Aspect-Sentiment-Analysis

This repository contains the source code for an Aspect Sentiment Analysis project. The project involves reading data from an XML file, splitting the data into ten subfiles, and subsequently training a model using nine of these files. Both the vectoriser and the model are saved for future use. The remaining XML file is used for testing purposes and performing an 10-fold cross-validation.

Please note that this code is case-specific and no considerations for scalability were made. This means that while the code should work well for the provided dataset, and the one-off processing tasks, it may not perform optimally for other datasets or under high throughput demands.

Code Structure

The code is initiated from main.py.

Requirements

To ensure all functionalities of the project run as expected, the following Python libraries need to be installed:

joblib==1.0.1
pandas==1.2.4
scikit-learn==0.24.2
nltk==3.6.2
gensim==4.3.1
beautifulsoup4==4.9.3

You can install these libraries by running:

pip install -r requirements.txt

Setting up the project

The project requires you to have XML files for training and testing. For this purpose, there is a function load_xmls in split.py that takes as argument the name of the XML file and splits it into 10 subfiles.

Training

The train.py script loads the 9 XML files, vectorises them, and fits them to a model. Both the vectoriser and the model are saved for testing.

Testing

The test.py script uses the remaining XML file, transforms it using the saved vectoriser, and fits it to the saved model.

Cross-Validation

The experiments.py script performs 10-fold cross-validation for various vectorisers and models, and returns the average F1 score along with the class individual score.

Executing the code

First populate you working directory with both dataset and python modules. Then, to run the code with the default configuration, type:

python main.py.

The running configuration has the following form:

python main.py --model_name [model_name] --vectoriser [vectoriser] --max_features [max_features] --min_df [min_df] --max_df [max_df]

So, running a custom configuration would look like:

python main.py --model_name random_forest --vectoriser word2vec --max_features 2000 --min_df 1 --max_df 0.5

main.py parameters:

The parameters for main.py are defined in utils.py.

• model_name: the name of the model to be used for training and testing; can be 'logistic_regression', 'random_forest', 'svm' and 'multinomial_nb'
• vectoriser: the name of the vectoriser to be used for training and testing; can be 'count_vectoriser', 'tfidf' and 'word2vec'
• max_features: the maximum number of features the vectorizer should consider; takes integer values
• min_df: the minimum document frequency for the Count Vectorizer; takes values from 0.0 to 1.0
• max_df: the maximum document frequency for the Count Vectorizer; takes values from 0.0 to 1.0

Note: All default running options, as well as the stemming\tokenisation\text-cleaning, can be found within the utils.py