Text Classification Model

Kashgari provides several models for text classification, All labeling models inherit from the ABCClassificationModel. You could easily switch from one model to another just by changing one line of code.

Available Models

Name info

Train basic classification model

Kashgari provides the basic intent-classification corpus for experiments. You could also use your corpus in any language for training.

# Load build-in corpus.
from kashgari.corpus import SMP2018ECDTCorpus

train_x, train_y = SMP2018ECDTCorpus.load_data('train')
valid_x, valid_y = SMP2018ECDTCorpus.load_data('valid')
test_x, test_y = SMP2018ECDTCorpus.load_data('test')

# Or use your own corpus
train_x = [['Hello', 'world'], ['Hello', 'Kashgari']]
train_y = ['a', 'b']

valid_x, valid_y = train_x, train_y
test_x, test_x = train_x, train_y

Then train our first model. All models provided some APIs, so you could use any labeling model here.

import kashgari
from kashgari.tasks.classification import BiLSTM_Model

import logging

model = BiLSTM_Model()
model.fit(train_x, train_y, valid_x, valid_y)

# Evaluate the model
model.evaluate(test_x, test_y)

# Model data will save to `saved_ner_model` folder

# Load saved model
loaded_model = BiLSTM_Model.load_model('saved_classification_model')

# To continue training, compile the newly loaded model first
model.fit(train_x, train_y, valid_x, valid_y)

That’s all your need to do. Easy right.

Text classification with transfer learning

Kashgari provides varies Language model Embeddings for transfer learning. Here is the example for BERT Embedding.

import kashgari
from kashgari.tasks.classification import BiGRU_Model
from kashgari.embeddings import BertEmbedding

import logging

bert_embed = BertEmbedding('<PRE_TRAINED_BERT_MODEL_FOLDER>')
model = BiGRU_Model(bert_embed, sequence_length=100)
model.fit(train_x, train_y, valid_x, valid_y)

You could replace bert_embedding with any Embedding class in kashgari.embeddings. More info about Embedding: LINK THIS.

Adjust model’s hyper-parameters

You could easily change model’s hyper-parameters. For example, we change the lstm unit in BiLSTM_Model from 128 to 32.

from kashgari.tasks.classification import BiLSTM_Model

hyper = BiLSTM_Model.default_hyper_parameters()
# {'layer_bi_lstm': {'units': 128, 'return_sequences': False}, 'layer_dense': {'activation': 'softmax'}}

hyper['layer_bi_lstm']['units'] = 32

model = BiLSTM_Model(hyper_parameters=hyper)

Use custom optimizer

Kashgari already supports using customized optimizer, like RAdam.

from kashgari.corpus import SMP2018ECDTCorpus
from kashgari.tasks.classification import BiLSTM_Model
# Remember to import kashgari before than RAdam
from keras_radam import RAdam

train_x, train_y = SMP2018ECDTCorpus.load_data('train')
valid_x, valid_y = SMP2018ECDTCorpus.load_data('valid')
test_x, test_y = SMP2018ECDTCorpus.load_data('test')

model = BiLSTM_Model()
# This step will build token dict, label dict and model structure
model.build_model(train_x, train_y, valid_x, valid_y)
# Compile model with custom optimizer, you can also customize loss and metrics.
optimizer = RAdam()

# Train model
model.fit(train_x, train_y, valid_x, valid_y)

Use callbacks

Kashgari is based on keras so that you could use all of the tf.keras callbacks directly with Kashgari model. For example, here is how to visualize training with tensorboard.

from tensorflow.python import keras
from kashgari.tasks.classification import BiGRU_Model
from kashgari.callbacks import EvalCallBack

import logging

model = BiGRU_Model()

tf_board_callback = keras.callbacks.TensorBoard(log_dir='./logs', update_freq=1000)

# Build-in callback for print precision, recall and f1 at every epoch step
eval_callback = EvalCallBack(kash_model=model,

          callbacks=[eval_callback, tf_board_callback])

Multi-Label Classification

Kashgari support multi-label classification, Here is how we build one.

