example, the rules-based system may establish metrics and key indicators for each of the themes and then quantitatively evaluate each on an industry-specific basis. For example, the theme of “water stress” could be measured by the water intensity of the beverage industry and its underlying industry constituents. Another example would be the percent of revenue derived from electric vehicle (EV) sales for an automobile company. These metrics, their historical statistical correlations with financial performance, and key performance indicators can serve as the basis to establish empirical evidence of a financial connection.

[0070] The rules-based system may include the database, where the database includes the evaluation of each industry based on their business model characteristics and the prevalence of their disclosures on the theme / topic, from a machine learning process. The machine learning process may use the Al system 40 to identify which themes have a higher propensity to be disclosed in the annual reports, quarterly reports, sustainability reports and other regulatory and non-regulatory public disclosures of corporate issuers in a given industry. For example, the SEC Form 10-K is a corporate disclosure with legal, regulatory, and audit information. A theme which is frequently discussed by the majority of companies in a given sector or industry as a financially relevant business risk or opportunity may be deemed to have a higher magnitude versus a theme which is rarely discussed by companies in a given industry.

[0071] In some embodiments, a survey of corporate finance and investment professionals on each of the magnitudes may be used as a verification of the magnitudes for each theme in each industry. As part of the survey, the investment professionals (who can be sector specialists with many years of experience researching a given industry) can systematically review and validate each magnitude in a “peer review” process. In other embodiments, the Al system 40 verifies the information.

[0072] A financial categories pane 104 / 106 on the UI can show, for example, time horizons for when various accounting and financial metrics are projected to be impacted by the sustainability themes. In the illustrated example, the financial categories 104 include assets, liabilities, revenues, expenses, and cash flow. In other embodiments, fewer, more,241603975172.8Docket No. 242206PCT and / or more granular accounting / financial metrics can be displayed. In the illustrated embodiment, five primary financial statement categories are shown, but there could be over 15 underlying accounting line items. The five primary financial statement categories displayed in the user interface can include, as shown in the example of Figs. 2a-2b, assets and liabilities (primary components of a balance sheet), revenues and expenses (primary components of an income statement), and cash flows (primary components of a cash flow statement.)

[0073] In one embodiment, the five primary financial statement categories are visually displayed, as shown in the example of Figures 2a-2b, although in various embodiments the database(s) of the back-end computer system 8 may store time horizons and connectivity logic for each sustainability theme for each industry for additional, more granular accounting and financial metrics, such as, for example, the 15 accounting and financial metrics shown in Table 4 below. The metrics listed in Table 4 are grouped into three groupings: Balance Sheet, Income Statement and Cash Flows. Further, the metrics listed in Table 4 can be grouped into the categories of assets (i.e., physical assets, intangible assets and financial assets); liabilities (current, legal and regulatory); revenues (price, volume and opportunities); expenses (COGS, operating and non-operating expenses); and cash flows (operating, investing and financing cash flows). These additional metrics may contain even more granular and distinct financial line items which serve as the ultimate connection for economic logic of each sustainability theme. As such, the metrics shown in Table 4, or other metrics or a subset of them, could be included in the field 104 of the UI shown in Figures 2a-2b. Also, as shown in the example of Figure 2a, the UI may include a summary of these metrics, e.g., assets, liabilities, revenues, expenses and cash flow. In such an embodiment, the three asset metrics (physical, intangible and financial) for a sustainability theme-industry pairing could be combined to generate a composite assets time horizon for sustainability theme-industry pairing, and so on for liabilities, revenues, expenses and cash flows. The time horizons for the granular metrics could be combined to generate the composite time horizon in different ways, such as using the shortest time horizon or a statistical measure such as a median or an average. Also, the connectivity logic pop-up window 114 could show connectivity for multiple, highly relevant accounting / financial metrics.251603975172.8Docket No. 242206PCTTable 4

[0074] Assets may be defined as any resource owned or controlled by a company or economic entity. Assets are a primary component of a balance sheet. Assets may include, but are not limited to, physical (e.g., property) and intangible (e.g., intellectual property) property.

[0075] Liabilities may be defined as a present obligation of a company or economic entity that will resort in a future outflow of resources. Liabilities are a primary component of a balance sheet. Liabilities may include, but are not limited to, legal (e.g., civil litigation) and / or regulatory (e.g., government legislation).

[0076] Revenue may be defined as the value of all sales of goods or services recognized by a company or economic entity in a given time period. Revenues are a primary component of an income statement. Revenues may include, but are not limited to, price (e.g., inflation) vs demand (e.g., volume).

[0077] Expenses may be defined as a cost or outlay that a company or economic entity incurs in the due course of its operations. Expenses are a primary component of an income statement. Expenses may include, but are not limited to, operating (e.g., utilities) versus nonoperating (e.g., restructuring).

[0078] Cash flow may be defined as the amount of cash that a company or economic entity generated or consumed during a given time period. Cash flows are a primary component of a cash flow statement. Cash flow may include, but are not limited to, operating (e.g., working capital) versus investing (e.g., capital expenditures).

[0079] The financial platform’s UI can display the time horizon 106 for each (high level261603975172.8Docket No. 242206PCT or granular) accounting category of each theme in a grid format, for example, as shown in the example of Figure 2a. The time horizon may be illustrated as short “S”, medium “M,” or long “L.” There may not be a time horizon for each accounting category of each theme. The long duration is longer than the medium duration, and the medium duration is longer than the short duration. The time horizons for each theme for each accounting / financial metric may be stored in the time horizon database 32 and may be determined by subject matter experts, rule-bases systems, and / or machine learning. Preferably, the time horizons are determined based on when a sustainability theme for a company or industry would manifest in a financial accounting metric.

[0080] The short time horizon may be defined as event-driven financial impacts which retain little warning before they occur. The short time horizon may be event-driven issues which tend to elicit a near-term business response by a company’s management. A short time horizon in the database and visualization is measured in business calendar quarters and may have a duration of one year or less.

[0081] The medium time horizon may be defined as business, economic or regulatory changes that manifest over definite periods of time and are more clearly defined. The medium time horizon may be business and regulatory changes which require process / product alterations and are in company management control. A medium time horizon in the database and visualization may be measured in calendar years. The medium time horizon may be greater than one year and less than five years. Three years may be considered the average.

[0082] The long time horizon may be defined as second order / voluntary market changes and are more evolutionary, systemic-oriented and retain many unknown factors. A long time horizon in the database and visualization may be measured in calendar years and may have a duration of greater than five years. Of course, in other embodiments, different horizon lengths and / or cut-offs could be used for the time horizons.

[0083] The time measurement system may establish when a given sustainability theme would manifest in a financial accounting impact. The time horizon values (e.g., L, M, S) clarify the financial relevance for a given industry. For example, certain sustainability themes will impact different industries over different time horizons.

[0084] Visualization of the time horizon 106 may help clarify the financial impact and range of practical applications. Establishing a time scale allows a user to adjust accounting271603975172.8Docket No. 242206PCT metric assumptions in a financial, actuarial, underwriting, or valuation model. For example, discounted cash flow (DCF), net present value (NPV), economic value add (EVA), dividend discount model (DDM), price and enterprise value comparable multiples (e.g. Price to Earnings and Enterprise Value to EBITDA), and other valuation methodologies utilize a minimum time horizon of one to five years of financial projections 408 for accounting categories such as revenues, expenses, profits and cash flows.

[0085] The connectivity logic field 108 can have interactive UI components, such as hyperlinks and other presentation mechanisms, that, when activated for a theme-industry pairing, provides, in a pop-up window 114 or slide-in panel for example, the connectivity logic for the theme-industry pairing to the accounting / financial metrics. The displayed connectivity logic can serve as a practical and evidenced-based explanation for how a sustainability theme impacts each accounting category in each industry. For example, if a user is reviewing the financial impact of the “climate opportunities” theme for an automobile company, the user may see the various stages and financial accounting impacts of pivoting its business lines towards hybrid or electric vehicles (EV). In this example, the user would see different scenarios including the establishment of EV intellectual property, building out EV production and infrastructure, and capturing share of a growing EV market. Each of these scenarios retains a comprehensive list of which of the 15 accounting categories would be affected. The connectivity logic for each of the 36 themes is transparent and visually displayed in the user interface. As shown in Figure 2b, a pop-up window 114 may display the complete connectivity logic. The logic for each theme-industry pairing can be stored in the data file 34 and may be determined by subject matter experts, rule-based systems, and / or machine learning, for example. In other embodiments, the Al computer system 40 is trained to derive the logic from documents in the centralized repository 22. For example, historical connectivity logic for theme-industry pairings can be used as training labels for a neural network that is trained, through supervised learning, to learn the connectivity logic for themeindustry pairings from the disclosure documents stored in the centralized repository. For example, logistic regression, support vector machine, random forests, and / or gradient boosting machine models could be used to learn the connectivity logic.

[0086] As shown in the example of Figures 2a-2b, the financial connectivity platform’s UI can also include interactive UI components (e.g., links), in the disclosure examples pane 112, that when activated show relevant corporate disclosures (e.g., SEC filing excerpts or281603975172.8Docket No. 242206PCT other regulatory filings in other countries) for how companies in the sector or industry report on the particular sustainability theme. The relevant corporate disclosures linked to in the field 112 for a theme-industry pairing may be determined by the Al computer system 40. For example, with reference to Figure 4, the Al computer system 40 may (i) learn a semantic web for the sustainability themes at step 62, (ii) vectorize the semantic web at step 64, (iii) vectorize paragraphs in source documents stored in the centralized repository 22 for companies at step 66, and (iv) compare the vectors for the paragraphs to the vectors for the semantic web at step 68 to identify the most relevant paragraphs in a company’s (or companies’) source documents to the sustainability themes.

