Hacker News Reader

Comment by pkd | Hacker News Reader

So I recently had a chance to use these APIs for a Microsoft hackathon project and wow are these APIs unstable right now.

I was trying to use their Linguistic Analysis API for POS tagging some text, and even for text as short as 500 characters, the API took over a minute to respond. If I bumped it upto a 1000 characters, the API would just error out and give a 500 status code without any explanation after keeping me hanging for over 5 minutes.

The same was the case with the speech to text API and some other ones as well.

As much as I appreciate there being something like this available for public use, these projects are effectively in alpha state and are not yet fit to be used in serious apps.

View on HN

Sorry to hear about the API instability. I am PM in Microsoft Speech Team and wanted to update that there was a defect in the Speech to Text JavaScript SDK which was fixed last week and there was no service outage. We are in preview now and have a 99.9% availability of our services.

Weird because POS tagging takes linear time in the number of words.

Anything that uses the standard decoding strategy (Viterbi) is quadratic in the number of "states" (i.e., tags), though, and that easily dominates.

(Nowadays, many people just use greedy---linear in the number of states---decoding because it rarely is much worse.)

That's assuming a lot... there are many, many ways to do part of speech tagging. You could imagine a slow implementation that exhaustively considered every possible POS tagging for an input sequence (would be O(k^N) where k is # of parts of speech)

There's not really that much assumption. Given the fact that good and simple linear time algorithms exist for decades it's very unlikely that they use some ridiculous permutation enumeration algorithm.

Even a Microsoft Research project has a fairly complex reductions implementation of a POS tagger that is blazingly fast and production ready (learning to search interface). [1]

[1]: https://github.com/JohnLangford/vowpal_wabbit/wiki

You could imagine a slow implementation that exhaustively considered every possible POS tagging for an input sequence (would be O(k^N) where k is # of parts of speech)

You could.. if you were crazy. No sane implementation would do that, and POS tagging isn't some brand new thing where people make that kind of mistake.

That seems.. like there is something wrong?

I run multiple POS taggers at decent scale, and I've never seen anything that slow.

Maybe they hadn't provisioned enough compute that day or something?

Reply on news.ycombinator.com