sujitpal / dl-models-for-qa Goto Github PK

Contributors

Stargazers

Forkers

wuzhongdehua leezqcst rafat-islam1186 pyx123 collawolley mdmustafizurrahman fanshaopu iamsile aarthiis sandy4321 praneetdutta dylan-fan tadaoyamaoka 2dpodcast glebalshanskii hualichenxi chen89 ghego nlpformyself andrejmuhic colinsongf ompanda thejanw stevenlol wolkerzheng wushicanasl liwzhi hitxujian benjamesbabala vforce825 zgsxwsdxg suraj-deshmukh domsia bhaveshoswal kieubinh sophieyoung717 iqbal-chowdhury steccami balaneshin r-wheeler ajoeajoe anjankdash ufukhurriyetoglu chunlinx hiteshn97 perryhau kanaadpathak shawnwuzh shubhampachori12110095 senior88oqz web199195 shubham150796 zzx2017 sabirdvd berryhn maulvi67 sulasen lianyi shdut whaozl tomarraj008 haif-liu fosterleejoe deceneu3 jinlongyu60 xiongnudahan simba2017 sougata09 zjms harishsdev iq-scm

model.add(Merge([qenc, aenc], mode="sum"))

Finally, we replace the cosine similarity with a MLP structure, whose input (282x2-dimension) is
the concatenation of question and answer embeddings, and the output is a single similarity score
and test the modified models by a variety of hidden layer size (100,500,1000). We observe that the
modified models not only get >10% accuracy decrease, but also converge much slower