Contrastive Learning

Hi, Based on my previous post, I looked into fine-tuning a bi-encoder via self-supervised learning. As a logical next step, I now want to fine-tune a cross-encoder for my specific task. This is not as easy as fine-tuning a bi-encoder, because cross-encoders are not designed to be trained with contrastive learning. Therefor I looked into four different approaches to fine-tune a cross-encoder for my specific task: BCE loss BCE loss with hard negatives InfoNCE loss Margin MSE loss BCE loss The simplest approach is to use binary cross-entropy loss to train the model. Here we have two labels, 1 for a positive pair and 0 for a negative pair and treat the relevance as a binary classification problem Passage Re-Ranking with BERT by Nogueira and Cho 2020. This can also be called a pointwise approach. For every query-document pair we calculate the logit, then apply a sigmoid activation twhich will represent the probability of relevance. ...

Hi, Based on my previous post, I need to write a correction. I was wrong. I previously suggested using embeddings directly from pre-trained models. That turns out to be a bad idea, because the training objectives are fundamentally different. As an alternative, I have now explored training an embedding model using Contrastive Learning. Please look at my notebook for the full code and details. The main reason you should not use raw embeddings from pre-trained models is the anisotropy problem. The paper How Contextual are Contextualized Word Representations? from Ethayarajh et al. 2019 shows that the embeddings are clustered in a narrow cone of the vector space, making them almost useless for differentiation. So, if you calculate the cosine similarity between two completely different sentences, the result will be around 0.80. ...