issues
search
chullhwan-song
/
Reading-Paper
151
stars
26
forks
source link
Convolutional Neural Networks for Soft-Matching N-Grams in Ad-hoc Search
#257
Open
chullhwan-song
opened
4 years ago
chullhwan-song
commented
4 years ago
http://www.cs.cmu.edu/~./callan/Papers/wsdm18-zhuyun-dai.pdf
chullhwan-song
commented
4 years ago
Abstract
KNRM :
https://github.com/chullhwan-song/Reading-Paper/issues/258
이 논문을 먼저 반드시 읽어야함. > 이 논문에서 Conv Embedding + N-gram 단위를 adding한 논문
Conv-KNRM
C
onvolutional
K
ernel-based
N
eural
R
anking
M
odel
KNRM 논문과 마찬가지로, soft matches by kernel pooling & learning-to-rank (Rankent)
다만, n-gram 단위 > n-gram soft matches
CONVOLUTIONAL N-GRAM RANKING
Fig.1에서 보듯이, 이전의 KNRM와 달리 KNRM에서는, w2v형태의 input에서 "Translation Matrix"으로 바로 진행했지만, 중간단계에서 Convolution network를 적용한 케이스
input : word단위의 w2v을 적용한 형태
convolutional layer
이는 convolution filters을 이용하여 n-grams단위로 계산 가능(T_q or T_d)하다.
sentence를 sliding window(h)개념으로 접근
i~h : sliding window
w:conv filter weights
F개의 filter를 적용하여 F scores를 낸다.
여기에, bias, activation function(relu)를 추가하면,
는 i번째 h-gram의 embedding이란 의미.
의 f번째 요소는 f번째 필터의 스코어란 의미.
we append up to h − 1 ‘
’ symbols for padding
n-gram단위로 보면,
그래서, query & document :
라고 표시가능
Convolution 연산은 text기반으로 볼때 n-gram형태로 계산~!!
convolution filter의 수와 n-gram수와 동일하다는 의미
cross-match layer
KNRM은 query word들과 document word들간의 similarity를 나타내는 "translation layer"의 translation matrix M 를 구하였다.
여기서도 마찬가지인데 단위가 word에서 n-gram 단위로 확장한다.
그래서, query & document n-gram pairs에 대한 similarity scores는,
통합한 embedding representation은 다른길이의 n-gram에 대한 cross-matching를 허락.
예를 들어, the query trigram “convolutional neural networks” and the document bigram “deep learning”.
최종적으로, translation matrices를 생성
Ranking with N-gram Translations
Kernel-pooling
soft matches위해 K Gaussian kernels 이용. > K different strength levels > soft-TF counts
자세한 내용은
https://github.com/chullhwan-song/Reading-Paper/issues/258
참조
soft-TF feature vector
counting-based pooling(== soft-TF count) methods have shown beer performance than score-based ones like mean-pooling or max-pooling
수식(5)는 한개의 word(or 하나의 gram)와 n-gram pairs를 가진 doc의 case
이를 확장하여, query h_q-grams & document h_d-grams > 잘 생각해보면 query의 n-gram(query n-gram)으로 확장한 케이스
그래서, 다음과 같이..
learning-to-rank (LeToR)
위에서 구한 soft-TF ranking features (
)에서 ranking score를 내기 위해서,
최종적으로 loss = pairwise learning-to-rank
실험
KNRM보다 높아졌음을 알수 있다.
구체적인 내용은 논문참조(~ing)
http://www.cs.cmu.edu/~./callan/Papers/wsdm18-zhuyun-dai.pdf