2017-05-25 12:18:02 +03:00
|
|
|
|
//- 💫 DOCS > USAGE > SPACY 101 > TRAINING
|
|
|
|
|
|
2017-06-01 12:53:16 +03:00
|
|
|
|
p
|
|
|
|
|
| spaCy's models are #[strong statistical] and every "decision" they make –
|
|
|
|
|
| for example, which part-of-speech tag to assign, or whether a word is a
|
|
|
|
|
| named entity – is a #[strong prediction]. This prediction is based
|
|
|
|
|
| on the examples the model has seen during #[strong training]. To train
|
|
|
|
|
| a model, you first need training data – examples of text, and the
|
|
|
|
|
| labels you want the model to predict. This could be a part-of-speech tag,
|
|
|
|
|
| a named entity or any other information.
|
|
|
|
|
|
|
|
|
|
p
|
|
|
|
|
| The model is then shown the unlabelled text and will make a prediction.
|
|
|
|
|
| Because we know the correct answer, we can give the model feedback on its
|
|
|
|
|
| prediction in the form of an #[strong error gradient] of the
|
|
|
|
|
| #[strong loss function] that calculates the difference between the training
|
|
|
|
|
| example and the expected output. The greater the difference, the more
|
|
|
|
|
| significant the gradient and the updates to our model.
|
|
|
|
|
|
|
|
|
|
+aside
|
|
|
|
|
| #[strong Training data:] Examples and their annotations.#[br]
|
|
|
|
|
| #[strong Text:] The input text the model should predict a label for.#[br]
|
|
|
|
|
| #[strong Label:] The label the model should predict.#[br]
|
|
|
|
|
| #[strong Gradient:] Gradient of the loss function calculating the
|
|
|
|
|
| difference between input and expected output.
|
|
|
|
|
|
|
|
|
|
+image
|
|
|
|
|
include ../../../assets/img/docs/training.svg
|
|
|
|
|
.u-text-right
|
|
|
|
|
+button("/assets/img/docs/training.svg", false, "secondary").u-text-tag View large graphic
|
|
|
|
|
|
|
|
|
|
p
|
|
|
|
|
| When training a model, we don't just want it to memorise our examples –
|
|
|
|
|
| we want it to come up with theory that can be
|
|
|
|
|
| #[strong generalised across other examples]. After all, we don't just want
|
|
|
|
|
| the model to learn that this one instance of "Amazon" right here is a
|
|
|
|
|
| company – we want it to learn that "Amazon", in contexts #[em like this],
|
|
|
|
|
| is most likely a company. That's why the training data should always be
|
|
|
|
|
| representative of the data we want to process. A model trained on
|
|
|
|
|
| Wikipedia, where sentences in the first person are extremely rare, will
|
|
|
|
|
| likely perform badly on Twitter. Similarly, a model trained on romantic
|
|
|
|
|
| novels will likely perform badly on legal text.
|
|
|
|
|
|
|
|
|
|
p
|
|
|
|
|
| This also means that in order to know how the model is performing,
|
|
|
|
|
| and whether it's learning the right things, you don't only need
|
|
|
|
|
| #[strong training data] – you'll also need #[strong evaluation data]. If
|
|
|
|
|
| you only test the model with the data it was trained on, you'll have no
|
|
|
|
|
| idea how well it's generalising. If you want to train a model from scratch,
|
|
|
|
|
| you usually need at least a few hundred examples for both training and
|
|
|
|
|
| evaluation. To update an existing model, you can already achieve decent
|
|
|
|
|
| results with very few examples – as long as they're representative.
|