semantics, piles, and clusters

As I approach the singularity (doing the whole spectral clustering thing on my own, rather than relying on genius kids for the heavy lifting), this caught my eye as a great rationale for doing the semantics-free work:

An interesting property of corpus-based theories of cognition (such as Latent Semantic Analysis) is that they cannot be tested independently of the corpus. Imagine that we collect a corpus, run and Single Value Decomposition on it, and use the resulting space to predict human similarity judgments between certain words. Imagine that the model does not explain the data very well. Is it that the model’s processes are unrealistic, or is it that the corpus is not very representative? In this situation, those two factors are confounded. A possible solution is to test the same model with different corpora and different tasks. If the models explains the judgments’ variance across different situations, we have more convincing evidence of the psychological reality of the model. (“Creating Your Own LSA Space,” Jose Quesada, Carnegie Mellon University 2002.)

The limitations of trying to work with the basic themes inherent in text are large; the complexity of the arbitrary patterns of using language don’t seem to lend themselves to having computers learn meanings. The Semantic Web cult assumed that a perfect set of taxonomies and folksonomies could be created so that a bunch of marked-up text could ‘know’ what it was about, and communicate that through a retrieval system. This has largely been a failure.

Chris Anderson’s article “The End of Theory” takes this to a overwrought extreme:

Google’s founding philosophy is that we don’t know why this page is better than that one: If the statistics of incoming links say it is, that’s good enough. No semantic or causal analysis is required. That’s why Google can translate languages without actually “knowing” them (given equal corpus data, Google can translate Klingon into Farsi as easily as it can translate French into German). And why it can match ads to content without any knowledge or assumptions about the ads or the content.

This is true insofar as the data is good and the systems work, but those conditions are rare, and despite all that data Google and other systems that analyze behavior patterns are still not very good (and the translations are really bad). And Google still renders its results in a long scroll. Whatever the intelligence behind it, there is still a person at the other end, doing most of the work to find the right item in a long, unorganized list.

Clustering as an interface, by contrast, doesn’t care about semantics, and doesn’t even try for a strict ranking. Groups and rough hierarchy fit human models of organization much better than a long list (much as piles remain the usual way people organize). Loose piles don’t have to be semantically understood, a set of items is easier to take in and choose from (with two dimensions — item and group — rather than one). And when the algorithm is based on purely on user activity analysis, a better interface for presenting results, solving the interface problem (and thus engendering and capturing more user interaction) is really solving the whole problem of giving people information in ways they can understand it and use it.

UPDATE: Taking this further, it’s been seen for a long time that changes in behavior often happen when a few people that are part of a small group cause that entire group to adopt the change (like buying a kind of shoe, or phone, etc.). This is called the “cluster effect“:

“The cluster effect is similar to (but not the same as) the network effect. It is similar in the sense that the price-independent preferences of both the market and its participants are based on each ones perception of the other rather than the market simply being the sum of all its participants actions as is usually the case. Thus, by being an effect greater than the sum of its causes, and as it occurs spontaneously, the cluster effect is a usually cited example of emergence.”

What better way to engender cluster effects and the large amount of significant social effects they have than to show people the clusters of activity around their interests?

Post a Comment

Your email is never shared. Required fields are marked *