4
$\begingroup$

I am interested in an unsupervised approach to training a POS-tagger.

Labeling is very difficult and I would like to test a tagger for my specific domain (chats) where users typically write in lower cases etc. If it matters, the data is mostly in German.

I read about about old techniques like HMM, but maybe there are newer and better ways?

Improve this question
$\endgroup$
1
  • $\begingroup$ Did you try spaCy on your data? What were the issues? $\endgroup$
    – Adam Bittlingmayer
    Commented Jul 19, 2018 at 6:06

4 Answers 4

Reset to default
3
$\begingroup$

There are no unsupervised methods to train a POS-Tagger that have similar performance to human annotations or supervised methods.

The current state-of-the-art supervised methods for training POS-Tagger are Long short-term memory (LSTM) neural networks.

Improve this answer
$\endgroup$
2
1
$\begingroup$

There is no genuinely unsupervised method for POS tagging; we can think of it as, Parts of speech are inferred by us, with rules defined by the specific language being tagged. There is no mathematical "notion" for a part of speech that we can conclude given some text without any predefined rule established empirically (Which is why it is not genuinely unsupervised).

A weakly-supervised approach: Estimate the hidden state parameters of HMM using the Baum-Welch Algorithm.

And other is to implement a Maximum Entropy Model utilizing Beam Search, with rules established empirically(hence, not truly unsupervised)

Improve this answer
$\endgroup$
1
$\begingroup$

Very interested to hear what you need a tagger for in the context of chatbots?

Maybe you need just a stemmer - to produce 'base form' for an inflected word?

In that case, you can check this.

Improve this answer
$\endgroup$
2
  • $\begingroup$ like for NER. But indeed, I wanted to use POSs for a lemmatizer: github.com/WZBSocialScienceCenter/germalemma and spaCy's lemmatizer for german is very bad. Do you know about the accuarcy of this stemmer? Sounds interesting. $\endgroup$
    – Tido
    Commented Jul 17, 2018 at 18:38
  • $\begingroup$ Snowball is a stemmer so it does not produce lemma (base form) but a kind of 'simplified form' common to all inflected forms. That can lead to a non-existing forms like 'company' -> 'compani but it's enough to have a common form - common dictionary identifier for all inflected forms - for further processing, e.g. text classification. But it may be a bit awkward when you want to visualize results. Having looked to Snowball for German, it does the same - tries to cut off typical suffixes for conjugation and declination et al. E.g. Hunde -> Hund but also Katze, Katzen -> Katz. $\endgroup$
    – MkL
    Commented Jul 18, 2018 at 16:47
0
$\begingroup$

Fortunately, you don't need unsupervised methods for PoS tagging for most languages, especially for German. There are semi or "weakly" supervised methods like mentioned old HMM/EM approaches, however there is new and quite fresh solution with Error-Correcting Output-Code classification: Weakly supervised POS tagging without disambiguation.

Of course the accuracy of fully supervised methods like LSTM is far far better from semi supervised, but due to known issues of fully supervised methods (eg. lot of manual work) people still try to find lazy approaches. Excellent accuracy always cause higher costs.

Improve this answer
$\endgroup$

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Not the answer you're looking for? Browse other questions tagged or ask your own question.