DaanMatch project seeks to: 1) Enable community meet administrative requirements for government mandated Corporate Social Responsibility (CSR) programs 2) Utilize big data to help donors intelligently target India's development goals using the framework provided by the United Nation's Sustainable Development Goals (SDGs) 3) Utilize advanced technology to increase the reach and impact of funding for global development. The initial research has been carried out in the context of India. India is a good test bed due to their ambitious government mandated CSR program. This paper describes one module of the DaanMatch system developed with ontology and knowledge graph technology: OWL, SPARQL, Protégé, Cellfie, the AllegroGraph graph database product and Python APIs from Franz Inc. The DaanMatch Knowledge Graph module (DaanKG) provides an ontology that models the UN Sustainable Development Goals, NGOs, and CSR programs as well as the records, people, and deliverables associated with NGO projects. The DaanMatch system demonstrates a new paradigm that utilizes technology to radically reinvent the funding and monitoring of projects enabling more focus on helping those in need rather than on administrative overhead.
India suffers from staggering inequality.2 Despite steady economic growth the gap between those with and those without has continued to expand. While India’s richest 1% hold over 40% share of the nation’s wealth, the half of the population at the lower end controls just 6%. [1] Non-profit organizations work to provide for those in need but struggle for resources. With the introduction of Section 135 of the Companies Act, 2013, India became the first country to mandate Corporate Social Responsibility (CSR). This legislation was intended to provide much-needed resources to
Organizations (NGOs) to compensate for dwindling government programs. [2] Ten years after going into effect, the mandate has not lived up to India’s aspirations. [3]
2023 Copyright for this paper by its authors. CEUR
ceur-ws.org
Annual
Compared to large NGOs that incur costs of travel, accommodations, high-priced consultants, and related expenses, grassroots NGOs have lower operating costs. [4] [5] Smaller, local NGOs are often run and manned by volunteers and spend more of their budget on direct services, to help those in need. In addition, because they tend to have close ties to the local culture, they often can provide better services for those in need. Despite this, the current system greatly favors larger NGOs, especially for corporate CSR programs that need to limit the time they spend finding, funding, and monitoring NGOs. [4]
The Companies Act is an admirable attempt to direct funding to NGOs. In practice, however, challenges faced by the corporations, the NGOs, and the government, have restricted program effectiveness overall, left billions of dollars in available aid untouched, and have left a vast number of effective grassroots NGOs neglected. [3] [6] [7]
DaanMatch is a system designed to address these issues utilizing technologies such as machine learning and knowledge graphs. DaanMatch seeks to democratize transparency and ease the burden of compliance for both Donors and NGOs, making formal funding easier and more equitable. The goals of DaanMatch are to: 1. Increase the pool of NGOs that are eligible for funding by helping NGOs use their data to show transparency and impact easily, using evidence in lieu of narrative and paperwork. 2. Save donors time and money by helping them connect with projects that address sustainable development goals, align with donor values, and are implemented by legitimate, effective organizations. 3. Provide better metrics on how effective CSR programs are in achieving the UN Sustainable
Development Goals.
The DaanMatch system consists of two main modules (see Figure 1): the CSR module, is focused on CSR managers who wish to analyze data about NGOs and how they relate to SDGs in specific areas. The NGO module utilizes technologies such as speech to text, video capture, image capture, and metadata to simplify the process for NGOs to receive and demonstrate compliance with CSR grants. As a result, the NGO module must run on mobile devices. Many NGO users in rural areas, especially smaller NGOs, use their phones as their only platform to access the Internet and applications. In addition, the vision of DaanMatch: to utilize technology to radically simplify the compliance process, requires that NGO users have their devices with them when they perform the services stipulated in project grants. The solution to obstacles facing NGOs were revealed during an unrelated volunteer opportunity with the Human Rights Center Lab at UC Berkeley. Monitoring events at middle east hotspots on social media for potential human rights violations, the project used metadata, reverse image and satellite imagery to remotely evaluate and substantiate claims. We theorized that similar methods could be used to verify work done by NGOs. I.e., video activities such as food distribution, sanitation development, etc. Our idea is to automate these processes for NGO data collection and apply these technologies to social impact and global development, streamlining the effort for transparency and easing reporting for grassroots NGOs.
