I haven't come across many blog posts comparing "standard" approaches to text classification. In this (soon-to-be) notebook, I'm going to evaluate logistic regression (a standard industry workhorse) against a ConvNet (the "new kid" baseline) on Twitter sentiment analysis.
(This work was inspired by some experiments I was doing on search query classification with some data that I can't share.)
- scikit-learn pipeline with a char-based CountVectorizer (3,4,5,6-grams) and LogisticRegression (with xval'd regularizer)
- basic ConvNet adhering closely to the architecture in Kim (2014) with filters of size 3,4,5,6 (100 each)
One of my pet peeves about posts like this is the lack of detail w.r.t. things like system hardware/architecture, training time, etc. as well as the simplicity of the eval metrics (usually just raw accuracy).