148
Applications in Natural Language Processing
TABLE 5.8
Quantization results of Bi-ColBERT.
Model
MRR@10
BERTbase
16.7
BERTlarge
19.8
ColBERT
32.8
Bi-ColBERT
31.7
In summary, this paper’s contributions can be concluded as: (1) The first work to binarize
ColBERT. (2) A semantic diffusion method to hedge the information loss against embedding
binarization. (3) An approximation of Unit Impulse Function [18] for more accurate gradient
estimation.