NLP: Dari Teks ke Sentiment Classifier

import re from collections import Counter # =========================== # 1. Tokenisasi sederhana # =========================== def tokenize(text): """Pecah teks jadi list kata (lowercase, no punctuation)""" text = text.lower() # lowercase text = re.sub(r'[^a-zA-Z\s]', '', text) # hapus punctuation return text.split() # pecah by spasi review = "This movie was absolutely amazing! Best film of 2026." tokens = tokenize(review) print(tokens) # ['this', 'movie', 'was', 'absolutely', 'amazing', # 'best', 'film', 'of', ''] # =========================== # 2. Bangun Vocabulary # =========================== # Kumpulkan semua kata dari seluruh dataset all_reviews = [ "This movie was great", "Terrible film, waste of time", "Amazing acting and story", "Boring and predictable", ] # Hitung frekuensi setiap kata word_counts = Counter() for review in all_reviews: word_counts.update(tokenize(review)) # Buat mapping: kata → angka vocab = {'<PAD>': 0, '<UNK>': 1} # Token khusus for word, count in word_counts.most_common(5000): vocab[word] = len(vocab) print(f"Vocabulary size: {len(vocab)}") # Vocabulary size: 16 # =========================== # 3. Encode: kata → angka # =========================== def encode(text, vocab, max_len=50): tokens = tokenize(text) ids = [vocab.get(t, vocab['<UNK>']) for t in tokens] # Padding: pastikan panjang sama ids = ids[:max_len] # truncate jika terlalu panjang ids += [0] * (max_len - len(ids)) # pad dengan 0 jika terlalu pendek return ids encoded = encode("This movie was great", vocab) print(encoded[:10]) # [5, 3, 6, 7, 0, 0, 0, 0, 0, 0] # ↑this ↑movie ↑was ↑great ↑...padding...

import torch import torch.nn as nn # Embedding layer: lookup table kata → vektor vocab_size = 5000 # Ukuran vocabulary embed_dim = 100 # Dimensi embedding per kata embedding = nn.Embedding(vocab_size, embed_dim, padding_idx=0) # ↑ padding_idx=0 → vektor untuk <PAD> selalu nol # Input: batch of token IDs input_ids = torch.tensor([[42, 7, 156, 89, 0], # review 1 [23, 45, 0, 0, 0]]) # review 2 (shorter, padded) # Output: setiap ID → vektor 100-dimensi embedded = embedding(input_ids) print(embedded.shape) # torch.Size([2, 5, 100]) # ↑batch=2 ↑seq_len=5 ↑embed_dim=100 # Setiap kata sekarang punya "alamat" di ruang 100-dimensi!

import torch import torch.nn as nn class SentimentLSTM(nn.Module): def __init__(self, vocab_size, embed_dim, hidden_dim, num_classes): super().__init__() # Layer 1: Embedding (kata → vektor) self.embedding = nn.Embedding( vocab_size, embed_dim, padding_idx=0 ) # Layer 2: LSTM (proses urutan kata) self.lstm = nn.LSTM( input_size=embed_dim, # Ukuran embedding hidden_size=hidden_dim, # Ukuran hidden state num_layers=2, # 2 layer LSTM (stacked) batch_first=True, # Input shape: [batch, seq, feature] bidirectional=True, # Baca maju DAN mundur dropout=0.3 # Dropout antar layer LSTM ) # Layer 3: Classifier head self.dropout = nn.Dropout(0.5) self.fc = nn.Linear(hidden_dim * 2, num_classes) # ↑ ×2 karena bidirectional (forward + backward) def forward(self, x): # x shape: [batch, seq_len] (token IDs) embedded = self.embedding(x) # → [batch, seq_len, embed_dim] lstm_out, (hidden, cell) = self.lstm(embedded) # hidden shape: [num_layers*2, batch, hidden_dim] # Ambil hidden state terakhir (forward + backward) hidden_fwd = hidden[-2] # Last forward layer hidden_bwd = hidden[-1] # Last backward layer hidden_cat = torch.cat([hidden_fwd, hidden_bwd], dim=1) # → [batch, hidden_dim * 2] output = self.dropout(hidden_cat) output = self.fc(output) # → [batch, num_classes] return output # Inisialisasi model model = SentimentLSTM( vocab_size=25000, # 25K kata paling sering embed_dim=100, # Embedding 100-dimensi hidden_dim=256, # LSTM hidden 256 num_classes=2 # Positif atau Negatif ) total = sum(p.numel() for p in model.parameters()) print(f"Parameters: {total:,}") # Parameters: 3,456,002 # Training (sama pattern seperti Part 2-4!) loss_fn = nn.CrossEntropyLoss() optimizer = torch.optim.Adam(model.parameters(), lr=0.001) # Hasil setelah 5 epoch pada IMDb dataset: # Epoch 1 | Loss: 0.6234 | Acc: 72.4% # Epoch 3 | Loss: 0.3821 | Acc: 84.1% # Epoch 5 | Loss: 0.2456 | Acc: 87.3% # Prediksi: # "This movie was absolutely terrible" → Negatif (96.2%) # "Best film I've seen this year!" → Positif (93.8%)

# pip install transformers from transformers import pipeline # Load pre-trained sentiment classifier — SATU BARIS! classifier = pipeline("sentiment-analysis") # Prediksi langsung — tanpa training! result = classifier("This movie was absolutely fantastic!") print(result) # [{'label': 'POSITIVE', 'score': 0.9998}] result = classifier("Terrible acting, boring plot.") print(result) # [{'label': 'NEGATIVE', 'score': 0.9994}] # Batch prediction results = classifier([ "I loved every minute of this film", "Waste of time and money", "Decent movie, nothing special" ]) for r in results: print(f"{r['label']:8s} ({r['score']:.2%})") # POSITIVE (99.98%) # NEGATIVE (99.89%) # POSITIVE (74.32%)

from transformers import ( AutoTokenizer, AutoModelForSequenceClassification, Trainer, TrainingArguments ) # Load pre-trained BERT + tokenizer model_name = "bert-base-uncased" tokenizer = AutoTokenizer.from_pretrained(model_name) model = AutoModelForSequenceClassification.from_pretrained( model_name, num_labels=2 # Positif / Negatif ) # Tokenize dataset Anda def tokenize_fn(examples): return tokenizer( examples["text"], padding="max_length", truncation=True, max_length=256 ) # Training arguments args = TrainingArguments( output_dir="./results", num_train_epochs=3, per_device_train_batch_size=16, learning_rate=2e-5, # LR kecil untuk fine-tune! evaluation_strategy="epoch", ) # Train! trainer = Trainer( model=model, args=args, train_dataset=train_dataset, # Dataset Anda eval_dataset=eval_dataset, ) trainer.train() # Hasil: 93-95% akurasi setelah fine-tune 3 epoch! # Jauh lebih baik dari LSTM (87%) dengan effort minimal.