Tackling Entity Resolution Across Resource Directory Data Sets: Lessons From the ServiceNet Team
Contributors:
David Botos, Head of Engineering – Connect 211
Greg Bloom, Director – Open Referral
On Open Referral’s blog, we recently posted about our Resource Record-Matcher: an open source AI-powered directory data collaboration tool.
As I wrote there, “At Connect211, we work with human service referral providers in more than twenty states to ‘orchestrate’ resource data pipelines, and support directory services. In areas where we serve multiple clients, and/or help them collaborate with their partners, we also facilitate collaboration across redundant yet siloed directory systems.”
We’ve seen many reasons that directories might choose to collaborate, including:
- Providing broader geographic coverage for directory and referral services
- Coordinating work to reduce maintenance costs
- Crowdsourcing quality-improvement feedback
To collaborate, directories need to compare their data, rapidly assessing where they overlap and what unique resources or features they offer. That’s where Connect 211’s Record Matcher tool comes in. Using a process called “entity resolution”, the Record Matcher walks users through verifying which Organizations and Services are the same between directories.
Record Matcher reduces the time required to compare data sets by days or even weeks. I think that’s pretty exciting, and I want to share a bit more about how we’ve done it.
The challenge: “sameness” is a matter of opinion.
We’ve discovered that the question “Are these entities the same?” is usually a matter of opinion. Take organizations, for example. It’s easy to think of them as recognizable brands with official names and even tax identification numbers that make them unique and easy to distinguish. However, the same organization might be referred to by different names in different contexts.
Furthermore:
- Large organizations may have sub-organizational entities that are essentially their own institutions.
- Government agencies don’t have unique IDs, and also frequently oversee programs or departments that overlap in irregular ways.
- Local data stewards sometimes separate organizations into separate entities based on regional offices or contacts.
These are just a few reasons there may not be a single right answer to the question of “Are these the same?”
We predict that effective entity-resolution and data-collaboration tools will facilitate the gradual alignment of perspectives on what constitutes a unique entity, though such perspectives may never completely converge. Legitimate differences of opinion about the boundaries around entities will remain. Our goal is not to eliminate opinions by establishing an objective truth, but rather to improve our capacity to cooperate and manage divergent opinions.
The role of (legally) objective facts in entity resolution.
Some kinds of data really are objective: legally verifiable facts. Phone numbers, URLs, email addresses, physical addresses, and external organization identifiers are all formally knowable in that they are governmentally known, rather than matters of perspective which people might reasonably disagree about. As such, these legally objective facts play a key role in determining what entities are the “same.”1 Combined with semantic (AI) comparisons, we can find duplicates even among records with very different names.
How to resolve services that are the same.
The most important entities that we have focused on resolving are services: Vital resources in our communities. There is a wide array of opinions about how to define “service.” For example, some directories define services in a very granular way, while others bundle related concepts together in longer descriptions. This results in the same cluster of activities defined as a single record in one system and five records in another.
To resolve services, we look for four categories of things they have in common:
- A parent organization
- Objectively factual data like phone numbers, addresses, etc
- Categories and taxonomies, also known as Service Terms for many 211s
- Semantic similarity in key fields like names and descriptions
#1 is actually not as useful as you’d expect, because organization names can vary significantly. And organization and service resolutions may align unevenly based on presentational requirements: sometimes organizations partner with each other, for instance, or are co-located, in ways that matter more or less to different kinds of users. It’s messy! We’re working to improve this, and in the meantime, our clients’ top priority is enhancing the quality of service-level information, which their users really need.
So our current method primarily leverages #2, and then draws inference from #3 and #4.2
Resolving similar categories from different taxonomies.
Categories, like those used in taxonomic schemas, provide highly valuable signals for comparing and resolving services; when two services from different sources share related taxonomic terms or categories, there is a stronger basis for comparison. Of course, the categories themselves need to be mapped to each other first.
Unlike other entities studied, taxonomies (also known as “indexing systems”) do not have internal duplication. Within a single taxonomy, each term is (at least in principle) distinct, which makes resolving related taxonomies across different sources more of a crosswalk process.
