Opinion Integration Through Opinion Integration Through Semi ‐ ‐ supervised Topic Modeling supervised Topic Modeling Semi Yue Lu and Chengxiang Zhai University of Illinois at Urbana ‐ Champaign 1
Why Opinion Integration? Why Opinion Integration? • Web 2.0 � huge amount of opinions • What have been said about Hillary Clinton? 190,451 posts How to digest all? 4,773,658 results 2
Two Kinds of Opinions Two Kinds of Opinions 4,773,658 results 190,451 posts How to benefit from both? Ordinary opinions Expert opinions •Forum discussions •CNET editor’s review •Blog articles •Wikipedia article •fragmental •Well ‐ structured •Hard to access •Easy to access •Represent the majority •Maybe biased •Up to date •Outdated soon 3
Research Questions Research Questions • How do we formalize the problem of opinion integration? • How do we solve the problem in a general way? • How do we evaluate it? 4
Problem Definition Problem Definition Integrated Summary Output Input Similar Supplementary Topic: opinions opinions Review Aspects iPod cute… tiny… ..thicker.. Design Design Battery last many die out Design Battery Design Expert Battery Battery hrs soon review with Price.. Price.. Price Price aspects could afford still it expensive Extra Aspects Weblogs … easy to use… iTunes …better to extend.. warranty 5
Challenges Challenges 2. How to distinguish 2. How to distinguish similar opinions with similar opinions with Integrated Summary supplementary ones? supplementary ones? 1. How 1. How to align to align Similar Supplementary opinions opinions opinions opinions to expert Review Aspects to expert cute… tiny… ..thicker.. Design aspects Design aspects Battery last many die out or extra Battery or extra aspects? hrs soon aspects? Price Price could afford still it expensive Extra Aspects … easy to use… iTunes 3. How to extract 3. How to extract representative representative …better to extend.. warranty opinions with support? opinions with support? 6
Two Major Steps Two Major Steps • Step 1: opinion sentences retrieval General Weblogs Weblogs on iPod Query = “iPod” • Step 2: opinion integration using probabilistic topic models (3 subtasks) 7
Subtask 1: Aspect Alignment Subtask 1: Aspect Alignment Align opinion sentences to aspects Review Aspects cute… tiny… ..thicker.. Design Design Battery last many hrs die out soon Battery Price Price could afford it still expensive Weblogs Extra Aspects … easy to use… iTunes …better to extend.. warranty 8
Subtask 2: Opinions Separation Subtask 2: Opinions Separation Separate sim opinions from supp ones Review Aspects cute… tiny… ..thicker.. Design Design Battery last many hrs die out soon Battery Price could afford it still expensive Price Extra Aspects … easy to use… iTunes …better to extend.. warranty 9
Subtask 3: Opinion Summary Subtask 3: Opinion Summary Summarize each block with representative sentences and support Review Aspects cute… tiny… ..thicker.. Design Design Battery last many hrs die out soon Battery Representative Opinion (RO) Price could afford it still expensive Price Extra Aspects • Representative sentence: … easy to use… iTunes • Support = 3 …better to extend.. warranty 10
Basic PLSA: Generation Process Basic PLSA: Generation Process [Hofmann 99], [Zhai et al. 04] Topic model = unigram language model Document = multinomial distribution Topics Generate a word Generate a word battery 0.3 θ 1 π d1 in a document life 0.2.. in a document θ 2 design 0.1 π d2 screen 0.05 w … λ B π dk θ k price 0.2 purchase 0.15 B Collection Is 0.05 the 0.04 background a 0.03 .. 11
Basic PLSA: Estimation Basic PLSA: Estimation [Hofmann 99], [Zhai et al. 04] Document Generate a word Generate a word Topics in a document in a document battery ? θ 1 π d1 life ? ? θ 2 design ? π d2 screen ? w ? … Log-likelihood of Log-likelihood of λ B π dk ? θ k the collection price ? the collection purchase ? Collection B Is ? background the ? a ? Estimated with Maximum Likelihood Estimator (MLE) through an EM algorithm 12
Basic PLSA: Problem? Basic PLSA: Problem? Extracted topics may Extracted topics may not align with expert not align with expert Design.. Design.. review aspects Expert review review aspects Battery.. Battery.. with aspects Price.. Price.. … … Solution: conjugate priors Semi-supervised PLSA Basic PLSA Basic PLSA Weblogs iPod nano iPod shuffle on iPod iPod touch … 13