As a result, the NGO module requires a web-based client and more conventional technology that will enable NGO users to easily record appropriate media and upload it into the DaanMatch database with the appropriate metadata. Figure 2 shows a screen from the NGO mobile GUI. NGO users can use this to document activities by taking videos, recording audio, using speech to text, etc. The information is automatically tagged with appropriate metadata indicating where, when, and by whom the media was captured and linking it with the appropriate project (grant) that the event was associated with. The CSR module on the other hand, is divided into two different distributions:
interested users. This is the version that is the short-term focus. 2. A proprietary distribution for corporations that utilize DaanMatch to manage their CSR program. This is the long-term vision and will include data specific to each NGO and corporation that utilizes DaanMatch to fund and monitor projects. This data will include the project status, the SDGs that it addresses, the media that document how the NGO fulfilled the requirements of the project, etc.
Distribution 1 provides an open-source platform for all users to browse the data collected by the
DaanMatch team about NGOs. This data has been “scraped” from various public web sites such as
GuideStar.[7] In addition, it is supplemented with Linked Data3 about geographic locations in India as
well as other public data that may be useful to CSR managers, CSR executives, UN workers,
economists, and academic researchers. Distribution 2 will utilize all the knowledge in distribution 1
supplemented with specific data for each CSR program. This module is more appropriate for complex
models that are best supported by formal ontologies and knowledge graphs. This module is more recent
and is built on an ontology that models the United Nations Sustainable Development Goals (SDG). This
is the module that will be the focus of this paper.
3 As of t
Section 2 describes the various ontologies and tools used to develop the knowledge graph system. It includes specific examples of how the formal ontology model can facilitate the DaanMatch goals. Section 3 describes the development process for this module. Section 4 discusses future plans and how our experience reflects on the use of formal ontologies for real world data and problems.
This section will describe the DaanMatch CSR module and how it is designed around formal ontologies. The project is taking an Agile [8] approach to development. The current ontology and code can be found at: https://github.com/mdebellis/Daan_Knowledge_Graph. 2.1.
The data model for NGOs and other tables utilized in the NGO module are replicated as data properties in the CSR ontology. Super properties are utilized to organize the data properties so that it is easy to manage changes made to the relational data model and to upload the data from the relational database into the knowledge graph. For example, there are tables in the relational database called NGOBackground, NGOContact, and NGOFinance. The various columns associated with each table are replicated as data properties in the ontology and are made sub-properties of the super-properties: ngoBackgroundProperty, ngoContactProperty and ngoFinanceProperty. Figure 3 shows a screen capture from the Protégé ontology editor with some of these super-properties expanded. • •
Uploading the data with minimal transformation makes the upload process less error prone and more maintainable.
Uploading data with minimal transformation provides an audit trail showing where information in the ontology originated. This is especially important for our problem because our vision is that recording media and metadata that records specific instances of
This makes the uploading of data from the relational database very straight forward. The domain for each of the subproperties is defined on the super-property and inherited by each sub-property. In some cases, such as NGOFinance, there is a class that corresponds to the table. In other cases, such as NGOContact, the data in the database (name, phone, email) applies to other classes in the ontology that are not modeled in the database. In those cases, the domain for the super-property is a higher level class than the table such as the Prov:Agent class for NGOContact.
Transform, and Load (ETL) paradigm into Extract, Load, and Transform (ELT). [9] This is becoming common in many uses of ontologies for real data. The reasons for • • • • •
Prov:Activity will significantly reduce and eventually eliminate standard bureaucratic paperwork. As part of transforming the data, the Python functions will correct datatype and other errors as much as possible. At the same time, correcting such data must be flagged for potential auditing because it may indicate fraud or errors in the data.
