Image Created by Canva AI

Simplifying Sentiment Analysis

Bringing Machine Learning to Customer Reviews: A Lightweight Tool for Business Insights

Nowadays, in the digital age, customer feedback can be used by businesses to find valuable information. However, reading hundreds or thousands of reviews to understand customer satisfaction is overwhelming and use a lot of time. This is where Natural Language Processing (NLP) and machine learning (ML) offer a solution.

To confront this challenge, I designed a web application that automatically classifies customer reviews as positive or negative based on a combination of NLP and ML methods. The objective of this project was to provide a tool for users, with or without technical knowledge, to interpret customer feedback quickly, intuitively, and reliably in seconds.

This project was not just an academic project; it was an effort to create real value. The value of the tool is that an accessible form of sentiment analysis will allow small business owners and consumers to draw inferences from customer feedback and create a sense of awareness regarding the opportunities and changes that exist when AI-driven applications are used in daily decision making.

Why Sentiment Analysis

Machine learning and natural language processing have always fascinated me because they make it much easier for us to understand human language. Today, sentiment analysis stands out as the most popular method for natural language processing. The concept has gained a good reputation across many industries, such as retail, health, education, etc., and shapes how organizations can improve their service.

Nowadays, understanding the opinion of the customer is more important than ever before. Despite the size and scope of an organization, sentiment analysis provides an easy way to transform messy, unstructured data into clean, insightful opinions. As I built the app, I strengthened my technical skills and knowledge and developed a web application anyone can use regardless of their technical background.

About the Dataset

For this project I have sourced two of the most popular datasets from existing data sources (Kaggle 2019 and Kaggle 2020). The two datasets that were utilized were from the Amazon review dataset and the IMDV review dataset and include a total of 74,999. The source of a large, balanced dataset is essential in the development and evaluation of the model.

Data Pre-processing

To feed the dataset into the ML model, I have pre-processed it. It involves removing irrelevant data (punctuation, comma, suffix, prefix, etc.) using stemming and stop word removal. Furthermore, I have used tf-idf vectorizer to extract features.TF-IDF vectorizers highlighted positive sentiment with terms such as “ excellent,” “ extraordinary,” and “ truly outstanding,” while adverse opinion was characterized by terms such as “ garbage,” “not worthy,” and “doesn’t last.” The completion of these data preprocessing steps has resulted in my models being able to focus on relevant patterns in the data, instead of false patterns associated with noise.

Model Implementation

The model implementation can be demonstrated as representing the architecture of the model on training data for training the model. For this step, I have divided the dataset into two parts.

Train-test split method has been used to divide the dataset into two parts 80% (59,999) and 20% (15,000) dataset. 80% dataset was used for model training, and 20% dataset was utilized to evaluate the final model as part of the whole product evaluation.

Once dataset partitioning is done, Linear SVM, Logistic Regression and Multinomial Naïve Bayes algorithms applied into it. Fig 1 shows the workflow of the algorithm after dividing the dataset.

Fig 1:Procedure of splitting the dataset into training and testing (Dagar, Kajal, and Bhatia 2021)

Final Model Selection

My next step was to apply three ML algorithms in terms of accuracy, speed, and generalizability: Logistic Regression, Random Forest, and Support Vector Machine (SVM).

According to the 8-fold cross-validation approach, 3 models — Logistic Regression, Support Vector Machine (SVM), and Multinomial Naive Bayes provided over 80% accuracy for the data, indicating that none of the models were necessarily overfitted or underfitted.

Evaluation results:

· The highest cross-validation accuracy was achieved with Logistic Regression.

· Linear SVM trailed closely with only 0.28% less accuracy than Logistic Regression

· As compared to Logistic Regression, Multinomial Naive Bayes’ accuracy was 7.06% lower.

All three models performed well during the training phase; however, the final web application requires a model that is both reliable and accurate. Although both Random Forest and SVM provided reasonable results for the dataset, Logistic Regression outperformed them. As well as performing was much faster at training than either Random Forest or SVM. Thus, I chose Logistic Regression as the model in the final web application since it was the simplest and most methodologically efficient.

