SLIDE 1
ADVISOR: JIA-LING, KOH SOURCE: AAAI-19 SPEAKER: WEI, LAI DATE: 2020/5/25
OUTLINE
03
INTRODUCTION EXPERIMENTS METHOD CONCLUSION
02 04 01
2
Towards Personalized Review Summarization via User-Aware Sequence - - PowerPoint PPT Presentation
Towards Personalized Review Summarization via User-Aware Sequence - - PowerPoint PPT Presentation
Towards Personalized Review Summarization via User-Aware Sequence Network ADVISOR: JIA-LING, KOH SOURCE: AAAI-19 SPEAKER: WEI, LAI DATE: 2020/5/25 OUTLINE 01 03 INTRODUCTION EXPERIMENTS 02 04 METHOD CONCLUSION 2 4 User-aware
SLIDE 2
SLIDE 3
4
SLIDE 4
INTRODUCTION
Review summarization aims to generate a condensed summary for a review or multiple reviews. Different users may:
- Care about different contents
- Have their own writing styles
User-aware Sequence Network (USN)
3
SLIDE 5
METHOD
User-aware Sequence Network (USN)
5
SLIDE 6
Problem Formulation
User
- Review
- Summary
π£ π¦ π§ {π¦!, π¦",β¦ , π¦%} {π§!, π§",β¦ , π§&}
οΌm
) (
πΎπ πΎπ
Source vocabulary (#:30,000) Target vocabulary (#:30, 000)
6
SLIDE 7
Sequence-to-sequence attentional model
Get To The Point: Summarization with Pointer-Generator Networks
7
SLIDE 8
8
SLIDE 9
review embedding (dim=128) single-layer encoder hidden state (dim=256)
User-aware Encoder
9
SLIDE 10
User-aware Encoder
πππ’π# = π(π
$ β#; π£ + π$)
ββ²# = β# β πππ’π# User Selection Strategy
(dim=128) User embedding (dim=256) Encoder hidden state
10
SLIDE 11
Context vector
11
SLIDE 12
Context Vector
Decoder hidden state π%,# = π'tanh(π
(ββ²# + π)(π‘% + π()
π%,# = exp(π%,#) β# exp(π%,#) πβ²% = =
#
π½%,#ββ²#
(dim=256) 12
SLIDE 13
Vocabulary state
13
SLIDE 14
Vocabulary State
Decoder hidden state π»π User-specific vocabulary memory (K most user-specific words)
π%,$ = π'tanh(π
*π$ + π)*π‘% + π*)
πΎ%,$ = exp(π%,$) β$ exp(π%,$) π% = =
$
πΎ%,$π$
14
SLIDE 15
Vocabulary distribution
15
SLIDE 16
Vocabulary Distribution
π»π π»π
π β²% = π
+) π)%; π‘%; π£; π% + π+)
πβ²,-. = π‘πππ’πππ¦(π
- π )% + π-)
User Prediction Strategy User Memory Prediction Strategy
Readout state
16
SLIDE 17
Final distribution
17
SLIDE 18
Pointer-generator Model
Get To The Point: Summarization with Pointer-Generator Networks
18
SLIDE 19
Vocabulary Distribution
π
*/0 = π(π */[π)%; π‘%; π%] + π*/)
User Memory Generation Strategy π π₯ = 1 β π*/0 π),-. π₯ + π*/0 =
$:2'34
πΎ%,$
19
SLIDE 20
Loss
π = β 1 π =
%35 6
ππππ
,-.(π§%)
20
SLIDE 21
EXPERIMENTS
Data
Training: 536255-(5000+5000) = 526255 Validation: 5000 Test: 5000 21
SLIDE 22
Comparison Methods
- Lead-1 : extractive the first sentence
- LexRank
: extractive approach based on PageRank algorithm
- S2S+Att : abstractive model
- SEASS : selective network + S2S + Att in sentence summarization
- PGN : copy mechanism + S2S + Att in document summarization
22
SLIDE 23
EXPERIMENTS
Results
23
SLIDE 24
EXPERIMENTS
Aspect-level Coverage ho hote tel
24
SLIDE 25
EXPERIMENTS
Aspect-level Coverage
25
SLIDE 26
EXPERIMENTS
Different User-based Strategies
- 1. User selection
- 2. User prediction
- 3. User memory prediction
- 4. User memory generation
26
SLIDE 27
EXPERIMENTS
User-specific vocabulary size
27
SLIDE 28
EXPERIMENTS
User-based Selective Gate Visualization
- Output summary: excellent service, comfy be
- True Summary: excellent service, very comfortable bed
28
SLIDE 29
EXPERIMENTS
Case study
29
SLIDE 30
Conclusion
- Propose a User-aware Sequence Network (USN) to consider user
information into personalized review summarization.
- Propose 4 user-based strategies.
- Construct a new dataset (Trip).
- Future work: extend model to the multi-review scenario.
30