Two-Layer Mutually Reinforced Random Walk
for Improved Multi-Party Meeting Summarization
Yun-Nung (Vivian) Chen and Florian Metze
1. Summary
Idea:
o Important utterances are topically similar to each other
o Utterances similar to important speakers should be more important
Approach for extractive summary
o Construct a two-layer graph to represent 1) the utterance nodes in utterance-layer
2) the speaker nodes in speaker-layer
o Mutually propagate importance scores via
within-layer edges and between-layer edges
• Basic Idea: high importance means
Utterances with higher original score
Utterances topically/lexically similar to the indicative utterances
Utterances similar important speakers’
utterances
5. Experiments
• Graph-based approaches can improve speech summarization performance
• Two-layer approaches involving speaker information can get further improvement
• Topical similarity is more robust to recognition errors
better for ASR transcripts
• Lexical similarity is more accurate when absence of errors
better for manual transcripts
• Our proposed approaches achieve more than 7% relative improvement compared to the baseline
6. Conclusions
• Dataset: 10 meetings from CMU Speech Group, #Speaker: 6 (total), 2-4 (each), WER = 44%
• Parameter setting: α = 0.9, summary ratio = 30%
3. Two-Layer Mutually Reinforced Random Walk
• Speaker-Layer
o Node: speakers in a document
combine all utterances from the
same speaker as the speaker node
o Edge weight (red, green): TF-IDF cosine similarity
2. Graph Construction
The utterances topically similar to more
important utterances should be more important
F-measure
Topic Model
(ex. PLSA, LDA)
Multi-Party Meeting
Corpus
ASR Manual
Mutually Reinforced Random Walk
Re-Rank
U5 93 U2 88 U3 75
:
Summary
Latent Topic Entropy
(Baseline)
U5 98 U4 85 U3 73
:
Summary
Baseline Part
Re-Rank Part
Flowchart of proposed approach
U1
U2
U3 U4
U5
U6
U7
Utterance-Layer
S2
Speaker-Layer
S1 S3
• Utterance-Layer
o Node: utterances in a document
o Edge weight (blue):
topical/lexical similarity
The utterances from the similar speaker can partially share the importance
• Similarity Matrix
LUU: utterance-to-utterance relation (topical/lexical similarity) LSS: speaker-to-speaker relation (TF-IDF cosine similarity)
LUS: utterance-to-speaker relation (TF-IDF cosine similarity) LSU: speaker-to-utterance relation (TF-IDF cosine similarity)
• Two-Layer MRRW-BP (Between-Layer Propagation)
utterance score at (t+1)-th iteration
original importance of utterance
scores propagated from speaker-layer
speaker score at (t+1)-th iteration
equal weight scores propagated from utterance-layer
U1
U2
U3 U4
U5
U6
U7 Utterance-Layer
S2 Speaker-Layer
S1 S3
• Two-Layer MRRW-WBP (Within- and Between-Layer Propagation)
scores propagated from speaker-layer then propagated within utterance-layer
scores propagated from utterance-layer then propagated within speaker-layer
U1
U2
U3 U4
U5
U6
U7
Utterance-Layer
S2 Speaker-Layer
S1 S3
Utterance node U can get higher score when
Higher original importance
More speaker nodes similar to utterance U
Utterance node U can get higher score when
Higher original importance
More speaker nodes similar to utterance U
More important utterances similar to utterance U
46 46.5 47 47.5 48 48.5 49 49.5 50 50.5
Baseline: LTE RandomWalk (LexSim)
RandomWalk (TopicSim)
Two-Layer MRRW-BP
Two-Layer MRRW-WBP
(LexSim)
Two-Layer MRRW-WBP
(TopicSim)
ROUGE-1 (ASR)
Baseline: LTE RandomWalk (LexSim)
RandomWalk (TopicSim)
Two-Layer MRRW-BP
Two-Layer MRRW-WBP
(LexSim)
Two-Layer MRRW-WBP
(TopicSim)
ROUGE-L (ASR)
44 44.5 45 45.5 46 46.5 47 47.5 48 48.5 49
Baseline: LTE RandomWalk (LexSim)
RandomWalk (TopicSim)
Two-Layer MRRW-BP
Two-Layer MRRW-WBP
(LexSim)
Two-Layer MRRW-WBP
(TopicSim)
ROUGE-1 (Manual)
Baseline: LTE RandomWalk (LexSim)
RandomWalk (TopicSim)
Two-Layer MRRW-BP
Two-Layer MRRW-WBP
(LexSim)
Two-Layer MRRW-WBP
(TopicSim)
ROUGE-L (Manual)