nwf-nlp 0.1.1

NWF for NLP: Transformer encoders (BERT, DistilBERT) with (z, sigma) output for incremental learning, semantic search, text classification.

nwf-nlp

NWF for Natural Language Processing

nwf-nlp provides Transformer-based text encoders that produce semantic charges (z, sigma) for incremental learning, semantic search, and classification. Built on HuggingFace models (BERT, RoBERTa, DistilBERT) with a custom head for uncertainty estimation.

Features

• TransformerEncoder — wrapper over HuggingFace transformers (DistilBERT, BERT, RoBERTa)
• Output (z, sigma) — compatible with nwf-core Field and Mahalanobis search
• Incremental learning — add new categories without retraining the full model
• Semantic search — find documents by query using charge similarity
• Batch encoding — GPU support, configurable batch size
• Pooling — [CLS] or mean pooling

---

Installation

pip install nwf-core nwf-nlp

Requires: nwf-core>=0.2.3, torch, transformers, scikit-learn.

---

Quick Start

from nwf import Charge, Field
from nwf.nlp import TransformerEncoder

enc = TransformerEncoder("distilbert-base-uncased", latent_dim=64)
enc.fit(train_texts, epochs=3)
z, sigma = enc.encode("Some text")
charge = Charge(z=z, sigma=sigma)

field = Field()
for i, text in enumerate(texts):
    z, s = enc.encode(text)
    field.add(Charge(z=z, sigma=s), labels=[labels[i]], ids=[i])

---

API

TransformerEncoder

Parameter	Description
model_name	HuggingFace model: "distilbert-base-uncased", "bert-base-uncased", "roberta-base"
latent_dim	Output dimension of z
pooling	"cls" or "mean"
freeze_backbone	If True, only train the head (faster)
max_length	Max tokens (default 512)

Method	Description
fit(texts, epochs, batch_size, lr)	Train head on texts (unsupervised: Gaussian prior)
encode(texts, batch_size)	Returns (z, sigma) as numpy arrays

---

Examples

Install with examples: pip install nwf-nlp[examples]

Script	Description
20newsgroups.py	Incremental text classification: 3 categories, add sci.med without retraining

Run:

python examples/20newsgroups.py --epochs 2 --k 5
python examples/20newsgroups.py --save results/nlp.png

Notebook: notebooks/20newsgroups.ipynb

---

Application areas (сферы применения)

Area	Use case	Components
Incremental text classification	Add new categories without retraining	TransformerEncoder, Field, k-NN
Semantic search	Find documents by query in charge space	encode(query), Field.search
Topic modeling	Cluster documents by latent charges	z, sigma from TransformerEncoder

---

License

MIT

---

nwf-nlp (Русский)

NWF для обработки естественного языка

nwf-nlp предоставляет Transformer-энкодеры для текста с выходом семантических зарядов (z, sigma) для инкрементального обучения, семантического поиска и классификации.

Компоненты

• TransformerEncoder — обёртка над HuggingFace (DistilBERT, BERT, RoBERTa)
• Выход (z, sigma) — совместим с Field и поиском по Махаланобису
• Инкрементальность — добавление новых тем без переобучения
• Семантический поиск — поиск документов по запросу

Установка

pip install nwf-core nwf-nlp

Пример

from nwf.nlp import TransformerEncoder
from nwf import Charge, Field

enc = TransformerEncoder("distilbert-base-uncased", latent_dim=64)
enc.fit(тексты, epochs=3)
z, sigma = enc.encode("Текст для кодирования")
field.add(Charge(z=z, sigma=sigma), labels=[метка])

Лицензия

MIT