Test shows big data text analysis inconsistent, inaccurate
computerworld.com
computerworld.com
It's great to see physicists getting into other disciplines to show everyone else how wrong they are. As we all know, when you study physics and math, you're at the top of the reductionism pyramid, so you don't even really need to worry about having a background or experience in the thing you're studying. You just data at a formula, and if it doesn't work, you get to go to town berating professionals in the field using that formula. That's how it works, right?
Pre-processing your topic-modeling data is a pre-requisite to getting good results. This is fairly common knowledge. Not pre-processing is certainly the more naive approach. Computational linguists have backgrounds in things like syntax, semantics, and discourse so that they know what components of language to select for and how to format them without mis-representing the source data.
This article characterizes a straw man to tear down in the service of advertising a proprietary solution. It's basically an advertisement with a scientific reference for good measure. We call these "white papers" -- not serious journalism.
I can make these assertions from personal experience. I used latent Dirichlet allocation on high-fidelity natural language data to provide topic modeling for hundreds of thousands of wikis. We used the data for ad optimization as well as recommendations -- both of which provided statistically significant improvements in engagement. The approach worked. The recommendations were reproducible. I used all open-source software.
I guess I should have gotten into physics.
They merely showed that Latent Dirichlet Allocation is not the best method to solve the specific problem of classifying human language from unstructured text. New techniques will always come up to be more accurate and reproducible, and eventually become the new most used one. TopicMapping sounds promising. :)