Let’s assume we have a dataset like this.

x = [
   ['This','news','are' , 'very','well','organized'],
   ['Multi-class', 'classification', 'means', 'a', 'classification', 'task', 'with', 'more', 'than', 'two', 'classes']

y = [
   ['A', 'B'],
   ['B', 'C'],

Now we need to init a Processor and Embedding for our model, then prepare model and fit.

import logging
from kashgari.embeddings import BertEmbedding
from kashgari.tasks.classification import BiLSTM_Model


bert_embed = BertEmbedding('<PRE_TRAINED_BERT_MODEL_FOLDER>')

model = BiLSTM_Model(bert_embed, sequence_length=100, multi_label=True)
model.fit(x, y)

Customize your own model

It is very easy and straightforward to build your own customized model, just inherit the ABCEmbedding and implement the default_hyper_parameters() function and build_model_arc() function.

from typing import Dict, Any

from tensorflow import keras

from kashgari.tasks.classification.abc_model import ABCClassificationModel
from kashgari.layers import L

import logging

class DoubleBLSTMModel(ABCClassificationModel):
    """Bidirectional LSTM Sequence Labeling Model"""

    def default_hyper_parameters(cls) -> Dict[str, Dict[str, Any]]:
        Get hyper parameters of model
            hyper parameters dict
        return {
            'layer_blstm1': {
                'units': 128,
                'return_sequences': True
            'layer_blstm2': {
                'units': 128,
                'return_sequences': False
            'layer_dropout': {
                'rate': 0.4
            'layer_time_distributed': {},
            'layer_output': {


    def build_model_arc(self):
        build model architectural
        output_dim = len(self.processor.label2idx)
        config = self.hyper_parameters
        embed_model = self.embedding.embed_model

        # Define your layers
        layer_blstm1 = L.Bidirectional(L.LSTM(**config['layer_blstm1']),
        layer_blstm2 = L.Bidirectional(L.LSTM(**config['layer_blstm2']),

        layer_dropout = L.Dropout(**config['layer_dropout'],

        layer_time_distributed = L.Dense(output_dim, **config['layer_output'])

        # You need to use this actiovation layer as final activation
        # to suppor multi-label classification
        layer_activation = self._activation_layer()

        # Define tensor flow
        tensor = layer_blstm1(embed_model.output)
        tensor = layer_blstm2(tensor)
        tensor = layer_dropout(tensor)
        tensor = layer_time_distributed(tensor)
        output_tensor = layer_activation(tensor)

        # Init model
        self.tf_model = keras.Model(embed_model.inputs, output_tensor)

model = DoubleBLSTMModel()
model.fit(train_x, train_y, valid_x, valid_y)

Short Sentence Classification Performance

We have run the classification tests on SMP2018ECDTCorpus. Here is the full code: colab link


  • EPOCHS = 30



  • BATCH_SIZE = 64

Embedding Model Best F1-Score Best F1 @ epochs
0 RoBERTa-wwm-ext BiLSTM_Model 92.89 15
1 RoBERTa-wwm-ext BiGRU_Model 94.57 10
2 RoBERTa-wwm-ext CNN_Model 92.95 12
3 RoBERTa-wwm-ext CNN_Attention_Model 92.07 3
4 RoBERTa-wwm-ext CNN_GRU_Model 89.56 22
5 RoBERTa-wwm-ext CNN_LSTM_Model 90.9 26
6 Bert-Chinese BiLSTM_Model 93.74 4
7 Bert-Chinese BiGRU_Model 93.12 13
8 Bert-Chinese CNN_Model 92.95 13
9 Bert-Chinese CNN_Attention_Model 92.04 8
10 Bert-Chinese CNN_GRU_Model 92.88 8
11 Bert-Chinese CNN_LSTM_Model 91.15 24
12 Bare BiLSTM_Model 81.96 11
13 Bare BiGRU_Model 82.86 9
14 Bare CNN_Model 86.61 11
15 Bare CNN_Attention_Model 78.84 12
16 Bare CNN_GRU_Model 66.14 26
17 Bare CNN_LSTM_Model 48.13 29