Menjelajahi Embedding Putar TF-Hub CORD-19

Lihat di TensorFlow.org Jalankan di Google Colab Lihat di GitHub Unduh buku catatan Lihat model TF Hub

The CORD-19 Swivel teks modul dari TF-Hub embedding ( https://tfhub.dev/tensorflow/cord-19/swivel-128d/3 ) dibangun untuk para peneliti dukungan menganalisis teks bahasa alami terkait dengan COVID-19. Embeddings ini dilatih pada judul, penulis, abstrak, teks tubuh, dan judul referensi dari artikel di CORD-19 dataset .

Dalam colab ini kita akan:

  • Analisis kata-kata yang mirip secara semantik di ruang embedding
  • Latih pengklasifikasi pada kumpulan data SciCite menggunakan penyematan CORD-19

Mempersiapkan

import functools
import itertools
import matplotlib.pyplot as plt
import numpy as np
import seaborn as sns
import pandas as pd

import tensorflow as tf

import tensorflow_datasets as tfds
import tensorflow_hub as hub

from tqdm import trange

Analisis embeddings

Mari kita mulai dengan menganalisis embedding dengan menghitung dan memplot matriks korelasi antara istilah yang berbeda. Jika embedding berhasil menangkap arti dari kata-kata yang berbeda, vektor embedding dari kata-kata yang mirip secara semantik harus berdekatan. Mari kita lihat beberapa istilah terkait COVID-19.

# Use the inner product between two embedding vectors as the similarity measure
def plot_correlation(labels, features):
  corr = np.inner(features, features)
  corr /= np.max(corr)
  sns.heatmap(corr, xticklabels=labels, yticklabels=labels)

# Generate embeddings for some terms
queries = [
  # Related viruses
  'coronavirus', 'SARS', 'MERS',
  # Regions
  'Italy', 'Spain', 'Europe',
  # Symptoms
  'cough', 'fever', 'throat'
]

module = hub.load('https://tfhub.dev/tensorflow/cord-19/swivel-128d/3')
embeddings = module(queries)

plot_correlation(queries, embeddings)

png

Kita dapat melihat bahwa embedding berhasil menangkap arti dari istilah yang berbeda. Setiap kata mirip dengan kata lain dari clusternya (yaitu "coronavirus" sangat berkorelasi dengan "SARS" dan "MERS"), sedangkan kata tersebut berbeda dari istilah cluster lainnya (yaitu kesamaan antara "SARS" dan "Spanyol" adalah mendekati 0).

Sekarang mari kita lihat bagaimana kita bisa menggunakan embeddings ini untuk menyelesaikan tugas tertentu.

SciCite: Klasifikasi Intensi Kutipan

Bagian ini menunjukkan bagaimana seseorang dapat menggunakan penyematan untuk tugas-tugas hilir seperti klasifikasi teks. Kami akan menggunakan dataset SciCite dari TensorFlow dataset ke maksud kutipan mengklasifikasikan dalam makalah akademis. Diberi kalimat dengan kutipan dari makalah akademis, klasifikasikan apakah maksud utama kutipan adalah sebagai informasi latar belakang, penggunaan metode, atau hasil perbandingan.

builder = tfds.builder(name='scicite')
builder.download_and_prepare()
train_data, validation_data, test_data = builder.as_dataset(
    split=('train', 'validation', 'test'),
    as_supervised=True)

Mari kita lihat beberapa contoh berlabel dari set pelatihan

Melatih pengklasifikasi niat citaton

Kami akan melatih classifier pada para dataset SciCite menggunakan Keras. Mari kita buat model yang menggunakan embeddings CORD-19 dengan lapisan klasifikasi di atasnya.

Hyperparameter

Model: "sequential"
_________________________________________________________________
 Layer (type)                Output Shape              Param #   
=================================================================
 keras_layer (KerasLayer)    (None, 128)               17301632  
                                                                 
 dense (Dense)               (None, 3)                 387       
                                                                 
=================================================================
Total params: 17,302,019
Trainable params: 387
Non-trainable params: 17,301,632
_________________________________________________________________

Latih dan evaluasi modelnya

Mari kita latih dan evaluasi model untuk melihat kinerja pada tugas SciCite

EPOCHS = 35
BATCH_SIZE = 32

history = model.fit(train_data.shuffle(10000).batch(BATCH_SIZE),
                    epochs=EPOCHS,
                    validation_data=validation_data.batch(BATCH_SIZE),
                    verbose=1)

