Summarizing Contrastive Viewpoints in Opinionated Text

Michael J. Paul, ChengXiang Zhai, Roxana GirjuEMNLP’10Speaker: Hsin-Lan, WangDate: 2010/12/07

2

Outline Introduction Modeling Viewpoints

Topic-Aspect Model Features

Multi-Viewpoint Summarization Comparative LexRank Summary Generation

Experiment and Evaluation Conclusion

3

Introduction The amount of opinionated text available online has been growing rapidly. In this paper, we study how to summarize opinionated text in a such a way that highlights contrast between multiple viewpionts.

4

Introduction Generate two types of multi-view

summaries: macro multi-view summary

Contains multiple sets of sentences, each representing a different viewpoint.

micro multi-view summary Contains a set of pairs of contrastive

sentences.

5

Modeling Viewpoints Challenge: to model and extract

viewpoints which are hidden in text.

Solve: Topic-Aspect Model (TAM)

6

Modeling Viewpoints TAM

7

Modeling Viewpoints Features

Words baseline approach do not do any stop word removal stemming

Dependency Relations use Stanford parser full-tuple: rel(a,b) split-tuple: rel(a,*), rel(*,b)

8

Modeling Viewpoints Features

Negation Rel(wi, wj), if either wi or wj is negated, then we simply rewrite it as .

Polarity use Subjectivity Clues lexicon amod(idea, good)→amod(idea,+) and amod(*,good) →rel(a, － ).

9

Modeling Viewpoints Features

Generalized Relations use Stanford dependencies Rewrite rel(a,b) as Rrel(a,b).

10

Multi-Viewpoint Summarization Comparative LexRank

Make it favor jumping to a good representative excerpt x of any viewpoint v. Make it favor jumping between two excerpts that can potentially form a good contrastive pair.

11

Multi-Viewpoint Summarization Comparative LexRank

12

Multi-Viewpoint Summarization Summary Generation

Macro contrastive summarization Using the random walk stationary distribution across all of the data to rank the excerpts. Separate the top ranked excerpts into two disjoint sets. Remove redundancy and produce the summary.

Micro contrastive summarization Consist of a pair (xi,xj) with the pairwise relevance score. Rank these pairs and remove redundancy.

13

Experiments and Evaluation Experimental Setup

First dataset: 948 verbatim responses to a Gallup phone survey about the 2010 U.S. healthcare bill. Second dataset: use the Bitterlemons corpus, a collection of 594 editorials about the Israel-Palestine conflict.

14

Experiments and Evaluation Stage One: Modeling Viewpoints

15

Experiments and Evaluation Stage Two: Summarizing

Viewpoints Gold Standard Summaries

Gallup healthcare poll

16

Experiments and Evaluation Stage Two: Summarizing Viewpoints

Baseline Approaches Graph-based algorithms

When λ=1, the random walk model only transitions to sentences within the same viewpoint. The modified algorithm produces the same ranking as the unmodified LexRank.

Model-based algorithms Compare against the approach of Lerman and McDonald.

17

Experiments and Evaluation Stage Two: Summarizing Viewpoints

Metrics using the standard ROUGE evaluation metric For evaluating the macro-level summaries:

For evaluating the micro-level summaries:

18

Experiments and Evaluation Stage Two: Summarizing

Viewpoints Evaluation Results

19

Experiments and Evaluation Unsupervised Summarization

Bitterlemons corpus (without a gold set) Asked 8 people to guess if each viewpoint’s summary was written by Israeli or Palestinian authors.

20

Experiments and Evaluation Unsupervised Summarization

Macro-level summaries: Correctly labeled 78% of the summary

sets. Micro-level summaries:

Many of the sentences are mislabeled, and the ones that are correctly labeled are not representative of the collection.

21

Conclusion Present steps toward a two-stage

system that can automatically extract and summarize viewpoints in opinionated text.

First: the accuracy of clustering documents by viewpoint can be enhanced by using rich dependency features.

22

Conclusion Second: use Comparative LexRank to generate contrastive summaries both at the macro and micro level.