Building a Grammar Correction Python App with Gramformer and Gradio

It is challenging to write English sentences without making grammatical mistakes. Many people struggle a lot with writing the correct vocabularies and tenses. <!--more--> When you think of grammar correction web applications, the grammarly app always comes to mind.

Grammarly is a service that runs on your web browser or in desktop applications such as Microsoft Word. When you type something, it helps you correct grammatical mistakes.

In some instances, it suggests new sentences for you. The problem with services like this is that they are not open-source and do not have free APIs. You, therefore, need to pay a subscription fee to fix advanced grammatical mistakes.

In this tutorial, we will build a service similar to Grammarly using an open-source framework called Gramformer.

This build will help us have an idea of how services such as Grammarly work. We will give a sentence with grammatical mistakes to the model and it will produce a corrected version of the sentence.

Table of contents #

• Prerequisites
• Introduction
• Installing Gramformer and importing dependencies
• Using the Gramformer model for grammar correction
• Deploying the model as an application using Gradio
• Wrapping up
• Further reading

Prerequisites #

To follow along with this tutorial, a reader needs to be familiar with:

• The Python programming language.
• Google colab or Jupyter Notebook.

Note that we will be using Google Colab in this tutorial.

Introduction #

Gramformer is an open-source framework for detecting, highlighting, and correcting grammar mistakes on natural language text.

It is a relatively new framework and is currently available on GitHub. You can pass a sentence to this framework and it would improve it for you.

The framework consists of three models; error correction, error detection, and error highlighter model.

At the time of writing this tutorial, only the error detector and highlighter models are available.

The fine-tuning for these models is done on relatively smaller models with not-so-much data due to limited computing budgets.

Installing Gramformer and importing dependencies #

PyTorch is a core dependency that you should install before working with Gramformer.

On your browser, navigate to PyTorch's website and click on the Install button. Depending on the OS you're using, click on PyTorch build -> package -> programming language -> compute platform, select your preferences and then run the install command.

For our case, we've chosen the LTS (1.8.2) PyTorch build, Linux OS, Pip package, Python programming language, CUDA 11.1 compute platform (because we're using Google Colab's GPU). Don't forget to add an exclamation mark ! before pip.

!pip3 install torch==1.8.2+cu111 torchvision==0.9.2+cu111 torchaudio==0.8.2 -f https://download.pytorch.org/whl/lts/1.8/torch_lts.html

Once that is done, go ahead and install Gramformer using the command below:

!pip3 install pip==20.1.1 

!pip3 install -U git+https://github.com/PrithivirajDamodaran/Gramformer.git

Finally, we need to import Gramformer into our project.

from gramformer import Gramformer

Using the Gramformer model for grammar correction #

Instantiating the Gramformer model #

gf = Gramformer(models=1, use_gpu=False) #1=corrector, 2=detector

Currently, Gramformer has two models; the error detector and the error corrector model. Therefore, value 1 represents the corrector, while 2 represents the detector.

Here's the result obtained from running the command above:

[Gramformer] Grammar error correct/highlight model loaded..

It has successfully loaded the error correct/highlight model.

Correcting sentences #

Let's start by accessing the correct() function from Gramformer.

gf.correct()

Using an example, let's try and use Gramformer to correct a sentence.

gf.correct('I taken dinner')

Here's our output:

[('I took dinner.', -30.599159240722656)]

We can see that the model corrected our sentence. You can change the input sentence into another grammatically incorrect sentence of your choice and try it out.

Make sure the input sentence is a string. The output sentence is automatically stored in a list.

If you want to perform corrections on a bunch of sentences, you can run a loop. We create a list called sentences to store our grammatical incorrect sentences as shown:

sentences = [
                'My name a Lilian',
                'I love watching a action movie',
                'I has a laptop',
                'My phone battery a missing'
]

We pass these sentences through a loop and churn out the improved version of the sentences.

for sentence in sentences:
    results = gf.correct(sentence)
    print(results[0])

Here's our output:

('My name is Lilian.', -28.28873062133789)
('I love watching an action movie.', -33.1871337890625)
('I have a laptop.', -25.25513458251953)
('My phone battery is missing!', -36.07805252075195)

Deploying the model as an application using Gradio #

Let's now put it all together using the library called Gradio.

Gradio is a lightweight library that allows you to build user interfaces for your machine learning apps.

!pip install gradio

def correct(sentence):
    res = gf.correct(sentence) 
    return res[0] 

The above function encapsulates all of our Gramformer code into a single application.

app_inputs = gr.inputs.Textbox(lines=3, placeholder="Enter a grammatically incorrect sentence here...")

Note that there are different types of inputs for Gradio. Textbox is just one of them. Refer to Gradio's documentation to learn more.

interface = gr.Interface(fn=correct, 
                        inputs=app_inputs,
                         outputs='text', 
                        title='Hi there, I\'m Gramformer')

The interface class requires three arguments:

• fn: A machine learning function to run. For our case, it's correct.
• inputs: The format for the inputs. For our case, it's app_inputs.
• outputs: The format for our output. For our case, it's text.

Finally, to launch Gradio, we use the launch() method, as shown below:

interface.launch()

An interesting thing to note is that once you launch it, it generates an external link that allows it to run as a web app. For our case:

Running on External URL: https://23225.gradio.app

Please find the complete code for this tutorial here.

Conclusion #

In this tutorial, we have learned how to build an end-to-end grammar correction web application using Gramformer.

However, this simple build is only suitable for experimental purposes, if you need to build one for production, a lot more work needs to be done.

Since the Gramformer model is still new, you can refer to the official documentation regularly for updates.

Further reading #

• Gramformer
• Gradio
• PyTorch
• Complete code

Happy coding!

---

Peer Review Contributions by: Willies Ogola