Background

This project is my attempt at leveraging the power of sentence embedding (e.g. BERT). I decided to use sentence embedding to design a search engine. The project is divided into two main components:

1. Using textual information provided in the MyAnimeList such as synopsis, genre, demographic, and etc. and transforming them into vector representation and storing them into vector database (since the size of the data will be small, we will store them as Numpy file).
2. Take in query by the user and convert them into vector representation and match them to top k anime in the vector database using Bi-Encoder (explanation below) then we will narrow down the matches using Cross-Encoder. The ranking of the matches is decided with the combination of score, popularity and favourites using geometric mean.

Architecture

• Bi-Encoder: all-distilroberta-v1
• Cross-Encoder: ms-marco-MiniLM-L-6-v2
• Summarizer: Few texts are shortened in the preprocessing step (c) Falconsai/text_summarization to ensure the input is under the context-length limit of the Bi-Encoder and the Cross-Encoder.
• Vector Database: Numpy

Use Cases

What I’ve explored is not limited to textual data, we can also use them for images, audio, and even the combination of different data types here are a few ways we can use the same concept:

1. Multimedia (text, audio, image) Retrieval: Match query (text) to text, audio, image. Or even match image/audio to text, audio, image. You get the point.
2. Question Answering: Provide accurate answers to user queries based on document content.
3. Recommendation Systems: Suggest items based on user preferences and similar items.
4. Content Personalization: Tailor content recommendations based on user interests and behaviour.
5. Duplicate Detection: Find similar or duplicate documents within a dataset.
6. Anything that involves matching: Literally anything or at least the majority of tasks that involve the matching process can be done using encoding and vector database.