Dwell in the Beginning: How Language Models Embed Long Documents for Dense Retrieval
This addresses a problem for researchers and practitioners in information retrieval by revealing incremental biases in how models process long documents.
The study investigated positional biases in Transformer-based models for text representation learning, finding that models generate embeddings that better capture early contents of input documents, with fine-tuning exacerbating this effect.
This study investigates the existence of positional biases in Transformer-based models for text representation learning, particularly in the context of web document retrieval. We build on previous research that demonstrated loss of information in the middle of input sequences for causal language models, extending it to the domain of representation learning. We examine positional biases at various stages of training for an encoder-decoder model, including language model pre-training, contrastive pre-training, and contrastive fine-tuning. Experiments with the MS-MARCO document collection reveal that after contrastive pre-training the model already generates embeddings that better capture early contents of the input, with fine-tuning further aggravating this effect.