Epoch 1/35
257/257 [==============================] - 3s 7ms/step - loss: 0.9244 - accuracy: 0.5924 - val_loss: 0.7915 - val_accuracy: 0.6627
Epoch 2/35
257/257 [==============================] - 2s 5ms/step - loss: 0.7097 - accuracy: 0.7152 - val_loss: 0.6799 - val_accuracy: 0.7358
Epoch 3/35
257/257 [==============================] - 2s 7ms/step - loss: 0.6317 - accuracy: 0.7551 - val_loss: 0.6285 - val_accuracy: 0.7544
Epoch 4/35
257/257 [==============================] - 2s 5ms/step - loss: 0.5938 - accuracy: 0.7687 - val_loss: 0.6032 - val_accuracy: 0.7566
Epoch 5/35
257/257 [==============================] - 2s 5ms/step - loss: 0.5724 - accuracy: 0.7750 - val_loss: 0.5871 - val_accuracy: 0.7653
Epoch 6/35
257/257 [==============================] - 2s 5ms/step - loss: 0.5580 - accuracy: 0.7825 - val_loss: 0.5800 - val_accuracy: 0.7653
Epoch 7/35
257/257 [==============================] - 2s 5ms/step - loss: 0.5484 - accuracy: 0.7870 - val_loss: 0.5711 - val_accuracy: 0.7718
Epoch 8/35
257/257 [==============================] - 2s 5ms/step - loss: 0.5417 - accuracy: 0.7896 - val_loss: 0.5648 - val_accuracy: 0.7806
Epoch 9/35
257/257 [==============================] - 2s 5ms/step - loss: 0.5356 - accuracy: 0.7902 - val_loss: 0.5628 - val_accuracy: 0.7740
Epoch 10/35
257/257 [==============================] - 2s 7ms/step - loss: 0.5313 - accuracy: 0.7903 - val_loss: 0.5581 - val_accuracy: 0.7849
Epoch 11/35
257/257 [==============================] - 2s 5ms/step - loss: 0.5277 - accuracy: 0.7928 - val_loss: 0.5555 - val_accuracy: 0.7838
Epoch 12/35
257/257 [==============================] - 2s 5ms/step - loss: 0.5242 - accuracy: 0.7940 - val_loss: 0.5528 - val_accuracy: 0.7849
Epoch 13/35
257/257 [==============================] - 2s 5ms/step - loss: 0.5215 - accuracy: 0.7947 - val_loss: 0.5522 - val_accuracy: 0.7828
Epoch 14/35
257/257 [==============================] - 2s 5ms/step - loss: 0.5190 - accuracy: 0.7961 - val_loss: 0.5527 - val_accuracy: 0.7751
Epoch 15/35
257/257 [==============================] - 2s 5ms/step - loss: 0.5176 - accuracy: 0.7940 - val_loss: 0.5492 - val_accuracy: 0.7806
Epoch 16/35
257/257 [==============================] - 2s 5ms/step - loss: 0.5154 - accuracy: 0.7978 - val_loss: 0.5500 - val_accuracy: 0.7817
Epoch 17/35
257/257 [==============================] - 2s 5ms/step - loss: 0.5136 - accuracy: 0.7968 - val_loss: 0.5488 - val_accuracy: 0.7795
Epoch 18/35
257/257 [==============================] - 2s 5ms/step - loss: 0.5127 - accuracy: 0.7967 - val_loss: 0.5504 - val_accuracy: 0.7838
Epoch 19/35
257/257 [==============================] - 2s 5ms/step - loss: 0.5111 - accuracy: 0.7970 - val_loss: 0.5470 - val_accuracy: 0.7860
Epoch 20/35
257/257 [==============================] - 2s 5ms/step - loss: 0.5101 - accuracy: 0.7972 - val_loss: 0.5471 - val_accuracy: 0.7871
Epoch 21/35
257/257 [==============================] - 2s 5ms/step - loss: 0.5082 - accuracy: 0.7997 - val_loss: 0.5483 - val_accuracy: 0.7784
Epoch 22/35
257/257 [==============================] - 2s 5ms/step - loss: 0.5077 - accuracy: 0.7995 - val_loss: 0.5471 - val_accuracy: 0.7860
Epoch 23/35
257/257 [==============================] - 2s 5ms/step - loss: 0.5064 - accuracy: 0.8012 - val_loss: 0.5439 - val_accuracy: 0.7871
Epoch 24/35
257/257 [==============================] - 2s 5ms/step - loss: 0.5057 - accuracy: 0.7990 - val_loss: 0.5476 - val_accuracy: 0.7882
Epoch 25/35
257/257 [==============================] - 2s 5ms/step - loss: 0.5050 - accuracy: 0.7996 - val_loss: 0.5442 - val_accuracy: 0.7937
Epoch 26/35
257/257 [==============================] - 2s 5ms/step - loss: 0.5045 - accuracy: 0.7999 - val_loss: 0.5455 - val_accuracy: 0.7860
Epoch 27/35
257/257 [==============================] - 2s 5ms/step - loss: 0.5032 - accuracy: 0.7991 - val_loss: 0.5435 - val_accuracy: 0.7893
Epoch 28/35
257/257 [==============================] - 2s 5ms/step - loss: 0.5034 - accuracy: 0.8022 - val_loss: 0.5431 - val_accuracy: 0.7882
Epoch 29/35
257/257 [==============================] - 2s 5ms/step - loss: 0.5025 - accuracy: 0.8017 - val_loss: 0.5441 - val_accuracy: 0.7937
Epoch 30/35
257/257 [==============================] - 2s 5ms/step - loss: 0.5017 - accuracy: 0.8013 - val_loss: 0.5463 - val_accuracy: 0.7838
Epoch 31/35
257/257 [==============================] - 2s 5ms/step - loss: 0.5015 - accuracy: 0.8017 - val_loss: 0.5453 - val_accuracy: 0.7871
Epoch 32/35
257/257 [==============================] - 2s 5ms/step - loss: 0.5011 - accuracy: 0.8014 - val_loss: 0.5448 - val_accuracy: 0.7915
Epoch 33/35
257/257 [==============================] - 2s 5ms/step - loss: 0.5006 - accuracy: 0.8025 - val_loss: 0.5432 - val_accuracy: 0.7893
Epoch 34/35
257/257 [==============================] - 2s 5ms/step - loss: 0.5005 - accuracy: 0.8008 - val_loss: 0.5448 - val_accuracy: 0.7904
Epoch 35/35
257/257 [==============================] - 2s 5ms/step - loss: 0.4996 - accuracy: 0.8016 - val_loss: 0.5448 - val_accuracy: 0.7915