[0087] The semantic web 62 can be implemented with or as a knowledge graph, a semantic network, or an ontology, for example. The Al computer system 40 can learn the semantic web, for example, from knowledge base 60 for the various ESG sustainability themes. The knowledge base 60 can include, for example: the dictionary definitions for the sustainability themes (described above); positive and negative examples of entities and relationships for the sustainability themes; a keyword dictionary for the sustainability themes; the taxonomic mapping 82 (see Fig. 5); and / or a proprietary ontology (e.g., expressed in a formal language, such as OWL or RDF) for the themes. The knowledge base 60 can be stored as a database of the Al computer system 40.

[0088] The process of creating the semantic web (e.g., knowledge graph) can involve entity extraction (identifying entities from the dictionary entities), relationship extraction (determining relationships between entities), property extraction (extracting properties associated with entities), and graph construction (e.g., nodes represent entities and edges represent relationships). The Al computer system 40 can employ natural language processing (NLP) for named entity recognition, dependency parsing, and semantic analysis to extract information from dictionary entries. The Al computer system 40 can also employ machine learning algorithms to learn patterns and relationships in the dictionary data.

[0089] In some embodiments, the system may incorporate a semantic web, knowledge graph, or ontology structure that represents relationships among themes, sub-themes, industries, entities, and domain-specific concepts. The semantic web may be used in several ways to enhance processing by the large language model (LLM). For example, the processor may use the semantic web to enrich or contextualize prompts provided to the LLM, to select or weight theme definitions during similarity scoring, or to supply structured knowledge that291603975172.8Docket No. 242206PCT supplements the LLM’s embedding or inference processes. In other embodiments, the semantic web may itself be embedded using a graph-embedding model, and the resulting vectors may be combined with LLM-generated text embeddings through concatenation, learned weighting, or late-fusion scoring. The system may further use the semantic web to guide downstream score aggregation or connectivity logic generation by constraining or interpreting how themes relate to financial statement metrics. These different modes of integration may be used individually or in combination, enabling the system to operate with LLM-only embeddings, semantic-web-augmented embeddings, or hybrid embeddings derived from both sources.

[0090] The Al computer system 40 can also vectorize the semantic web for the sustainability themes at step 64. This can involve representing the semantic web’s entities and relationships as numerical vectors in a high-dimensional space. A vector embedding is a mathematical representation of data, such as entities in the knowledge graph or paragraphs in a company’s source documents, in the form of a high-dimensional vector (an ordered list of numbers). The goal of embeddings is to capture the semantic meaning or properties of the data in such a way that similar items (in terms of meaning or function) are represented by vectors that are close together in the vector space, while dissimilar items are farther apart. In various embodiments, an LLM, such as OpenAI’s Ada or BERT-based models, is used to generate embeddings. The vector embeddings for the semantic web can be stored in a vector database 42A.

[0091] At step 66, the Al computer system 40 can also compute vector embeddings for paragraphs of company source documents stored in the centralized repository 22 (e.g., a data lake), with the paragraph vector embeddings stored in a vector database 42B. These paragraph vector embeddings can be used to identify, via a vector search at step 68, the most relevant corporate disclosures for the sustainability themes and the industries. The paragraph vector embeddings can be generated in a manner similar to how the semantic web vector embeddings are generated, e.g., with an LLM, such as OpenAI’s Ada or BERT-based models.

[0092] The vector database 42B also preferably stores metadata for each paragraph vector embedding, such as, for example, the company in whose source documents the paragraph was found, the date of the source document, the type of source document (e.g., SEC annual report or other corporate filings in other countries), etc. The vector database 42A, B may be301603975172.8Docket No. 242206PCT implemented with in-memory data structure (such as arrays or hash maps), databases (e.g., SQL or NoSQL databases), file systems, Hadoop Distributed File System, an in-memory keyvalue store, such as Redis, and / or cloud-based storage. The vector database 42A, B may be part of the vector database 42 shown in Figure 1.

[0093] To generate the vector embeddings from the source documents, the Al computer system 40 may split paragraphs in the source documents and then compute vector embeddings for each paragraph. The Al computer system 40 may use various text processing techniques to split the paragraphs in the source documents. For example, the Al computer system 40 may use sentence segmentation algorithms to identify the end of sentences within paragraphs to identify (and split) a paragraph in the source document. Also, the Al computer system 40 could analyze whitespace patterns to identify (or confirm existence of) paragraphs in the source document.

[0094] Alternatively or additionally, the Al computer system 40 could use machine learning models to detect paragraphs. For example, the Al computer system 40 could have a machine learning model (e.g., such as recurrent neural network, a long short-term memory model, or a transformer model), trained via a labeled data set where paragraphs are clearly marked, to predict paragraph boundaries. The Al computer system 40 could also employ feature extraction - extracting features like text length, presence of newline characters, punctuation, and whitespace patterns - to help the model learn paragraph boundaries.

[0095] To elaborate on the vector embeddings for either the semantic web or the source document paragraphs, the Al computer system 40 may also use, in various embodiments, pretrained language models to compute the vector embeddings. The pre-trained language models may comprise, for example, a BERT (Bidirectional Encoder Representations from Transformers) model (including variants of BERT models, such as RoBERTa or DistilBERT) or a GPT (Generative Pre-trained Transformer) series model. The Al computer system 40 can convert the semantic web elements and / or source document paragraphs into sequences of tokens using a tokenizer associated with the chosen language model. The tokens can then be converted into input IDs, attention masks, and / or segment IDs, which are input to the language model to obtain hidden state vectors for the paragraphs. For example, the hidden state from the final layer of the model can be used as the hidden state vector. Then, pooling can be used to reduce the high-dimensional representations of individual tokens into a fixedsized vector. For BERT-like models, the [CLS] token’s hidden state can be extracted, or the311603975172.8Docket No. 242206PCT hidden states of all tokens can be averaged, to obtain a fixed-size vector representing the entire paragraph. For GPT-like models, the hidden state of the final token can be used or an average of the hidden states of all tokens. The resulting vector from the pooling step can be the embedding, whether for the semantic web element or source document paragraph, as the case may be. These vectors can be used for vector searches, as described herein. Each vector may have a dimension range in the hundreds or even thousands of elements.

[0096] The vector search process 68 may generate the connectivity logic. The vector search process 68 can use a similarity metric, such as cosine similarity, to, for example, identify the most relevant paragraphs in a company’s source documents to each ESG or sustainability theme based on the vectors for the sustainability themes’ semantic web. The most relevant paragraphs can be the ones linked to in the disclosure example pane 112 of the user interface of Figure 2A. For example, the Al computer system 40 can determine the N most relevant paragraphs for each sustainability theme. It can also employ a threshold cutoff; that is, for example, only if the similarity metric for a paragraph is greater than or equal to a minimum threshold is the paragraph linked to in the disclosure example pane 112 of the user interface of Figure 2A. The Al computer system 40 can, therefore, score the paragraphs of corporate issuer disclosures from government databases (e.g. SEC EDGAR), or any other governmental or non-governmental sources including website disclosures, social media and / or aggregated news reports, transcripts of public statements, copies and summarizations of correspondence. The Al computer system 40 can score the paragraphs for thousands of companies and index them based on the metadata by company, sector, industry, country, region and date / year inside the database in the document processing stage. The scoring of the paragraphs enables the Al computer system to generate the connectivity logic, mapping issues from paragraphs pertaining to the theme to specific accounting and financial metrics (e.g., assets, liabilities, non-operating expenses, COGS, operating expenses, CapEx, etc.). To ensure comprehensiveness, the connectivity logic for the themes can be considered from multiple stakeholder perspectives 86, such as suppliers, customers, owners, regulators and employees of companies in the relevant sector / industry group (e.g., “SCORE” framework in Figure 8).

[0097] Referring still to Figure 4, the vector searching capability could also be used for bespoke research. For example, a user of the FCP 8 at the client device 12 could submit a query, via the web server 16, to the FCP 8 for corporate source documents relevant to a321603975172.8Docket No. 242206PCT particular economic or sustainability theme. Because the stored source documents are tagged in the centralized repository 22 with accompanying metadata, such as company, industry, sector, date, type of filing, a user could submit queries for source documents from a particular company, or in a particular industry or sector, in a particular time period, and / or in a particular type of source document, that are most relevant to a particular sustainability theme, and the vector searching capability could search the corporate source documents for relevant documents. As an example, the user could ask, for Company X in the automobiles and components industry, for relevant corporate filings of Company X related to the GHG emissions theme. The Al computer system 40 could then search Company X’s annual report (e.g. SEC 10-K) or other filings and disclosures for paragraphs relevant to the GHG emissions theme, based on the paragraph vectors for Company X’s 10-K filings and the semantic web vectors related to the GHG emissions theme to identify, as an example, as most financially relevant a paragraph in a 10-K filing from any given year for Company X that states: “JFe regularly evaluate our current and future product plans and strategies for compliance with fuel economy and GHG regulations. In the years ended December 31, 2024, 2023 and 2022, we paid $2.0 billion, $0.5 billion and $1.0 billion to purchase credits to facilitate our compliance with emissions regulations. Compliance-related costs of $1.0 billion, $0.7 billion and $0.5 billion were recorded in cost of sales in the years ended December 31, 2024, 2023 and 2022. ” That way, a user of the FCP 8 could identify easily and efficiently how a company, in the relevant sector or industry, reports on a particular theme, in an evidence-based and financially connected context.

[0098] In a related manner, the FCP 8 could also be used to generate portfolio reports that, for example, determine a portfolio’s exposure to any or all of the sustainability themes and its corresponding financial accounting connection. For example, the user could upload to the Al computer system 40 the holdings in a portfolio or index. The Al computer system 40 could then determine the industries of the holdings in the portfolio and then determine the magnitude for each of the different industries in the portfolio for each of the sustainability themes, as described above. The Al computer system 40 could also compute composite, weighted magnitudes for each of the sustainability themes and their financial accounting category connections based on the holdings in the portfolio or index. For example, the high, medium and low magnitudes could be mapped to numerical values (such as 5, 3, 1 for example) and the magnitude for a particular industry in the portfolio could be scaled by the percentage, by value, of that particular industry in the portfolio. For example, if the331603975172.8Docket No. 242206PCT magnitude of the GHG emissions theme is High for the automobiles and components grouping, and a High magnitude is mapped to a value of 5, and the automobiles and components makes up 4% of the portfolio or index by value, then the weighted score for the automobiles and components grouping in the portfolio for the GHG emissions theme could be computed as (5 x 0.04) = 0.2. Then, all of the weighted GHG emissions theme scores across the industries in the portfolio could be summed to compute a composite GHG emissions theme score for the portfolio, which can then be mapped backed to a high, medium or low magnitude accounting and / or financial metrics, for example. In a similar manner, portfolio composite time horizons could be computed for how each sustainability theme will impact high or low level accounting and / or financial metrics.

