Making decisions¶

Loading the model¶

from zeropoint import ZeroPoint

zp = ZeroPoint()

Load it once and reuse it. Loading checks every model file against its sha256 and takes about a second. ZeroPoint("my-model") loads a model you adapted and saved yourself.

Choosing among options¶

d = zp.decide("my card was swallowed by the cash machine",
              ["block a card", "check the balance", "order a new card"])
d.choice        # the chosen option, or None

The options are any short texts: intents, queue names, categories. ZeroPoint reads the text and all the options together, so the wording of every option matters. Write them as a person would read them ("Billing and Payments", "order a new card"), not as codes ("BIL_02").

The order of the options does not change the result.
A list can be of any length. Options are read in blocks of 20 with the text, and all probabilities come from one distribution over the whole list.
An empty option, or the same option twice, raises ValueError.

What a decision holds¶

Field	What it holds
`choice`	the chosen option, or `None` when none is the most probable answer
`confidence`	probability of `choice` (of none when `choice` is `None`)
`p_none`	probability that none of the options fits
`probabilities`	probability of each option, in the order you gave them; with `p_none` they add up to 1
`depths`	how strongly the text belongs to each option (0: not at all)
`text_truncated`	`True` if the text was longer than 192 word pieces and its end was ignored
`options_truncated`	options longer than 16 word pieces, whose end was ignored
`blocks`	how many reads of 20 options the list needed
`latency_ms`	time the decision took

Setting a threshold¶

By default ZeroPoint returns its most probable answer. To act only when it is confident enough, compare confidence with your own threshold:

d = zp.decide(text, options)
if d.choice is None or d.confidence < 0.8:
    send_to_person(text)
else:
    route(text, d.choice)

The right threshold depends on what a wrong action costs. Choose it on a few hundred of your own labelled texts: for each threshold, count how many cases would be acted on and how many of those would be wrong.

Many texts at once¶

decisions = zp.decide_many(texts, ["transfer", "balance", "pay bill"])        # one list for all texts
decisions = zp.decide_many(texts, [options_1, options_2, options_3])          # one list per text

Each text is computed on its own, so every result is exactly what decide gives for that text alone.

Languages¶

ZeroPoint reads 25 languages written in the Latin script: Basque, Catalan, Croatian, Czech, Danish, Dutch, English, Estonian, Finnish, French, Galician, German, Hungarian, Irish, Italian, Latvian, Lithuanian, Maltese, Polish, Portuguese, Romanian, Slovak, Slovenian, Spanish and Swedish (zp.languages).

A text or an option written mostly in another script raises UnsupportedLanguage:

from zeropoint import UnsupportedLanguage

try:
    d = zp.decide(text, options)
except UnsupportedLanguage:
    send_to_person(text)

Other languages in the Latin script, such as Turkish or Vietnamese, are not detected: ZeroPoint answers, but it was not trained or tested on them. Its training texts are in English and Spanish; the other languages are read through the multilingual encoder it is built on.