IR-Project

The purpose of this project was to create a search engine based on Elasticsearch for retrieving AI-generated images based on their prompts.

In addition, we have used machine learning models to perform sentiment analysis on the prompts, to allow the user to tune their search results based on emotion labels.

We have compared the performace of CNN, BERT and pQRNN based models for this task.

For details and references, see report/IR_report.pdf

Students:

• Mohamed Darkaoui
• Viktor Hura
• Mounir Madmar

Requirements

• python
• pip
• cuda installion of pytorch and tensorflow (optional)

All the required libraries can be found in requirements.txt

Structure

report/IR_report.pdf holds the project report

data holds all the external data that we used, some data is missing from this repo because it's too big, but README.md and .gitignore files should point you where to get it.

results holds data and models generated by our code

src hold all the source code

It's split up in two parts, classification and search + retrieval + interface

Classification

1. Preprocess the data by running the script in src/preprocessing directory, more info in the corresponding README
2. (Optional) generate word vectors by using the scripts in src/generate_wordvectors, more info in the corresponding README
3. Train models, validate models, and label our lexica data in src/classification, more info in the corresponding README