Semi ‐ ‐ supervised PLSA supervised PLSA Semi Maximum Likelihood Maximum Likelihood Estimation (MLE) Estimation (MLE) Add Dirichlet priors Add Dirichlet priors Document Topics battery θ 1 r 1 π d1 life 1 - λ B Maximum A Posterior Maximum A Posterior design θ 2 (MAP) Estimation r 2 (MAP) Estimation π d2 screen w … λ B π dk θ k B Collection • Confidence in priors Is 0.05 • Regularization the 0.04 background a 0.03 .. 14
Subtask 1: Aspect Alignment Subtask 1: Aspect Alignment Review Aspects cute… tiny… ..thicker.. Design Design Battery last many hrs die out soon Battery Price Price could afford it still expensive Weblogs Extra Aspects … easy to use… iTunes …better to extend.. warranty 15
Subtask 1: Aspect Alignment Subtask 1: Aspect Alignment Align to Align to review aspects Document review aspects “aspect words” Topics (nouns) battery θ 1 r 1 π d1 Opinion sentence aligned to life 1 - λ B the most relevant aspect design θ 2 r 2 π d2 screen w … λ B π dk θ k B Collection Is 0.05 Discover extra the 0.04 background Discover extra a 0.03 .. aspects aspects 16
Subtask 2: Opinions Separation Subtask 2: Opinions Separation Review Aspects cute… tiny… ..thicker.. Design Design Battery last many hrs die out soon Battery Price could afford it still expensive Price Extra Aspects … easy to use… iTunes …better to extend.. warranty 17
Subtask 2: Opinions Separation Subtask 2: Opinions Separation Sub-Collection: Battery Document “opinion words” Topics about Battery long θ sim r 1 π d1 1 - λ B many θ supp π d2 Similar opinions w Similar opinions λ B Supplementary Supplementary opinions opinions B Collection Is 0.05 the 0.04 background a 0.03 .. 18
Subtask 3: Opinion Summary Subtask 3: Opinion Summary Review Aspects cute… tiny… ..thicker.. Design Design Battery last many hrs die out soon Battery Representative Opinion (RO) Price could afford it still expensive Price Extra Aspects • Representative sentence: … easy to use… iTunes • Support = 3 …better to extend.. warranty 19
Subtask 3: Opinion Summary Subtask 3: Opinion Summary Sub-Collection: A Block Document RO clusters Centroid Sentence 1 θ 1 π d1 1 - λ B θ 2 Centroid Sentence 2 π d2 w … … λ B π dk θ k B Support = cluster size Collection Is 0.05 the 0.04 background a 0.03 .. 20
Experiment Setup Experiment Setup • Expert review data: Topic Source # words # aspects iPhone CNET 4434 19 Barack Obama Wikipedia 312 14 • Ordinary opinion data: Topic Query Terms # articles # sentences iPhone iPhone 552 3000 Barack Obama Barack+Obama 639 1000 21
Results: Product (iPhone iPhone) ) Results: Product ( • Opinion Integration with review aspects Review article Similar opinions Supplementary opinions You can make N/A … methods for unlocking the emergency calls, but iPhone have emerged on the Unlock/hack you can't use any Internet in the past few weeks, iPhone other functions… although they involve tinkering Confirm the Activation with the iPhone hardware… opinions from the review rated battery life of 8 iPhone will Feature Playing relatively high bitrate hours talk time , 24 Up to 8 Hours of Talk VGA H.264 videos, our iPhone hours of music Time , 6 Hours of lasted almost exactly 9 freaking playback, 7 hours of Internet Use, 7 Hours hours of continuous playback video playback, and 6 of Video Playback or with cell and WiFi on (but hours on Internet use. 24 Hours of Audio Bluetooth off). Additional info Playback under real usage Battery 22
Results: Product (iPhone iPhone) ) Results: Product ( • Opinions on extra aspects support Supplementary opinions on extra aspects 15 You may have heard of iASign … an iPhone Dev Wiki tool that Another way to allows you to activate your phone without going through the activate iPhone iTunes rigamarole. 13 Cisco has owned the trademark on the name " iPhone " since 2000, when it acquired InfoGear Technology Corp., which iPhone trademark originally registered the name. originally owned 13 With the imminent availability of Apple's uber cool iPhone, a by Cisco look at 10 things current smartphones like the Nokia N95 have been able to do for a while and that the iPhone can't currently A better choice for match ... smart phones? 23
Results: Product (iPhone iPhone) ) Results: Product ( • Support statistics for review aspects People care Controversy: about price activation requires contract with AT&T People comment a lot about the unique wi-fi feature 24
Recommend
More recommend
