70 likes | 86 Views
Explore how a novel approach matches questions to pre-answered ones using NLP techniques for accurate results. With a focus on semantic expansion and phrase extraction, this system improves query results by leveraging vast pre-answered question repositories.
E N D
Question Answering via Question-to-Question Mapping Tait Larson Johnson Gong Josh Daniel
Overview • QA systems on the web currently are popular • Common QA systems extract facts from the web, match natural language questions to facts • Our approach • Take pre-answered questions, and match questions to questions • Hopefully useful because of the vast amount of pre-answered questions available on the web • Google answers • Yahoo answers • Lawguru.com • FAQs • Do this via several NLP techniques, primarily focused around query expansion using Wordnet and language model
Query Expansion • POS tagging – preprocess • Search through domain of semantically similar sentences • Goal: Generate phrases that will identify semantically equivalent questions in our corpus • Semantic expansion • Language Model for pruning • Prune incorrect word sense • Trained on question repository • “Can I get into Stanford?” -> “Butt I get into Stanford?” • Phrase Extraction
Information Retrieval • Different from traditional IR • Bigrams • Index • Query • No stop words • No stemming • Why? • These choice emphasize semantic structure of question
Results • Us vs Yahoo • Test questions from “unresolved” Yahoo questions • Metric - Mean Reciprocal Rank • We only index questions, Yahoo indexes answers also
Example Results • A sharp pain in the center of the chest breastbone area? • Keep getting a throbbing pain in the middle of my rib cage . any idea what it could be? • Do Bush baked beans give you gas • Do baked beans make you fart???? yes/no? • Why i sweat and how can i stop this problem • How can I stop sweating? I sweat more when it’s cold…