Text2Text Generations using HuggingFace Model

Text2text generation is a technique in Natural Language Processing (NLP) that allows us to transform input text into a different, task-specific output. It covers any task where an input sequence is transformed into another, context-dependent output. Tasks include:

• Translation: Converting text between languages.
• Summarization: Condensing verbose content.
• Paraphrasing: Restating with different wording.
• Question Answering: Extracting/generating answers from context.
• Sentiment Classification: Assigning sentiment labels.
• Question Generation: Creating questions from statements.

Modern transformer models make this process both flexible and highly effective, especially when we use Hugging Face Transformers library.

Step-by-Step Implementation

Let's see the implementation of text2text generator using HuggingFace Model,

Step 1: Install Required Libraries

Use below command to install Hugging Face Transformers library

!pip install transformers

Step 2: Import and Prepare the Model

• Loads the tokenizer and model for google/flan-ul2.
• Moves the model to GPU if available for much faster processing.

import torch
from transformers import AutoTokenizer, AutoModelForSeq2SeqLM

model_name = 'google/flan-ul2'
tokenizer = AutoTokenizer.from_pretrained(model_name)
model = AutoModelForSeq2SeqLM.from_pretrained(model_name)

device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
model = model.to(device)

Output:

Other models available are as follows:

• t5-base (balanced model size and accuracy)
• google/flan-t5-small
• google/flan-t5-xxl (instruction-tuned, better at following diverse prompts)
• facebook/bart-large (Great for summarization, paraphrasing and some translation tasks.)
• facebook/mbart-large-50-many-to-many-mmt (Handles many languages; requires more setup.)
• facebook/bart-large-cnn (Specialized for summarization)
• Vamsi/T5_Paraphrase_Paws (Paraphrasing)
• pszemraj/grammar-correction (Grammar correction)
• google/flan-ul2 (Very strong all-purpose instruction-following model)

Step 3: Use the generate_text Utility Function

def generate_text(task_prefix, input_text, max_length=64, num_beams=4):
    prompt = f"{task_prefix}: {input_text}"
    inputs = tokenizer(prompt, return_tensors="pt").to(device)
    output_ids = model.generate(
        **inputs,
        max_length=max_length,
        num_beams=num_beams,
        early_stopping=True,
        no_repeat_ngram_size=2
    )
    return tokenizer.decode(output_ids[0], skip_special_tokens=True)

Step 4: Give prompts and Check Result

1. Summarization

summary = generate_text(
    task_prefix="summarize",
    input_text="Natural language processing is a rapidly growing field enabling machines to understand and interact with human language."
)
print(summary)

Output:

Natural language processing enables machines to understand human language.

2. Translation

french_translation = generate_text(
    task_prefix="translate French to English",
    input_text="La bibliothèque HuggingFace Transformers met à disposition de puissants modèles AI, permettant ainsi une expérimentation et un déploiement rapides."
)
print(french_translation)

Output:

translate French to English: The HuggingFace Transformers library offers powerful AI models, allowing rapid experimentation and deployment.

3. Paraphrasing

paraphrase = generate_text(
    task_prefix="paraphrase",
    input_text="This library makes advanced AI accessible to everyone."
)
print(paraphrase)

Output:

This library allows everyone to use advanced AI.

4. Question-Generation

question = generate_text(
    task_prefix="generate question",
    input_text="The mitochondria is the powerhouse of the cell."
)
print(question)

Output:

What is the powerhouse of the cell?

5. Sentiment Analysis

result = generate_text(
    task_prefix="sst2 sentence",
    input_text="I had an amazing experience with the new smartphone"
)
print(result)

Output :

positive

We ca see that our Text2text generation model using hugging face is working fine.