[0099] Figure 6 illustrates a user interface for the financial connectivity platform, according to an embodiment of the present disclosure. The financial connectivity platform may be the financial connectivity platform 8 described in Figure 1.

[0100] The user interface 400 may display connectivity logic for an industry or company. An industry pane 402 may be displayed which includes the industry, such as automobiles. The industry pane 402 may also include a specific company within the industry, such as a specific global automobile manufacturer (e.g., Automotive Company A). The underlying logic of the financial connectivity platform links the sustainability themes 404 to the financial metrics 406 most affected by each individual theme 404. Multiple themes may be shown in a list or the most financially relevant themes may be shown, such as the five or ten most impacted themes. Any number of themes may be listed. For each individual themes 404, the financial metrics 406 associated with the individual theme may be displayed alongside the actual historical values or future projections of the specific financial metric 408.

[0101] For example, the theme, climate opportunities, is shown in the user interface 404. Climate opportunities impacts revenue and growth 408 and are listed under the financial metrics column 406. The financial connectivity platform includes connectivity logic for how an industry or company’s revenue connects 410 to the climate opportunities theme. The financial connectivity platform can upload and review financial statements and underlying accounting categories to determine and align the connectivity between the theme and financial metrics that are impacted for that industry or company. The process for reviewing and aligning financial accounting line items may be the process shown in Figure 10.341603975172.8Docket No. 242206PCT

[0102] The financial accounting information column 408 includes the accounting information that corresponds to the financial metric 406 listed. The financial accounting information may also include historical data and forward projections for a pre-determined number of years. For example, the financial accounting information 408 displays three years of historical information for each financial metric 406 and two years of forward-looking projections according to market consensus estimates. The accounting information may be sourced from corporate disclosures or accounting statements for the corresponding company listed in the industry panel 402. The financial accounting information may also be user- created inputs from a proprietary financial model or valuation model. The financial accounting information may be displayed in multiple currencies (e.g., U.S. dollars $, Japanese yen, euros, pounds sterling, etc.)

[0103] The financial connectivity platform generates the connectivity logic 410 based on the themes, financial metrics impacted by the theme, and the accounting information from the financial statement. The connectivity logic connects the financial metric to a time horizon for each theme. For example, the electric vehicle market presents opportunities for medium term growth which impacts an automobile company’s revenue and growth prospects. The connectivity logic pane 410 also includes the metrics from corporate disclosures alongside each of the financial metrics listed.

[0104] The connectivity logic 410 may include the direct connection to the accounting theme, the time horizon, and a message detailing the connectivity between the theme and the financial metric. The message may be synthesized from corporate disclosures or other sources, such as the historical financial accounting records used for the accounting information column 408. The connectivity logic 410 may include a message detailing firsthand engagement or direct interaction with the specific company or its industry peers as it pertains to the theme or financial metric.

[0105] The user interface 400 includes SCORE logic 412 for each theme. The SCORE logic 412 includes categories suppliers, customers, owners, regulators, and employees. The SCORE connectivity logic corresponds to each theme. The SCORE logic displays which of suppliers, customers, owners, regulators, and employees are impacted by the theme. For example, the theme climate opportunities primarily impacts customers, owners, and regulators in this example.351603975172.8Docket No. 242206PCT

[0106] In some embodiments, the user interface 400 may include a notes section where a user can type and store comments for each category within the connectivity logic 410, including, but not limited to revenue connectivity, expense connectivity, expense connectivity, and / or asset / CapEx connectivity. These notes may connect a user’s qualitative and quantitative thematic insights with the financial model metrics, investment thesis, valuation assumptions, connectivity logic, time horizon, and stakeholder impact.

[0107] The connectivity logic includes the time horizon for each theme. Figure 7 illustrates an example of financial connectivity logic for accounting metrics and time horizons, according to an embodiment of the present disclosure. The connectivity logic 500 illustrates the theme and the accounting connectivity at an intersection of financial magnitude 502 and time horizon 504.

[0108] For example, the financial connectivity logic map for automobile industry includes climate opportunities which impacts the assets and revenues at a high level in a medium time horizon.

[0109] The financial connectivity logic graph may include a predetermined number of themes that are most impacted. For example, the graph may include the top 10 themes for a given sector, industry or portfolio. Any number of themes may be shown.

[0110] The connectivity logic 500 also displays the accounting connectivity 506 for the theme displayed on the graph. The accounting connectivity 506 may be shown in different colors to coordinate the different accounting metrics that connect to the theme.[oni] The graph may include a scale in the x and y-axis that includes low, medium, or high. The time horizon and financial magnitude for each theme intersect at a point on the graph. The point on the graph contains the additional dimension defined by the legend.

[0112] Figure 8 illustrates the SCORE connectivity framework, according to an embodiment of the present disclosure. The user interface 600 illustrates a SCORE framework for a sustainability theme 602. The SCORE framework can be broken into forces 604, firm responses 606, and marketplaces responses 608. The SCORE framework defines market forces 604 which flow through suppliers, customers, owners, regulators and employees. The SCORE framework includes at least one of customers, regulatory and legal, geophysical shocks, ownership, operations, suppliers, employees, remedial, opportunities, or361603975172.8Docket No. 242206PCT reductions.

[0113] Forces may be an incentive applied to the customer. The firm responses may be the industry’s response to the force. The marketplace responses 608 may be the market’s response to the forces. The forces 604 may include customers, regulatory and legal, geophysical shocks, and ownership. The connectivity logic determines a connection between each of the categories and the sustainability theme.

[0114] The SCORE framework is then evaluated against the financial metrics 610, such as assets, liabilities, revenues, expenses, and capex. Other accounting categories or financial metrics may be displayed. The financial connectivity platform links the firm responses 606 and marketplace responses 608 to the accounting metrics 610 that are impacted. The financial connectivity platform 600 may include the time horizon for each response and the corresponding impacted financial metric. For example, the pollution theme 602 for the chemicals industry, a firm response 606 of changing suppliers or having to find alternative suppliers due to pollutant concerns impacts expenses 610 (e.g. cost of goods sold) in a medium term time horizon.

[0115] The SCORE framework 600 enables the forces on an industry and the corresponding responses to be analyzed in terms of accounting categories and financial metrics 610. Certain accounting / financial metrics will be impacted more heavily by the forces and subsequent responses than others. For example, the employees of a chemical company are not directly impacted by the forces of pollution, but operations are impacted by pollution. The impact on operations is financially connected to assets, revenues, expenses, and capex. Each of the financial connections may occur over differing time horizons, e.g. the impact to revenues is short time horizon and the impact to assets is medium time horizon.

[0116] The SCORE framework displays a message that explains the connection between the theme 602 and the specific force, firm response, or marketplace response. Some themes may not have a connection between the force or response.

[0117] Figure 9 illustrates a visualization of thematic prevalence, according to an embodiment of the present disclosure. The heatmap 700 illustrates the thematic prevalence for a chosen theme for a given industry group. The heatmap may be generated by the back- end computer system 10, as shown in Figure 1. For example, the heatmap 700 illustrates the chosen meta theme, social, for the automobiles industry. The heatmap is created based on a371603975172.8Docket No. 242206PCT source, such as corporate public disclosures. The source can also be any of the source documents described herein.

[0118] Each meta theme can be broken into underlying themes. The heatmap 700 visualizes the array of themes and the year. For each year, the themes may be ranked relative to each other from less prevalence to most prevalence.

[0119] In some embodiments, a user can toggle between different combinations of themes, industries, and sources. The user may select different themes, industries, or sources to change the heat map displayed.

[0120] The financial connectivity platform determines the prevalence for each of the themes for a number of years. The prevalence may be low, below average, above average, or high. Other statistical measures such as median, variance and standard deviation may be applied. The prevalence may be relative to the other themes for that year. The number of years displayed may differ based on user preferences. For example, five years are shown, but more or less may be used. The prevalence scale may be shown in different colors or different gradients.

[0121] The heatmap may include more than two dimensions. For example, the heatmap can include a dimension for each source document.

[0122] The heatmap may display a composite score for the variety of different sources. The back-end computer system may generate a composite score through dimensionality reducing algorithms such as, Principal Component Analysis (PCA) and t-Distributed Stochastic Neighbor Embedding (t-SNE).

[0123] Figure 10 illustrates the architecture for creating connectivity logic, according to an embodiment of the present disclosure. The system 800 includes at least one database storing source documents 802 and taxonomy theme definitions 804. The system 800 also includes at least one processor configured to execute (i) a document processing module 812 that obtains and segments the source documents, and (ii) an LLM-based similarity scoring module 806. In various embodiments, the LLM 806 may execute on the same processor and / or server as the document-processing module 812, or on a different set of processors (e.g., different set of servers) accessible over a network interface (e.g., an API endpoint). The LLM 806 can receive the segmented portions of the source documents 802 and theme381603975172.8Docket No. 242206PCT definitions 804 to compute similarity values as described below.