Transforming data that is in the knowledge graph allows the transformations to take advantage of axioms in the ontology and the reasoner. This simplifies the transformation process and allows more sophisticated kinds of transformations. [10]
Additional knowledge that is not required for the NGO module is added from ontologies based on the UN Sustainable Development Goals model and linked data sources such as Wikidata. All 3 ontologies were initially modeled in the Protégé ontology editor. In order to support the large amount of data required for the ontologies, the data is stored in a knowledge graph hosted in the AllegroGraph product from Franz Inc. Python is used to:
Transform the database data into a knowledge graph format. I.e., go from “string to things”. [11] Perform various manipulations and analysis of the data. E.g., link NGOs and CSRs to appropriate UN SDGs via text matching.
Present the information in the ontologies to CSR users in a GUI.
Grouping properties together in this way makes maintenance and testing significantly easier when managing large amounts of data. The same technique was used in [10] and several consulting projects that the lead author has done for industry clients. An example of the usefulness of this approach is that when the team realized that many of our classes should be replaced with classes from the Prov-O vocabulary [12] having the domains of these classes specified on super-properties made this change significantly easier than if they were defined on each data property. This is an example of how formal models can facilitate an Agile development process. 2.2.
In this section we describe the 3 main ontologies in the CSR module. In the current system these are all represented in one ontology and the boundaries between the (eventually separate) ontologies is defined by different namespaces such as Prov, SDG, and NGO. The ontologies are currently in one large ontology because as we were developing, we were making large changes to the design on a regular basis, and it was easier to make such changes with the entire model in one ontology. Now that the model has mostly been finalized, we will divide the large ontology into 3 different ontologies and different AllegroGraph repositories. This is discussed in section 3.
The most important metrics for describing and analyzing data about NGO and CSR programs comes from the United Nations Sustainable Development Goals. [13] The UN has a large amount of data available to describe these goals on their SDG portal. [14] The SDG ontology primarily consists of 3 classes: 1. SDGGoal. These are the 17 high level goals defined by the UN. For example: 4 Goal 1: End poverty in all its forms everywhere. 2. SDGTarget. Each goal has one or more targets. The SDG goals are high-level aspirational goals. The SDG targets are more concrete goals that relate to each high-level goal. For example, a target for Goal 1 is: Target 1.1: By 2030, eradicate extreme poverty for all people everywhere, currently measured as people living on less than $1.25 a day. 4 https://sdgs.un.org/goals/goal1 3. SDGIndicator. Each target has one or more indicators. These are concrete metrics that measure how well a nation or region is achieving the target. For example, the indicator for target 1.1 is: Indicator 1.1.1: Proportion of the population living below the international poverty line by sex, age, employment status and geographical location (urban/rural).
Due to the well thought out design and numbering of each goal, target, and indicator and the meta data provided by the UN in a spreadsheet on their SDG data portal it was extremely easy to model the SDGs in OWL. The process to develop this ontology was to: 1. Rework the metadata spreadsheet so that it was in the proper format to be uploaded by the
Cellfie plugin [15] for Protégé. 2. Use Cellfie to create the appropriate instances of each of the classes described above. In addition, metadata that described the identifier for each Goal, Target, and Indicator was added as an annotation property values (modeled via Dublin Core) for each new instance. E.g., the identifier for Goal 1 is the string “1.” The identifier for Target 1.1 is “1.1” and the identifier for Indicator 1.1.1 is “1.1.1”. 3. Use SPARQL to utilize the identifier keys to automatically create appropriate object property links between each Goal, Target, and Indicator. For example, the SPARQL code to create the hasIndicator values from a target to its indicator is shown in Figure 4. This code processes each SDGGoal and finds any instances of SDGTarget that begin with the same identifier as the goal. If the target has the same prefix as the goal, then the goal hasTarget of the target. E.g., it will find that the identifier for target 1.1 begins with the identifier for goal 1 and hence will add a property value illustrating that the goal hasTarget for that target.
The UNSDG ontology was developed before this project and is available as a separate ontology at [16].
