Rename SimilarityMatrix to StackedSimilarityMatrix, which is intended for use with multiple channels (eg BERT layers)
Add new SimilarityMatrix class that accepts negative indices for OOV terms. Cosine embedding similarity is used for positive indices, while negative indices are exact matched.
Rename existing EmbedText extractor to SlowEmbedText
Add a new EmbedText extractor that produces output suitable for new SimilarityMatrix class. This is much more efficient, because we do not need to generate embeddings for OOV terms. Switch rerankers that use static embeddings (DRMM*, KNRM, PACRR) to use new extractor by default.
SimilarityMatrixtoStackedSimilarityMatrix, which is intended for use with multiple channels (eg BERT layers)SimilarityMatrixclass that accepts negative indices for OOV terms. Cosine embedding similarity is used for positive indices, while negative indices are exact matched.EmbedTextextractor toSlowEmbedTextEmbedTextextractor that produces output suitable for newSimilarityMatrixclass. This is much more efficient, because we do not need to generate embeddings for OOV terms. Switch rerankers that use static embeddings (DRMM*, KNRM, PACRR) to use new extractor by default.Closes #63 .