Post on 20-Mar-2016
description
Data Frame Augmentation of Free Form Queries for Constraint Based Document Filtering
Andrew Zitzelberger
Problem
Constraint Based Queries
Queries
Test Queries 1) Find me a Wii game. 2) Find me a Honda for under 15 thousand dollars. 3) Roller Coaster more than 150 feet high 4) mountains at least 15K feet 5) games under $25 6) mountains less than 4 km 7) ps games < $40 8) coasters longer than 1000 feet 9) car for under 5 grand newer than 1990 with less than 115K miles 10) more than 15K miles under 5 grand newer than 2004
Keywords + Semantics
• Semantic queries are computationally expensive
• Keyword queries are fast and simpleo People are used to keyword queries
• Synergistic solution:o extract numerical constraints from the queryo use keywords to quickly narrow the search spaceo use constraints as a filter
Data Frames
Price internal representation: Double external representation: \$[1-9]\d{0,2}(,\d{3})*|... ... right units: (K)?\s*(cents|dollars|[Gg]rand|...) canonicalization method: toUSDollars comparison methods: LessThan(p1: Price, p2: Price) returns (Boolean) external representation: (less than|<|under|...)\s*{p2}|... ... end
Data Frame Library
Free Form Query
• Car under 6 grand newer than 1990 with less than 115K miles
Step 1: Condition Extraction
• Car under 6 grand newer than 1990 with less than 115K miles
• Extracted Conditionso (Price < 6000)o (Year > 1990)o (Distance < 115000)
Step 2: Remove Condition Values
• Car under newer than with less than
Step 3: Remove Stopwords
• Car
Step 4: Perform Keyword Search
Step 5: Filter Document on Constraints
• Keep page if every constraint is satisfied by at least one extracted value
Experimental Setup
• 300 web documentso 100 car+trucks pages from http://provo.craigslist.orgo 100 video gaming pages from http://provo.craigslist.orgo 50 mountain pages from http://en.wikipedia.orgo 50 roller coaster pages from http://en.wikipedia.org
• 10 querieso 8 with usable conditions
• 2 data setso test-developmento blind test
Results Summary
• Precision increase for 56% of queries o 75% for test-dev, 50% for blind-test
• Precision never worse than keyword query• Most effective for short, focused documents
Precision@3/Query Type Keyword Queries Reduced Queries Data Frame Augmented Queries
Dev-Test Queries 33% 40% 60%
Blind-Test Queries 50% 46% 63%
Overall 42% 43% 62%
Discussion
• Issues:1.inadequate narrowing or ranking of search space2.noise caused by other numbers
Distance < 115000
Future Work
• Scalabilityo Indexing data frame extracted terms
• Precision vs Recall trade-offs
• Pay-as-you-go search construction
Related Work
• Question-Answering Systems
• Keyword search over databases and semantic stores
Questions?
Results (Test-Dev Set)
Query Keyword Condition Removed Keyword
Data Frame Augmentation
Find me a Wii game. 0.33 0.33 0.33
Find me a Honda for under 15 thousand dollars. 0.67 1.00 1.00
roller coaster more than 150ft high 0.33 0.33 0.67
mountains at least 15K ft 1.00 0.67 1.00
games under $25 0.00 0.33 0.67
mountains less than 4 km 0.00 0.00 0.33
ps games < 40 bucks 0.33 0.00 0.33
coasters longer than 1000 feet 0.33 1.00 1.00
car for under 6 grand newer than 1990 with less than 115K miles
0.33 0.33 0.67
more than 15K miles under 10 grand newer than 2000
0.00 0.00 0.00
Results (Blind Test Set)
Query Keyword Condition Removed Keyword
Data Frame Augmentation
Find me a Wii game. 0.67 0.67 0.67
Find me a Honda for under 15 thousand dollars. 0.67 1.00 1.00
roller coaster more than 150ft high 0.67 0.67 0.67
mountains at least 5K ft 0.33 0.33 0.67
games under $25 0.67 0.67 1.00
mountains less than 4 km 0.00 0.00 0.00
ps games < 40 bucks 0.33 0.33 0.33
coasters longer than 1000 feet 0.67 0.67 0.67
car for under 6 grand newer than 1990 with less than 115K miles
0.67 0.00 1.00
more than 15K miles under 10 grand newer than 2000
0.33 0.33 0.33