The geography of India includes states, territories, cities, wards, towns, and villages. Rather than recreating this information manually, we chose to utilize Linked Data [17] and download the appropriate classes and instances from Wikidata. This is a large enough amount of data of its own that it merits its own ontology. In addition to the basic geographic information such as latitude, longitude, and population, we hope to incorporate additional information from Wikidata and other linked data sources that will further allow users to perform sophisticated analysis of the CSR data in relation to the UN SDG metrics as well as specific metrics that may relate to each corporation’s CSR vision.
The NGO Ontology has the majority of the classes. These are classes to model NGOs, Corporations and their CSR programs, grants that have been allocated from a CSR to an NGO, media (Prov:Entity) that are used to document the fulfillment of a grant, and other miscellaneous classes and properties that describe the NGO/CSR process and data.
We considered using an Upper Model; however, we did not see any tangible benefit from using any of the popular Upper Models. In addition, there was significant reuse potential in the Prov-O ontology. Prov-O provides: “…the foundation to implement provenance applications in different domains that can represent, exchange, and integrate provenance information generated in different systems and under different contexts”. [12] This is directly relevant to our long term goals of using media (instances of the Prov:Entity class) to supplement and eventually mostly replace traditional documentation for NGO projects.
The two most significant ways that knowledge graph technology has been utilized are to: 1. Use text matching to align NGOs and CSRs with the UN goals that they are most focused on. Then utilize these links to the goals to match NGOs with CSR programs. 2. Create graphs that document how instances of Prov:Entity such as videos document the work done by an NGO to fulfill the goals of a specific project.
These directly map to the goals discussed in the introduction. Alignment of CSRs and NGOs with the UN SDGs simplifies the process of CSRs finding NGOs that align with their vision. The knowledge graph connecting NGOs to Prov:Entities created on site and tagged with metadata helps to achieve the goal of radically simplifying the CSR funding process so that smaller, less computer literate NGOs will be able to show compliance with project grant requirements with minimal effort. These will be discussed next in sections 2.3 and 2.4. 2.3.
Figure 5 shows the simple SPARQL query that can match CSR programs and NGOs that have compatible SDG visions. This is a screen print from the AllegroGraph Gruff SPARQL query and visualization tool. Gruff can take the results of a SPARQL query and automatically create a graph that represents them. [19] In addition, the user can rework the graph in various ways such as selecting a specific node and having Gruff automatically generate a tree graph from that node.
An essential requirement to realizing the DaanMatch vision of using technology to significantly reduce the administrative burden on NGOs is to maintain the same or higher levels of verification with the new process.
In addition to capturing the relations among media used to document projects, the knowledge graph can facilitate auditing by modeling logic associated with the proper execution of a project. Such logic can be captured in axioms for relevant classes, rules in the Semantic Web Rule Language (SWRL), and SPARQL queries.
For example, a SWRL rule in the ontology captures the business logic that the creationDate for any media used to document a project must come after the startDate of the project: startDate(?p, ?psdt) ^ hasDocumentation(?p, ?doc) ^ creationDate(?doc, ?crdt) ^ swrlb:lessThan(?crdt, ?psdt) -> hasStatus(?p, RequiresAudit)
To test this logic, we created some test data (unrelated to any actual NGO or project) for a food distribution project in Pradesh. The documentation for this test project includes a video that was created before the start of the project. This causes the SWRL rule to fire and the RequresAudit to be added to the project status. One benefit of using SWRL is that explanations can automatically be generated for inferences. Figure 9 shows the explanation generated in Protégé when clicking on the RequiresAudit status for this project.
