Concepts are emergent effects

I started off reading Ryan Heuser’s post, Word Vectors in the Eighteenth Century, Episode 1: Concepts, and quickly followed a link to Michael Gavin’s post, The Arithmetic of Concepts: a response to Peter de Bolla. Blitzing on through, as I tend to do with these things, I came up against this in the final paragraph:

What’s more striking to me, though, is how commensurable the assumptions of vector-based semantics are with a theory of concepts that sees them, not as mental objects or meanings, per se, but instead as what de Bolla calls “the common unshareable of culture”: a pattern of word use that establishes networks across the language. A concept is not denoted by a word; rather, concepts are emergent effects of words’ co-occurrence with other words.

YES! But not without qualification.

Unfortunately this is not the time or place to explain what I mean, as that would require that I explain my ideas on semantics. That would require more than a blog post. Way ore.

I will remark, however, that the concept of a “word” is a tricky one. As generally used the concept encompasses meaning and syntactic affordances as well as orthography and sound. The computational techniques that Gavin is writing about, however, have no access whatever to meanings and syntactic affordances, just to orthography (in digital representation). But if you get a large corpus it is possible to analyze that corpus in a way that reveals traces, if you will, of word meanings from a mathematical model of the corpus.

But then, the young child learning language doesn’t have direct access to word meanings either. The child hears the words and has access to the context of utterance, which of course includes the person uttering the word. And out of that manages to extract meanings for words. The meanings initially extracted are crude, but they become refined over time.

Now, as I have explained more times than enough, early in my career I was knee deep in semantic models developed in computational linguistics, not the models Gavin mentions, but the semantic of cognitive networks of the 1970s and 1980s. For such models it was clear that the meaning of a node in the network was a function of its position in the network. A word’s signifier would be one node in the network and its signified would be, not some other node, but well, you know, it would be a function of the whole network. To a first approximation that’s what Gavin is saying.

He also says (quoting from a formal book review of his):

If concepts exist culturally as lexical networks rather than as expressions contained in individual texts, the whole debate between distant and close reading needs reframing. Conceptual history should be traceable using techniques like lexical mapping and supervised probabilistic topic modeling.

Yes. And I say something like that in a blog post on discourse and conceptual topology which I then included in a working paper: Toward a Computational Historicism: From Literary Networks to the Autonomous Aesthetic. When I turned to computational linguistics in the mid-1970s I was pursuing “close” reading, not “distant” reading. But this work of Gavin’s sure has a familiar ring.

The fact is that while computational critics have been very careful to avoid any use or even mention of computational theories of mind and thinking, they work they do seems inevitably to point back to those ideas from the 1970s and 1980s. And why should we be surprised about that?

If the eye were not sun-like, the sun’s light it would not see. 

– Johann Wolfgang von Goethe