Natural language processing (NLP) is a subfield of computer science and artificial intelligence that uses machine learning to enable computers to understand and communicate with human language1.
In this repository, we'll implement the following NLP tasks:
- Sentiment Analysis
- Entity Extraction
- Text Summarization
We'll use the following publicly available dataset, from Amazon food reviews.
The data dictionary is as follows:
| Column Name | Description | Data Type |
|---|---|---|
| Id | Row ID | int64 |
| ProductId | Unique identifier for Product | object |
| UserId | Unique identifier for User | object |
| ProfileName | Profile name of the user | object |
| HelpfulnessNumerator | Number of users who found the review helpful | int64 |
| HelpfulnessDenominator | Number of users who indicated wether they found the review helpful or not | int64 |
| Score | Rating between 1 and 5 | int64 |
| Time | Timestamp for the review | int64 |
| Summary | Brief summary of the review | object |
| Text | Full review | object |
Follow the notebook located on the jupyter_notebooks directory. The main finding is with regards to the class balances of the review Score:
As seen in the graph above, the score of 5 is by far the most popular, compared to the other scores.
1. Balancing Data
As noted in the EDA, there is a class imbalance in the Score of the reviews, so we'll address it by:
- Mapping the score from 1-5 to 0-2 (bad, neutral, and good respectively)
- Remove duplicate reviews
- Downsampling the category with the highest review
2. Text Cleaning
In this step we'll remove text that doesn't convey any meaningful information such as
- HTML tags
- URLs
- Excessive whitespace
Note that at this point we're not removing any punctuation, numbers, or special symbols. I want to leave the text human-readable prior to the tokenization step.
3. Tokenization
We'll use the spaCy library to perform:
- tokenization
- stop word and punctuation removal
- lemmatization
In this section I want to try different approaches to perform a sentiment analysis (predict if the text conveys positive, neutral, or negative sentiment) on the reviews. We'll implement and compare the following models.
- Bag of words model with Count Vectorizer
- TFID
- LSTM
- Other pre-trained models
- Finalize selecting all the model evaluation metrics
- Modelling with TFID
- Modelling with Pre-trained models