An Honest Guy on AI IV: Entity Resolution, Pragmatism, and the “Black Box” Problem
If you have a low tolerance for buzzwords, you’re in the right place. This interview with Carlos Azeglio, Basis Technology’s senior product manager, gives an insider’s perspective on AI and its impact on industry without utopian visions, apocalyptic scenarios, or vague terminology. (For those in risk technology, make sure to check out the last two questions).
Could you tell us a little about yourself and your background in AI?
I’ve been involved in the data aggregation and entity resolution space for two decades now. I actually started my career as a librarian specializing in cataloging, indexing, reference, and research. The dotcom boom, fortunately for me, created a high demand for these skills.
My interaction with AI started around 6 years ago working with machine learning models for entity resolution platforms. For all those unfamiliar with entity resolution, it’s the process of connecting and linking real-world entities (people, places, organizations, etc.) that exist in structured and unstructured data to a specific knowledge base.
A quick example: If you match a person to an existing entity record in Wikipedia or some other source, you’ve just performed entity resolution.
Previously, all other solutions I worked with were heavily rule-based based approaches. These rule-based methods were very good at establishing high precision results but at a high cost in terms of resolution exhaust requiring manual intervention. The lessons I learned from my first AI experience were…
- Building machine learning models required lots of training data.
- The results were amazing…but sometimes difficult to explain.
- AI works best as a hybrid approach: You can’t disregard rules altogether.
- Yes, you can automate at much higher rates, but some level of manual remediation is always necessary.
- No matter how tight the curation rules are, you’d be surprised how subjective entity matching can be from a human perspective.
Often underestimated, entity resolution is the foundation of good data analytics and science. By uniting the available reference information around people, places, and organization, entity resolution creates a single version of the truth–the necessary foundation of any trustworthy analytical work.
AI has become a key technology in entity resolution, and its importance in the space is only growing. Because of the sudden and extreme rise in the need for quality data vetting and enrichment, businesses have no choice but to start embracing AI for core entity integration and unification tasks.
How would you describe AI?
AI is really an augmentation layer applied to tasks requiring constant decision making.
The more sophisticated the AI system, the faster the processing speed and the better it can handle complex decisions. The more extensible the AI platform, the more practical and applicable it becomes. These qualities are in particularly high demand in highly regulated industries, where performance, as well as auditability, are mandatory.
What do you think about the technology, where’s been and where’s going?
Over the last five to seven years, businesses have been getting their feet wet when it comes to AI. They’ve begun acclimating to AI’s benefits and implementing at the periphery of their organizations. The current trend is now to apply AI at deeper levels, integrating the technology into core business process workflows.
As far as the actual technology is concerned, there have been great strides made in areas requiring pattern recognition, knowledge graph generation, and NLP (Natural Language Processing).
Where’s the hype? What’s being oversold?
Often overused, “AI” means different things to different people. The technology is also often treated as though it were some kind of the panacea: AI will provide all answers and solve all problems. This kind of branding can lead to disappointment when ultra-high expectations meet the sobering reality of technological capability.
To a greater or lesser degree, this phenomenon is happening across industries.
I try to hold to a more pragmatic approach: Define a very narrow, realistic goal–focusing on augmenting instead of replacing human processes–and you’ll often find success.
Basis Tech recently produced a booklet entitled The Honest Guide to AI for Risk. Why does risk need AI? What are some major challenges facing this sector and what applications might AI have to them?
My background is primarily in compliance risk, and I believe AI has a lot to offer this sector. For professionals in this space, the list of operational challenges grows year after year. Regulatory guidance, penalties for infractions, and analytical tasks continue to pile up. Data complexity, multilingualism, and structural diversity are all on the (exponential) rise. Their job has become more difficult on every dimension.
Fortunately for this space, AI technology can help address each of these challenges–especially with NLP. Existing NLP-driven applications can help compliance professionals stay abreast of the relevant policy changes, enable scalability, handle esoteric content, and quickly adapt to ever-changing criminal tactics.
In short, it’s a good fit.
What are some of the major roadblocks to AI integration in this space and how can they be overcome?
One of the main obstacles to deeper AI integration is the inscrutability of an AI application’s decision-making process. This is known as the “black box” problem. In highly regulated sectors, being able to explain and audit actions is a requirement in most–if not all–use cases. This being the case, the most successful AI-centric products and services will include auditability and traceability from the ground up…instead of just as an add-on.
Inertia is another notable barrier…especially when it comes to the transition from purely human to augmented systems. Processes that have traditionally relied on manual workflows for remediation and vetting can be remarkably resistant to change. However, I believe the second obstacle can be overcome once the benefits of AI are understood.