The project to date has been completely driven by volunteers. Work on the NGO
The initial SPARQL queries to connect SDG Goals, Targets, and Indicators were done with the Snap
SPARQL Protégé plugin. However, the SPARQL implementations in Protégé are primarily intended as
an introduction to SPARQL, not for project work. In addition, Protégé itself is a modeling tool not a
database. Hence, to support the large amount of data, more complex SPARQL queries, knowledge
graph visualization, and utilization of Python the AllegroGraph graph database product from Franz Inc.
was utilized. [20] In J
The immediate future milestones for the project are: 1. Integrate data from Wikidata and replace many of the current properties in the Geography ontology with properties from GeoSPARQL.5 We initially created properties such as containedIn and contains to model relations between states/territories and cities/villages/districts before realizing that GeoSPARQL models these relations in a standard way and provides additional capabilities that will be useful in the future. 2. Create a more sophisticated statistical algorithm to match NGOs and CSRs to SDGs. The current algorithm is fairly simple and only matches the highest-level UN SDG goals. We are developing a more sophisticated statistical algorithm that we believe will both be more accurate and will map to targets and indicators as well as the high-level goals. 3. Release an open-source system that integrates data about NGOs and CSRs using the UN Sustainable Development Goals model. This release will include a GUI that allows nontechnical users to generate SPARQL queries by filling out a form that describes parameters such as the SDG metrics they want to search for, the size of the NGO, locations, etc. This GUI has already been developed using the QT GUI Python library. We have a small amount of additional work to do to write the appropriate SPARQL query based on the parameters provided by the user.
The planned August 2023 release will be the most comprehensive data source on NGOs in India because it integrates data from all of the largest public NGO databases, provides significant cleanup of the data, and supplements it via SPARQL queries and the UN SDG ontology. The user friendly GUI will provide the formal rigor and reasoning of the ontologies to non-technical users.
This project illustrates the benefit of an agile, reuse-based development model when using ontologies for real problems. This is a different model than the model used to create repositories of reusable vocabularies which relies on an overarching upper model to ensure consistency across the vocabularies. The DaanMatch KG project to date has had the equivalent of less than 6 person months 5 https://www.ogc.org/standard/geosparql/ of development but has created a system that demonstrates the goals of the DaanMatch project to utilize the UN sustainable development goals to match NGOs with CSRs and to replace traditional documentation with media and metadata recorded in real time as part of the execution of an NGO project. It does this with real world data collected by the DaanMatch team as well as test data generated by the KG team. The relational data model of the NGO module and the ontology model of the CSR module have been aligned via data properties as described in section 2.1 and the integration of the most recent data collected and generated by the NGO module into the CSR module is complete.
The Agile approach to ontology development (also demonstrated in [10]) utilizes SPARQL to automate many tasks traditionally done by hand. Rather than focus on a top-down methodology that assumes that the “correct” model can be developed before any data has been imported, the Agile approach recognizes that good design is both a bottom up and top down process. [8] I.e., the structure of data from systems that will populate the knowledge graph often impacts the design of the ontology as much as the domain and reusable vocabularies.
In addition, we are currently working with a data science graduate student from UC Berkeley who is accessing various possible uses of machine learning to further help realize the DaanMatch vision. The first result of this collaboration was a more sophisticated statistical algorithm for matching organizations to SDGs. Possible future opportunities include: • •
Utilize deep learning models such as ChatGPT to automate the creation of various forms required for NGOs to receive funding grants. As described above, many of the smaller NGOs, while providing excellent services, lack the business skills to negotiate the existing bureaucratic process.
Utilize ML to analyze the content of Prov:Entities (e.g., video and audio files) to determine if they may require human auditing due to potential fraud or error in documentation. One of the biggest issues that we anticipate with the long term DaanMatch vision is change management. People don’t like change. They especially don’t like change when it involves bureaucratic procedures. The case for corporations to adopt the DaanMatch model will be significantly aided if it can be demonstrated that the new approach not only makes life easier for NGOs so that they can focus more of their time providing services while also providing equal or superior levels of auditability with less effort via the OWL reasoner, SPARQL engine, and Machine Learning.
The project demonstrates how to use semantic technology for the benefit of society to achieve a high level of productivity, flexibility, and integration with traditional systems and data.
This work was conducted using the Protégé resource, which is supported by grant GM10331601 from the National Institute of General Medical Sciences of the United States National Institutes of Health. Thanks to Franz Inc. (http://www.allegrograph.com) and its help with AllegroGraph and Gruff.
D. Tripathi, Interviewee, CEO Assam State Disaster Management Authority. [Interview]. [Online].