Join us on Monday to hear about some cool stuff you can do with Hidden Markov Models. Elias Ponvert will present his thesis research on unsupervised parsing. He'll talk about how he solved a (relatively) difficult NLP problem with (relatively) simple machine learning and hacks.
Unsupervised parsing is automating the discovery of grammatical structure of natural language sentences without human curated rules or annotations. He came to focus on Unsupervised Partial Parsing, the problem of learning to find names, interesting phrases and chunks of text.
It turns out:
1. Predicting phrases & chunks accurately is arguably much of what current state-of-the-art unsupervised parsing systems do
2. You can predict these phrases from raw text surprisingly accurately using very simple rules
3. Hidden Markov Models (and a variation thereof) do a way better job than those simple rules
He'll also show how combining a sequence of HMM partial parsers can be used to construct full grammatical constituent structure predictions. This technique is (still) the best performing unsupervised parsing system when evaluated against standard data sets in English, German and Chinese.
=== After the presentation, we'll have a time slot for anyone who would like to present to the group about their ML project in order to get ideas or feedback. ===
Metered parking downtown is free after 6pm on Mondays. Food and drinks at 6:30; speaker presentation at 7:00.