[실험]CHAN-DST 논문 정리 #47
dkswndms4782
started this conversation in
Experiments
Replies: 0 comments
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
-
CHAN-DST
: dislogue history에서 relevant context를 찾기 위함.
→ 각 턴의 발화로부터 word-level 관련 정보 검색
→ contextual representation으로 encode
→ 모든 context표현을 turn-level관련 정보로 집계한 후 word-level 정보와 합친 output 생성.
Definition
: corresponding value of the slot
slot : concatenation of a domain name and a slot name
Contextual Hierarchical Attention Network
1. Sentence Encoder
utterance encoder
→ [CLS] : 문장의 representation들을 합치기위해 사용 (to aggregate the whole representation of a sentence)
→ [SEP] : 문장의 끝을 나타내기위해 사용.
= {} (system response)
t : dialogue turn
( : contextual word representations)
slot-value encoder
=
2. Slot-Word Attention
3. Context Encoder
4. Slot-Turn Attention
5. Global-Local Fusion Gate
=
=
State Transition Prediction
Adaptive Objective
slot-level difficulty
if ;
→ slot s 가 slot s'보다 더 어려운 것.
→ : slot-level difficulty
sample-level difficulty
→ Suppose there are two samples {} and {}.
→ 만약 former confidence 가 latter보다 더 낮다면, 첫번째 sample이 두번째보다 더 어려운 것.
→ : sample level difficulty
: hyper-parameter
Optimization
- At the fine-tuning phase, we adopt the adaptive objective to fine-tune DST task as following:
Beta Was this translation helpful? Give feedback.
All reactions