from matplotlib import pyplot as plt
def display_training_curves(training, validation, title, subplot):
  if subplot%10==1: # set up the subplots on the first call
    plt.subplots(figsize=(10,10), facecolor='#F0F0F0')
    plt.tight_layout()
  ax = plt.subplot(subplot)
  ax.set_facecolor('#F8F8F8')
  ax.plot(training)
  ax.plot(validation)
  ax.set_title('model '+ title)
  ax.set_ylabel(title)
  ax.set_xlabel('epoch')
  ax.legend(['train', 'valid.'])

display_training_curves(history.history['accuracy'], history.history['val_accuracy'], 'accuracy', 211)
display_training_curves(history.history['loss'], history.history['val_loss'], 'loss', 212)

png

Evaluasi modelnya

Dan mari kita lihat bagaimana kinerja modelnya. Dua nilai akan dikembalikan. Loss (angka yang mewakili kesalahan kami, nilai yang lebih rendah lebih baik), dan akurasi.

results = model.evaluate(test_data.batch(512), verbose=2)

for name, value in zip(model.metrics_names, results):
  print('%s: %.3f' % (name, value))

4/4 - 0s - loss: 0.5357 - accuracy: 0.7891 - 441ms/epoch - 110ms/step
loss: 0.536
accuracy: 0.789

Kita dapat melihat bahwa kerugian dengan cepat berkurang sementara terutama akurasi meningkat dengan cepat. Mari kita plot beberapa contoh untuk memeriksa bagaimana prediksi berhubungan dengan label yang sebenarnya:

prediction_dataset = next(iter(test_data.batch(20)))

prediction_texts = [ex.numpy().decode('utf8') for ex in prediction_dataset[0]]
prediction_labels = [label2str(x) for x in prediction_dataset[1]]

predictions = [
    label2str(x) for x in np.argmax(model.predict(prediction_texts), axis=-1)]


pd.DataFrame({
    TEXT_FEATURE_NAME: prediction_texts,
    LABEL_NAME: prediction_labels,
    'prediction': predictions
})

Kita dapat melihat bahwa untuk sampel acak ini, model sering kali memprediksi label yang benar, menunjukkan bahwa model tersebut dapat menyematkan kalimat ilmiah dengan cukup baik.

Apa berikutnya?

Sekarang setelah Anda mengetahui lebih banyak tentang embeddings CORD-19 Swivel dari TF-Hub, kami mendorong Anda untuk berpartisipasi dalam kompetisi CORD-19 Kaggle untuk berkontribusi dalam memperoleh wawasan ilmiah dari teks akademik terkait COVID-19.