SkimLit Project: Replacing pretrained token embeddings with custom token embeddings actually improved results!

[sgkouzias]

By restructuring the token model as:

token_inputs = Input(shape=(1,), dtype=tf.string, name='token_inputs')
token_vectors = text_vectorizer(token_inputs)
token_embeddings = token_embed(token_vectors)
x = Conv1D(64, kernel_size=5, padding='same', activation='relu')(token_embeddings)
token_outputs = GlobalAveragePooling1D()(x)
token_model = tf.keras.Model(inputs=token_inputs, outputs=token_outputs)

I get the following results:

{'accuracy': 87.16735072156759,
'precision': 0.8721944632776405,
'recall': 0.8716735072156759,
'f1': 0.8692547073478577}