Consensus on how terms are mapped varies, and we must design for multiple concurrent answers. That said, broad adoption of standard taxonomic mappings across major indexing systems is one of the best ways to improve entity resolution.
Lessons learned resolving Organizations.
Initially, we tried to match organizations as a precursor and aid for matching services. Two things changed our minds:
- As mentioned earlier, organizations were harder to resolve than expected. Their boundaries (whether by domain, description, or geography) vary widely between directories.
- We eventually discovered we could resolve services without first resolving organizations.
We decided to punt organization resolution for this round of development, but not before mapping out plans for the future.
The best approach is to create “global” organization records that require legal identifiers such as FEINs (where possible), which we can then associate with subsidiary organizations, like an umbrella. This approach would preserve the fidelity of individual data sources while also aligning them globally. Whether there is value in hard-aligning services to those mapped organizations is an open question in our team. Regardless, this method would balance the benefits of aggregation while preventing the collapse of localized distinctions for organizations.
Humans Still Need to Make the Decisions
All of this modeling and mechanical analysis is, at best, only useful to the extent that it helps knowledgeable humans make faster and more effective decisions. It’s actually people who resolve the identities, and our tools just help them do that more efficiently. Humans remain in the loop for critical steps:
- Verify or correct our tool’s proposed pair-wise groupings. AI does get it wrong sometimes. So our tools are designed to present possible matches for human review, which requires way less effort than doing the comparison manually.
- Prioritize data in the final, reconciled result based on the target audience’s subjective requirements. Only humans know what’s important.
Humans are also irreplaceable (it turns out) for gathering resource information in the first place – which is to say: as impressive as our AI tooling is, it depends on the expertise and relationships that our clients offer their communities, by which they aggregate directory data. But that’s for another article.
What Comes Next?
After identifying which entities are the same (and verifying the results), the next step is to reconcile groups of the same entities so that the resulting outputs, such as a resource directory database, don’t display duplicates to users. This is a key step in synthesizing diverse data sources while preserving consistency and usability for people seeking information.
Reconciliation is a complicated topic with many potential solutions that are outside the scope of this article. If you would like to learn more, we will describe a solution that we are currently developing in the next blog post: A Minimal Approach to Deduplicating Services in Resource Directories.
The State Of This Project
Through pilot projects in Washington State and Illinois, we have progressed at a breakneck pace since 2022. Early prototypes were simple proof-of-concepts. As of 2026, we are running a beautiful user interface integrated with the enterprise-grade data orchestration from Connect 211. We have also started releasing our code:
- The model we use for entity resolution is available on Hugging Face (https://huggingface.co/connect211/HSDS-Dedupe-LightGBM-v1.2)
- The Python Package that embeds that into our orchestration pipeline is available for reference as well (https://github.com/211-Connect/hsds-entity-resolution)
Work is ongoing as we shift from experimentation into enterprise data federation and collaboration infrastructure. Our years of pilots have informed a clearer understanding of how to turn multiple overlapping data sources into a scalable system.
As work continues, we are actively looking for:
- More pilot opportunities
- Funders
- Collaborators
We are unlocking the ability to create supply chains for resource data and reducing the number of overlapping data silos that currently dot our resource directory landscape. This may be relevant to Rural Health Transformation Projects, bringing more accurate directory data to sparsely served, less-connected areas. The ripple effect is huge.
Please reach out to sky@connect211.com to learn more and collaborate.
- Formatting specifics (like “Street” vs “St,” or whether a URL has “www” or not) might still vary across datasets, but for trained AI models, these differences are increasingly trivial to resolve. ↩︎
- One might think we could resolve organizations first and then use that as a stronger signal for services. Experience suggests it’s not that simple. Organization and service resolutions may align unevenly based on presentational requirements. Further, resolving organizations first adds an additional burden to humans involved in this process. We found it more expedient and effective to proceed directly to service resolution at this time. ↩︎