Embed Size (px)
Citation preview
DYNAMIC TIME-AWARE ATTENTION TO SPEAKER ROLES AND CONTEXTS FOR
SPOKEN LANGUAGE UNDERSTANDINGPo-Chun Chen*,Ta-Chung Chi*, Shang-Yu Su*, Yun-Nung (Vivian) Chen
*Three authors contribute this work equally.
https://github.com/MiuLab/Time-SLU
Summary
o Approach: an end-to-end attentional role-based contextual model that leverages
various content-aware and time-aware attention mechanisms
o Experiment: impressive improvement on a benchmark multi-domain dialogue dataset
o Result: temporal information is important when modeling history utterances for
language understanding
The Proposed Approach: Time-Aware Attention to Roles & Contexts
Code Available:
https://github.com/MiuLab/Time-SLU
➢Task Definition
o Language understanding for human-human
dialogues between tourists and guides (DSTC4)
➢Motivation
o Human-human dialogues contain multiple
reasoning steps
o Additional attention mechanism like role or
temporal information may be useful
➢Method: Time-Aware Attention to Speaker Roles and Contexts
o Modeling by the fixed time-aware attention and other learnable attention mechanisms
Experiments and Discussions
Conclusions
➢Setup
o Dataset: DSTC4 35 human-human dialogues
o Evaluation metrics: F1 for multi-label classification
➢Experimental Results
o Time-aware attention models significantly outperform the baselines
o The attention mechanisms provide improvement for LU
o Role-level attention requires content-related information to achieve
better performance
➢Discussion
o Guide results are consistently better than tourist results
o The reason may be that the guide has similar behavior patterns (e.g.
providing information and confirming questions) while the user has
more diverse intentions
➢Result
o The model achieves state-of-the-art performance
Dense LayerΣ
Current Utterance
wt wt+1 wT… …
Dense Layer
LU
Role History Summary
u2
u6
User (Tourist)
u4 Agent (Guide)
u1
u7 Current
Attentional Contextual Model
Sentence AttentionContent Time
u3
u5
Role Attention
➢Role-Level Attention
o Different speaker roles behave differently
o Role-level attention is based on how
much to address on different speaker
roles’ contexts
➢Time-Aware Attention
o Recent utterance contains more
relevant information
➢Content-Aware Attention
o The semantic relation decides
where and how much the model
should focus on given the contexts
➢End-to-End Training for Contextual Language Understanding
o BLSTM-encoded current utterance concatenated with the history vector for multi-label intent prediction
o All encoders, prediction models, and attention weights (except time-aware attention) can be
automatically learned in an end-to-end manner
miu
Attention Mechanism
Attention Type
• Content-Aware 𝛼𝐶
• Time-Aware 𝛼𝑇
Attention Level
• Sentence 𝛼𝑢𝑖• Role 𝛼𝑟𝑖
distance between u and
the current utterance
Leveraging different types and
levels of attention can improve
language understanding
Time-aware attention brings significant improvement for
language understanding
you were saying that you wanted to come to singapore? FOL_CONFIRM
Uh yes
I'm actually planning to visit uh on august
FOL_CONFIRM
Guide (Agent)
Tourist (User)
Task : Language Understanding
(User Intents) RES_WHEN
uh maybe can I have a little bit more details like uh when will you be coming
QST_WHENand like who will you be coming with QST_WHO
Content-Aware Time-Aware Both
Sentence Role Both
60
62
64
66
68
70
72
74
76
Baseline w/oContext
w/Context
