Simple and powerful NLP framework, build your state-of-art model in 5 minutes for named entity recognition (NER), part-of-speech tagging (PoS) and text classification tasks.
Kashgare is:
- Human-friendly. Kashgare's code is straightforward, well documented and tested, which makes it very easy to understand and modify.
- Powerful and simple. Kashgare allows you to apply state-of-the-art natural language processing (NLP) models to your text, such as named entity recognition (NER), part-of-speech tagging (PoS) and classification.
- Keras based. Kashgare builds directly on Keras, making it easy to train your models and experiment with new approaches using different embeddings and model structure.
- Easy to fine-tune. Kashgare build-in pre-trained BERT and Word2vec embedding models, which makes it very simple to fine-tune your model based on this embeddings.
- Fully scalable. Kashgare provide a simple, fast, and scalable environment for fast experimentation.
- Embedding support
- Classic word2vec embedding
- BERT embedding
- GPT-2 embedding
- Sequence(Text) Classification Models
- CNNModel
- BLSTMModel
- CNNLSTMModel
- AVCNNModel
- KMaxCNNModel
- RCNNModel
- AVRNNModel
- DropoutBGRUModel
- DropoutAVRNNModel
- Sequence(Text) Labeling Models (NER, PoS)
- CNNLSTMModel
- BLSTMModel
- BLSTMCRFModel
- Model Training
- Model Evaluate
- GPU Support / Multi GPU Support
- Customize Model
| Task | Language | Dataset | Score | Detail |
|---|---|---|---|---|
| Named Entity Recognition | Chinese | People's Daily Ner Corpus | 92.20 (F1) | 基于 BERT 的中文命名实体识别 |
- Migrate to tf.keras
- ELMo Embedding
- Pre-trained models
- More model structure
Here is a set of quick tutorials to get you started with the library:
There are also articles and posts that illustrate how to use Kashgari:
The project is based on Keras 2.2.0+ and Python 3.6+, because it is 2019 and type hints is cool.
pip install kashgari
# CPU
pip install tensorflow==1.12.0
# GPU
pip install tensorflow-gpu==1.12.0lets run a text classification with CNN model over SMP 2017 ECDT Task1.
>>> from kashgari.corpus import SMP2017ECDTClassificationCorpus
>>> from kashgari.tasks.classification import CNNLSTMModel
>>> x_data, y_data = SMP2017ECDTClassificationCorpus.get_classification_data()
>>> x_data[0]
['你', '知', '道', '我', '几', '岁']
>>> y_data[0]
'chat'
# provided classification models `CNNModel`, `BLSTMModel`, `CNNLSTMModel`
>>> classifier = CNNLSTMModel()
>>> classifier.fit(x_data, y_data)
_________________________________________________________________
Layer (type) Output Shape Param #
=================================================================
input_1 (InputLayer) (None, 10) 0
_________________________________________________________________
embedding_1 (Embedding) (None, 10, 100) 87500
_________________________________________________________________
conv1d_1 (Conv1D) (None, 10, 32) 9632
_________________________________________________________________
max_pooling1d_1 (MaxPooling1 (None, 5, 32) 0
_________________________________________________________________
lstm_1 (LSTM) (None, 100) 53200
_________________________________________________________________
dense_1 (Dense) (None, 32) 3232
=================================================================
Total params: 153,564
Trainable params: 153,564
Non-trainable params: 0
_________________________________________________________________
Epoch 1/5
1/35 [..............................] - ETA: 32s - loss: 3.4652 - acc: 0.0469
...
>>> x_test, y_test = SMP2017ECDTClassificationCorpus.get_classification_data('test')
>>> classifier.evaluate(x_test, y_test)
precision recall f1-score support
calc 0.75 0.75 0.75 8
chat 0.83 0.86 0.85 154
contacts 0.54 0.70 0.61 10
cookbook 0.97 0.94 0.95 89
datetime 0.67 0.67 0.67 6
email 1.00 0.88 0.93 8
epg 0.61 0.56 0.58 36
flight 1.00 0.90 0.95 21
...from kashgari.embeddings import GPT2Embedding
from kashgari.tasks.classification import CNNLSTMModel
from kashgari.corpus import SMP2017ECDTClassificationCorpus
gpt2_embedding = GPT2Embedding('<path-to-gpt-model-folder>', sequence_length=30)
model = CNNLSTMModel(gpt2_embedding)
train_x, train_y = SMP2017ECDTClassificationCorpus.get_classification_data()
model.fit(train_x, train_y)from kashgari.embeddings import BERTEmbedding
from kashgari.tasks.classification import CNNLSTMModel
from kashgari.corpus import SMP2017ECDTClassificationCorpus
bert_embedding = BERTEmbedding('bert-base-chinese', sequence_length=30)
model = CNNLSTMModel(bert_embedding)
train_x, train_y = SMP2017ECDTClassificationCorpus.get_classification_data()
model.fit(train_x, train_y)from kashgari.embeddings import WordEmbeddings
from kashgari.tasks.classification import CNNLSTMModel
from kashgari.corpus import SMP2017ECDTClassificationCorpus
bert_embedding = WordEmbeddings('sgns.weibo.bigram', sequence_length=30)
model = CNNLSTMModel(bert_embedding)
train_x, train_y = SMP2017ECDTClassificationCorpus.get_classification_data()
model.fit(train_x, train_y)from kashgari.embeddings import BERTEmbedding
from kashgari.tasks.classification import CNNLSTMModel
train_x, train_y = prepare_your_classification_data()
# build model with embedding
bert_embedding = BERTEmbedding('bert-large-cased', sequence_length=128)
model = CNNLSTMModel(bert_embedding)
# or without pre-trained embedding
model = CNNLSTMModel()
# Build model with your corpus
model.build_model(train_x, train_y)
# Add multi gpu support
model.build_multi_gpu_model(gpus=8)
# Train, 256 / 8 = 32 samples for every GPU per batch
model.fit(train_x, train_y, batch_size=256)Thanks for your interest in contributing! There are many ways to get involved; start with the contributor guidelines and then check these open issues for specific tasks.
This library is inspired by and references following frameworks and papers.
- flair - A very simple framework for state-of-the-art Natural Language Processing (NLP)
- anago - Bidirectional LSTM-CRF and ELMo for Named-Entity Recognition, Part-of-Speech Tagging
- Chinese-Word-Vectors
