How Do Data Ontologies Organize Knowledge?

TL;DR

Data ontologies organize large amounts of information by defining entities, classifications, and relationships that humans and computers can understand. They make collective knowledge easier to create and use at scale, but effective systems still require human representations and ground truth. Simple tasks can rely on agreement, while complex tasks such as translation and clinical decisions require more sophisticated evaluation.

Transcript

Hi, and welcome to the a16z podcast. In this episode, we talk about collective intelligence, human computation, and really the mapping out of knowledge, with data ontologies in particular, and how data ontologies enable scalable knowledge creation, both in philosophical terms and also in a very real practical way in terms of, for example, a doctor ... Read More

Key Insights

• A data ontology is a structured way to organize large amounts of information by describing entities, their classifications, and their relationships. Wikipedia, Google Search, taxonomies, folder structures, Siri, and Waze illustrate different ways that knowledge or events can be arranged.
• Ontology originated as the philosophical study of being and what is real. In practical information systems, it becomes a mutually understandable representation of the world, connecting the limits of language with the categories and relationships used to encode knowledge.
• Common ontologies are necessary for coordination among people as well as communication between people and computers. Without shared structures, information exchanged across languages, expertise levels, or types of agents cannot be organized consistently enough to support large-scale collaboration.
• Computers require explicit relationships because they have not lived in the human world. A computer does not inherently know that a cat is an animal or that an animal is a thing, so an ontology supplies the hierarchy needed to interpret such concepts.
• Human involvement is necessary for creating representations that remain meaningful and interpretable to people. Machine learning can group data automatically, but human perspectives identify which entities, distinctions, and relationships matter within everyday perceptual experience.
• Ground truth is a major human contribution to deep learning systems because accurate algorithms need many reference answers. Comparing human ground truth with computer-generated classifications can expose assumptions supported by human experience but not by the available data.
• reCAPTCHA converted a routine security task into human computation for book digitization. Words that computers could not decipher were presented as CAPTCHA challenges, and matching answers from 10 people were treated as ground truth for recognizing the scanned text.
• Complex data requires more sophisticated validation than simple agreement. Duolingo could not expect identical translations from 10 bilingual contributors, so participants translated sentences, voted on other translations, and corrected answers when necessary to evaluate multiple potentially valid formulations.

Questions & Answers

Q: What is a data ontology and how does it work?

A data ontology is a way to impose order on large amounts of information by defining entities, classifying them, and describing how they relate to one another. It creates a shared representation that people and computers can interpret. Taxonomies, digital folder structures, Wikipedia, Google Search, Siri, and Waze all demonstrate aspects of ontology-based organization.

Q: Why are data ontologies important for computers?

Data ontologies give computers explicit structures for understanding concepts and relationships that humans often take for granted. Computers have not lived in the human world, so they do not inherently know that a cat belongs to the category of animals or that an animal is a thing. Ontologies encode these relationships in an organized, usable form.

Q: How do shared ontologies improve human coordination?

Shared ontologies give people a common framework for describing and organizing information. This framework reduces confusion when participants have different languages, levels of technical expertise, experiences, or assumptions about the world. By making categories and relationships mutually understandable, ontologies help diverse people coordinate flexibly and contribute to scalable knowledge projects through the internet.

Q: Why do machine learning systems still need humans?

Machine learning systems need humans to create representations that are intelligible and relevant to human concerns. Human perception determines which distinctions and relationships matter in everyday life. Humans also provide the ground truth required to train accurate deep learning systems. Without human input, automatically generated groupings may be difficult to interpret or may not reflect useful human priorities.

Q: How did reCAPTCHA help digitize books?

reCAPTCHA used words from scanned book pages that computers could not recognize reliably as challenges for internet users. People deciphered those words while completing CAPTCHA tasks, such as when buying tickets. By gathering answers from multiple users, the project turned a widespread security interaction into human computation that supported the recognition and digitization of book text.

Q: How did reCAPTCHA determine ground truth?

reCAPTCHA established ground truth by presenting the same word to 10 different people and comparing their answers. When all 10 participants agreed, the shared response was considered the correct transcription. This method worked well because identifying a pictured word generally has a single expected answer that can be tested through direct agreement among independent contributors.

Q: Why is validating translations harder than recognizing words?

Translation is harder to validate because a sentence can have several correct translations. When 10 people translate the same passage, their wording may differ even when their answers are valid, so exact agreement is not a reliable standard. Duolingo addressed this by letting people vote on translations and revise submissions they considered incorrect or improvable.

Q: How can data ontologies support healthcare decisions?

Data ontologies can create a common structure through which physicians, other healthcare roles, and machines exchange and interpret clinical knowledge. HumanDX applies an open-system approach in which people post clinical cases and multiple physicians independently attempt to solve them. Such structured collective knowledge can help machines augment one level of care with capabilities associated with the next level.

Summary & Key Takeaways

• Data ontologies provide shared structures for representing entities, classifications, and relationships. Examples include taxonomies, folder systems, Wikipedia, Google Search, Siri, and Waze. These structures help computers interpret a world they have not directly experienced and help people coordinate despite differences in language, expertise, perception, and background.

• Human participation remains essential because machine-generated groupings may not be intelligible or useful to people. Humans supply representations based on lived priorities and create the ground truth needed by deep learning systems. Comparing human classifications with computer outputs can also reveal which assumptions come from experience rather than the available data.

• Ground-truth methods must match the complexity of the task. reCAPTCHA could establish answers by asking 10 people to identify the same word and accepting agreement. Translation produced many valid alternatives, so Duolingo required voting and correction. Healthcare presents another complex domain where shared ontologies can support collaboration among physicians, machines, and different clinical roles.