狠狠撸

Entity Summarization with User Feedback
Qingxia Liu1, Yue Chen1, Gong Cheng1, Evgeny Kharlamov2,3, Junyou Li1, and Yuzhong Qu1
1 National Key Laboratory for Novel Software Technology, Nanjing University, China
2 Department of Informatics, University of Oslo, Norway
3 Bosch Center for Artificial Intelligence, Robert Bosch GmbH, Germany

■ RDF Dataset: T
● triple t∈T: <subj, pred, obj>
■ Entity Description: Desc(e)
● Desc(e) = {t∈T: subj(t)=e or obj(t)=e}
● t∈Desc(e): <e, prop, value>
■ Entity Summarization: S(e, k)
● S?Desc(e) , |S|≤k
2020.06 2
Entity Summarization
Solson
Publications
Rich Buckler
Organization
Gary Brosky
Brooklyn
Company
comics
United States-themed
superheroes “Solson Publications”
“1986”
“1987”
type
type
headquarters
industry
label
subject defunct
founded
founder
keyPerson

■ Goal
● to satisfy users’ information needs
■ Limitations
● have large diff from user-created (F1<0.6)
=> do not meet users’ expectations
● summaries are static
=> cannot adjust to users
2020.06 3
Entity Summarization Methods
A lack of mechanisms for improving an entity summary
when its quality could not satisfy users’ information needs.

■ Our proposal: cooperation between summarizer and user
● involve users in the summarization process
● ask users for feedback
● utilize user’s feedback when computing summaries
2020.06 4
Adding User in the Loop
Summarizer
Summarizer
User

■ Research Challenges
[C1] Cooperative process of 2 agents
– How to represent the process?
[C2] Combining results of actions of 2 agents
– How to represent interdependence between summary and feedback?
2020.06 5
Adding User in the Loop
Summarizer
Summarizer
User

■ Cooperation Process
● Summarizer: presents (current) summary Si
● User: crosses off an irrelevant triple as negative feedback fi
● Summarizer: replaces it with a new triple ri ,
presents an improved summary Si+1
● User: … (repeated)
2020.06 6
Cross-Replace Scenario
f1
r1
replaced by
f0 r0
replaced by
S0 S1 S2
Solson Publications

■ [C1] Representation of the Cross-Replace Scenario
● Markov Decision Process (MDP) based modeling
● Summarizer = reinforcement learning agent
2020.06 7
Our Approach: DRESSED
Solson Publications
- Label: “Solson Publications”
- Key person: Rich Bucker
- Subject: United State-themed superheroes
- Extinction year: “1987”
- Founded by: Gary Brodsky
- Industry: Comics
- Name: “Solson Publications”
- Founding year: “1986”
…
Feedback
Replace
Current Summary
Candidate Facts
Setting:
? agent: summarizer
? environment: Desc(e), user

■ [C2] Representation of Triple Interdependence
● Policy Network
– Encoding triples
– Encoding sets of triples
– Encoding triple interdependence
and scoring candidates
– Selecting replacement triple
2020.06 8

■ Policy Learning
● Learning task: to find a policy that maximize the expected reward
● REINFORCE: a standard policy gradient method in reinforcement learning
– to maximize the expected reward
2020.06 9

■ Entity Summarization
(not utilize user feedback)
● FACES-E: rank triples and select top-k as summary
■ Interactive Document Summarization
(Re-compute doc summaries utilizing user feedback)
● IPS
– positive feedback
■ Interactive Document Retrieval
(Re-rank retrieved documents based on user feedback)
● NRF
– negative relevance feedback
– query-based
● PDGD -L, -N
– positive feedback
– SOTA online learning to rank
– utilize user feedback for model selection
(not as input of the model)
2020.06 10
Baselines Adapted for Evaluation
query
model
generate
update
model
generate
update

■ Setting
● 24 participants
● train: Entity Summarization Benchmark (ESBM)
● test: re-sampled entities from
DBpedia and LinkedMDB
■ Results
● DRESSED: in general best-performing
● with FACES-E and DRESSED
– users stop quickly: I < 4
– obtain reasonably good results: Q_stop > 4
– replacement triples selected by DRESSED
significantly better than the ones of FACES-E
2020.06 11
Experiment 1: User Study
Metrics:
? I = number of iterations
? Qrplc = user rating of suggested
replacements (1-5)
? Qstop = user rating of the final
summary after all iterations (1-5)

Experiment 2: Offline Evaluation
■ Benchmarks
● ESBM (ESBM-D, ESBM-L), FED
■ Metrics
● NDCF: for summary sequence
– Normalized Discounted Cumulative F1
● NDCG: for replacement sequence
– Normalized Discounted Cumulative Gain
■ Results
? DRESSED outperforms
? all baselines
? on all datasets
? DRESSED significantly outperforms
? in most cases
2020.06 12

■ Iterations
● DRESSED is consistently
above all the baselines
● DRESSED better exploits
early feedback to quickly
improve computed
summaries
2020.06 13
Experiment 2: Offline Evaluation

■ Our Approach: DRESSED
● showed better performance than baselines
● has the potential to replace static entity cards
■ Future Work
● cross-replace scenario
– can be extended to support other scenarios:
? multi-feedback, without replacement, positive feedback
● extend the scope of entity summary
– deal with paths, more complex structures, RDF sentence
2020.06 14
Conclusion and Future Work

■ Contributions
● The first research effort to improve entity summarization with user feedback.
● A representation of entity summarization with iterative userfeedback.
– cross-replace scenario, MDP
● A representation of set of triples and their interdependence as a novel DNN.
– DRESSED, solve by RL
● The first empirical study of entity summarization with user feedback.
– based on real users and simulated users
■ DRESSED
● Deep Reinforced Entity Summarization with uSer fEedback
● GitHub Repository: nju-websoft/DRESSED
2020.06 15
Take-home Message

Thank you !
Questions ?
2020.06 16

狠狠撸

Entity Summarization with User Feedback (ESWC 2020)

More Related Content

Entity Summarization with User Feedback (ESWC 2020)