[0124] The source documents can comprise corporate disclosures and filings for publicly- listed or private companies in any country globally, news, unstructured text, audio, video, and other publicly available information. For example, the source documents could include Securities and Exchange Commission (SEC) filings or similar filings in other countries for a specific corporate issuer, such as annual reports, 10-K filings, 8-K filings or supplemental filings, which can be indexed in the centralized repository 22 by sector, industry, company, year and date. The source documents could also comprise, in various embodiments and / or to the extent available, CSR (corporate sustainability) reports, published patents and patent applications, government reports such as EPA (Environmental Protection Agency) reports, non-governmental organization reports such as CDP (carbon disclosure project) reports, SBTi (Science-Based Targets initiative) reports, company website disclosures, social media, aggregated news reports, transcripts of public statements, and / or copies and summarizations of correspondence. Thus, the source documents may include different sources and different modalities (e.g., text, video, audio).

[0125] The document processing module 812 may divide each source document for the company for a time period into separate portions, such as sentences, paragraphs, sections and / or entire documents, depending on the embodiment. These portions are supplied to the LLM 806 for per-portion similarity scoring against the taxonomy theme definitions 804.

[0126] The LLM 806 may compute, for each separate portion, a similarity value between content of the separate portion and each of the theme definitions 804, thereby generating pertheme portion scores for the one or more source documents for the time period. For example, the LLM 806 can scores each source document for one company (in a particular sector) for the given time period. The LLM may compute the per-theme similarity scores for each portion of the source document 802 based on the corresponding theme definition. The scoring process for the source document 802 may be repeated for each theme. As elaborated further below, the process may also be repeated for other companies in the same sector, for additional time periods, and for other sectors or grouping classifications.

[0127] A processor(s) of the system 800 may aggregate the per-theme portion scores output by the LLM 806 using a score-aggregation function to produce a theme score for each theme for the company for the time period. Examples of score-aggregation functions include391603975172.8Docket No. 242206PCT mean pooling, max pooling, top-K pooling, global average pooling, sum pooling, weighted scoring, or other statistical or machine-leaming-based aggregation techniques. The LLM 806 computes the similarity values for the individual portions, and the processor aggregates those values to generate the overall theme scores.

[0128] In some embodiments, the processor(s) may be configured to, at step 808, cause the LLM 806 to compute the similarity values for each separate portion by generating embedding vectors for each separate portion of the source document and for the theme definitions 804. The LLM 806 may embed both the portions and the theme definitions into a shared vector space, and the processor may compute the similarity value between a portion and a theme definition using a cosine similarity or other vector-space similarity metric applied to their embedding vectors.

[0129] The processor may execute instructions, at step 810, to aggregate the per-theme portion scores for each source document to compute an aggregate score for the source document as a whole for each theme. The aggregation may use a score-aggregation function such as maximum pooling, mean pooling, top-K pooling, weighted pooling, global average pooling, or another statistical or machine-leaming-based function, thereby producing a document-level theme score for the particular time period. The document-level theme scores may then be aggregated across all source documents for a company to produce companylevel theme scores. The aggregation process may be repeated for additional companies within a sector and for additional sectors and / or industries.

[0130] In some embodiments, connectivity logic at the sector, industry, and theme level is generated, at step 812, by analyzing the company -level theme scores for companies within the sector to determine how each theme relates to financial-statement metrics across the sector and / or industry. The connectivity logic may be produced automatically by the processor using predefined rules or machine-learning models, or it may be generated or validated by a subject-matter expert with domain knowledge.

[0131] In some embodiments, the database further stores source documents for each company in a sector and industry for the time period. Then at least one processor further executes instructions to cause the LLM to compute per-theme portion scores for each company in the sector for the time period based on the source documents for each company for the time period stored in the database. The LLM aggregates the per-theme portions scores401603975172.8Docket No. 242206PCT for each company in the sector to produce theme scores for the time period for each company in a sector. The LLM aggregates the theme scores for each company in the sector for the time period to produce sector-level descriptive statistics and inter-company comparison metrics, for the time period, and for the sector / industry, based on the theme scores.

[0132] In that connection, the system may maintain a scoring datastore that stores intermediate and final results generated throughout the processing pipeline. The datastore may store, for each document portion, the embedding vectors and the per-theme portion similarity scores computed by the LLM. The datastore may further store document-level theme scores, company-level theme scores for each time period, and sector-level descriptive statistics such as means, standard deviations, percentiles, normalized scores, and intercompany comparison metrics. These stored values may be indexed by company, sector, industry, theme, document identifier, and time period, and may be retrieved by downstream modules or user-interface components for time-series analysis, visualization, or generation of connectivity logic. The scoring datastore may be implemented using a relational database, a key-value store, a columnar analytics engine, or a vector database supporting metadata indexing.

[0133] In various embodiments, the architecture of Figure 10 may be implemented using at least one processor that coordinates the operation of multiple software components, which may include document-processing modules, embedding or similarity-scoring services, and one or more LLM -based agents. The processor may interface with local or remote data stores to obtain source documents and theme definitions and may invoke a document-processing component to prepare content for analysis. The system may employ an agent-orchestration layer in which specialized agents — such as a retrieval agent, an embedding agent, a scoring agent, or an aggregation agent — cooperate to process the documents, compute similarity values, and produce intermediate and final scoring results. These agents may execute on the same processor, on different processors within a distributed system, or on remote systems accessed through APIs, such as cloud-hosted LLM services or model-serving endpoints.

[0134] Then at least one processor may manage the workflow among these agents, including sending requests to an LLM or embedding service, receiving the model outputs, and directing downstream modules or agents to compute aggregated scores or higher-level analytics. In some implementations, the entire workflow may run locally on a single processor using a locally hosted model, while in other implementations the processor may act411603975172.8Docket No. 242206PCT primarily as a client in a client-server configuration, delegating model inference to a remote LLM service and performing aggregation or further analysis locally. The same architecture may support rule-based, statistical, or machine-learning-based components as well as human- in-the-loop review or refinement by a subject-matter expert. This flexible arrangement enables the system to operate in centralized, distributed, agent-orchestrated, or hybrid environments while remaining consistent with implementations in which the functions are carried out by at least one processor.

[0135] In various embodiments, at least one processor may interact with the large LLM using different execution arrangements. The LLM may be hosted locally on the same physical machine as the processor, with the model weights loaded into memory and executed by local compute resources. In other embodiments, the LLM may be deployed on one or more remote servers or cloud-based endpoints, and the processor may invoke the LLM through a network interface using API calls (e.g., HTTPS or gRPC requests), thereby operating in a client-server configuration. The system may further employ an agent-based orchestration layer in which specialized software agents manage subtasks such as document segmentation, embedding generation, LLM inference requests, and score aggregation. In such embodiments, the processor coordinates agent activities, including routing of document portions to an LLM service, receiving LLM outputs, and invoking downstream scoring or aggregation modules. These execution models may be used individually or in combination, enabling the system to operate in local, distributed, cloud-based, or hybrid environments while remaining consistent with implementations in which the described functions are carried out by at least one processor.

[0136] The process of scoring may be repeated for each company in a sector, industry or other grouping classification to compute sector or industry-specific descriptive statistics (mean, median, standard deviation, percentiles, etc.) and company comparisons (z-scores, ranks, sector normalized scores).

[0137] The scoring process may be done over time to get comparative trends over time. For example, the scoring may determine year-over-year changes, rolling averages, etc. The scoring process may use historical source documents from different years to determine the scores for each year. This may be repeated for each sector, industry, other grouping method.

[0138] Figure 3 is a diagram of the Al computer system 40 according to various421603975172.8Docket No. 242206PCT embodiments. The Al computer system can be used to train machine learning models to compute the vectors (for semantic web and / or source document paragraphs) and then to score the paragraphs. Once trained, the Al computer system 40 can be used to implement the models to generate the relevant corporate disclosure excerpts for field 112 in the UI 100.

[0139] The Al computer system 40 can also be trained, as described above, to identify relevant paragraphs in a source document for a company in a particular grouping that are relevant to the themes. For example, the Al computer system 40 could be trained with historical source documents, with paragraphs having labels for particular theme-industry grouping pairings. Based on this training, the Al computer system 40 can infer the paragraph(s) in a new source document (e.g., a new corporate filing) for a company in a particular industry that are relevant to the themes. For example, in the training, the Al computer system 40 learns the keywords and phrases that companies in a grouping use for the individual themes. Having learned the keywords and phrases for theme-industry grouping pairings, the Al computer system 40 can identify the relevant paragraph(s) addressing particular themes in a new source document for a company in a known industry grouping. A text classification machine learning model could be used by the Al computer system 40 to infer the applicability to the themes of a source document. The text classification model could comprise, for example, transformer based model (e.g., BERT, RoBERTa, GPT), recurrent neural networks (e.g., Long Short-Term Memory (LSTM) or Gated Recurrent Units (GRUs) models), or convolutional neural networks.

[0140] The Al computer system 40 can also be used to identify how the disclosure by a company, or companies in a grouping, have changed over time for a particular sustainability theme(s) by searching for relevant disclosures for the theme(s) in several discrete time periods (e.g., year after year for a span of years). For example, the Al computer system 40 can be used to create a custom scoring system based on the prevalence of a specific theme being disclosed in each industry or grouping. The FCP’s user interface 100 can show, in various embodiments, a gauge or graph (or some other visual indicator) that shows the rate of change (historical time series) in disclosures for a particular theme by a specific company or industry / grouping. Similarly, the FCP’s user interface could also show a flag which signals if a company is currently a leader (e.g. above industry or grouping average and / or median) or laggard (e.g. below industry or grouping average and / or median) as it pertains to the level of transparency on a given theme relative to its industry peers.431603975172.8Docket No. 242206PCT

[0141] Referring back to Figure 3, the illustrated Al computer system 40 may comprise multiple processor units 302A-B that each comprises, in the illustrated embodiment, multiple (N) sets of processor cores 304A-N. Each processor unit 302A-B may comprise on-board memory (ROM or RAM) (not shown) and off-board memory 306A-B. The on-board memory may comprise primary, volatile and / or non-volatile, storage (e.g., storage directly accessible by the processor cores 304A-N). The off-board memory 306A-B may comprise secondary, non-volatile storage (e.g., storage that is not directly accessible by the processor cores 304A-N), such as ROM, HDDs, SSD, flash, etc. The processor cores 304A-N may be CPU cores, GPU cores and / or Al accelerator cores. GPU cores operate in parallel (e.g., a general -purpose GPU (GPGPU) pipeline) and, hence, can typically process data more efficiently that a collection of CPU cores, but all the cores of a GPU execute the same code at one time. Al accelerators are a class of microprocessor designed to accelerate artificial neural networks. They typically are employed as a co-processor in a device with a host CPU 310 as well. An Al accelerator typically has tens of thousands of matrix multiplier units that operate at lower precision than a CPU core, such as 8-bit precision in an Al accelerator versus 64-bit precision in a CPU core.

