🦥 LazyNLP - label data and train models with low effort

This library allows you to label data using zeroshot and train simple classifiers - without the need to do anything!

This library is for :

• ✅ lazy people
• ✅ fast prototyping or quick experiments
• ❌ super accurate results
• ❌ production-grade models
• ❌ every use-case

What you need

To use this library, all you need to do is:

• Have a list containing texts
• Have a second list containing all the possible labels you want to assign
• (When using from GitHub)pip install -r requirements.txt

Quickstart

First, install LazyNLP with pip install lazy-nlp.

Then you could use LazyNLP like this:

# import LazyNLP
from lazy_nlp import LazyNLP

# load in some data set
df = pd.read_csv("some_text_dataset.csv")

# convert the text column to a list
sentences = df["text"].to_list()

# write a list with all possible labels
labels = ["positive", "neutral", "negative"]

# LazyNLP will handle the rest for you
lnlp = LazyNLP()
model, encoder = lnlp.run(sentences, labels)

Save a model

The result of .run will be a pytorch model and a labelencoder. You can save these by using lnlp.save(model, encoder)

Using the model

After you saved the model and encoder, you can simply predict on new data like this:

# pass a list to .predict with a list of new data. You need to .save the model and encoder before!
preds = lnlp.predict(["This is a new sentence"])

LazyNLP steps

LazyNLP consists of three steps: zeroshot labeling, embedding and model training. With .run you trigger all of these steps at once, but you can also use LazyNLP only for the zeroshot, embedding or model training component idividually.

Zeroshot

# provide a list of texts you want to label as well as a list of all potential labels
your_texts = ["This is a bad sentence.", "This is another sentence.", "More sentences!", "I, too, am a sentence", "This is a good sentence."]
your_labels = ["negative", "neutral", "positive"]

# returns a list of zeroshot labels
zeroshot_labels = lnlp.zeroshot(your_texts, your_labels)

Embedding

your_texts = ["This is a bad sentence.", "This is another sentence.", "More sentences!", "I, too, am a sentence", "This is a good sentence."]

# returns the embedded texts
embeddings = lnlp.embed(your_texts)

Model training

If you have embedded texts as well as some labels, you can then train a model like this:

# returns a pytorch model and a label encoder
model, encoder = lnlp.classify(embeddings, your_labels)

The model is extremly simple: is a MLP with only one hidden layer. Nothing fancy at all, but usually enough for working with high quality embeddings.