Sure, I'd say you can do better than wit.ai, too :). I don't think their architecture is super sophisticated.

So, one awkwardness is that currently spaCy's parser is pretty crap on imperatives (e.g. "Go to the zoo"), because it has almost none of these in its training data. An awkward hack, if you know your data is always imperative, is to actually just transform it into a declarative sentence ("You must go to the zoo"). The syntax is analysed better this way. Hopefully soon I'll finish adding this transformation during the training process, so that you don't have to do that. But for now bear that limitation in mind.

If your problem is always such that you receive a single piece of text, and must learn a single label, then I recommend using a linear model and adding a grab-bag of features that mostly consists of sections of the dependency parse.

For example:
Input: Set the volume to zero
Intent: mute_volume
Parse: http://spacy.io/displacy/?full=Set%20the%20volume%20to%20zero.

Example feature extraction:

>>> from spacy.en import English
>>> nlp = English()
>>> doc = nlp(u'Set the volume to zero.')
>>> for word in doc:
...   print(word.text, word.head.text, word.dep_)
... 
Set Set ROOT
the volume det
volume Set dobj
to volume prep
zero to pobj
. Set punct

>>> for word in doc:
...   for i, right in enumerate(word.rights):
...     print(word.text, i, right.tag_, right.dep_)
... 
Set 0 NN dobj
Set 1 . punct
volume 0 IN prep
to 0 CD pobj

>>> for word in doc:
...   for i, left in enumerate(word.lefts):
...     print(word.text, i, left.tag_, left.dep_)
... 
volume 0 DT det

For early experiments, I would make the features string-concatenations, and use spacy.strings.StringStore to map them to sequential integer IDs, so that it's easy to play with an external machine learning library. Once you want better performance, I would switch that part of the code to Cython, and make an integer array of the feature, and then hash it. You can see an example of this in the averaged perceptron code I use for spaCy, which is in the library thinc.

If your data has more complicated labellings, it's probably worth using something like a CRF. Personally I always prefer beam-search and structured perceptron, because it's more efficient and more flexible. But, it starts to depend on the specifics.

Uh oh!

Can I train my own model for NER with just spaCy? #144

Description

Metadata

Metadata

Assignees

Labels

Type

Projects

Milestone

Relationships

Development

Issue actions