[0142] In various embodiments, the different processor cores 304 may train and / or implement different networks or subnetworks or components. For example, in one embodiment, the cores of the first processor unit 302 A may implement the paragraph splitting algorithm and the second processor unit 302B may implement paragraph vector embedding computations. Yet another set of processor units may learn how to classify the text of the source documents as pertaining to the relevant sustainability themes. One or more host processors 310 may coordinate and control the processor units 302A-B.

[0143] In other embodiments, the system 8 could be implemented with one processor unit 302. In embodiments where there are multiple processor units, the processor units could be co-located or distributed. For example, the processor units 302 may be interconnected by data networks, such as a LAN, WAN, the Internet, etc., using suitable wired and / or wireless data communication links. Data may be shared between the various processing units 302 using suitable data links, such as data buses (preferably high-speed data buses) or network links (e.g., Ethernet).

[0144] Figure 11 shows a transformer based LLM model, which is a type of neural network architecture designed for natural language processing (NLP) tasks and which could441603975172.8Docket No. 242206PCT be used for the LLM. The key components of a transformer model include an encoderdecoder architecture, a self-attenuation mechanism, a feed-forward neural network (FFNN), positional encoding, and layer normalization and residual connections.

[0145] Tokens are the basic units of input that a transformer network processes. They can represent words, sub-words, characters, or other elements, depending on the tokenization method used. Tokens are discrete elements that the model uses to understand and generate text. They are numerical representations of the textual data that the model processes. Tokens come from a process called tokenization, which converts raw text into a sequence of tokens. After tokenization, tokens are usually converted into numerical representations (token IDs) that the model can process. This is done using a vocabulary, which maps each token to a unique integer. These token IDs are then transformed into dense vectors (embeddings) that capture semantic information.

[0146] Embeddings are high-dimensional representations where tokens with similar meanings are located close to each other in the vector space. The encoder 912 can take an input sequence and transforms it into a sequence of continuous representations. An input embedding layer 114 can convert words or tokens into dense vectors of fixed size. Positional encoding 916 can add information about the position of each token in the sequence since the model does not inherently understand order. These encodings are added to the input embeddings. A self-attention mechanism 918 allows the model to focus on different parts of the input sequence when encoding a particular token. It calculates a weighted sum of the input values, where the weights are determined by the similarity between the token being processed and other tokens in the sequence.

[0147] A feed-forward neural network (FFNN) 920 can apply a two-layer fully connected network to the output of the self-attention mechanism. Each sub-layer (e.g., self-attention and FFNN) is followed by a residual connection (adding the input of the sub-layer to its output) and layer normalization to stabilize and speed up training.

[0148] The decoder 922 takes the encoder’ s output and generates the output sequence, one token at a time. Similar to the encoder, an output embedding layer 923 and positional encoding 924 converts output tokens into dense vectors and adds positional information. A masked self attention mechanism 926 ensures that the prediction for a particular token451603975172.8Docket No. 242206PCT depends only on the known outputs up to that position (i.e., the model cannot "cheat" by looking ahead).

[0149] An encoder-decoder attention layer 928 allows the decoder to focus on relevant parts of the input sequence (the encoder's output) when generating each token of the output sequence. An output / decoder feed-forward neural network (FFNN) 930, similar to the encoder FFNN 920, applies a two-layer fully connected network to the output of the attention mechanism 928. Residual connections and layer normalization can be applied in the same manner as in the encoder.

[0150] Encodings in the transformer refer to the representations of tokens at various stages. The input embeddings comprise initial dense vector representations of the input tokens. The positional encodings are added to input embeddings to incorporate position information. Contextualized encodings are the representations produced by the self-attention and FFNN layers, which capture the context of each token in the sequence.

[0151] Attention allows the model to focus on different parts of the sequence when processing a token. It can involve Query (Q), Key (K), and Value (V) matrices that are derived

[0152] from the input embeddings by multiplying them with learned weight matrices. Scaled dotproduct attention can calculate attention scores by taking the dot product of the Query and Key matrices, scaling them down, and applying a softmax function to get the attention weights. These weights are then used to compute a weighted sum of the Value matrix, producing the attention output.

[0153] The softmax function 932 can convert the attention scores into probabilities, ensuring that they sum to one. In the context of attention, the softmax function 932 ensures that the attention weights highlight the most relevant tokens while maintaining a probabilistic interpretation.

[0154] The LLM could also be adapted for a particular domain or context, such as a domain(s) specific to the enterprise, via fine tuning of the LLM, which adjusts the pre-trained LLM' s weights using domain-specific data to make it more effective for particular applications. Fine tuning can involve collecting a large and diverse dataset relevant to the specific domain or context (e.g., companies, industries, and sustainability related themes).461603975172.8Docket No. 242206PCTFor example, for a financial services application, materials describing the financial services and / or the products of the financial services could be used. This adaptation training data can be tokenized into smaller units (tokens) that the LLM can process. The tokenization of the adaptation training data can use the same tokenization method as the base model of the LLM. The fine-tuning process can involve supervised fine-tuning (e.g. labeled data) where possible. The model is then trained on the domain-specific data, typically under supervised learning techniques. Fine-tuning can be done using frameworks like Hugging Face's Transformers library, TensorFlow, or PyTorch. The fine tuning can involve conventional hyperparameter adjustments and validation of the model's performance.

[0155] The LLM can generate text (e.g., code in the target programming language) using a sophisticated next-word prediction mechanism. The model can be trained on a vast dataset of text from various sources. During training, it learns patterns, structures, and the statistical relationships between words and phrases in the text. This training process involves adjusting the model's parameters to minimize the error in predicting the next word in a sequence of text. When given a prompt and / or initial text, the model analyzes the context using its learned patterns. It takes into account the words and phrases that have already been provided to understand the context and meaning. Based on the context, the model generates a probability distribution over the potential next words. It uses this distribution to predict the most likely next word. This process is repeated word by word to generate coherent and contextually relevant text (e.g., software code). The model can use different strategies to choose the next word from the probability distribution. Common strategies include greedy sampling (choosing the word with the highest probability), top-k sampling (limiting the choices to the top k most probable words and sampling from them), top-p (nucleus) sampling (choosing words from the smallest set whose cumulative probability exceeds a certain threshold (p )), and / or temperature (adjusting the randomness of the predictions, where a lower temperature makes the model more conservative, while a higher temperature makes it more creative and diverse). The model repeats the process, using the newly generated word as part of the context for predicting the next word, continuing this until the desired length of text is generated or until it encounters a stopping condition (like a specific token indicating the end).

[0156] The LLM could be tailored to the domain of the enterprise. For example, the enterprise could be a financial services firm and the LLM' s could be tailored to the domain471603975172.8Docket No. 242206PCT of financial products or some other suitable domain(s). To tailor the LLM to the domain of the enterprise, the model must be adapted to understand the enterprise's specialized terminology, workflows, regulations, and data structures. This process involves training and fine-tuning the model using domain-specific data, ensuring that it can generate insights, answer questions, and assist in decision-making with a high degree of accuracy and relevance. This process can involve the step of data collection and preprocessing, in which the enterprise gathers extensive datasets, including, in the example of a financial services firm, financial reports, regulatory filings, transaction records, market data, internal policies, and customer interactions. These datasets can be cleaned, formatted, and tokenized to be compatible with the LLM' s training framework. If the enterprise handles sensitive financial data, privacy -preserving techniques, such as differential privacy and data anonymization, may be applied to protect confidential information.

[0157] Once the data is prepared, the enterprise can proceed with fine-tuning the LLM to incorporate industry-specific and / or company-specific knowledge. Fine-tuning can involve training the model on labeled datasets where the expected outputs are known, enabling the model to improve its performance in particular enterprise-specific applications. This process typically uses deep learning frameworks such as TensorFlow or PyTorch and optimization algorithms like AdamW to adjust the model's parameters. The fine-tuning stage helps the LLM understand jargon of the enterprise, interpret complex documents, and generate responses that align with industry practices. To further improve performance, the enterprise could integrate retrieval-augmented generation (RAG) techniques, where the LLM retrieves and references external financial data sources in real-time. This allows the model to incorporate up-to-date information, regulatory changes, or company reports into its responses, improving accuracy and reducing outdated or incorrect outputs.

[0158] In various implementations, each functional component of the software system comprises an identifiable unit of the source code for the software system that performs a distinct function within the software system, and where each functional component operates within a defined scope and adheres to syntactic and semantic rules of its respective programming language. At least one of the functional components can be a UI, a method, a program in a specific programming language, a web service, a stored procedure, or a table.

[0159] In various implementations, the programming-language-specific prompts direct the LLM to generate the plurality of labeled graph nodes as JSON text in a JSON text file. In481603975172.8Docket No. 242206PCT various implementations, the LLM comprises a transformer. In various implementations, the method further comprises training the LLM by tailoring the LLM to a domain of the enterprise. The enterprise can comprise a financial services firm and the domain comprises financial products.

