Sentiment Analysis Model for Reviews
Budget: $10 – $30 USD
I have a growing collection of customer reviews and need a sentiment analysis module that pulls actionable insights from them. The core objective is natural language understanding, so I want more than a simple positive-negative label; the model should capture nuanced sentiment while staying robust to noisy, user-generated text.
Here is what I have in mind:
• Data work. Clean the raw review text with standard NLP preprocessing—tokenisation, stemming, stop-word removal—while retaining enough context for good signal. Term frequency–based features (TF, TF-IDF) are the starting point, and I’m open to word embeddings if they genuinely improve accuracy.
• Modelling. Build and compare several supervised and unsupervised approaches. Decision trees and random forests must be part of the supervised set, alongside whichever additional classifiers you believe will raise performance. On the unsupervised side, clustering or topic discovery that feeds back into the sentiment pipeline is encouraged.
• Evaluation & delivery. I’ll need the full, well-commented Python code (ideally in Jupyter notebooks), a short technical report explaining data preparation, model choice, hyper-parameters and metrics, plus instructions for retraining on new review dumps. I’ll consider the job complete once the model reaches reliable precision/recall on a held-out test set and the notebook runs end-to-end on my machine.
Stack notes: I currently work with Python 3.9, scikit-learn, NLTK and pandas, so please stick to that ecosystem unless there’s a compelling reason to introduce something else.
If this sounds like your kind of project, I’m keen to get started right away.
Here is what I have in mind:
• Data work. Clean the raw review text with standard NLP preprocessing—tokenisation, stemming, stop-word removal—while retaining enough context for good signal. Term frequency–based features (TF, TF-IDF) are the starting point, and I’m open to word embeddings if they genuinely improve accuracy.
• Modelling. Build and compare several supervised and unsupervised approaches. Decision trees and random forests must be part of the supervised set, alongside whichever additional classifiers you believe will raise performance. On the unsupervised side, clustering or topic discovery that feeds back into the sentiment pipeline is encouraged.
• Evaluation & delivery. I’ll need the full, well-commented Python code (ideally in Jupyter notebooks), a short technical report explaining data preparation, model choice, hyper-parameters and metrics, plus instructions for retraining on new review dumps. I’ll consider the job complete once the model reaches reliable precision/recall on a held-out test set and the notebook runs end-to-end on my machine.
Stack notes: I currently work with Python 3.9, scikit-learn, NLTK and pandas, so please stick to that ecosystem unless there’s a compelling reason to introduce something else.
If this sounds like your kind of project, I’m keen to get started right away.