Seq2SeqModel_LSTM

Text Summarization Model with Seq2Seq and LSTM

This model is a sequence-to-sequence (seq2seq) model for text summarization. It uses a bidirectional LSTM encoder and an LSTM decoder to generate summaries from input articles. The model was trained on a dataset with sequences of length up to 800 tokens.

Dataset

CNN-DailyMail News Text Summarization from kaggle

Model Architecture

Encoder

• Input Layer: Takes input sequences of length max_len_article.
• Embedding Layer: Converts input sequences into dense vectors of size 100.
• Bidirectional LSTM Layer: Processes the embedded input, capturing dependencies in both forward and backward directions. Outputs hidden and cell states from both directions.
• State Concatenation: Combines forward and backward hidden and cell states to form the final encoder states.

Decoder

• Input Layer: Takes target sequences of variable length.
• Embedding Layer: Converts target sequences into dense vectors of size 100.
• LSTM Layer: Processes the embedded target sequences using an LSTM with the initial states set to the encoder states.
• Dense Layer: Applies a Dense layer with softmax activation to generate the probabilities for each word in the vocabulary.

Model Summary

Total params: 94,724,060

Trainable params: 94,724,058

Non-trainable params: 0

Training

The model was trained on a dataset with sequences of length up to 800 tokens using the following configuration:

• Optimizer: Adam
• Loss Function: Categorical Crossentropy
• Metrics: Accuracy

Training Loss and Validation Loss

Test Loss

Usage -- I will update this soon

To use this model, you can load it using the Hugging Face Transformers library:

from transformers import TFAutoModel

model = TFAutoModel.from_pretrained('your-model-name')

from transformers import AutoTokenizer, TFAutoModelForSeq2SeqLM

tokenizer = AutoTokenizer.from_pretrained('your-model-name')
model = TFAutoModelForSeq2SeqLM.from_pretrained('your-model-name')

article = "Your input text here."
inputs = tokenizer.encode("summarize: " + article, return_tensors="tf", max_length=800, truncation=True)
summary_ids = model.generate(inputs, max_length=150, min_length=40, length_penalty=2.0, num_beams=4, early_stopping=True)
summary = tokenizer.decode(summary_ids[0], skip_special_tokens=True)

print(summary)