[0160] The software for the various computer systems described herein and other computer functions described herein may be implemented in computer software using any suitable computer programming language such as .NET, C, C++, Python, and using conventional, functional, or object-oriented techniques. Programming languages for computer software and other computer-implemented instructions may be translated into machine language by a compiler or an assembler before execution and / or may be translated directly at run time by an interpreter. Examples of assembly languages include ARM, MIPS, and x86; examples of high level languages include Ada, BASIC, C, C++, C#, COBOL, Fortran, Java, Lisp, Pascal, Object Pascal, Haskell, ML; and examples of scripting languages include Bourne script, JavaScript, Python, Ruby, Lua, PHP, and Perl.

[0161] In one general aspect, therefore, the present invention is directed to computer systems and computer-implemented methods for providing a financial connectivity platform that connects sustainability-related themes to financial statement metrics for companies in a sector. In one embodiment, the computer system comprises a back-end computer system and a user computing device that is in communication with the back-end computer system. The user computing device is configured to render, based on data served by the back-end computer system, graphical user interfaces that display, for a specified company that belongs to a sector within a set of predetermined sectors, for each of the sustainability -related themes in a taxonomy of sustainability-related themes: a magnitude indicator indicative of a magnitude of impact of the sustainability -related theme on the financial statement metrics for companies in the sector; a time horizon indicator indicative of a time horizon for the sustainability -related theme to impact the sector for each of financial statement metrics; a first set of interactive UI components that, when activated, display economic scenarios and financial connectivity logic for the sustainability-related theme for the sector; and a second set of interactive UI components that, when activated, display one or more excerpts of source documents for companies in the sector. The excerpts show how the companies in the sector address the sustainability-related theme.491603975172.8Docket No. 242206PCT

[0162] The method according to various embodiments, comprises the step of rendering, based on data served by the back-end computer system, graphical user interfaces that display, for a specified company that belongs to a sector within a set of predetermined sectors, for each of the sustainability-related themes in a taxonomy of sustainability-related themes: a magnitude indicator indicative of a magnitude of impact of the sustainability -related theme on the financial statement metrics for companies in the sector; and a time horizon indicator indicative of a time horizon for the sustainability -related theme to impact the sector for each of the financial statement metrics. The method also comprises the step of detecting, via the graphical user interfaces, activation of a first set of interactive UI components. The method also comprises the step of, in response to detecting the activation, displaying, by the graphical user interfaces, economic scenarios and financial connectivity logic for the sustainability- related theme for the sector. The method also comprises the step of detecting, via the graphical user interfaces, activation of a second set of interactive UI components. And the method further comprises the step of, in response to detecting activation of the second set of interactive UI components, displaying, via the graphical user interfaces, one or more excerpts of source documents for companies in the sector, where the excerpts show how the companies in the sector address the sustainability -related theme.

[0163] In various implementations, back-end computer system comprises an Al computer system that is configured to: learn a knowledge representation for the taxonomy of sustainability -related themes, where, for each sustainability-related theme in the taxonomy, the knowledge representation captures concepts related to the sustainability -related theme; generate vector embeddings for concepts in the knowledge representation; generate vector embeddings for passages or paragraphs of source documents for companies in the sector; and score the paragraphs of the source documents for each of the sustainability-related themes based on a similarity between the vector embeddings for the paragraphs and the vector embeddings for the concepts in the knowledge representation.

[0164] In various implementations, the back-end computer system is further configured to identify, based on the scores, a subset of paragraphs having highest scores for each sustainability-related theme, and to cause the user computing device to display at least an excerpt from at least one of the identified paragraphs in the second set of interactive UI components.501603975172.8Docket No. 242206PCT

[0165] In various implementations, the sector comprises an industry group of companies that share similar business model dynamics.

[0166] In various implementations, the financial statement metrics comprise financial and accounting metrics.

[0167] In various implementations, the financial statement metrics comprise two or more of assets, liabilities, revenues, expenses, and cash flows.

[0168] In various implementations, the graphical user interfaces comprise web pages.

[0169] In various implementations, the sustainability-related themes comprise environmental, social, governance and strategy themes.

[0170] In various implementations, the back-end computer system further comprises digital taxonomic mappings stored in a database and that map, for each of the predetermined sectors, the taxonomy of sustainability -related themes to the financial connectivity logic.

[0171] In various implementations, the digital taxonomic mappings map for each sector, the taxonomy of sustainability -related themes to the financial connectivity logic from perspectives of at least one of suppliers, customers, owners, regulators or employees of companies in each of the predetermined sectors.

[0172] In various implementations, the back-end computer system is further configured to identify passages or paragraphs in the source documents having a score above a threshold value for each of the sustainability-related themes.

[0173] In various implementations, the connectivity logic indicates how the sustainability -related theme impacts the financial statement metrics of companies in the sector.

[0174] In another general aspect, the present invention is directed to an Al computer systems and computer-implemented methods for implementing a machine-learning model configured to learn relationships between sustainability -related themes and corporatedisclosure text. The Al computer system comprises one or more processing cores; and computer memory in communication with the one or more processing cores, the computer memory stores computer instructions that when executed by the one or more processing cores511603975172.8Docket No. 242206PCT cause the one or more processing cores to: learn a knowledge representation for a taxonomy of sustainability-related themes, wherein, for each sustainability-related theme in the taxonomy, the knowledge representation captures concepts related to the sustainability- related theme; generate vector embeddings for concepts in the knowledge representation; generate vector embeddings for paragraphs in a source document for a company; and score the paragraphs of the source document for each of the sustainability-related themes based on a similarity between the vector embeddings for the paragraphs and the vector embeddings for the concepts in the knowledge representation.

[0175] The method, according to various embodiments, comprises the steps of: learning, by a computer system that comprises one or more processors, a knowledge representation for a taxonomy of sustainability -related themes, where, for each sustainability-related theme in the taxonomy, the knowledge representation captures concepts related to the sustainability- related theme; generating, by the computer system, vector embeddings for concepts in the knowledge representation; generating, by the computer system, vector embeddings for paragraphs in a source document for a company; and scoring, by the computer system, the paragraphs of the source document for each of the sustainability -related themes based on a similarity between the vector embeddings for the paragraphs and the vector embeddings for the concepts in the knowledge representation.

[0176] In various implementations, the computer memory stores further instructions that when executed by the one or more processing cores cause the one or more processing cores to identify paragraphs in the source document having a score above a threshold value for each of the sustainability-related themes.

[0177] In various implementations, the computer memory stores further instructions that when executed by the one or more processing cores cause the one or more processing cores to generate and store a digital taxonomic mapping that maps, for each sector in a set of predetermined sectors, the taxonomy of the sustainability -related themes to financial connectivity logic.

[0178] In various implementations, the digital taxonomic mappings map, for each sector within the set of predetermined sectors, the taxonomy of sustainability -related themes to the financial connectivity logic from a perspective of at least one of a supplier, a customer, an owner, a regulator or an employee of companies in the sector.521603975172.8Docket No. 242206PCT

[0179] In yet another general aspect, the present invention is directed to computer systems and computer-implemented methods for producing theme scores, for each of a plurality of predetermined sustainability -related themes, for a company. In various embodiments, the computer system comprises at least one database storing one or more source documents for a company; and sustainability -related theme definitions. The computer system further comprises at least one processor configured to execute instructions to: cause the document processing module to divide each of the one or more source documents for the company for a time period into separate portions; and cause a large language model (LLM) to: compute, for each separate portion, a similarity value between content of the separate portion and each of the theme definitions to generate per-theme portion scores for the one or more source documents for the time period; and aggregate the per-theme portion scores for the one or more source documents using a score-aggregation function to produce a theme score for each theme for the company for the time period.

[0180] A method according to such an embodiment comprises the steps of: storing, in a computer database of a computer system, one or more source documents for a company and sustainability-related theme definitions, wherein the computer system comprises one or more processors; dividing, by the computer system, each of the one or more source documents for the company for a time period into separate portions; computing, , by the computer system, for each separate portion, a similarity value between content of the separate portion and each of the theme definitions to generate per-theme portion scores for the one or more source documents for the time period; and aggregating, by the computer system, the per-theme portion scores for the one or more source documents using a score-aggregation function to produce a theme score for each theme for the company for the time period.

[0181] In various implementations, the at least one processor further executes instructions to aggregate each theme score for the one or more source documents for the time period to produce a connectivity logic between each theme and the company for the time period.

[0182] In various implementations, the connectivity logic indicates how the sustainability-related theme impacts financial statement metrics of the company.

[0183] In various implementations, the database further stores source documents for each company in a sector for the time period. And the at least one processor further executes instructions to cause the LLM to: compute per-theme portion scores for each company in the531603975172.8Docket No. 242206PCT sector for the time period based on the source documents for each company for the time period stored in the database; aggregate, via a score aggregation function, the per-theme portions scores for each company in the sector, to produce theme scores for the time period for each company in a sector; and aggregate the theme scores for each company in the sector for the time period to produce sector-level descriptive statistics and inter-company comparison metrics, for the time period, for the sector, based on the theme scores.

[0184] In various implementations, the at least one processor further executes instructions to: aggregate the themes scores for each company in the sector over a plurality of time periods; and determine connectivity logic for each theme at company and sector levels over the plurality of time periods.

[0185] In various implementations, the score aggregation function comprises one of mean pooling, max pooling, top-K pooling, global average pooling, sum pooling, or weighted scoring.

[0186] In various implementations, at least one processor is configured to cause the LLM to compute the similarity values for the separate portions by generating embedding vectors for each separate portion and for the theme definitions, and computing the similarity value between each separate portion and each theme definition based on a cosine similarity of the embedding vector for the separate portion and the embedding vector for the theme definition.

[0187] In various implementations, the at least one processor comprises a first set of one or more processors and a second set of one or more processors; the second set of one or more processors executes the LLM; the first set of one or more processors are configured to transmit the separate portions to second set of processors and to receive the similarity values from the second set of processors; and the computer-implemented system further comprises a network interface configured to communicatively couple the first and second sets of one or more processors.

