最简单的方法/黑匣子根据现有的（标记的）数据集为短帖子建议标签？

Question

We have comments of ~50–300 chars pre-tagged with multiple topics like “music”, “tech” as well as particular films, artists etc.

We want to train an algorithm to autotag future comments. We'll manually tweak suggestions to improve accuracy and manually add many more tags (eg, new artists) over time. Posts will have one or many tags.

What's the simplest way to start this? I'm looking for something as simple as adding content and tag 1, tag 2... , automatically training, and then later giving it text to get back a list of suggested tags (preferably with confidence %).

We will end up with thousands of tags, and potentially 100k+ posts.

I've played around with a few things (naive bayes, LDA) but I feel there must be something simpler for such a common and simple use case. Perhaps a library or SaaS to make it this straightforward.

Answer 1

Consider support vector machines - with a preliminary feature extraction made of stemming , stop-words removal , n-gramming (especially skip-ngramming may provide a substantial boost at a cost).

Beware : no matter how brilliant a classifier, a bad feature extraction will render it dumb in no time.

Porter stemmer

snowball stemming