Sponsored Search Using Landing Pages for Sponsored Search Ad - PDF document

2/21/2012 Sponsored Search Using Landing Pages for Sponsored Search Ad Selection Yejin Choi†, Marcus Fontoura‡, Evgeniy Gabrilovich‡, Vanja Josifovski‡, Mauricio Mediano‡, and Bo Pang‡ † Cornell University ‡ Yahoo! Research Web eb Sear earch ch Spons ponsor ored ed Sear earch ch Exact Match Problem – Vocabulary Mismatch Bid Phrase Sailing vacation Bid Phrase Sailing vacation Cruise ? Advanced Match Advanced Match with Ad + Landing Pages Sailing vacation Cruise ? Sailing vacation Cruise ? Bid Phrase Bid Phrase Title Huge Cruise Discounts Title Huge Cruise Discounts Short Huge Cruise Discounts. Low Price Short Huge Cruise Discounts. Low Price Description Guarantee. Book Online and Save. Description Guarantee. Book Online and Save. Display URL www.crowncruisevacations.com Display URL www.crowncruisevacations.com  Ad retrieval as Web retrieval Landing Page – less explored resource E.g., Ribeiro-Neto et al. (2005), Broder et al. (2008) E.g., Ribeiro-Neto et al. (2005), Murdock et al. (2007) for Content Match, but not for Sponsored search 1

2/21/2012 Benefit of Analyzing Landing Pages • Reducing the vocabulary gap – between the ad and the query • Detecting Ad spam How to make the best use of How to make the best use of noisy Landing Pages? noisy Landing Pages?  Extractive Summarization Task  Extractive Summarization Task  What is a good summary?  What is a good summary? Two Hypotheses: 1. Summary in the context of advertisement intent  In-Context Term Selection 2. Intrinsic summary, independent from an ad  Out-of-Context Term Selection Extractive summarization of Landing Pages Extractive summarization of Landing Pages  In-context Term Selection  In-context Term Selection – Using Ad – Using Ad Bid Phrase Cruise Title Huge Cruise Discounts Short Huge Cruise Discounts. Low Price  In-context Term Selection  In-context Term Selection Description Guarantee. Book Online and Save. – Using Ad + Enriched Ad Context – Using Ad + Enriched Ad Context  Out-of-context Term Selection  Out-of-context Term Selection 2

2/21/2012 In-Context Term Selection In-Context Term Selection Extract Relevant Regions (RR) : Extract Relevant Regions (RR) : 1. Build Ad Vector – TF-IDF vector of an Ad • Ad = (bid phrase, title, short description) 2. Locate Candidate Regions – Text span of landing pages in [-5, +5] window 3. Select Relevant Regions – Cosine Similarity (Ad Vector, Candidate Region) > d Ad Vector Ad Vector a := TF-IDF vector of an Ad a := TF-IDF vector of an Ad Bid Bid Cruise Cruise Phrase Phrase Title Huge Cruise Discounts Title Huge Cruise Discounts Short Huge Cruise Discounts. Short Huge Cruise Discounts. Descript Low Price Guarantee. Descript Low Price Guarantee. ion Book Online and Save. ion Book Online and Save. Relevant Regions Cosine similarity (RR) (Ad vector, candidate region) > d Ad Vector Ad Vector a := TF-IDF vector of an Ad a := TF-IDF vector of an Ad Bid Bid Cruise Cruise Phrase Phrase Title Huge Cruise Discounts Title Huge Cruise Discounts Short Huge Cruise Discounts. Short Huge Cruise Discounts. Descript Low Price Guarantee. Descript Low Price Guarantee. ion Book Online and Save. ion Book Online and Save. 3

2/21/2012 In-Context Term Selection Extractive summarization of Landing Pages with Enriched Ad Context  In-context Term Selection 1. Construction of Co-occurrence Vectors – from Ad Corpus ( half million ads) – Using Ad  In-context Term Selection Bid Cruise Bid Phrase Cruise Bid Phrase – Using Ad + Enriched Ad Context Cruise Bid Title Huge Cruise Discounts Phrase Cruise Title Huge Cruise Discounts Phrase Short Huge Cruise Discounts. Title Huge Cruise Discounts Short Huge Cruise Discounts. Title Descript Huge Cruise Discounts Low Price Guarantee. Short Huge Cruise Discounts.  Out-of-context Term Selection Descript Low Price Guarantee. ion Book Online and Save. Short Huge Cruise Discounts. Descript Low Price Guarantee. ion Book Online and Save. Descript Low Price Guarantee. ion Book Online and Save. ion Book Online and Save. Co- occurrence Vector for “mattress” Co- occurrence Vector for “mattress”  queen(5.64), shopping(2.2), brand(2.5) , store(2.44)  queen(5.64) , shopping(2.2), brand(2.5) , store(2.44) serta(7.64), sealy(7.79), visco(7.75), platform(4.74), serta(7.64), sealy(7.79) , visco(7.75), platform(4.74), products(1.94), cover(4.1), outlet(3.46), allergen(6.63), etc products(1.94), cover(4.1), outlet(3.46), allergen(6.63), etc  weighted by Point-wise Mutual Information (PMI)  weighted by Point-wise Mutual Information (PMI)  Provides related or relevant words  Provides related or relevant words – Not synonyms – Not synonyms In-Context Term Selection Co-occurrence Vectors for Ad with Enriched Ad Context 1. Construction of Co-occurrence Vectors  How to use co-occurrence vectors to enrich ad – from Ad Corpus context?  One co-occurrence vector for each word in an Ad. 2. Composition of Co-occurrence vectors  Need to combine multiple co-occurrence vectors – In order to make Enriched Ad Context – “ Compositional Semantic Vectors (CSV) ” 4