Final Model Evaluation:

An evaluative analysis of the final Logistic Regression model was carried out using a confusion matrix and classification metrics. Results showed the following:

Fig 2: Accuracy score

· Correct prediction: Predicted 7054 negative values and 5704 positive values correctly.

· Incorrect prediction: 997 false positive and 1245 false negative respectively.

Fig 3: Confusion matrix of logistic Regression model

Fig 4: Classification Report of Logistic Regression

Fig 5: Visualization of Accurate sentiment prediction

According to these metrics and graph, the model adapts well to varying data sets and offers a reliable basis for real-world applications.

Web Application Development

The application was created on python and Flask as the back end and HTML and CSS as the front end. A server constructed using Flask became a lightweight solution that was helpful to accept user reviews, process the received review data, and provide predictions based on the trained model.

Key Features:

· Simple, user-friendly interface

· Onboarding tips for First-time user’s

· Troubleshooting tips for users

· The reviews can be pasted into the box or uploaded as Excel files

· Analyse sentiments quickly

· Visualize Excel format data in pie charts based on categories such as positive or negative

· Pivot charts download option

Functionality Walkthrough

The app works by using user reviews and delivers the sentiment output very simply. Below are the screenshots of the user interface, and the examples of reviews that were analyzed and output by the app.

User Interface:

The interface has a very straightforward layout, which has fields for input on one page, and the resultant output on another page of the application. The sentiment analysis is displayed in an easy-to-read format through either labeling on a chart or providing a chart.

Fig 6:Web User Interface of starting page in the web application

Fig 7: Notification for using 25 character while writing review

Fig 8: Notification for mandatory filling the input field

Analysis Demonstration:

Whenever users make reviews, the application will take the text and look at it as a whole, classifying all the reviews as a positive, negative sentiment. The output is very easy to understand and organizes a set of sentiment scores for individual or aggregate reviews across multiple reviews.

Fig 9: Single review analysis

Fig 10: Multiple review analysis

Fig 11: File review visualization

Test and Integration

The importance of the testing and integration had a major contribution in assessing system reliability and performance had. The application validated via unit testing, beginning with data input types and concluding with model outputs.

Fig 12: Web application Testing

Integration testing included functional testing of the Flask web application and data pipeline. Simulating user interactions, validating route responses, and identifying any functional issues were achieved with the Flask test client. Throughout the testing process, Google Collab was used for model behaviour analysis and PyCharm for end-to-end application integration, resulting in a robust and user-friendly sentiment analysis tool

What Is the Real-World Relevance and Impact?

A significant part of the project is the application of AI in solving the problems of the real world. Customers’ voices can be accurately analysed by businesses and organizations to improve their customer service efforts. Without a programming background, researchers and educators can easily and quickly analyse bulk text. By analysing themes and patterns, customer experience teams can detect common problems, monitor sentiment, and improve their responses.

In addition to its technological capabilities, the tool represents the benefits of accessible AI, which are both new and usable, a combination that is truly valuable. This tool bridges the gap between advanced machine learning models and the everyday person by making sentiment analysis accessible to everyone, embodying the growing demand for AI solutions that are accessible, which adhere to effective user-centred design principles, resulting in a positive and sustainable impact in industries that make high-stakes decisions and contribute to the economy on a commercial and social level.

What’s next for the project

There are many ways in which the project can be advanced, like adding neutral classifications and weightings in order to ensure that it is kept up to date. Adding real-time review scraping and prediction from Amazon, Shein, or Yelp could be another improvement, as well as developing a dashboard for deep learners and using open AI for the future.

As of now, this application demonstrates a simple, easy-to-use, and great example of how artificial intelligence can be integrated into day-to-day life to make it more enjoyable.

Conclusion

It has been an incredible experience, not only in tackling a technical problem, but also in connecting AI and providing substantive value. By utilizing NLP, ML, and some good web development practices, I have been able to extract insights from unstructured text reviews as well as provide some substantive and consumable information for decision-making. Whether you are in education, a small business, or inquisitive about this whole AI technology, I hope you have found the sense of how small a tool can make a big difference.

If you have any questions you would like to ask me then you may email me or go to my GitHub Account.

Thanks for reading! I will be writing even more on projects tied to AI and how it intersects with the tangible world. Also in case you wish to keep updated, I am looking forward to your connections and following me on LinkedIn and Medium!