100 likes | 114 Views
This paper discusses applying diversity metrics to address the problem of redundant refinements in web search. The study evaluates the utility of different refinements and explores the impact of increased diversity. The findings highlight the importance of incorporating popularity, diversity, length, and category classification in refinement selection.
E N D
CS224N 2008 Tague Griffith, Jan Pfeifer Applying Diversity Metrics to Improve the Selection of Web Search Term Refinements
Problem Redundant refinements in a limited space Technical senses dominate others: Java island vs Java programming language Amazon river/rain forest vs Amazon the company What happens with too much diversity Amazon grill houston Embraer ERJ 145 Amazon
CBC Word Sense Similarity Similarity of terms measured by feature vectors Features are a combination of co-occurring words with their syntactic context “wine”: [“sip _”+“Verb-Object”, ...] Data from Wikipedia corpus Problems: Little overlap between web data and Wikipedia data Hyponym siblings too similar, but good refinements “planet jupiter” and “planet earth”
Web Semantic Similarity Similarity as a function of web search engines results Maximum Marginal Relevance greedy algorithm MMR=argmax_x { (1-a)popularity(x) + (a)diversity(x) } x = candidate refinement popularity(x) given by recent search logs diversity(x) given by overlapping search results Clustering of terms demonstrates validity
Tools: demo http://abstract.homelinux.org:9240/janpf/fp/diversity_demo.php?term=target
AB Editorial Test 0.0, 0.3 and 0.8 diversity Evaluate utility of refinements Scale: definitely better, slightly better, same 17 editors Mixed results, with high variability
Results Problems with increased diversity: Editor penalized long refinements Spam and adult terms have “artificial” diversity in web semantic More mixed language results Esoteric refinements Refinement selection should include: Popularity feature Diversity feature Length feature Category classification feature (spam, adult, etc.)