2/21/2012 Compositional Semantic Vector (CSV) Compositional Semantic Vector (CSV) • Composition of Vector Space Models • Composition of Vector Space Models – Mitchell and Lapata (2008) – Mitchell and Lapata (2008) – Compositional Semantics (Montague, 1973) – Compositional Semantics (Montague, 1973) “Principle of Compositionality” “Principle of Compositionality” – Given a set of co-occurrence vectors – The composed meaning can be obtained by Compositional Semantic Vector (CSV) Compositional Semantic Vector (CSV) 1. Component-wise 1. Component-wise vector addition: vector addition: 2. Component-wise 2. Component-wise vector multiplication: vector multiplication: 3. Component-wise vector multiplication with smoothing: “ cruise ” = “ cruise ” = (…, (…, ship, ship, sailing, sailing, vacation, vacation, beach , beach , …, …,  Two different meanings Tom, ( Polysemy ) of “cruise” ! Katie, Holmes, scientology, …) 5

2/21/2012 “ cruise ” = “ sailing ” = “ cruise ” = “ sailing ” = “ cruise & sailing ” = (…, (…, (…, (…, ship, yacht, ship, yacht, (…, sailing, vacation, sailing, vacation, ship, vacation, adventure, vacation, adventure, yacht, beach , beach, beach , beach, vacation, …, charters, …, charters, beach, Tom, Islands, Tom, Islands, Islands, Katie, Caribbean, Katie, Caribbean, Caribbean, Holmes, …) Holmes, …) …) scientology , scientology , …) …) In-Context Term Selection with Enriched Ad Context Cosine Similarity (Ad vector, candidate region) > d 1. Construction of Co-occurrence Vectors – from Ad Corpus 2. Composition of Co-occurrence vectors Ad Vector a := TF-IDF vector of an Ad – In order to make Enriched Ad Context – “ Compositional Semantic Vectors (CSV) ” Bid Cruise 3. Extract Relevant Regions using Phrase – Ad Vector Title Huge Cruise Discounts Short Huge Cruise Discounts. – “ Compositional Semantic Vectors (CSV) ” Descript Low Price Guarantee. ion Book Online and Save. Ad Retrieval System Cosine Similarity (Ad vector, candidate region) + Cosine Similarity (CSV, candidate region) > d+ • Based on Broder et al. (2008) csv : c ompositional semantic vector of an Ad Ad Vector a := TF-IDF vector of an Ad • Terms in Relevant Regions as extra features Bid Cruise Phrase Title Huge Cruise Discounts Short Huge Cruise Discounts. Descript Low Price Guarantee. ion Book Online and Save. 6

2/21/2012 Data Data • Test: • Test: • Human judgment – 22500 query-ad pairs – 22500 query-ad pairs – Perfect (10.0) – Excellent (7.0) • Development: • Development: – Good (3.0) – 3600 query-ad pairs – 3600 query-ad pairs – Fair (0.5) • Excluding query-ad pairs • Excluding query-ad pairs – Bad (0.0) – without valid landing pages – without valid landing pages – URL queries – URL queries Evaluation NDCG @ 1 * 0.62 + * + • Discounted Cumulative Gain (DCG) 0.61 * 0.6 • Normalized DCG (NDCG) 0.59 0.58 0.57 0.56 0.55 Baseline RR RR- csv (Ʃ) RR-csv ( Π) RR-csv ( Π) (smoothed) + : statistically significant gain over baseline (Wilcoxon test) * : statistically significant gain over baseline (paired t-test) Extractive summarization of Landing Pages NDCG @ 3 0.515 * +  In-context Term Selection 0.51 – Using Ad 0.505 0.5  In-context Term Selection 0.495 – Using Ad + Enriched Ad Context 0.49 Baseline RR RR- csv (Ʃ) RR-csv ( Π) RR-csv ( Π)  Out-of-context Term Selection (smoothed) + : statistically significant gain over baseline (Wilcoxon test) * : statistically significant gain over baseline (paired t-test) 7