DSSM

  1. introduction
    1. two line for extending latent semantic models
      1. clickthrought data
      2. deep atuo-encoders
    2. text hashing. This hashing is not for security, while word embedding or sentence embedding.
    3. translation model for IR, directely translate doc into query. suboptimal
  2. Question
    1. target is for re-construction of document term vectors, rather than differentiating the relevant docs from irrelevant ones for a given query.
    2. small vocabulary size, only 2000 frequent words
  3. loss
    1. loss funcion

Web Search - 图1

Web Search - 图2 is a smoothing factor in the softmax fomulatation, it is for the held-out data set.

MIX

Question:

  1. after uni-conv, bi-conv and tir-conv, the sentence is transformed into a one-demension embedding?
  2. are there 9*3 feature maps in weight channels layer?

Screenshot_20210219_105257.png