[0188] In some embodiments, the financial connectivity platform is implemented as a computer-executed vector-processing pipeline configured to transform unstructured textual disclosure data into fixed-dimension semantic vectors using a machine-implemented embedding model. The embedding model executes on a processor and applies tokenization, attention-based neural transformations, and numerical projection operations to generate highdimensional vectors representing semantic properties of paragraphs and theme definitions.541603975172.8Docket No. 242206PCTThis vectorization is performed using optimized linear-algebra kernels that employ matrix multiply-accumulate instructions (e.g., fused multiply-add) to reduce compute latency. Executing the pipeline in this manner yields a technical effect of reducing inference time and memory footprint during large-scale similarity search across millions of candidate disclosure paragraphs.

[0189] The platform may further maintain an index structure (e.g., Hierarchical Navigable Small World (HNSW) graph, product quantization, or other approximate nearest neighbor index) in a vector database to accelerate similarity-search queries. These index structures are stored in machine-addressable memory in a format optimized for cache-aligned traversal. By arranging vector nodes with locality -preserving mappings, the system achieves a technical effect of minimizing random memory accesses and improving cache-hit rates, thereby materially reducing processor cycles required to identify paragraphs relevant to a given sustainability theme.

[0190] In some embodiments, the platform uses an internally maintained knowledge graph that encodes nodes for sustainability themes, sub-themes, regulatory constructs, accounting metrics, and industry classifications. The knowledge graph is stored in a machine- readable semantic encoding (e.g., RDF triples or serialized graph adjacency structures).When the platform computes embeddings, it integrates the embeddings of graph entities with embeddings of disclosure text via a vector-fusion operation executed by the processor. This integration produces a technical effect of improving retrieval precision through structural disambiguation, allowing the system to distinguish otherwise semantically similar terminology across industries.

[0191] The system may automatically normalize disclosure text using a computer- implemented text-processing module that performs sentence segmentation, punctuation disambiguation, Unicode normalization, and stop-word filtering. This module applies deterministic algorithms (e.g., finite-state automata and rule-based token filters) prior to vectorization. The normalization process yields a technical effect of reducing noise in the embedding space, thereby reducing downstream computational overhead for machinelearning components.

[0192] The similarity-scoring engine may apply a multi-stage scoring process comprising: (i) a coarse ANN (approximate nearest-neighbor) search, (ii) a refined cosine-551603975172.8Docket No. 242206PCT similarity computation, and (iii) an adaptive thresholding mechanism executed by the processor. This staged architecture provides the technical effect of lowering end-to-end latency for identifying disclosure paragraphs relevant to a given sustainability theme, especially when processing millions of vectors stored in the centralized repository.

[0193] In some embodiments, the system maintains incremental update caches that store recently computed theme-to-paragraph mappings and partial ANN graph updates. These caches allow the platform to update the vector index in real time as new disclosure documents are ingested without reconstructing the entire index. This produces a technical effect of improving computer functionality by reducing reconstruction time and minimizing processor load during large-scale ingestion events.

[0194] The platform may store time-horizon assignments, magnitude scores, and connectivity logic in a computer-readable data structure that encodes these attributes as machine-interpretable tuples or multi-dimensional tensors. By organizing these values in alignment with memory page boundaries, the system achieves a technical effect of enabling rapid sequential access during UI rendering and machine-learning inference, thereby reducing memory fragmentation.

[0195] In embodiments where a transformer-based LLM is used, the platform may invoke GPU- or TPU-accel erated matrix-multiplication kernels for attention computation and feed-forward layers. Offloading these operations to specialized hardware results in a technical effect of reducing compute time for generating embeddings, enabling near-real-time updates to thematic relevance and disclosure extraction at scale.

[0196] Where taxonomic mappings between sustainability themes and accounting metrics are produced using machine-learning models, the processor executes feature-learning algorithms (e.g., gradient-boosting machines, graph neural networks, or transformers configured for relation extraction). The models analyze encoded representations of theme definitions and disclosure paragraphs to generate updates to taxonomic mappings. This yields a technical effect of automatically maintaining up-to-date relational structures without manual intervention, enabling improved consistency and accuracy in the connectivity logic.

[0197] The UI data sent to client devices may be pre-computed and server-side rendered by the back-end computer system. The server compresses rendered JSON payloads using dictionary-based compression algorithms (e.g., Brotli or zstd) prior to transmitting them via561603975172.8Docket No. 242206PCTHTTPS. This produces a technical effect of reducing network bandwidth consumption and decreasing end-user latency, especially when transmitting large multidimensional data structures representing theme-metric grids.

[0198] The disclosure ingestion subsystem employs a streaming parser that processes incoming PDF or HTML filings using buffer-level extraction, optical character recognition (OCR) when necessary, and real-time vector generation. This streaming approach provides a technical effect of reducing end-to-end processing time compared with batch-mode ingestion, enabling more timely updates to the system’s connectivity logic.

[0199] In some embodiments, the platform stores text, numerical values, metadata, and connectivity logic in compressed columnar formats (e.g., Apache Parquet). Chunked encoding and dictionary compression reduce disk I / O, providing a technical effect of improving memory efficiency and retrieval speed during large-scale analytical queries performed by the Al computer system.

[0200] By integrating vector-search architectures, semantic-web augmentation, machine- readable taxonomies, GPU-accelerated embedding computation, and cache-coherent index structures, the described platform provides technical improvements to the functioning of the underlying computer systems, beyond merely presenting ESG-related content. The system enables faster similarity retrieval, lower processor load, reduced memory bandwidth consumption, and more accurate machine-interpretable mappings from textual disclosure to accounting-metric impacts.

[0201] Having thus described several aspects and embodiments of the technology of this application, it is to be appreciated that various alterations, modifications, and improvements will readily occur to those of ordinary skill in the art. Such alterations, modifications, and improvements are intended to be within the scope of the technology described in the application. It is, therefore, to be understood that the foregoing embodiments are presented by way of example only and that, within the scope of the appended claims and equivalents thereto, inventive embodiments may be practiced otherwise than as specifically described. In addition, any combination of two or more features, systems, articles, materials, and / or methods described herein, if such features, systems, articles, materials, and / or methods are not mutually inconsistent, is included within the scope of the present disclosure.571603975172.8

Claims

Docket No. 242206PCTCLAIMSWhat is claimed is:

1. A computer system for providing a financial connectivity platform that connects sustainability-related themes to financial statement metrics for companies in a sector, the computer system comprising: a back-end computer system ; and a user computing device in communication with the back-end computer system, wherein the user computing device is configured to render, based on data served by the back-end computer system, graphical user interfaces that display, for a specified company that belongs to a sector within a set of predetermined sectors, for each of the sustainability- related themes in a taxonomy of sustainability -related themes: a magnitude indicator indicative of a magnitude of impact of the sustainability -related theme on the financial statement metrics for companies in the sector; a time horizon indicator indicative of a time horizon for the sustainability-related theme to impact the sector for each of the financial statement metrics; a first set of interactive UI components that, when activated, display economic scenarios and financial connectivity logic for the sustainability-related theme for the sector; and a second set of interactive UI components that, when activated, display one or more excerpts of source documents for companies in the sector, wherein the excerpts show how the companies in the sector address the sustainability -related theme.

2. The computer system of claim 1, wherein the back-end computer system comprises an Al computer system that is configured to: learn a knowledge representation for the taxonomy of sustainability-related themes, wherein, for each sustainability-related theme in the taxonomy, the knowledge representation captures concepts related to the sustainability -related theme; generate vector embeddings for concepts in the knowledge representation; generate vector embeddings for passages or paragraphs of source documents for companies in the sector; and score the paragraphs of the source documents for each of the sustainability-related themes based on a similarity between the vector embeddings for the paragraphs and the vector embeddings for the concepts in the knowledge representation.581603975172.8Docket No. 242206PCT3. The computer system of claim 2, wherein the back-end computer system is further configured to identify, based on the scores, a subset of paragraphs having highest scores for each sustainability-related theme, and to cause the user computing device to display at least an excerpt from at least one of the identified paragraphs in the second set of interactive UI components.

4. The computer system of claim 1, wherein the sector comprises an industry group of companies that share similar business model dynamics.

5. The computer system of claim 1, wherein the financial statement metrics comprise financial and accounting metrics.

6. The computer system of claim 1, wherein the financial statement metrics comprise two or more of assets, liabilities, revenues, expenses, and cash flows.

7. The computer system of claim 1, wherein the graphical user interfaces comprise web pages.

8. The computer system of claim 1, wherein the sustainability-related themes comprise environmental, social, governance and strategy themes.

9. The financial connectivity platform of claim 1, wherein the back-end computer system further comprises digital taxonomic mappings stored in a database and that map, for each of the predetermined sectors, the taxonomy of sustainability -related themes to the financial connectivity logic.

10. The financial connectivity platform of claim 9, wherein the digital taxonomic mappings map for each sector, the taxonomy of sustainability -related themes to the financial connectivity logic from perspectives of at least one of suppliers, customers, owners, regulators or employees of companies in each of the predetermined sectors.

11. The financial connectivity platform of claim 2, wherein the back-end computer system is further configured to identify passages or paragraphs in the source documents591603975172.8Docket No. 242206PCT having a score above a threshold value for each of the sustainability -related themes.

12. The financial connectivity platform of claim 1, wherein the connectivity logic indicates how the sustainability-related theme impacts the financial statement metrics of companies in the sector.

13. An Al computer system implementing a machine-learning model configured to learn relationships between sustainability-related themes and corporate-disclosure text, the Al computer system comprising: one or more processing cores; and computer memory in communication with the one or more processing cores, wherein the computer memory stores computer instructions that when executed by the one or more processing cores cause the one or more processing cores to: learn a knowledge representation for a taxonomy of sustainability-related themes, wherein, for each sustainability -related theme in the taxonomy, the knowledge representation captures concepts related to the sustainability-related theme; generate vector embeddings for concepts in the knowledge representation; generate vector embeddings for paragraphs in a source document for a company; and score the paragraphs of the source document for each of the sustainability -related themes based on a similarity between the vector embeddings for the paragraphs and the vector embeddings for the concepts in the knowledge representation.

