[CNN-DSSM/CLSM]A Latent Semantic Modelwith Convolutional-Pooling Structure for Information Retrieval

CLSM uses a fixed size sliding window to capture local context information, and a max pooling layer to capture global context information.

Architecture

Letter-trigram based Word-n-gramRepresentation

word hashing和DSSM的有些不同

letter-trigram layer:30k*3=90k, word hashing之后每个单词的纬度为30k,然后3-gram拼接为90k

Loss Function

paper