Radical Saussurean semantics?

I have been following Benjamin Schmidt‘s posts (and here and here) about word-embedding models (WEMs). I don’t claim to have any grasp of the underlying mathematics, but the results are very interesting – I’m planning to download the R-package and do some playing with text that I am working on with Brian Zuccala (when I have time [he said optimistically]).

Here I just wanted to draw attention to one foundational aspect of this approach. Schmidt comments:

The question that word embedding models ask is: what if we could model all relationship between words as spatial ones? Or put another way: how can we reduce words into a field where they are purely defined by their relations?

This seems to me to be a very Saussurean approach to semantics – each word is defined by its place in the system, and that is in turn defined by the relationships (especially the differences) between the target word and all other words. The problem for Saussurean semantics has been the scale of the task of establishing all those relationships, except in circumscribed domains such as pronouns. But if that task can be handled by machine learning procedures, then suddenly this is a viable approach! Of course there are problems: the learning data will only ever be a snapshot; very rare words may not occur; are numerical estimates of relationships the one’s we really want (and I’m sure there are others I’m not thinking of yet). Using a very large corpus should give some traction on these sort of problems, but even acknowledging them, these methods seem like potentially a huge advance for empirical semantics.


Leave a Reply

Fill in your details below or click an icon to log in:

WordPress.com Logo

You are commenting using your WordPress.com account. Log Out / Change )

Twitter picture

You are commenting using your Twitter account. Log Out / Change )

Facebook photo

You are commenting using your Facebook account. Log Out / Change )

Google+ photo

You are commenting using your Google+ account. Log Out / Change )

Connecting to %s

%d bloggers like this: