Learning Effective Representations for Retrieval using Self-Distillation with Adaptive Relevance Margins

---

Overview

This is the code repository for the paper "Learning Effective Representations for Retrieval using Self-Distillation with Adaptive Relevance Margins".

Representation-based retrieval models, so-called bi-encoders, estimate the relevance of a document to a query by calculating the similarity of their respective embeddings. Current state-of-the-art bi-encoders are trained using an expensive training regime involving knowledge distillation from a teacher model and extensive batch-sampling techniques. Instead of relying on a teacher model, we contribute a novel parameter-free loss function for self-supervision that exploits the pre-trained text similarity capabilities of the encoder model as a training signal, eliminating the need for batch sampling by performing implicit hard negative mining. We explore the capabilities of our proposed approach through extensive ablation studies, demonstrating that self-distillation can match the effectiveness of teacher-distillation approaches while requiring only a fraction of the data and compute.

Supplementary data (TREC-format run files for all final trained models) is hosted on Zenodo.

Project Organization

├── Dockerfile         <- Dockerfile with all dependencies for reproducible execution
├── LICENSE            <- License file
├── Makefile           <- Makefile with commands to reproduce artifacts (data + models)
├── README.md          <- The top-level README for project
├── configs            <- Configuration files for model and sweep parameters
├── data               <- Data folder; will be populated by data scripts
├── main.py            <- Main Lightning CLI entrypoint
├── requirements.txt   <- Dependencies
├── scripts            <- Scripts to automate single tasks (data parsing, sweep agents, ...)
├── setup.py           <- Makes project pip installable (pip install -e .) so src can be imported
└── src                <- Model source code

Replication

Data, model training, and evaluation is replicable with make targets:

$ make
Available rules:

requirements        Install Python Dependencies 
data-train          Download and preprocess train dataset 
data-eval           Download and preprocess eval datasets
fit                 Run the training process
eval                Run eval process 
clean               Delete all compiled Python files 

These can be run in the given order to fully replicate the experimental pipeline.

Each training run from the paper can be executed with its given config file in configs with the following command:

python3 main.py fit -c <path-to-config-file>

---