14. The Al computer system of claim 13, wherein the computer memory stores further instructions that when executed by the one or more processing cores cause the one or more processing cores to identify paragraphs in the source document having a score above a threshold value for each of the sustainability -related themes.

15. The Al computer system of claim 13, wherein the computer memory stores further instructions that when executed by the one or more processing cores cause the one or more processing cores to generate and store a digital taxonomic mapping that maps, for each sector in a set of predetermined sectors, the taxonomy of the sustainability -related themes to financial connectivity logic.

16. The Al computer system of claim 15, wherein the digital taxonomic mappings map,601603975172.8Docket No. 242206PCT for each sector within the set of predetermined sectors, the taxonomy of sustainability -related themes to the financial connectivity logic from a perspective of at least one of a supplier, a customer, an owner, a regulator or an employee of companies in the sector.

17. A computer-implemented system comprising: at least one database storing: one or more source documents for a company; and sustainability-related theme definitions; and at least one processor configured to execute instructions to: cause the document processing module to divide each of the one or more source documents for the company for a time period into separate portions; and cause a large language model (LLM) to: compute, for each separate portion, a similarity value between content of the separate portion and each of the theme definitions to generate per-theme portion scores for the one or more source documents for the time period; and aggregate the per-theme portion scores for the one or more source documents using a score-aggregation function to produce a theme score for each theme for the company for the time period.

18. The computer-implemented system of claim 17, wherein the at least one processor further executes instructions to aggregate each theme score for the one or more source documents for the time period to produce a connectivity logic between each theme and the company for the time period.

19. The computer-implemented system of claim 18, wherein the connectivity logic indicates how the sustainability-related theme impacts financial statement metrics of the company.

20. The computer-implemented system of claim 18, wherein: the database further stores source documents for each company in a sector for the time period; and the at least one processor further executes instructions to cause the LLM to:611603975172.8Docket No. 242206PCT compute per-theme portion scores for each company in the sector for the time period based on the source documents for each company for the time period stored in the database; aggregate, via a score aggregation function, the per-theme portions scores for each company in the sector, to produce theme scores for the time period for each company in a sector; and aggregate the theme scores for each company in the sector for the time period to produce sector-level descriptive statistics and inter-company comparison metrics, for the time period, for the sector, based on the theme scores.

21. The computer-implemented system of claim 20, wherein the at least one processor further executes instructions to: aggregate the themes scores for each company in the sector over a plurality of time periods; and determine connectivity logic for each theme at company and sector levels over the plurality of time periods.

22. The computer-implemented system of claim 20, wherein the score aggregation function comprises one of mean pooling, max pooling, top-K pooling, global average pooling, sum pooling, or weighted scoring.

23. The computer-implemented system of claim 17, wherein at least one processor is configured to cause the LLM to compute the similarity values for the separate portions by generating embedding vectors for each separate portion and for the theme definitions, and computing the similarity value between each separate portion and each theme definition based on a cosine similarity of the embedding vector for the separate portion and the embedding vector for the theme definition.

24. The computer-implemented system of claim 17, wherein: the at least one processor comprises a first set of one or more processors and a second set of one or more processors; the second set of one or more processors executes the LLM;621603975172.8Docket No. 242206PCT the first set of one or more processors are configured to transmit the separate portions to second set of processors and to receive the similarity values from the second set of processors; and the computer-implemented system further comprises a network interface configured to communicatively couple the first and second sets of one or more processors.

25. A method for providing a financial connectivity platform that connects sustainability- related themes to financial statement metrics for companies in a sector, the method comprising: rendering, based on data served by a back-end computer system, graphical user interfaces that display, for a specified company that belongs to a sector within a set of predetermined sectors, for each of the sustainability -related themes in a taxonomy of sustainability- related themes: a magnitude indicator indicative of a magnitude of impact of the sustainability-related theme on the financial statement metrics for companies in the sector; and a time horizon indicator indicative of a time horizon for the sustainability -related theme to impact the sector for each of the financial statement metrics; detecting, via the graphical user interfaces, activation of a first set of interactive UI components; in response to detecting the activation, displaying, by the graphical user interfaces, economic scenarios and financial connectivity logic for the sustainability -related theme for the sector; detecting, via the graphical user interfaces, activation of a second set of interactive UI components; and in response to detecting activation of the second set of interactive UI components, displaying, via the graphical user interfaces, one or more excerpts of source documents for companies in the sector, wherein the excerpts show how the companies in the sector address the sustainability-related theme.

26. The method of claim 25, further comprising: learning, by the back-end computer system, a knowledge representation for the taxonomy of sustainability-related themes, wherein, for each sustainability -related theme in the taxonomy, the knowledge representation captures concepts related to the sustainability- related theme;631603975172.8Docket No. 242206PCT generating, by the back-end computer system, vector embeddings for concepts in the knowledge representation; generating, by the back-end computer system, vector embeddings for passages or paragraphs of source documents for companies in the sector; and scoring, by the back-end computer system, the paragraphs of the source documents for each of the sustainability -related themes based on a similarity between the vector embeddings for the paragraphs and the vector embeddings for the concepts in the knowledge representation.

27. The method of claim 26, further comprising: identifying, , by the back-end computer system, based on the scores, a subset of paragraphs having highest scores for each sustainability-related theme; and causing a user computing device, that is in communication with the back-end computer system, to display at least an excerpt from at least one of the identified paragraphs in the second set of interactive UI components.

28. The method of claim 25, wherein the sector comprises an industry group of companies that share similar business model dynamics.

29. The method of claim 25, wherein the financial statement metrics comprise financial and accounting metrics.

30. The method of claim 25, wherein the financial statement metrics comprise two or more of assets, liabilities, revenues, expenses, and cash flows.

31. The method of claim 25, wherein the graphical user interfaces comprise web pages.

32. The method of claim 25, wherein the sustainability-related themes comprise environmental, social, governance and strategy themes.

33. The method of claim 25, further comprising storing, by the back-end computer system, digital taxonomic mappings in a database that maps, for each of the predetermined sectors, the taxonomy of sustainability-related themes to the financial connectivity logic.641603975172.8Docket No. 242206PCT34. The method of claim 33, wherein the digital taxonomic mappings map for each sector, the taxonomy of sustainability -related themes to the financial connectivity logic from perspectives of at least one of suppliers, customers, owners, regulators or employees of companies in each of the predetermined sectors.

35. The method of claim 26, further comprising identifying, by the back-end computer system, passages or paragraphs in the source documents having a score above a threshold value for each of the sustainability-related themes.

36. The method of claim 25, wherein the connectivity logic indicates how the sustainability-related theme impacts the financial statement metrics of companies in the sector.

37. A method implementing a machine-learning model configured to learn relationships between sustainability-related themes and corporate-disclosure text, the method comprising: learning, by a computer system that comprises one or more processors, a knowledge representation for a taxonomy of sustainability-related themes, wherein, for each sustainability-related theme in the taxonomy, the knowledge representation captures concepts related to the sustainability-related theme; generating, by the computer system, vector embeddings for concepts in the knowledge representation; generating, by the computer system, vector embeddings for paragraphs in a source document for a company; and scoring, by the computer system, the paragraphs of the source document for each of the sustainability-related themes based on a similarity between the vector embeddings for the paragraphs and the vector embeddings for the concepts in the knowledge representation.

38. The method of claim 37, further comprising identifying, by the computer system, paragraphs in the source document having a score above a threshold value for each of the sustainability-related themes.

39. The method of claim 37, further comprising generating, by the computer system, and storing a digital taxonomic mapping that maps, for each sector in a set of predetermined sectors, the taxonomy of the sustainability -related themes to financial connectivity logic.651603975172.8Docket No. 242206PCT40. The method of claim 39, wherein the digital taxonomic mappings map, for each sector within the set of predetermined sectors, the taxonomy of sustainability -related themes to the financial connectivity logic from a perspective of at least one of a supplier, a customer, an owner, a regulator or an employee of companies in the sector.

41. A method compri sing : storing, in a computer database of a computer system, one or more source documents for a company and sustainability-related theme definitions, wherein the computer system comprises one or more processors; dividing, by the computer system, each of the one or more source documents for the company for a time period into separate portions; computing, , by the computer system, for each separate portion, a similarity value between content of the separate portion and each of the theme definitions to generate per-theme portion scores for the one or more source documents for the time period; and aggregating, by the computer system, the per-theme portion scores for the one or more source documents using a score-aggregation function to produce a theme score for each theme for the company for the time period.

42. The method of claim 41, further comprising aggregating, by the computer system, each theme score for the one or more source documents for the time period to produce a connectivity logic between each theme and the company for the time period.

43. The method of claim 42, wherein the connectivity logic indicates how the sustainability-related theme impacts financial statement metrics of the company.

44. The method of claim 42, comprising: storing, in the computer database, source documents for each company in a sector for the time period; computing, by the computer system, per-theme portion scores for each company in the sector for the time period based on the source documents for each company for the time period;661603975172.8Docket No. 242206PCT aggregating, by the computer system, via a score aggregation function, the per-theme portions scores for each company in the sector, to produce theme scores for the time period for each company in a sector; and aggregating, by the computer system, the theme scores for each company in the sector for the time period to produce sector-level descriptive statistics and inter-company comparison metrics, for the time period, for the sector, based on the theme scores.

45. The method of claim 44, further comprising: aggregating, by the computer system, the themes scores for each company in the sector over a plurality of time periods; and determining, by the computer system, connectivity logic for each theme at company and sector levels over the plurality of time periods.

46. The method of claim 44, wherein the score aggregation function comprises one of mean pooling, max pooling, top-K pooling, global average pooling, sum pooling, or weighted scoring.

47. The method of claim 41, further comprising computing, by the computer system, the similarity values for the separate portions by generating embedding vectors for each separate portion and for the theme definitions, and computing the similarity value between each separate portion and each theme definition based on a cosine similarity of the embedding vector for the separate portion and the embedding vector for the theme definition.671603975172.8

Images