mvr

Configuration, model, and tokenizer for MVR (Multi-View Representation) type models.

MVR (Multi-View Representation) is an information retrieval approach that strikes a balance between the simplicity of single-vector models and the complexity of token-level architectures by representing a document using a small, fixed number of dense vectors. Instead of compressing an entire passage into one embedding or storing a separate vector for every single word. MVR generates multiple distinct representations of the text, with each vector capturing a different semantic facet or topic. During a search, the query is compared against these multiple document views to find the strongest match.

Originally proposed in Multi-View Document Representation Learning for Open-Domain Dense Retrieval.

Classes

MvrConfig([query_length, doc_length, ...])	Configuration class for a MVR model.
MvrModel(config, *args, **kwargs)	MVR model for multi-view representation learning.
MvrTokenizer(*args[, query_length, ...])