Semantic and Epistemic Networks

I am very interested in modeling a person’s network of ideas. What key concepts or values particularly motivate their thinking and how are those ideas connected?

I see this task as being particularly valuable in understanding and improving civil and political discourse. In this model, dialogue can be seen as an informal and iterative process through which people think about how their own ideas are connected, reason with each other about what ideas should be connected, and ultimately revise (or don’t) their way of thinking by adding or removing idea nodes or connections between them.

This concept of knowledge networks – epistemic networks – has been used by David Williamson Shaffer to measure the development of students’ professional knowledge; eg their ability to “think like an engineer” or “think like an urban planner.” More recently, Peter Levine has advanced the use of epistemic networks in “moral mapping” – modeling a person’s values and ways of thinking.

This work has made valuable progress, but a critical question remains: just what is the best way to model a person’s epistemic network? Is there an unbiased way to determine the most critical nodes? Must we rely on a given person’s active reasoning to determine the links? In the case of multi-person exchanges, what determines if two concepts are the “same”? Is semantic similarity sufficient, or must individuals actively discuss and determine that they do each indeed mean the same thing? If you make adjustments to a visualized epistemic network following a discussion, can we distinguish between genuine changes in view from corrections due to accidental omission?

Questions and challenges abound.

But these problems aren’t necessarily insurmountable.

As a starting place, it is helpful to think about semantic networks. In the 1950s, Richard H. Richens original proposed semantic networks as a tool to aid in machine translation.

“I refer now to the construction of an interlingua in which all the structural peculiarities of the base language are removed and we are left with what I shall call a ‘semantic net’ of ‘naked ideas,'” he wrote. “The elements represent things, qualities or relations…A bond points from a thing to its qualities or relations, or from a quality or relation to a further qualification.”

Thus, from its earliest days, semantic networks were seen as somewhat synonymous with epistemic networks: words presumably represent ideas, so it logically follows that a network of words is a network of ideas.

This may well be true, but I find it helpful to separate the two ideas. A semantic network is observed; an epistemic network is inferred.

That is, through any number of advanced Natural Language Processing algorithms, it is essentially possible to feed text into a computer and have it return of network of words which are connected in that text.

You can imagine some simple algorithms for accomplishing this: perhaps two words are connected if they co-occur in the same sentence or paragraph. Removing stop words prevents your retrieved network from being over connected by instances of “the” or “a.” Part-of-speech tagging – a relatively simple task thanks to huge databanks of tagged corpora – can bring an additional level of sophistication. Perhaps we want to know which subjects are connected to which objects. And there are even cooler techniques relying on probabilistic models or projections of the corpus into k-space, where k is the number of unique words.

These models typically assume some type of unobserved data – eg, we observe a list of words and use that to discover the unobserved connections – but colloquially speaking, semantic networks are observed in the sense that they can be drawn out directly from a text. They exist in some indirect but concrete way.

And while it seems fair to assume that words do indeed have meaning, it still takes a bit of a leap to take a semantic network as synonymous with an epistemic network.

Consider an example: if we were to take some great novel and cleverly reduce it to a semantic network, would the resulting network illustrate exactly what the author was intending?

The fact that it’s even worth asking that question to me indicates that the two are not intrinsically one and the same.

Arguably, this is fundementally a matter of degrees. It seems reasonable to say that, unless our algorithm was terribly off, the semantic network can tell us something interesting and worthwhile about the studied text. Yet it seems like a stretch to claim that such a simplistic representation could accurately and fully capture the depth of concepts and connections an author was seeking to convey.

If that were the case, we could study networks instead of reading books and – notably – everyone would agree on their meaning.

A semantic network, then, can be better considered as a representation of an epistemic network. It takes reason and judgement to interpret a semantic network epistemically.

Perhaps it is sufficient to be aware of the gap between these two – to know that interpreting a semantic network epistemically necessarily means introducing bias and methodological subjectivity.

But I wonder if there’s something better we can do to model this distinction – some better way to capture the complex, dynamic, and possibly conflicting essence of a more accurately epistemic network.


Leave a Reply

Your email address will not be published. Required fields are marked *

This site uses Akismet to reduce spam. Learn how your comment data is processed.