Sentiment Analysis with BERT using HuggingFace and StreamLit

Sentiment analysis plays a crucial role in understanding the opinions of people and enhancing customer experience for businesses. Being one of the key aspects of Natural Language Processing(NLP), sentiment analysis is a process of analyzing digital text to determine its emotional states such as positive, negative or neutral. Here, we dive deeper into implementing a pre-trained BERT model and deploying it using the StreamLit framework from the HuggingFace(HF) hub.

What is BERT?

BERT stands for Bidirectional Encoder Representations from Transformers and was introduced under the research paper titled "BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding" by Google's AI research team in 2018. It transformed the AI language processing and the way machines understood text. Unlike the traditional models, BERT analyses text bi-directionally to understand the context from the preceding and following words. With the help of a two-step process involving pre-training and fine-tuning BERT can predict missing words in sentences and adapt to various language tasks. Thus making it versatile for different use cases such as sentiment analysis, QnA, Name Entity Recognition(NER) and many more.

Implementation of Sentiment Analysis

As we have gained an overview of the BERT model, let's begin with the implementation of the same.

Working with data and dependencies

The first step is to activate the GPU on the Google Colab notebook and install the necessary dependencies, such as datasets, transformers, and huggingface_hub, to facilitate model training and deployment. Next, we download the IMDB movie review dataset and split it into test and train data, which will then be tokenized using the DistilBERT tokenizer.

Working with the DistilBERT Model

Now that the data is ready, it's time to begin training of BERT model for sentiment analysis. As the BERT model is bulky and quite slow, DistilBERT is used for this project due to the limitations of computational resources in Colab.

DistilBERT

DistilBERT is the distilled version of the BERT model proposed by HuggingFace trained in the technique of knowledge distillation. This allows the model to retain the performance of the original BERT model while having fewer parameters.

Training

We will be using classification tasks for sentiment analysis that enable the transfer of knowledge from DistilBERT to the custom model.

We import the AutoModelForSequenceClassification class from the transformers library. This class is designed for sequence classification tasks and is compatible with models like BERT.
We create an instance of the model by calling AutoModelForSequenceClassification.from_pretrained("distilbert-base-uncased", num_labels=2). Here, "distilbert-base-uncased" refers to the pre-trained DistilBERT model, and num_labels=2 specifies that we are performing binary sentiment classification (positive or negative sentiment).
We import TrainingArguments and Trainer API from the transformers library. The TrainingArguments class allows us to define training-related settings.
We set repo_name to the desired name of the repository where the trained model will be stored.
We create an instance of TrainingArguments by specifying various training-related settings such as learning rate, batch sizes, gradient accumulation, number of epochs, and more. These settings impact the training process and performance.
We set push_to_hub=True to indicate that we want to push the trained model to the Hugging Face model hub.
We initiate the training process using the .train() method of the trainer instance.
The model begins training based on the specified training arguments and the provided datasets.
The training loop involves forward and backward passes, gradient updates, and optimization steps.

Evaluation

We define a custom function compute_metrics that takes the evaluation predictions (eval_pred) as input.
Inside the function, we load the accuracy and F1 score metrics using load_metric("accuracy") and load_metric("f1").
We extract the logits and labels from eval_pred and calculate the predictions by selecting the index with the maximum value along the last axis (using np.argmax(logits, axis=-1)).
We calculate the accuracy and F1 score by passing the predictions and labels to the respective metric calculators and returning them as a dictionary.
After training, we evaluate the model's performance using the .evaluate() method of the trainer instance.
This step calculates and displays metrics like accuracy and F1 score based on the evaluation dataset.

Deploying the trained model

While the model can achieve high accuracy it is time for it to be deployed for use by people. When working with the deployment of models HuggingFace offers multiple ways to do so. In this tutorial, we will learn about deploying model through the following methods available on HF:

Hosted Inference API
Streamlit app using Spaces

Let's dive deep into the deployment phase:

Hosted Inference API

After pushing the trained model to HF Hub it becomes publicly accessible to the community. We import the pipeline class from the transformers library to simply process performing NLP tasks using pre-trained models. Next, we create a new pipeline sentiment_model where we specify the model's name or identifier.

Finally, we use the sentiment_model pipeline to analyze sentiments on two phrases: "I love this movie" and "This movie sucks!" The model processes these inputs and provides sentiment predictions, which can include labels (e.g., positive-1/negative-0) with prediction accuracy.

Streamlit app with Spaces

The new Hugging Face (HF) update enables model deployment to production.
To create an accessible Streamlit app, start by creating a new Space through the Spaces tab on the HuggingFace hub.
During this process, select the suitable Space SDK from available options, including Streamlit, Gradio, Docker, and Static.
Since we're focusing on Streamlit, select it as the SDK.
A default app.py is created, which can be customized to suit our needs for Streamlit app deployment.
To ensure proper functionality, add a new file named requirements.txt in the Files tab to install the necessary Python dependencies torch and transformers.
Customize the default app.py file to create your Streamlit web app.
With these few lines of code, your web app is ready for use.
To access the app, navigate to the App tab of the Space where it's hosted.

These steps will guide you through the process of deploying your Streamlit app and making it accessible via the Hugging Face model hub.

Embedding Streamlit app

Additionally, if one thinks of adding the web app to their website or blog. One can do it by clicking the "Embed this Space" option which allows either to directly embed or Embed with iFrames.

With this, we end the tutorial on how to implement sentiment analysis by fine-tuning BERT using HuggingFace and deploying through Streamlit. The links to the code for implementation of the same have been provided below:

Sentiment Analysis Made Easy

Table of contents